xref: /openbmc/linux/drivers/iommu/amd/iommu.c (revision af9b2ff010f593d81e2f5fb04155e9fc25b9dfd0)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2007-2010 Advanced Micro Devices, Inc.
4  * Author: Joerg Roedel <jroedel@suse.de>
5  *         Leo Duran <leo.duran@amd.com>
6  */
7 
8 #define pr_fmt(fmt)     "AMD-Vi: " fmt
9 #define dev_fmt(fmt)    pr_fmt(fmt)
10 
11 #include <linux/ratelimit.h>
12 #include <linux/pci.h>
13 #include <linux/acpi.h>
14 #include <linux/pci-ats.h>
15 #include <linux/bitmap.h>
16 #include <linux/slab.h>
17 #include <linux/debugfs.h>
18 #include <linux/scatterlist.h>
19 #include <linux/dma-map-ops.h>
20 #include <linux/dma-direct.h>
21 #include <linux/iommu-helper.h>
22 #include <linux/delay.h>
23 #include <linux/amd-iommu.h>
24 #include <linux/notifier.h>
25 #include <linux/export.h>
26 #include <linux/irq.h>
27 #include <linux/msi.h>
28 #include <linux/irqdomain.h>
29 #include <linux/percpu.h>
30 #include <linux/io-pgtable.h>
31 #include <linux/cc_platform.h>
32 #include <asm/irq_remapping.h>
33 #include <asm/io_apic.h>
34 #include <asm/apic.h>
35 #include <asm/hw_irq.h>
36 #include <asm/proto.h>
37 #include <asm/iommu.h>
38 #include <asm/gart.h>
39 #include <asm/dma.h>
40 
41 #include "amd_iommu.h"
42 #include "../dma-iommu.h"
43 #include "../irq_remapping.h"
44 
45 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
46 
47 #define LOOP_TIMEOUT	100000
48 
49 /* IO virtual address start page frame number */
50 #define IOVA_START_PFN		(1)
51 #define IOVA_PFN(addr)		((addr) >> PAGE_SHIFT)
52 
53 /* Reserved IOVA ranges */
54 #define MSI_RANGE_START		(0xfee00000)
55 #define MSI_RANGE_END		(0xfeefffff)
56 #define HT_RANGE_START		(0xfd00000000ULL)
57 #define HT_RANGE_END		(0xffffffffffULL)
58 
59 #define DEFAULT_PGTABLE_LEVEL	PAGE_MODE_3_LEVEL
60 
61 static DEFINE_SPINLOCK(pd_bitmap_lock);
62 
63 LIST_HEAD(ioapic_map);
64 LIST_HEAD(hpet_map);
65 LIST_HEAD(acpihid_map);
66 
67 const struct iommu_ops amd_iommu_ops;
68 
69 static ATOMIC_NOTIFIER_HEAD(ppr_notifier);
70 int amd_iommu_max_glx_val = -1;
71 
72 /*
73  * general struct to manage commands send to an IOMMU
74  */
75 struct iommu_cmd {
76 	u32 data[4];
77 };
78 
79 struct kmem_cache *amd_iommu_irq_cache;
80 
81 static void detach_device(struct device *dev);
82 static int domain_enable_v2(struct protection_domain *domain, int pasids);
83 
84 /****************************************************************************
85  *
86  * Helper functions
87  *
88  ****************************************************************************/
89 
get_acpihid_device_id(struct device * dev,struct acpihid_map_entry ** entry)90 static inline int get_acpihid_device_id(struct device *dev,
91 					struct acpihid_map_entry **entry)
92 {
93 	struct acpi_device *adev = ACPI_COMPANION(dev);
94 	struct acpihid_map_entry *p;
95 
96 	if (!adev)
97 		return -ENODEV;
98 
99 	list_for_each_entry(p, &acpihid_map, list) {
100 		if (acpi_dev_hid_uid_match(adev, p->hid,
101 					   p->uid[0] ? p->uid : NULL)) {
102 			if (entry)
103 				*entry = p;
104 			return p->devid;
105 		}
106 	}
107 	return -EINVAL;
108 }
109 
get_device_sbdf_id(struct device * dev)110 static inline int get_device_sbdf_id(struct device *dev)
111 {
112 	int sbdf;
113 
114 	if (dev_is_pci(dev))
115 		sbdf = get_pci_sbdf_id(to_pci_dev(dev));
116 	else
117 		sbdf = get_acpihid_device_id(dev, NULL);
118 
119 	return sbdf;
120 }
121 
get_dev_table(struct amd_iommu * iommu)122 struct dev_table_entry *get_dev_table(struct amd_iommu *iommu)
123 {
124 	struct dev_table_entry *dev_table;
125 	struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
126 
127 	BUG_ON(pci_seg == NULL);
128 	dev_table = pci_seg->dev_table;
129 	BUG_ON(dev_table == NULL);
130 
131 	return dev_table;
132 }
133 
get_device_segment(struct device * dev)134 static inline u16 get_device_segment(struct device *dev)
135 {
136 	u16 seg;
137 
138 	if (dev_is_pci(dev)) {
139 		struct pci_dev *pdev = to_pci_dev(dev);
140 
141 		seg = pci_domain_nr(pdev->bus);
142 	} else {
143 		u32 devid = get_acpihid_device_id(dev, NULL);
144 
145 		seg = PCI_SBDF_TO_SEGID(devid);
146 	}
147 
148 	return seg;
149 }
150 
151 /* Writes the specific IOMMU for a device into the PCI segment rlookup table */
amd_iommu_set_rlookup_table(struct amd_iommu * iommu,u16 devid)152 void amd_iommu_set_rlookup_table(struct amd_iommu *iommu, u16 devid)
153 {
154 	struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
155 
156 	pci_seg->rlookup_table[devid] = iommu;
157 }
158 
__rlookup_amd_iommu(u16 seg,u16 devid)159 static struct amd_iommu *__rlookup_amd_iommu(u16 seg, u16 devid)
160 {
161 	struct amd_iommu_pci_seg *pci_seg;
162 
163 	for_each_pci_segment(pci_seg) {
164 		if (pci_seg->id == seg)
165 			return pci_seg->rlookup_table[devid];
166 	}
167 	return NULL;
168 }
169 
rlookup_amd_iommu(struct device * dev)170 static struct amd_iommu *rlookup_amd_iommu(struct device *dev)
171 {
172 	u16 seg = get_device_segment(dev);
173 	int devid = get_device_sbdf_id(dev);
174 
175 	if (devid < 0)
176 		return NULL;
177 	return __rlookup_amd_iommu(seg, PCI_SBDF_TO_DEVID(devid));
178 }
179 
to_pdomain(struct iommu_domain * dom)180 static struct protection_domain *to_pdomain(struct iommu_domain *dom)
181 {
182 	return container_of(dom, struct protection_domain, domain);
183 }
184 
alloc_dev_data(struct amd_iommu * iommu,u16 devid)185 static struct iommu_dev_data *alloc_dev_data(struct amd_iommu *iommu, u16 devid)
186 {
187 	struct iommu_dev_data *dev_data;
188 	struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
189 
190 	dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
191 	if (!dev_data)
192 		return NULL;
193 
194 	spin_lock_init(&dev_data->lock);
195 	dev_data->devid = devid;
196 	ratelimit_default_init(&dev_data->rs);
197 
198 	llist_add(&dev_data->dev_data_list, &pci_seg->dev_data_list);
199 	return dev_data;
200 }
201 
search_dev_data(struct amd_iommu * iommu,u16 devid)202 static struct iommu_dev_data *search_dev_data(struct amd_iommu *iommu, u16 devid)
203 {
204 	struct iommu_dev_data *dev_data;
205 	struct llist_node *node;
206 	struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
207 
208 	if (llist_empty(&pci_seg->dev_data_list))
209 		return NULL;
210 
211 	node = pci_seg->dev_data_list.first;
212 	llist_for_each_entry(dev_data, node, dev_data_list) {
213 		if (dev_data->devid == devid)
214 			return dev_data;
215 	}
216 
217 	return NULL;
218 }
219 
clone_alias(struct pci_dev * pdev,u16 alias,void * data)220 static int clone_alias(struct pci_dev *pdev, u16 alias, void *data)
221 {
222 	struct amd_iommu *iommu;
223 	struct dev_table_entry *dev_table;
224 	u16 devid = pci_dev_id(pdev);
225 
226 	if (devid == alias)
227 		return 0;
228 
229 	iommu = rlookup_amd_iommu(&pdev->dev);
230 	if (!iommu)
231 		return 0;
232 
233 	amd_iommu_set_rlookup_table(iommu, alias);
234 	dev_table = get_dev_table(iommu);
235 	memcpy(dev_table[alias].data,
236 	       dev_table[devid].data,
237 	       sizeof(dev_table[alias].data));
238 
239 	return 0;
240 }
241 
clone_aliases(struct amd_iommu * iommu,struct device * dev)242 static void clone_aliases(struct amd_iommu *iommu, struct device *dev)
243 {
244 	struct pci_dev *pdev;
245 
246 	if (!dev_is_pci(dev))
247 		return;
248 	pdev = to_pci_dev(dev);
249 
250 	/*
251 	 * The IVRS alias stored in the alias table may not be
252 	 * part of the PCI DMA aliases if it's bus differs
253 	 * from the original device.
254 	 */
255 	clone_alias(pdev, iommu->pci_seg->alias_table[pci_dev_id(pdev)], NULL);
256 
257 	pci_for_each_dma_alias(pdev, clone_alias, NULL);
258 }
259 
setup_aliases(struct amd_iommu * iommu,struct device * dev)260 static void setup_aliases(struct amd_iommu *iommu, struct device *dev)
261 {
262 	struct pci_dev *pdev = to_pci_dev(dev);
263 	struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
264 	u16 ivrs_alias;
265 
266 	/* For ACPI HID devices, there are no aliases */
267 	if (!dev_is_pci(dev))
268 		return;
269 
270 	/*
271 	 * Add the IVRS alias to the pci aliases if it is on the same
272 	 * bus. The IVRS table may know about a quirk that we don't.
273 	 */
274 	ivrs_alias = pci_seg->alias_table[pci_dev_id(pdev)];
275 	if (ivrs_alias != pci_dev_id(pdev) &&
276 	    PCI_BUS_NUM(ivrs_alias) == pdev->bus->number)
277 		pci_add_dma_alias(pdev, ivrs_alias & 0xff, 1);
278 
279 	clone_aliases(iommu, dev);
280 }
281 
find_dev_data(struct amd_iommu * iommu,u16 devid)282 static struct iommu_dev_data *find_dev_data(struct amd_iommu *iommu, u16 devid)
283 {
284 	struct iommu_dev_data *dev_data;
285 
286 	dev_data = search_dev_data(iommu, devid);
287 
288 	if (dev_data == NULL) {
289 		dev_data = alloc_dev_data(iommu, devid);
290 		if (!dev_data)
291 			return NULL;
292 
293 		if (translation_pre_enabled(iommu))
294 			dev_data->defer_attach = true;
295 	}
296 
297 	return dev_data;
298 }
299 
300 /*
301 * Find or create an IOMMU group for a acpihid device.
302 */
acpihid_device_group(struct device * dev)303 static struct iommu_group *acpihid_device_group(struct device *dev)
304 {
305 	struct acpihid_map_entry *p, *entry = NULL;
306 	int devid;
307 
308 	devid = get_acpihid_device_id(dev, &entry);
309 	if (devid < 0)
310 		return ERR_PTR(devid);
311 
312 	list_for_each_entry(p, &acpihid_map, list) {
313 		if ((devid == p->devid) && p->group)
314 			entry->group = p->group;
315 	}
316 
317 	if (!entry->group)
318 		entry->group = generic_device_group(dev);
319 	else
320 		iommu_group_ref_get(entry->group);
321 
322 	return entry->group;
323 }
324 
pci_iommuv2_capable(struct pci_dev * pdev)325 static bool pci_iommuv2_capable(struct pci_dev *pdev)
326 {
327 	static const int caps[] = {
328 		PCI_EXT_CAP_ID_PRI,
329 		PCI_EXT_CAP_ID_PASID,
330 	};
331 	int i, pos;
332 
333 	if (!pci_ats_supported(pdev))
334 		return false;
335 
336 	for (i = 0; i < 2; ++i) {
337 		pos = pci_find_ext_capability(pdev, caps[i]);
338 		if (pos == 0)
339 			return false;
340 	}
341 
342 	return true;
343 }
344 
345 /*
346  * This function checks if the driver got a valid device from the caller to
347  * avoid dereferencing invalid pointers.
348  */
check_device(struct device * dev)349 static bool check_device(struct device *dev)
350 {
351 	struct amd_iommu_pci_seg *pci_seg;
352 	struct amd_iommu *iommu;
353 	int devid, sbdf;
354 
355 	if (!dev)
356 		return false;
357 
358 	sbdf = get_device_sbdf_id(dev);
359 	if (sbdf < 0)
360 		return false;
361 	devid = PCI_SBDF_TO_DEVID(sbdf);
362 
363 	iommu = rlookup_amd_iommu(dev);
364 	if (!iommu)
365 		return false;
366 
367 	/* Out of our scope? */
368 	pci_seg = iommu->pci_seg;
369 	if (devid > pci_seg->last_bdf)
370 		return false;
371 
372 	return true;
373 }
374 
iommu_init_device(struct amd_iommu * iommu,struct device * dev)375 static int iommu_init_device(struct amd_iommu *iommu, struct device *dev)
376 {
377 	struct iommu_dev_data *dev_data;
378 	int devid, sbdf;
379 
380 	if (dev_iommu_priv_get(dev))
381 		return 0;
382 
383 	sbdf = get_device_sbdf_id(dev);
384 	if (sbdf < 0)
385 		return sbdf;
386 
387 	devid = PCI_SBDF_TO_DEVID(sbdf);
388 	dev_data = find_dev_data(iommu, devid);
389 	if (!dev_data)
390 		return -ENOMEM;
391 
392 	dev_data->dev = dev;
393 	setup_aliases(iommu, dev);
394 
395 	/*
396 	 * By default we use passthrough mode for IOMMUv2 capable device.
397 	 * But if amd_iommu=force_isolation is set (e.g. to debug DMA to
398 	 * invalid address), we ignore the capability for the device so
399 	 * it'll be forced to go into translation mode.
400 	 */
401 	if ((iommu_default_passthrough() || !amd_iommu_force_isolation) &&
402 	    dev_is_pci(dev) && pci_iommuv2_capable(to_pci_dev(dev))) {
403 		dev_data->iommu_v2 = iommu->is_iommu_v2;
404 	}
405 
406 	dev_iommu_priv_set(dev, dev_data);
407 
408 	return 0;
409 }
410 
iommu_ignore_device(struct amd_iommu * iommu,struct device * dev)411 static void iommu_ignore_device(struct amd_iommu *iommu, struct device *dev)
412 {
413 	struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
414 	struct dev_table_entry *dev_table = get_dev_table(iommu);
415 	int devid, sbdf;
416 
417 	sbdf = get_device_sbdf_id(dev);
418 	if (sbdf < 0)
419 		return;
420 
421 	devid = PCI_SBDF_TO_DEVID(sbdf);
422 	pci_seg->rlookup_table[devid] = NULL;
423 	memset(&dev_table[devid], 0, sizeof(struct dev_table_entry));
424 
425 	setup_aliases(iommu, dev);
426 }
427 
amd_iommu_uninit_device(struct device * dev)428 static void amd_iommu_uninit_device(struct device *dev)
429 {
430 	struct iommu_dev_data *dev_data;
431 
432 	dev_data = dev_iommu_priv_get(dev);
433 	if (!dev_data)
434 		return;
435 
436 	if (dev_data->domain)
437 		detach_device(dev);
438 
439 	dev_iommu_priv_set(dev, NULL);
440 
441 	/*
442 	 * We keep dev_data around for unplugged devices and reuse it when the
443 	 * device is re-plugged - not doing so would introduce a ton of races.
444 	 */
445 }
446 
447 /****************************************************************************
448  *
449  * Interrupt handling functions
450  *
451  ****************************************************************************/
452 
dump_dte_entry(struct amd_iommu * iommu,u16 devid)453 static void dump_dte_entry(struct amd_iommu *iommu, u16 devid)
454 {
455 	int i;
456 	struct dev_table_entry *dev_table = get_dev_table(iommu);
457 
458 	for (i = 0; i < 4; ++i)
459 		pr_err("DTE[%d]: %016llx\n", i, dev_table[devid].data[i]);
460 }
461 
dump_command(unsigned long phys_addr)462 static void dump_command(unsigned long phys_addr)
463 {
464 	struct iommu_cmd *cmd = iommu_phys_to_virt(phys_addr);
465 	int i;
466 
467 	for (i = 0; i < 4; ++i)
468 		pr_err("CMD[%d]: %08x\n", i, cmd->data[i]);
469 }
470 
amd_iommu_report_rmp_hw_error(struct amd_iommu * iommu,volatile u32 * event)471 static void amd_iommu_report_rmp_hw_error(struct amd_iommu *iommu, volatile u32 *event)
472 {
473 	struct iommu_dev_data *dev_data = NULL;
474 	int devid, vmg_tag, flags;
475 	struct pci_dev *pdev;
476 	u64 spa;
477 
478 	devid   = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
479 	vmg_tag = (event[1]) & 0xFFFF;
480 	flags   = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
481 	spa     = ((u64)event[3] << 32) | (event[2] & 0xFFFFFFF8);
482 
483 	pdev = pci_get_domain_bus_and_slot(iommu->pci_seg->id, PCI_BUS_NUM(devid),
484 					   devid & 0xff);
485 	if (pdev)
486 		dev_data = dev_iommu_priv_get(&pdev->dev);
487 
488 	if (dev_data) {
489 		if (__ratelimit(&dev_data->rs)) {
490 			pci_err(pdev, "Event logged [RMP_HW_ERROR vmg_tag=0x%04x, spa=0x%llx, flags=0x%04x]\n",
491 				vmg_tag, spa, flags);
492 		}
493 	} else {
494 		pr_err_ratelimited("Event logged [RMP_HW_ERROR device=%04x:%02x:%02x.%x, vmg_tag=0x%04x, spa=0x%llx, flags=0x%04x]\n",
495 			iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
496 			vmg_tag, spa, flags);
497 	}
498 
499 	if (pdev)
500 		pci_dev_put(pdev);
501 }
502 
amd_iommu_report_rmp_fault(struct amd_iommu * iommu,volatile u32 * event)503 static void amd_iommu_report_rmp_fault(struct amd_iommu *iommu, volatile u32 *event)
504 {
505 	struct iommu_dev_data *dev_data = NULL;
506 	int devid, flags_rmp, vmg_tag, flags;
507 	struct pci_dev *pdev;
508 	u64 gpa;
509 
510 	devid     = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
511 	flags_rmp = (event[0] >> EVENT_FLAGS_SHIFT) & 0xFF;
512 	vmg_tag   = (event[1]) & 0xFFFF;
513 	flags     = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
514 	gpa       = ((u64)event[3] << 32) | event[2];
515 
516 	pdev = pci_get_domain_bus_and_slot(iommu->pci_seg->id, PCI_BUS_NUM(devid),
517 					   devid & 0xff);
518 	if (pdev)
519 		dev_data = dev_iommu_priv_get(&pdev->dev);
520 
521 	if (dev_data) {
522 		if (__ratelimit(&dev_data->rs)) {
523 			pci_err(pdev, "Event logged [RMP_PAGE_FAULT vmg_tag=0x%04x, gpa=0x%llx, flags_rmp=0x%04x, flags=0x%04x]\n",
524 				vmg_tag, gpa, flags_rmp, flags);
525 		}
526 	} else {
527 		pr_err_ratelimited("Event logged [RMP_PAGE_FAULT device=%04x:%02x:%02x.%x, vmg_tag=0x%04x, gpa=0x%llx, flags_rmp=0x%04x, flags=0x%04x]\n",
528 			iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
529 			vmg_tag, gpa, flags_rmp, flags);
530 	}
531 
532 	if (pdev)
533 		pci_dev_put(pdev);
534 }
535 
536 #define IS_IOMMU_MEM_TRANSACTION(flags)		\
537 	(((flags) & EVENT_FLAG_I) == 0)
538 
539 #define IS_WRITE_REQUEST(flags)			\
540 	((flags) & EVENT_FLAG_RW)
541 
amd_iommu_report_page_fault(struct amd_iommu * iommu,u16 devid,u16 domain_id,u64 address,int flags)542 static void amd_iommu_report_page_fault(struct amd_iommu *iommu,
543 					u16 devid, u16 domain_id,
544 					u64 address, int flags)
545 {
546 	struct iommu_dev_data *dev_data = NULL;
547 	struct pci_dev *pdev;
548 
549 	pdev = pci_get_domain_bus_and_slot(iommu->pci_seg->id, PCI_BUS_NUM(devid),
550 					   devid & 0xff);
551 	if (pdev)
552 		dev_data = dev_iommu_priv_get(&pdev->dev);
553 
554 	if (dev_data) {
555 		/*
556 		 * If this is a DMA fault (for which the I(nterrupt)
557 		 * bit will be unset), allow report_iommu_fault() to
558 		 * prevent logging it.
559 		 */
560 		if (IS_IOMMU_MEM_TRANSACTION(flags)) {
561 			/* Device not attached to domain properly */
562 			if (dev_data->domain == NULL) {
563 				pr_err_ratelimited("Event logged [Device not attached to domain properly]\n");
564 				pr_err_ratelimited("  device=%04x:%02x:%02x.%x domain=0x%04x\n",
565 						   iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid),
566 						   PCI_FUNC(devid), domain_id);
567 				goto out;
568 			}
569 
570 			if (!report_iommu_fault(&dev_data->domain->domain,
571 						&pdev->dev, address,
572 						IS_WRITE_REQUEST(flags) ?
573 							IOMMU_FAULT_WRITE :
574 							IOMMU_FAULT_READ))
575 				goto out;
576 		}
577 
578 		if (__ratelimit(&dev_data->rs)) {
579 			pci_err(pdev, "Event logged [IO_PAGE_FAULT domain=0x%04x address=0x%llx flags=0x%04x]\n",
580 				domain_id, address, flags);
581 		}
582 	} else {
583 		pr_err_ratelimited("Event logged [IO_PAGE_FAULT device=%04x:%02x:%02x.%x domain=0x%04x address=0x%llx flags=0x%04x]\n",
584 			iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
585 			domain_id, address, flags);
586 	}
587 
588 out:
589 	if (pdev)
590 		pci_dev_put(pdev);
591 }
592 
iommu_print_event(struct amd_iommu * iommu,void * __evt)593 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
594 {
595 	struct device *dev = iommu->iommu.dev;
596 	int type, devid, flags, tag;
597 	volatile u32 *event = __evt;
598 	int count = 0;
599 	u64 address;
600 	u32 pasid;
601 
602 retry:
603 	type    = (event[1] >> EVENT_TYPE_SHIFT)  & EVENT_TYPE_MASK;
604 	devid   = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
605 	pasid   = (event[0] & EVENT_DOMID_MASK_HI) |
606 		  (event[1] & EVENT_DOMID_MASK_LO);
607 	flags   = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
608 	address = (u64)(((u64)event[3]) << 32) | event[2];
609 
610 	if (type == 0) {
611 		/* Did we hit the erratum? */
612 		if (++count == LOOP_TIMEOUT) {
613 			pr_err("No event written to event log\n");
614 			return;
615 		}
616 		udelay(1);
617 		goto retry;
618 	}
619 
620 	if (type == EVENT_TYPE_IO_FAULT) {
621 		amd_iommu_report_page_fault(iommu, devid, pasid, address, flags);
622 		return;
623 	}
624 
625 	switch (type) {
626 	case EVENT_TYPE_ILL_DEV:
627 		dev_err(dev, "Event logged [ILLEGAL_DEV_TABLE_ENTRY device=%04x:%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x]\n",
628 			iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
629 			pasid, address, flags);
630 		dump_dte_entry(iommu, devid);
631 		break;
632 	case EVENT_TYPE_DEV_TAB_ERR:
633 		dev_err(dev, "Event logged [DEV_TAB_HARDWARE_ERROR device=%04x:%02x:%02x.%x "
634 			"address=0x%llx flags=0x%04x]\n",
635 			iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
636 			address, flags);
637 		break;
638 	case EVENT_TYPE_PAGE_TAB_ERR:
639 		dev_err(dev, "Event logged [PAGE_TAB_HARDWARE_ERROR device=%04x:%02x:%02x.%x pasid=0x%04x address=0x%llx flags=0x%04x]\n",
640 			iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
641 			pasid, address, flags);
642 		break;
643 	case EVENT_TYPE_ILL_CMD:
644 		dev_err(dev, "Event logged [ILLEGAL_COMMAND_ERROR address=0x%llx]\n", address);
645 		dump_command(address);
646 		break;
647 	case EVENT_TYPE_CMD_HARD_ERR:
648 		dev_err(dev, "Event logged [COMMAND_HARDWARE_ERROR address=0x%llx flags=0x%04x]\n",
649 			address, flags);
650 		break;
651 	case EVENT_TYPE_IOTLB_INV_TO:
652 		dev_err(dev, "Event logged [IOTLB_INV_TIMEOUT device=%04x:%02x:%02x.%x address=0x%llx]\n",
653 			iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
654 			address);
655 		break;
656 	case EVENT_TYPE_INV_DEV_REQ:
657 		dev_err(dev, "Event logged [INVALID_DEVICE_REQUEST device=%04x:%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x]\n",
658 			iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
659 			pasid, address, flags);
660 		break;
661 	case EVENT_TYPE_RMP_FAULT:
662 		amd_iommu_report_rmp_fault(iommu, event);
663 		break;
664 	case EVENT_TYPE_RMP_HW_ERR:
665 		amd_iommu_report_rmp_hw_error(iommu, event);
666 		break;
667 	case EVENT_TYPE_INV_PPR_REQ:
668 		pasid = PPR_PASID(*((u64 *)__evt));
669 		tag = event[1] & 0x03FF;
670 		dev_err(dev, "Event logged [INVALID_PPR_REQUEST device=%04x:%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x tag=0x%03x]\n",
671 			iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
672 			pasid, address, flags, tag);
673 		break;
674 	default:
675 		dev_err(dev, "Event logged [UNKNOWN event[0]=0x%08x event[1]=0x%08x event[2]=0x%08x event[3]=0x%08x\n",
676 			event[0], event[1], event[2], event[3]);
677 	}
678 
679 	/*
680 	 * To detect the hardware errata 732 we need to clear the
681 	 * entry back to zero. This issue does not exist on SNP
682 	 * enabled system. Also this buffer is not writeable on
683 	 * SNP enabled system.
684 	 */
685 	if (!amd_iommu_snp_en)
686 		memset(__evt, 0, 4 * sizeof(u32));
687 }
688 
iommu_poll_events(struct amd_iommu * iommu)689 static void iommu_poll_events(struct amd_iommu *iommu)
690 {
691 	u32 head, tail;
692 
693 	head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
694 	tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
695 
696 	while (head != tail) {
697 		iommu_print_event(iommu, iommu->evt_buf + head);
698 		head = (head + EVENT_ENTRY_SIZE) % EVT_BUFFER_SIZE;
699 	}
700 
701 	writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
702 }
703 
iommu_handle_ppr_entry(struct amd_iommu * iommu,u64 * raw)704 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
705 {
706 	struct amd_iommu_fault fault;
707 
708 	if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
709 		pr_err_ratelimited("Unknown PPR request received\n");
710 		return;
711 	}
712 
713 	fault.address   = raw[1];
714 	fault.pasid     = PPR_PASID(raw[0]);
715 	fault.sbdf      = PCI_SEG_DEVID_TO_SBDF(iommu->pci_seg->id, PPR_DEVID(raw[0]));
716 	fault.tag       = PPR_TAG(raw[0]);
717 	fault.flags     = PPR_FLAGS(raw[0]);
718 
719 	atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
720 }
721 
iommu_poll_ppr_log(struct amd_iommu * iommu)722 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
723 {
724 	u32 head, tail;
725 
726 	if (iommu->ppr_log == NULL)
727 		return;
728 
729 	head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
730 	tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
731 
732 	while (head != tail) {
733 		volatile u64 *raw;
734 		u64 entry[2];
735 		int i;
736 
737 		raw = (u64 *)(iommu->ppr_log + head);
738 
739 		/*
740 		 * Hardware bug: Interrupt may arrive before the entry is
741 		 * written to memory. If this happens we need to wait for the
742 		 * entry to arrive.
743 		 */
744 		for (i = 0; i < LOOP_TIMEOUT; ++i) {
745 			if (PPR_REQ_TYPE(raw[0]) != 0)
746 				break;
747 			udelay(1);
748 		}
749 
750 		/* Avoid memcpy function-call overhead */
751 		entry[0] = raw[0];
752 		entry[1] = raw[1];
753 
754 		/*
755 		 * To detect the hardware errata 733 we need to clear the
756 		 * entry back to zero. This issue does not exist on SNP
757 		 * enabled system. Also this buffer is not writeable on
758 		 * SNP enabled system.
759 		 */
760 		if (!amd_iommu_snp_en)
761 			raw[0] = raw[1] = 0UL;
762 
763 		/* Update head pointer of hardware ring-buffer */
764 		head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
765 		writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
766 
767 		/* Handle PPR entry */
768 		iommu_handle_ppr_entry(iommu, entry);
769 
770 		/* Refresh ring-buffer information */
771 		head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
772 		tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
773 	}
774 }
775 
776 #ifdef CONFIG_IRQ_REMAP
777 static int (*iommu_ga_log_notifier)(u32);
778 
amd_iommu_register_ga_log_notifier(int (* notifier)(u32))779 int amd_iommu_register_ga_log_notifier(int (*notifier)(u32))
780 {
781 	iommu_ga_log_notifier = notifier;
782 
783 	/*
784 	 * Ensure all in-flight IRQ handlers run to completion before returning
785 	 * to the caller, e.g. to ensure module code isn't unloaded while it's
786 	 * being executed in the IRQ handler.
787 	 */
788 	if (!notifier)
789 		synchronize_rcu();
790 
791 	return 0;
792 }
793 EXPORT_SYMBOL(amd_iommu_register_ga_log_notifier);
794 
iommu_poll_ga_log(struct amd_iommu * iommu)795 static void iommu_poll_ga_log(struct amd_iommu *iommu)
796 {
797 	u32 head, tail;
798 
799 	if (iommu->ga_log == NULL)
800 		return;
801 
802 	head = readl(iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
803 	tail = readl(iommu->mmio_base + MMIO_GA_TAIL_OFFSET);
804 
805 	while (head != tail) {
806 		volatile u64 *raw;
807 		u64 log_entry;
808 
809 		raw = (u64 *)(iommu->ga_log + head);
810 
811 		/* Avoid memcpy function-call overhead */
812 		log_entry = *raw;
813 
814 		/* Update head pointer of hardware ring-buffer */
815 		head = (head + GA_ENTRY_SIZE) % GA_LOG_SIZE;
816 		writel(head, iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
817 
818 		/* Handle GA entry */
819 		switch (GA_REQ_TYPE(log_entry)) {
820 		case GA_GUEST_NR:
821 			if (!iommu_ga_log_notifier)
822 				break;
823 
824 			pr_debug("%s: devid=%#x, ga_tag=%#x\n",
825 				 __func__, GA_DEVID(log_entry),
826 				 GA_TAG(log_entry));
827 
828 			if (iommu_ga_log_notifier(GA_TAG(log_entry)) != 0)
829 				pr_err("GA log notifier failed.\n");
830 			break;
831 		default:
832 			break;
833 		}
834 	}
835 }
836 
837 static void
amd_iommu_set_pci_msi_domain(struct device * dev,struct amd_iommu * iommu)838 amd_iommu_set_pci_msi_domain(struct device *dev, struct amd_iommu *iommu)
839 {
840 	if (!irq_remapping_enabled || !dev_is_pci(dev) ||
841 	    !pci_dev_has_default_msi_parent_domain(to_pci_dev(dev)))
842 		return;
843 
844 	dev_set_msi_domain(dev, iommu->ir_domain);
845 }
846 
847 #else /* CONFIG_IRQ_REMAP */
848 static inline void
amd_iommu_set_pci_msi_domain(struct device * dev,struct amd_iommu * iommu)849 amd_iommu_set_pci_msi_domain(struct device *dev, struct amd_iommu *iommu) { }
850 #endif /* !CONFIG_IRQ_REMAP */
851 
amd_iommu_handle_irq(void * data,const char * evt_type,u32 int_mask,u32 overflow_mask,void (* int_handler)(struct amd_iommu *),void (* overflow_handler)(struct amd_iommu *))852 static void amd_iommu_handle_irq(void *data, const char *evt_type,
853 				 u32 int_mask, u32 overflow_mask,
854 				 void (*int_handler)(struct amd_iommu *),
855 				 void (*overflow_handler)(struct amd_iommu *))
856 {
857 	struct amd_iommu *iommu = (struct amd_iommu *) data;
858 	u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
859 	u32 mask = int_mask | overflow_mask;
860 
861 	while (status & mask) {
862 		/* Enable interrupt sources again */
863 		writel(mask, iommu->mmio_base + MMIO_STATUS_OFFSET);
864 
865 		if (int_handler) {
866 			pr_devel("Processing IOMMU (ivhd%d) %s Log\n",
867 				 iommu->index, evt_type);
868 			int_handler(iommu);
869 		}
870 
871 		if ((status & overflow_mask) && overflow_handler)
872 			overflow_handler(iommu);
873 
874 		/*
875 		 * Hardware bug: ERBT1312
876 		 * When re-enabling interrupt (by writing 1
877 		 * to clear the bit), the hardware might also try to set
878 		 * the interrupt bit in the event status register.
879 		 * In this scenario, the bit will be set, and disable
880 		 * subsequent interrupts.
881 		 *
882 		 * Workaround: The IOMMU driver should read back the
883 		 * status register and check if the interrupt bits are cleared.
884 		 * If not, driver will need to go through the interrupt handler
885 		 * again and re-clear the bits
886 		 */
887 		status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
888 	}
889 }
890 
amd_iommu_int_thread_evtlog(int irq,void * data)891 irqreturn_t amd_iommu_int_thread_evtlog(int irq, void *data)
892 {
893 	amd_iommu_handle_irq(data, "Evt", MMIO_STATUS_EVT_INT_MASK,
894 			     MMIO_STATUS_EVT_OVERFLOW_MASK,
895 			     iommu_poll_events, amd_iommu_restart_event_logging);
896 
897 	return IRQ_HANDLED;
898 }
899 
amd_iommu_int_thread_pprlog(int irq,void * data)900 irqreturn_t amd_iommu_int_thread_pprlog(int irq, void *data)
901 {
902 	amd_iommu_handle_irq(data, "PPR", MMIO_STATUS_PPR_INT_MASK,
903 			     MMIO_STATUS_PPR_OVERFLOW_MASK,
904 			     iommu_poll_ppr_log, amd_iommu_restart_ppr_log);
905 
906 	return IRQ_HANDLED;
907 }
908 
amd_iommu_int_thread_galog(int irq,void * data)909 irqreturn_t amd_iommu_int_thread_galog(int irq, void *data)
910 {
911 #ifdef CONFIG_IRQ_REMAP
912 	amd_iommu_handle_irq(data, "GA", MMIO_STATUS_GALOG_INT_MASK,
913 			     MMIO_STATUS_GALOG_OVERFLOW_MASK,
914 			     iommu_poll_ga_log, amd_iommu_restart_ga_log);
915 #endif
916 
917 	return IRQ_HANDLED;
918 }
919 
amd_iommu_int_thread(int irq,void * data)920 irqreturn_t amd_iommu_int_thread(int irq, void *data)
921 {
922 	amd_iommu_int_thread_evtlog(irq, data);
923 	amd_iommu_int_thread_pprlog(irq, data);
924 	amd_iommu_int_thread_galog(irq, data);
925 
926 	return IRQ_HANDLED;
927 }
928 
amd_iommu_int_handler(int irq,void * data)929 irqreturn_t amd_iommu_int_handler(int irq, void *data)
930 {
931 	return IRQ_WAKE_THREAD;
932 }
933 
934 /****************************************************************************
935  *
936  * IOMMU command queuing functions
937  *
938  ****************************************************************************/
939 
wait_on_sem(struct amd_iommu * iommu,u64 data)940 static int wait_on_sem(struct amd_iommu *iommu, u64 data)
941 {
942 	int i = 0;
943 
944 	while (*iommu->cmd_sem != data && i < LOOP_TIMEOUT) {
945 		udelay(1);
946 		i += 1;
947 	}
948 
949 	if (i == LOOP_TIMEOUT) {
950 		pr_alert("Completion-Wait loop timed out\n");
951 		return -EIO;
952 	}
953 
954 	return 0;
955 }
956 
copy_cmd_to_buffer(struct amd_iommu * iommu,struct iommu_cmd * cmd)957 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
958 			       struct iommu_cmd *cmd)
959 {
960 	u8 *target;
961 	u32 tail;
962 
963 	/* Copy command to buffer */
964 	tail = iommu->cmd_buf_tail;
965 	target = iommu->cmd_buf + tail;
966 	memcpy(target, cmd, sizeof(*cmd));
967 
968 	tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
969 	iommu->cmd_buf_tail = tail;
970 
971 	/* Tell the IOMMU about it */
972 	writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
973 }
974 
build_completion_wait(struct iommu_cmd * cmd,struct amd_iommu * iommu,u64 data)975 static void build_completion_wait(struct iommu_cmd *cmd,
976 				  struct amd_iommu *iommu,
977 				  u64 data)
978 {
979 	u64 paddr = iommu_virt_to_phys((void *)iommu->cmd_sem);
980 
981 	memset(cmd, 0, sizeof(*cmd));
982 	cmd->data[0] = lower_32_bits(paddr) | CMD_COMPL_WAIT_STORE_MASK;
983 	cmd->data[1] = upper_32_bits(paddr);
984 	cmd->data[2] = lower_32_bits(data);
985 	cmd->data[3] = upper_32_bits(data);
986 	CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
987 }
988 
build_inv_dte(struct iommu_cmd * cmd,u16 devid)989 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
990 {
991 	memset(cmd, 0, sizeof(*cmd));
992 	cmd->data[0] = devid;
993 	CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
994 }
995 
996 /*
997  * Builds an invalidation address which is suitable for one page or multiple
998  * pages. Sets the size bit (S) as needed is more than one page is flushed.
999  */
build_inv_address(u64 address,size_t size)1000 static inline u64 build_inv_address(u64 address, size_t size)
1001 {
1002 	u64 pages, end, msb_diff;
1003 
1004 	pages = iommu_num_pages(address, size, PAGE_SIZE);
1005 
1006 	if (pages == 1)
1007 		return address & PAGE_MASK;
1008 
1009 	end = address + size - 1;
1010 
1011 	/*
1012 	 * msb_diff would hold the index of the most significant bit that
1013 	 * flipped between the start and end.
1014 	 */
1015 	msb_diff = fls64(end ^ address) - 1;
1016 
1017 	/*
1018 	 * Bits 63:52 are sign extended. If for some reason bit 51 is different
1019 	 * between the start and the end, invalidate everything.
1020 	 */
1021 	if (unlikely(msb_diff > 51)) {
1022 		address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
1023 	} else {
1024 		/*
1025 		 * The msb-bit must be clear on the address. Just set all the
1026 		 * lower bits.
1027 		 */
1028 		address |= (1ull << msb_diff) - 1;
1029 	}
1030 
1031 	/* Clear bits 11:0 */
1032 	address &= PAGE_MASK;
1033 
1034 	/* Set the size bit - we flush more than one 4kb page */
1035 	return address | CMD_INV_IOMMU_PAGES_SIZE_MASK;
1036 }
1037 
build_inv_iommu_pages(struct iommu_cmd * cmd,u64 address,size_t size,u16 domid,int pde)1038 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
1039 				  size_t size, u16 domid, int pde)
1040 {
1041 	u64 inv_address = build_inv_address(address, size);
1042 
1043 	memset(cmd, 0, sizeof(*cmd));
1044 	cmd->data[1] |= domid;
1045 	cmd->data[2]  = lower_32_bits(inv_address);
1046 	cmd->data[3]  = upper_32_bits(inv_address);
1047 	CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
1048 	if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
1049 		cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
1050 }
1051 
build_inv_iotlb_pages(struct iommu_cmd * cmd,u16 devid,int qdep,u64 address,size_t size)1052 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
1053 				  u64 address, size_t size)
1054 {
1055 	u64 inv_address = build_inv_address(address, size);
1056 
1057 	memset(cmd, 0, sizeof(*cmd));
1058 	cmd->data[0]  = devid;
1059 	cmd->data[0] |= (qdep & 0xff) << 24;
1060 	cmd->data[1]  = devid;
1061 	cmd->data[2]  = lower_32_bits(inv_address);
1062 	cmd->data[3]  = upper_32_bits(inv_address);
1063 	CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
1064 }
1065 
build_inv_iommu_pasid(struct iommu_cmd * cmd,u16 domid,u32 pasid,u64 address,bool size)1066 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, u32 pasid,
1067 				  u64 address, bool size)
1068 {
1069 	memset(cmd, 0, sizeof(*cmd));
1070 
1071 	address &= ~(0xfffULL);
1072 
1073 	cmd->data[0]  = pasid;
1074 	cmd->data[1]  = domid;
1075 	cmd->data[2]  = lower_32_bits(address);
1076 	cmd->data[3]  = upper_32_bits(address);
1077 	cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
1078 	cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
1079 	if (size)
1080 		cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
1081 	CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
1082 }
1083 
build_inv_iotlb_pasid(struct iommu_cmd * cmd,u16 devid,u32 pasid,int qdep,u64 address,bool size)1084 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, u32 pasid,
1085 				  int qdep, u64 address, bool size)
1086 {
1087 	memset(cmd, 0, sizeof(*cmd));
1088 
1089 	address &= ~(0xfffULL);
1090 
1091 	cmd->data[0]  = devid;
1092 	cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
1093 	cmd->data[0] |= (qdep  & 0xff) << 24;
1094 	cmd->data[1]  = devid;
1095 	cmd->data[1] |= (pasid & 0xff) << 16;
1096 	cmd->data[2]  = lower_32_bits(address);
1097 	cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
1098 	cmd->data[3]  = upper_32_bits(address);
1099 	if (size)
1100 		cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
1101 	CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
1102 }
1103 
build_complete_ppr(struct iommu_cmd * cmd,u16 devid,u32 pasid,int status,int tag,bool gn)1104 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, u32 pasid,
1105 			       int status, int tag, bool gn)
1106 {
1107 	memset(cmd, 0, sizeof(*cmd));
1108 
1109 	cmd->data[0]  = devid;
1110 	if (gn) {
1111 		cmd->data[1]  = pasid;
1112 		cmd->data[2]  = CMD_INV_IOMMU_PAGES_GN_MASK;
1113 	}
1114 	cmd->data[3]  = tag & 0x1ff;
1115 	cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
1116 
1117 	CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
1118 }
1119 
build_inv_all(struct iommu_cmd * cmd)1120 static void build_inv_all(struct iommu_cmd *cmd)
1121 {
1122 	memset(cmd, 0, sizeof(*cmd));
1123 	CMD_SET_TYPE(cmd, CMD_INV_ALL);
1124 }
1125 
build_inv_irt(struct iommu_cmd * cmd,u16 devid)1126 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
1127 {
1128 	memset(cmd, 0, sizeof(*cmd));
1129 	cmd->data[0] = devid;
1130 	CMD_SET_TYPE(cmd, CMD_INV_IRT);
1131 }
1132 
1133 /*
1134  * Writes the command to the IOMMUs command buffer and informs the
1135  * hardware about the new command.
1136  */
__iommu_queue_command_sync(struct amd_iommu * iommu,struct iommu_cmd * cmd,bool sync)1137 static int __iommu_queue_command_sync(struct amd_iommu *iommu,
1138 				      struct iommu_cmd *cmd,
1139 				      bool sync)
1140 {
1141 	unsigned int count = 0;
1142 	u32 left, next_tail;
1143 
1144 	next_tail = (iommu->cmd_buf_tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
1145 again:
1146 	left      = (iommu->cmd_buf_head - next_tail) % CMD_BUFFER_SIZE;
1147 
1148 	if (left <= 0x20) {
1149 		/* Skip udelay() the first time around */
1150 		if (count++) {
1151 			if (count == LOOP_TIMEOUT) {
1152 				pr_err("Command buffer timeout\n");
1153 				return -EIO;
1154 			}
1155 
1156 			udelay(1);
1157 		}
1158 
1159 		/* Update head and recheck remaining space */
1160 		iommu->cmd_buf_head = readl(iommu->mmio_base +
1161 					    MMIO_CMD_HEAD_OFFSET);
1162 
1163 		goto again;
1164 	}
1165 
1166 	copy_cmd_to_buffer(iommu, cmd);
1167 
1168 	/* Do we need to make sure all commands are processed? */
1169 	iommu->need_sync = sync;
1170 
1171 	return 0;
1172 }
1173 
iommu_queue_command_sync(struct amd_iommu * iommu,struct iommu_cmd * cmd,bool sync)1174 static int iommu_queue_command_sync(struct amd_iommu *iommu,
1175 				    struct iommu_cmd *cmd,
1176 				    bool sync)
1177 {
1178 	unsigned long flags;
1179 	int ret;
1180 
1181 	raw_spin_lock_irqsave(&iommu->lock, flags);
1182 	ret = __iommu_queue_command_sync(iommu, cmd, sync);
1183 	raw_spin_unlock_irqrestore(&iommu->lock, flags);
1184 
1185 	return ret;
1186 }
1187 
iommu_queue_command(struct amd_iommu * iommu,struct iommu_cmd * cmd)1188 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
1189 {
1190 	return iommu_queue_command_sync(iommu, cmd, true);
1191 }
1192 
1193 /*
1194  * This function queues a completion wait command into the command
1195  * buffer of an IOMMU
1196  */
iommu_completion_wait(struct amd_iommu * iommu)1197 static int iommu_completion_wait(struct amd_iommu *iommu)
1198 {
1199 	struct iommu_cmd cmd;
1200 	unsigned long flags;
1201 	int ret;
1202 	u64 data;
1203 
1204 	if (!iommu->need_sync)
1205 		return 0;
1206 
1207 	data = atomic64_add_return(1, &iommu->cmd_sem_val);
1208 	build_completion_wait(&cmd, iommu, data);
1209 
1210 	raw_spin_lock_irqsave(&iommu->lock, flags);
1211 
1212 	ret = __iommu_queue_command_sync(iommu, &cmd, false);
1213 	if (ret)
1214 		goto out_unlock;
1215 
1216 	ret = wait_on_sem(iommu, data);
1217 
1218 out_unlock:
1219 	raw_spin_unlock_irqrestore(&iommu->lock, flags);
1220 
1221 	return ret;
1222 }
1223 
iommu_flush_dte(struct amd_iommu * iommu,u16 devid)1224 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1225 {
1226 	struct iommu_cmd cmd;
1227 
1228 	build_inv_dte(&cmd, devid);
1229 
1230 	return iommu_queue_command(iommu, &cmd);
1231 }
1232 
amd_iommu_flush_dte_all(struct amd_iommu * iommu)1233 static void amd_iommu_flush_dte_all(struct amd_iommu *iommu)
1234 {
1235 	u32 devid;
1236 	u16 last_bdf = iommu->pci_seg->last_bdf;
1237 
1238 	for (devid = 0; devid <= last_bdf; ++devid)
1239 		iommu_flush_dte(iommu, devid);
1240 
1241 	iommu_completion_wait(iommu);
1242 }
1243 
1244 /*
1245  * This function uses heavy locking and may disable irqs for some time. But
1246  * this is no issue because it is only called during resume.
1247  */
amd_iommu_flush_tlb_all(struct amd_iommu * iommu)1248 static void amd_iommu_flush_tlb_all(struct amd_iommu *iommu)
1249 {
1250 	u32 dom_id;
1251 	u16 last_bdf = iommu->pci_seg->last_bdf;
1252 
1253 	for (dom_id = 0; dom_id <= last_bdf; ++dom_id) {
1254 		struct iommu_cmd cmd;
1255 		build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1256 				      dom_id, 1);
1257 		iommu_queue_command(iommu, &cmd);
1258 	}
1259 
1260 	iommu_completion_wait(iommu);
1261 }
1262 
amd_iommu_flush_tlb_domid(struct amd_iommu * iommu,u32 dom_id)1263 static void amd_iommu_flush_tlb_domid(struct amd_iommu *iommu, u32 dom_id)
1264 {
1265 	struct iommu_cmd cmd;
1266 
1267 	build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1268 			      dom_id, 1);
1269 	iommu_queue_command(iommu, &cmd);
1270 
1271 	iommu_completion_wait(iommu);
1272 }
1273 
amd_iommu_flush_all(struct amd_iommu * iommu)1274 static void amd_iommu_flush_all(struct amd_iommu *iommu)
1275 {
1276 	struct iommu_cmd cmd;
1277 
1278 	build_inv_all(&cmd);
1279 
1280 	iommu_queue_command(iommu, &cmd);
1281 	iommu_completion_wait(iommu);
1282 }
1283 
iommu_flush_irt(struct amd_iommu * iommu,u16 devid)1284 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1285 {
1286 	struct iommu_cmd cmd;
1287 
1288 	build_inv_irt(&cmd, devid);
1289 
1290 	iommu_queue_command(iommu, &cmd);
1291 }
1292 
amd_iommu_flush_irt_all(struct amd_iommu * iommu)1293 static void amd_iommu_flush_irt_all(struct amd_iommu *iommu)
1294 {
1295 	u32 devid;
1296 	u16 last_bdf = iommu->pci_seg->last_bdf;
1297 
1298 	if (iommu->irtcachedis_enabled)
1299 		return;
1300 
1301 	for (devid = 0; devid <= last_bdf; devid++)
1302 		iommu_flush_irt(iommu, devid);
1303 
1304 	iommu_completion_wait(iommu);
1305 }
1306 
iommu_flush_all_caches(struct amd_iommu * iommu)1307 void iommu_flush_all_caches(struct amd_iommu *iommu)
1308 {
1309 	if (iommu_feature(iommu, FEATURE_IA)) {
1310 		amd_iommu_flush_all(iommu);
1311 	} else {
1312 		amd_iommu_flush_dte_all(iommu);
1313 		amd_iommu_flush_irt_all(iommu);
1314 		amd_iommu_flush_tlb_all(iommu);
1315 	}
1316 }
1317 
1318 /*
1319  * Command send function for flushing on-device TLB
1320  */
device_flush_iotlb(struct iommu_dev_data * dev_data,u64 address,size_t size)1321 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1322 			      u64 address, size_t size)
1323 {
1324 	struct amd_iommu *iommu;
1325 	struct iommu_cmd cmd;
1326 	int qdep;
1327 
1328 	qdep     = dev_data->ats.qdep;
1329 	iommu    = rlookup_amd_iommu(dev_data->dev);
1330 	if (!iommu)
1331 		return -EINVAL;
1332 
1333 	build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1334 
1335 	return iommu_queue_command(iommu, &cmd);
1336 }
1337 
device_flush_dte_alias(struct pci_dev * pdev,u16 alias,void * data)1338 static int device_flush_dte_alias(struct pci_dev *pdev, u16 alias, void *data)
1339 {
1340 	struct amd_iommu *iommu = data;
1341 
1342 	return iommu_flush_dte(iommu, alias);
1343 }
1344 
1345 /*
1346  * Command send function for invalidating a device table entry
1347  */
device_flush_dte(struct iommu_dev_data * dev_data)1348 static int device_flush_dte(struct iommu_dev_data *dev_data)
1349 {
1350 	struct amd_iommu *iommu;
1351 	struct pci_dev *pdev = NULL;
1352 	struct amd_iommu_pci_seg *pci_seg;
1353 	u16 alias;
1354 	int ret;
1355 
1356 	iommu = rlookup_amd_iommu(dev_data->dev);
1357 	if (!iommu)
1358 		return -EINVAL;
1359 
1360 	if (dev_is_pci(dev_data->dev))
1361 		pdev = to_pci_dev(dev_data->dev);
1362 
1363 	if (pdev)
1364 		ret = pci_for_each_dma_alias(pdev,
1365 					     device_flush_dte_alias, iommu);
1366 	else
1367 		ret = iommu_flush_dte(iommu, dev_data->devid);
1368 	if (ret)
1369 		return ret;
1370 
1371 	pci_seg = iommu->pci_seg;
1372 	alias = pci_seg->alias_table[dev_data->devid];
1373 	if (alias != dev_data->devid) {
1374 		ret = iommu_flush_dte(iommu, alias);
1375 		if (ret)
1376 			return ret;
1377 	}
1378 
1379 	if (dev_data->ats.enabled)
1380 		ret = device_flush_iotlb(dev_data, 0, ~0UL);
1381 
1382 	return ret;
1383 }
1384 
1385 /*
1386  * TLB invalidation function which is called from the mapping functions.
1387  * It invalidates a single PTE if the range to flush is within a single
1388  * page. Otherwise it flushes the whole TLB of the IOMMU.
1389  */
__domain_flush_pages(struct protection_domain * domain,u64 address,size_t size,int pde)1390 static void __domain_flush_pages(struct protection_domain *domain,
1391 				 u64 address, size_t size, int pde)
1392 {
1393 	struct iommu_dev_data *dev_data;
1394 	struct iommu_cmd cmd;
1395 	int ret = 0, i;
1396 
1397 	build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1398 
1399 	for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1400 		if (!domain->dev_iommu[i])
1401 			continue;
1402 
1403 		/*
1404 		 * Devices of this domain are behind this IOMMU
1405 		 * We need a TLB flush
1406 		 */
1407 		ret |= iommu_queue_command(amd_iommus[i], &cmd);
1408 	}
1409 
1410 	list_for_each_entry(dev_data, &domain->dev_list, list) {
1411 
1412 		if (!dev_data->ats.enabled)
1413 			continue;
1414 
1415 		ret |= device_flush_iotlb(dev_data, address, size);
1416 	}
1417 
1418 	WARN_ON(ret);
1419 }
1420 
domain_flush_pages(struct protection_domain * domain,u64 address,size_t size,int pde)1421 static void domain_flush_pages(struct protection_domain *domain,
1422 			       u64 address, size_t size, int pde)
1423 {
1424 	if (likely(!amd_iommu_np_cache)) {
1425 		__domain_flush_pages(domain, address, size, pde);
1426 		return;
1427 	}
1428 
1429 	/*
1430 	 * When NpCache is on, we infer that we run in a VM and use a vIOMMU.
1431 	 * In such setups it is best to avoid flushes of ranges which are not
1432 	 * naturally aligned, since it would lead to flushes of unmodified
1433 	 * PTEs. Such flushes would require the hypervisor to do more work than
1434 	 * necessary. Therefore, perform repeated flushes of aligned ranges
1435 	 * until you cover the range. Each iteration flushes the smaller
1436 	 * between the natural alignment of the address that we flush and the
1437 	 * greatest naturally aligned region that fits in the range.
1438 	 */
1439 	while (size != 0) {
1440 		int addr_alignment = __ffs(address);
1441 		int size_alignment = __fls(size);
1442 		int min_alignment;
1443 		size_t flush_size;
1444 
1445 		/*
1446 		 * size is always non-zero, but address might be zero, causing
1447 		 * addr_alignment to be negative. As the casting of the
1448 		 * argument in __ffs(address) to long might trim the high bits
1449 		 * of the address on x86-32, cast to long when doing the check.
1450 		 */
1451 		if (likely((unsigned long)address != 0))
1452 			min_alignment = min(addr_alignment, size_alignment);
1453 		else
1454 			min_alignment = size_alignment;
1455 
1456 		flush_size = 1ul << min_alignment;
1457 
1458 		__domain_flush_pages(domain, address, flush_size, pde);
1459 		address += flush_size;
1460 		size -= flush_size;
1461 	}
1462 }
1463 
1464 /* Flush the whole IO/TLB for a given protection domain - including PDE */
amd_iommu_domain_flush_tlb_pde(struct protection_domain * domain)1465 void amd_iommu_domain_flush_tlb_pde(struct protection_domain *domain)
1466 {
1467 	domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1468 }
1469 
amd_iommu_domain_flush_complete(struct protection_domain * domain)1470 void amd_iommu_domain_flush_complete(struct protection_domain *domain)
1471 {
1472 	int i;
1473 
1474 	for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1475 		if (domain && !domain->dev_iommu[i])
1476 			continue;
1477 
1478 		/*
1479 		 * Devices of this domain are behind this IOMMU
1480 		 * We need to wait for completion of all commands.
1481 		 */
1482 		iommu_completion_wait(amd_iommus[i]);
1483 	}
1484 }
1485 
1486 /* Flush the not present cache if it exists */
domain_flush_np_cache(struct protection_domain * domain,dma_addr_t iova,size_t size)1487 static void domain_flush_np_cache(struct protection_domain *domain,
1488 		dma_addr_t iova, size_t size)
1489 {
1490 	if (unlikely(amd_iommu_np_cache)) {
1491 		unsigned long flags;
1492 
1493 		spin_lock_irqsave(&domain->lock, flags);
1494 		domain_flush_pages(domain, iova, size, 1);
1495 		amd_iommu_domain_flush_complete(domain);
1496 		spin_unlock_irqrestore(&domain->lock, flags);
1497 	}
1498 }
1499 
1500 
1501 /*
1502  * This function flushes the DTEs for all devices in domain
1503  */
domain_flush_devices(struct protection_domain * domain)1504 static void domain_flush_devices(struct protection_domain *domain)
1505 {
1506 	struct iommu_dev_data *dev_data;
1507 
1508 	list_for_each_entry(dev_data, &domain->dev_list, list)
1509 		device_flush_dte(dev_data);
1510 }
1511 
1512 /****************************************************************************
1513  *
1514  * The next functions belong to the domain allocation. A domain is
1515  * allocated for every IOMMU as the default domain. If device isolation
1516  * is enabled, every device get its own domain. The most important thing
1517  * about domains is the page table mapping the DMA address space they
1518  * contain.
1519  *
1520  ****************************************************************************/
1521 
domain_id_alloc(void)1522 static u16 domain_id_alloc(void)
1523 {
1524 	int id;
1525 
1526 	spin_lock(&pd_bitmap_lock);
1527 	id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1528 	BUG_ON(id == 0);
1529 	if (id > 0 && id < MAX_DOMAIN_ID)
1530 		__set_bit(id, amd_iommu_pd_alloc_bitmap);
1531 	else
1532 		id = 0;
1533 	spin_unlock(&pd_bitmap_lock);
1534 
1535 	return id;
1536 }
1537 
domain_id_free(int id)1538 static void domain_id_free(int id)
1539 {
1540 	spin_lock(&pd_bitmap_lock);
1541 	if (id > 0 && id < MAX_DOMAIN_ID)
1542 		__clear_bit(id, amd_iommu_pd_alloc_bitmap);
1543 	spin_unlock(&pd_bitmap_lock);
1544 }
1545 
free_gcr3_tbl_level1(u64 * tbl)1546 static void free_gcr3_tbl_level1(u64 *tbl)
1547 {
1548 	u64 *ptr;
1549 	int i;
1550 
1551 	for (i = 0; i < 512; ++i) {
1552 		if (!(tbl[i] & GCR3_VALID))
1553 			continue;
1554 
1555 		ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1556 
1557 		free_page((unsigned long)ptr);
1558 	}
1559 }
1560 
free_gcr3_tbl_level2(u64 * tbl)1561 static void free_gcr3_tbl_level2(u64 *tbl)
1562 {
1563 	u64 *ptr;
1564 	int i;
1565 
1566 	for (i = 0; i < 512; ++i) {
1567 		if (!(tbl[i] & GCR3_VALID))
1568 			continue;
1569 
1570 		ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1571 
1572 		free_gcr3_tbl_level1(ptr);
1573 	}
1574 }
1575 
free_gcr3_table(struct protection_domain * domain)1576 static void free_gcr3_table(struct protection_domain *domain)
1577 {
1578 	if (domain->glx == 2)
1579 		free_gcr3_tbl_level2(domain->gcr3_tbl);
1580 	else if (domain->glx == 1)
1581 		free_gcr3_tbl_level1(domain->gcr3_tbl);
1582 	else
1583 		BUG_ON(domain->glx != 0);
1584 
1585 	free_page((unsigned long)domain->gcr3_tbl);
1586 }
1587 
set_dte_entry(struct amd_iommu * iommu,u16 devid,struct protection_domain * domain,bool ats,bool ppr)1588 static void set_dte_entry(struct amd_iommu *iommu, u16 devid,
1589 			  struct protection_domain *domain, bool ats, bool ppr)
1590 {
1591 	u64 pte_root = 0;
1592 	u64 flags = 0;
1593 	u32 old_domid;
1594 	struct dev_table_entry *dev_table = get_dev_table(iommu);
1595 
1596 	if (domain->iop.mode != PAGE_MODE_NONE)
1597 		pte_root = iommu_virt_to_phys(domain->iop.root);
1598 
1599 	pte_root |= (domain->iop.mode & DEV_ENTRY_MODE_MASK)
1600 		    << DEV_ENTRY_MODE_SHIFT;
1601 
1602 	pte_root |= DTE_FLAG_IR | DTE_FLAG_IW | DTE_FLAG_V;
1603 
1604 	/*
1605 	 * When SNP is enabled, Only set TV bit when IOMMU
1606 	 * page translation is in use.
1607 	 */
1608 	if (!amd_iommu_snp_en || (domain->id != 0))
1609 		pte_root |= DTE_FLAG_TV;
1610 
1611 	flags = dev_table[devid].data[1];
1612 
1613 	if (ats)
1614 		flags |= DTE_FLAG_IOTLB;
1615 
1616 	if (ppr) {
1617 		if (iommu_feature(iommu, FEATURE_EPHSUP))
1618 			pte_root |= 1ULL << DEV_ENTRY_PPR;
1619 	}
1620 
1621 	if (domain->flags & PD_IOMMUV2_MASK) {
1622 		u64 gcr3 = iommu_virt_to_phys(domain->gcr3_tbl);
1623 		u64 glx  = domain->glx;
1624 		u64 tmp;
1625 
1626 		pte_root |= DTE_FLAG_GV;
1627 		pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1628 
1629 		/* First mask out possible old values for GCR3 table */
1630 		tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1631 		flags    &= ~tmp;
1632 
1633 		tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1634 		flags    &= ~tmp;
1635 
1636 		/* Encode GCR3 table into DTE */
1637 		tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1638 		pte_root |= tmp;
1639 
1640 		tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1641 		flags    |= tmp;
1642 
1643 		tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1644 		flags    |= tmp;
1645 
1646 		if (amd_iommu_gpt_level == PAGE_MODE_5_LEVEL) {
1647 			dev_table[devid].data[2] |=
1648 				((u64)GUEST_PGTABLE_5_LEVEL << DTE_GPT_LEVEL_SHIFT);
1649 		}
1650 
1651 		if (domain->flags & PD_GIOV_MASK)
1652 			pte_root |= DTE_FLAG_GIOV;
1653 	}
1654 
1655 	flags &= ~DEV_DOMID_MASK;
1656 	flags |= domain->id;
1657 
1658 	old_domid = dev_table[devid].data[1] & DEV_DOMID_MASK;
1659 	dev_table[devid].data[1]  = flags;
1660 	dev_table[devid].data[0]  = pte_root;
1661 
1662 	/*
1663 	 * A kdump kernel might be replacing a domain ID that was copied from
1664 	 * the previous kernel--if so, it needs to flush the translation cache
1665 	 * entries for the old domain ID that is being overwritten
1666 	 */
1667 	if (old_domid) {
1668 		amd_iommu_flush_tlb_domid(iommu, old_domid);
1669 	}
1670 }
1671 
clear_dte_entry(struct amd_iommu * iommu,u16 devid)1672 static void clear_dte_entry(struct amd_iommu *iommu, u16 devid)
1673 {
1674 	struct dev_table_entry *dev_table = get_dev_table(iommu);
1675 
1676 	/* remove entry from the device table seen by the hardware */
1677 	dev_table[devid].data[0]  = DTE_FLAG_V;
1678 
1679 	if (!amd_iommu_snp_en)
1680 		dev_table[devid].data[0] |= DTE_FLAG_TV;
1681 
1682 	dev_table[devid].data[1] &= DTE_FLAG_MASK;
1683 
1684 	amd_iommu_apply_erratum_63(iommu, devid);
1685 }
1686 
do_attach(struct iommu_dev_data * dev_data,struct protection_domain * domain)1687 static void do_attach(struct iommu_dev_data *dev_data,
1688 		      struct protection_domain *domain)
1689 {
1690 	struct amd_iommu *iommu;
1691 	bool ats;
1692 
1693 	iommu = rlookup_amd_iommu(dev_data->dev);
1694 	if (!iommu)
1695 		return;
1696 	ats   = dev_data->ats.enabled;
1697 
1698 	/* Update data structures */
1699 	dev_data->domain = domain;
1700 	list_add(&dev_data->list, &domain->dev_list);
1701 
1702 	/* Update NUMA Node ID */
1703 	if (domain->nid == NUMA_NO_NODE)
1704 		domain->nid = dev_to_node(dev_data->dev);
1705 
1706 	/* Do reference counting */
1707 	domain->dev_iommu[iommu->index] += 1;
1708 	domain->dev_cnt                 += 1;
1709 
1710 	/* Update device table */
1711 	set_dte_entry(iommu, dev_data->devid, domain,
1712 		      ats, dev_data->iommu_v2);
1713 	clone_aliases(iommu, dev_data->dev);
1714 
1715 	device_flush_dte(dev_data);
1716 }
1717 
do_detach(struct iommu_dev_data * dev_data)1718 static void do_detach(struct iommu_dev_data *dev_data)
1719 {
1720 	struct protection_domain *domain = dev_data->domain;
1721 	struct amd_iommu *iommu;
1722 
1723 	iommu = rlookup_amd_iommu(dev_data->dev);
1724 	if (!iommu)
1725 		return;
1726 
1727 	/* Update data structures */
1728 	dev_data->domain = NULL;
1729 	list_del(&dev_data->list);
1730 	clear_dte_entry(iommu, dev_data->devid);
1731 	clone_aliases(iommu, dev_data->dev);
1732 
1733 	/* Flush the DTE entry */
1734 	device_flush_dte(dev_data);
1735 
1736 	/* Flush IOTLB */
1737 	amd_iommu_domain_flush_tlb_pde(domain);
1738 
1739 	/* Wait for the flushes to finish */
1740 	amd_iommu_domain_flush_complete(domain);
1741 
1742 	/* decrease reference counters - needs to happen after the flushes */
1743 	domain->dev_iommu[iommu->index] -= 1;
1744 	domain->dev_cnt                 -= 1;
1745 }
1746 
pdev_iommuv2_disable(struct pci_dev * pdev)1747 static void pdev_iommuv2_disable(struct pci_dev *pdev)
1748 {
1749 	pci_disable_ats(pdev);
1750 	pci_disable_pri(pdev);
1751 	pci_disable_pasid(pdev);
1752 }
1753 
pdev_pri_ats_enable(struct pci_dev * pdev)1754 static int pdev_pri_ats_enable(struct pci_dev *pdev)
1755 {
1756 	int ret;
1757 
1758 	/* Only allow access to user-accessible pages */
1759 	ret = pci_enable_pasid(pdev, 0);
1760 	if (ret)
1761 		return ret;
1762 
1763 	/* First reset the PRI state of the device */
1764 	ret = pci_reset_pri(pdev);
1765 	if (ret)
1766 		goto out_err_pasid;
1767 
1768 	/* Enable PRI */
1769 	/* FIXME: Hardcode number of outstanding requests for now */
1770 	ret = pci_enable_pri(pdev, 32);
1771 	if (ret)
1772 		goto out_err_pasid;
1773 
1774 	ret = pci_enable_ats(pdev, PAGE_SHIFT);
1775 	if (ret)
1776 		goto out_err_pri;
1777 
1778 	return 0;
1779 
1780 out_err_pri:
1781 	pci_disable_pri(pdev);
1782 
1783 out_err_pasid:
1784 	pci_disable_pasid(pdev);
1785 
1786 	return ret;
1787 }
1788 
1789 /*
1790  * If a device is not yet associated with a domain, this function makes the
1791  * device visible in the domain
1792  */
attach_device(struct device * dev,struct protection_domain * domain)1793 static int attach_device(struct device *dev,
1794 			 struct protection_domain *domain)
1795 {
1796 	struct iommu_dev_data *dev_data;
1797 	struct pci_dev *pdev;
1798 	unsigned long flags;
1799 	int ret;
1800 
1801 	spin_lock_irqsave(&domain->lock, flags);
1802 
1803 	dev_data = dev_iommu_priv_get(dev);
1804 
1805 	spin_lock(&dev_data->lock);
1806 
1807 	ret = -EBUSY;
1808 	if (dev_data->domain != NULL)
1809 		goto out;
1810 
1811 	if (!dev_is_pci(dev))
1812 		goto skip_ats_check;
1813 
1814 	pdev = to_pci_dev(dev);
1815 	if (domain->flags & PD_IOMMUV2_MASK) {
1816 		struct iommu_domain *def_domain = iommu_get_dma_domain(dev);
1817 
1818 		ret = -EINVAL;
1819 
1820 		/*
1821 		 * In case of using AMD_IOMMU_V1 page table mode and the device
1822 		 * is enabling for PPR/ATS support (using v2 table),
1823 		 * we need to make sure that the domain type is identity map.
1824 		 */
1825 		if ((amd_iommu_pgtable == AMD_IOMMU_V1) &&
1826 		    def_domain->type != IOMMU_DOMAIN_IDENTITY) {
1827 			goto out;
1828 		}
1829 
1830 		if (dev_data->iommu_v2) {
1831 			if (pdev_pri_ats_enable(pdev) != 0)
1832 				goto out;
1833 
1834 			dev_data->ats.enabled = true;
1835 			dev_data->ats.qdep    = pci_ats_queue_depth(pdev);
1836 			dev_data->pri_tlp     = pci_prg_resp_pasid_required(pdev);
1837 		}
1838 	} else if (amd_iommu_iotlb_sup &&
1839 		   pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
1840 		dev_data->ats.enabled = true;
1841 		dev_data->ats.qdep    = pci_ats_queue_depth(pdev);
1842 	}
1843 
1844 skip_ats_check:
1845 	ret = 0;
1846 
1847 	do_attach(dev_data, domain);
1848 
1849 	/*
1850 	 * We might boot into a crash-kernel here. The crashed kernel
1851 	 * left the caches in the IOMMU dirty. So we have to flush
1852 	 * here to evict all dirty stuff.
1853 	 */
1854 	amd_iommu_domain_flush_tlb_pde(domain);
1855 
1856 	amd_iommu_domain_flush_complete(domain);
1857 
1858 out:
1859 	spin_unlock(&dev_data->lock);
1860 
1861 	spin_unlock_irqrestore(&domain->lock, flags);
1862 
1863 	return ret;
1864 }
1865 
1866 /*
1867  * Removes a device from a protection domain (with devtable_lock held)
1868  */
detach_device(struct device * dev)1869 static void detach_device(struct device *dev)
1870 {
1871 	struct protection_domain *domain;
1872 	struct iommu_dev_data *dev_data;
1873 	unsigned long flags;
1874 
1875 	dev_data = dev_iommu_priv_get(dev);
1876 	domain   = dev_data->domain;
1877 
1878 	spin_lock_irqsave(&domain->lock, flags);
1879 
1880 	spin_lock(&dev_data->lock);
1881 
1882 	/*
1883 	 * First check if the device is still attached. It might already
1884 	 * be detached from its domain because the generic
1885 	 * iommu_detach_group code detached it and we try again here in
1886 	 * our alias handling.
1887 	 */
1888 	if (WARN_ON(!dev_data->domain))
1889 		goto out;
1890 
1891 	do_detach(dev_data);
1892 
1893 	if (!dev_is_pci(dev))
1894 		goto out;
1895 
1896 	if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2)
1897 		pdev_iommuv2_disable(to_pci_dev(dev));
1898 	else if (dev_data->ats.enabled)
1899 		pci_disable_ats(to_pci_dev(dev));
1900 
1901 	dev_data->ats.enabled = false;
1902 
1903 out:
1904 	spin_unlock(&dev_data->lock);
1905 
1906 	spin_unlock_irqrestore(&domain->lock, flags);
1907 }
1908 
amd_iommu_probe_device(struct device * dev)1909 static struct iommu_device *amd_iommu_probe_device(struct device *dev)
1910 {
1911 	struct iommu_device *iommu_dev;
1912 	struct amd_iommu *iommu;
1913 	int ret;
1914 
1915 	if (!check_device(dev))
1916 		return ERR_PTR(-ENODEV);
1917 
1918 	iommu = rlookup_amd_iommu(dev);
1919 	if (!iommu)
1920 		return ERR_PTR(-ENODEV);
1921 
1922 	/* Not registered yet? */
1923 	if (!iommu->iommu.ops)
1924 		return ERR_PTR(-ENODEV);
1925 
1926 	if (dev_iommu_priv_get(dev))
1927 		return &iommu->iommu;
1928 
1929 	ret = iommu_init_device(iommu, dev);
1930 	if (ret) {
1931 		if (ret != -ENOTSUPP)
1932 			dev_err(dev, "Failed to initialize - trying to proceed anyway\n");
1933 		iommu_dev = ERR_PTR(ret);
1934 		iommu_ignore_device(iommu, dev);
1935 	} else {
1936 		amd_iommu_set_pci_msi_domain(dev, iommu);
1937 		iommu_dev = &iommu->iommu;
1938 	}
1939 
1940 	iommu_completion_wait(iommu);
1941 
1942 	return iommu_dev;
1943 }
1944 
amd_iommu_probe_finalize(struct device * dev)1945 static void amd_iommu_probe_finalize(struct device *dev)
1946 {
1947 	/* Domains are initialized for this device - have a look what we ended up with */
1948 	set_dma_ops(dev, NULL);
1949 	iommu_setup_dma_ops(dev, 0, U64_MAX);
1950 }
1951 
amd_iommu_release_device(struct device * dev)1952 static void amd_iommu_release_device(struct device *dev)
1953 {
1954 	struct amd_iommu *iommu;
1955 
1956 	if (!check_device(dev))
1957 		return;
1958 
1959 	iommu = rlookup_amd_iommu(dev);
1960 	if (!iommu)
1961 		return;
1962 
1963 	amd_iommu_uninit_device(dev);
1964 	iommu_completion_wait(iommu);
1965 }
1966 
amd_iommu_device_group(struct device * dev)1967 static struct iommu_group *amd_iommu_device_group(struct device *dev)
1968 {
1969 	if (dev_is_pci(dev))
1970 		return pci_device_group(dev);
1971 
1972 	return acpihid_device_group(dev);
1973 }
1974 
1975 /*****************************************************************************
1976  *
1977  * The next functions belong to the dma_ops mapping/unmapping code.
1978  *
1979  *****************************************************************************/
1980 
update_device_table(struct protection_domain * domain)1981 static void update_device_table(struct protection_domain *domain)
1982 {
1983 	struct iommu_dev_data *dev_data;
1984 
1985 	list_for_each_entry(dev_data, &domain->dev_list, list) {
1986 		struct amd_iommu *iommu = rlookup_amd_iommu(dev_data->dev);
1987 
1988 		if (!iommu)
1989 			continue;
1990 		set_dte_entry(iommu, dev_data->devid, domain,
1991 			      dev_data->ats.enabled, dev_data->iommu_v2);
1992 		clone_aliases(iommu, dev_data->dev);
1993 	}
1994 }
1995 
amd_iommu_update_and_flush_device_table(struct protection_domain * domain)1996 void amd_iommu_update_and_flush_device_table(struct protection_domain *domain)
1997 {
1998 	update_device_table(domain);
1999 	domain_flush_devices(domain);
2000 }
2001 
amd_iommu_domain_update(struct protection_domain * domain)2002 void amd_iommu_domain_update(struct protection_domain *domain)
2003 {
2004 	/* Update device table */
2005 	amd_iommu_update_and_flush_device_table(domain);
2006 
2007 	/* Flush domain TLB(s) and wait for completion */
2008 	amd_iommu_domain_flush_tlb_pde(domain);
2009 	amd_iommu_domain_flush_complete(domain);
2010 }
2011 
2012 /*****************************************************************************
2013  *
2014  * The following functions belong to the exported interface of AMD IOMMU
2015  *
2016  * This interface allows access to lower level functions of the IOMMU
2017  * like protection domain handling and assignement of devices to domains
2018  * which is not possible with the dma_ops interface.
2019  *
2020  *****************************************************************************/
2021 
cleanup_domain(struct protection_domain * domain)2022 static void cleanup_domain(struct protection_domain *domain)
2023 {
2024 	struct iommu_dev_data *entry;
2025 	unsigned long flags;
2026 
2027 	spin_lock_irqsave(&domain->lock, flags);
2028 
2029 	while (!list_empty(&domain->dev_list)) {
2030 		entry = list_first_entry(&domain->dev_list,
2031 					 struct iommu_dev_data, list);
2032 		BUG_ON(!entry->domain);
2033 		do_detach(entry);
2034 	}
2035 
2036 	spin_unlock_irqrestore(&domain->lock, flags);
2037 }
2038 
protection_domain_free(struct protection_domain * domain)2039 static void protection_domain_free(struct protection_domain *domain)
2040 {
2041 	if (!domain)
2042 		return;
2043 
2044 	if (domain->iop.pgtbl_cfg.tlb)
2045 		free_io_pgtable_ops(&domain->iop.iop.ops);
2046 
2047 	if (domain->id)
2048 		domain_id_free(domain->id);
2049 
2050 	kfree(domain);
2051 }
2052 
protection_domain_init_v1(struct protection_domain * domain,int mode)2053 static int protection_domain_init_v1(struct protection_domain *domain, int mode)
2054 {
2055 	u64 *pt_root = NULL;
2056 
2057 	BUG_ON(mode < PAGE_MODE_NONE || mode > PAGE_MODE_6_LEVEL);
2058 
2059 	spin_lock_init(&domain->lock);
2060 	domain->id = domain_id_alloc();
2061 	if (!domain->id)
2062 		return -ENOMEM;
2063 	INIT_LIST_HEAD(&domain->dev_list);
2064 
2065 	if (mode != PAGE_MODE_NONE) {
2066 		pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2067 		if (!pt_root) {
2068 			domain_id_free(domain->id);
2069 			return -ENOMEM;
2070 		}
2071 	}
2072 
2073 	amd_iommu_domain_set_pgtable(domain, pt_root, mode);
2074 
2075 	return 0;
2076 }
2077 
protection_domain_init_v2(struct protection_domain * domain)2078 static int protection_domain_init_v2(struct protection_domain *domain)
2079 {
2080 	spin_lock_init(&domain->lock);
2081 	domain->id = domain_id_alloc();
2082 	if (!domain->id)
2083 		return -ENOMEM;
2084 	INIT_LIST_HEAD(&domain->dev_list);
2085 
2086 	domain->flags |= PD_GIOV_MASK;
2087 
2088 	domain->domain.pgsize_bitmap = AMD_IOMMU_PGSIZES_V2;
2089 
2090 	if (domain_enable_v2(domain, 1)) {
2091 		domain_id_free(domain->id);
2092 		return -ENOMEM;
2093 	}
2094 
2095 	return 0;
2096 }
2097 
protection_domain_alloc(unsigned int type)2098 static struct protection_domain *protection_domain_alloc(unsigned int type)
2099 {
2100 	struct io_pgtable_ops *pgtbl_ops;
2101 	struct protection_domain *domain;
2102 	int pgtable;
2103 	int mode = DEFAULT_PGTABLE_LEVEL;
2104 	int ret;
2105 
2106 	/*
2107 	 * Force IOMMU v1 page table when iommu=pt and
2108 	 * when allocating domain for pass-through devices.
2109 	 */
2110 	if (type == IOMMU_DOMAIN_IDENTITY) {
2111 		pgtable = AMD_IOMMU_V1;
2112 		mode = PAGE_MODE_NONE;
2113 	} else if (type == IOMMU_DOMAIN_UNMANAGED) {
2114 		pgtable = AMD_IOMMU_V1;
2115 	} else if (type == IOMMU_DOMAIN_DMA || type == IOMMU_DOMAIN_DMA_FQ) {
2116 		pgtable = amd_iommu_pgtable;
2117 	} else {
2118 		return NULL;
2119 	}
2120 
2121 	domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2122 	if (!domain)
2123 		return NULL;
2124 
2125 	switch (pgtable) {
2126 	case AMD_IOMMU_V1:
2127 		ret = protection_domain_init_v1(domain, mode);
2128 		break;
2129 	case AMD_IOMMU_V2:
2130 		ret = protection_domain_init_v2(domain);
2131 		break;
2132 	default:
2133 		ret = -EINVAL;
2134 	}
2135 
2136 	if (ret)
2137 		goto out_err;
2138 
2139 	/* No need to allocate io pgtable ops in passthrough mode */
2140 	if (type == IOMMU_DOMAIN_IDENTITY)
2141 		return domain;
2142 
2143 	domain->nid = NUMA_NO_NODE;
2144 
2145 	pgtbl_ops = alloc_io_pgtable_ops(pgtable, &domain->iop.pgtbl_cfg, domain);
2146 	if (!pgtbl_ops) {
2147 		domain_id_free(domain->id);
2148 		goto out_err;
2149 	}
2150 
2151 	return domain;
2152 out_err:
2153 	kfree(domain);
2154 	return NULL;
2155 }
2156 
dma_max_address(void)2157 static inline u64 dma_max_address(void)
2158 {
2159 	if (amd_iommu_pgtable == AMD_IOMMU_V1)
2160 		return ~0ULL;
2161 
2162 	/* V2 with 4/5 level page table */
2163 	return ((1ULL << PM_LEVEL_SHIFT(amd_iommu_gpt_level)) - 1);
2164 }
2165 
amd_iommu_domain_alloc(unsigned type)2166 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
2167 {
2168 	struct protection_domain *domain;
2169 
2170 	/*
2171 	 * Since DTE[Mode]=0 is prohibited on SNP-enabled system,
2172 	 * default to use IOMMU_DOMAIN_DMA[_FQ].
2173 	 */
2174 	if (amd_iommu_snp_en && (type == IOMMU_DOMAIN_IDENTITY))
2175 		return NULL;
2176 
2177 	domain = protection_domain_alloc(type);
2178 	if (!domain)
2179 		return NULL;
2180 
2181 	domain->domain.geometry.aperture_start = 0;
2182 	domain->domain.geometry.aperture_end   = dma_max_address();
2183 	domain->domain.geometry.force_aperture = true;
2184 
2185 	return &domain->domain;
2186 }
2187 
amd_iommu_domain_free(struct iommu_domain * dom)2188 static void amd_iommu_domain_free(struct iommu_domain *dom)
2189 {
2190 	struct protection_domain *domain;
2191 
2192 	domain = to_pdomain(dom);
2193 
2194 	if (domain->dev_cnt > 0)
2195 		cleanup_domain(domain);
2196 
2197 	BUG_ON(domain->dev_cnt != 0);
2198 
2199 	if (!dom)
2200 		return;
2201 
2202 	if (domain->flags & PD_IOMMUV2_MASK)
2203 		free_gcr3_table(domain);
2204 
2205 	protection_domain_free(domain);
2206 }
2207 
amd_iommu_attach_device(struct iommu_domain * dom,struct device * dev)2208 static int amd_iommu_attach_device(struct iommu_domain *dom,
2209 				   struct device *dev)
2210 {
2211 	struct iommu_dev_data *dev_data = dev_iommu_priv_get(dev);
2212 	struct protection_domain *domain = to_pdomain(dom);
2213 	struct amd_iommu *iommu = rlookup_amd_iommu(dev);
2214 	int ret;
2215 
2216 	/*
2217 	 * Skip attach device to domain if new domain is same as
2218 	 * devices current domain
2219 	 */
2220 	if (dev_data->domain == domain)
2221 		return 0;
2222 
2223 	dev_data->defer_attach = false;
2224 
2225 	if (dev_data->domain)
2226 		detach_device(dev);
2227 
2228 	ret = attach_device(dev, domain);
2229 
2230 #ifdef CONFIG_IRQ_REMAP
2231 	if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
2232 		if (dom->type == IOMMU_DOMAIN_UNMANAGED)
2233 			dev_data->use_vapic = 1;
2234 		else
2235 			dev_data->use_vapic = 0;
2236 	}
2237 #endif
2238 
2239 	iommu_completion_wait(iommu);
2240 
2241 	return ret;
2242 }
2243 
amd_iommu_iotlb_sync_map(struct iommu_domain * dom,unsigned long iova,size_t size)2244 static void amd_iommu_iotlb_sync_map(struct iommu_domain *dom,
2245 				     unsigned long iova, size_t size)
2246 {
2247 	struct protection_domain *domain = to_pdomain(dom);
2248 	struct io_pgtable_ops *ops = &domain->iop.iop.ops;
2249 
2250 	if (ops->map_pages)
2251 		domain_flush_np_cache(domain, iova, size);
2252 }
2253 
amd_iommu_map_pages(struct iommu_domain * dom,unsigned long iova,phys_addr_t paddr,size_t pgsize,size_t pgcount,int iommu_prot,gfp_t gfp,size_t * mapped)2254 static int amd_iommu_map_pages(struct iommu_domain *dom, unsigned long iova,
2255 			       phys_addr_t paddr, size_t pgsize, size_t pgcount,
2256 			       int iommu_prot, gfp_t gfp, size_t *mapped)
2257 {
2258 	struct protection_domain *domain = to_pdomain(dom);
2259 	struct io_pgtable_ops *ops = &domain->iop.iop.ops;
2260 	int prot = 0;
2261 	int ret = -EINVAL;
2262 
2263 	if ((amd_iommu_pgtable == AMD_IOMMU_V1) &&
2264 	    (domain->iop.mode == PAGE_MODE_NONE))
2265 		return -EINVAL;
2266 
2267 	if (iommu_prot & IOMMU_READ)
2268 		prot |= IOMMU_PROT_IR;
2269 	if (iommu_prot & IOMMU_WRITE)
2270 		prot |= IOMMU_PROT_IW;
2271 
2272 	if (ops->map_pages) {
2273 		ret = ops->map_pages(ops, iova, paddr, pgsize,
2274 				     pgcount, prot, gfp, mapped);
2275 	}
2276 
2277 	return ret;
2278 }
2279 
amd_iommu_iotlb_gather_add_page(struct iommu_domain * domain,struct iommu_iotlb_gather * gather,unsigned long iova,size_t size)2280 static void amd_iommu_iotlb_gather_add_page(struct iommu_domain *domain,
2281 					    struct iommu_iotlb_gather *gather,
2282 					    unsigned long iova, size_t size)
2283 {
2284 	/*
2285 	 * AMD's IOMMU can flush as many pages as necessary in a single flush.
2286 	 * Unless we run in a virtual machine, which can be inferred according
2287 	 * to whether "non-present cache" is on, it is probably best to prefer
2288 	 * (potentially) too extensive TLB flushing (i.e., more misses) over
2289 	 * mutliple TLB flushes (i.e., more flushes). For virtual machines the
2290 	 * hypervisor needs to synchronize the host IOMMU PTEs with those of
2291 	 * the guest, and the trade-off is different: unnecessary TLB flushes
2292 	 * should be avoided.
2293 	 */
2294 	if (amd_iommu_np_cache &&
2295 	    iommu_iotlb_gather_is_disjoint(gather, iova, size))
2296 		iommu_iotlb_sync(domain, gather);
2297 
2298 	iommu_iotlb_gather_add_range(gather, iova, size);
2299 }
2300 
amd_iommu_unmap_pages(struct iommu_domain * dom,unsigned long iova,size_t pgsize,size_t pgcount,struct iommu_iotlb_gather * gather)2301 static size_t amd_iommu_unmap_pages(struct iommu_domain *dom, unsigned long iova,
2302 				    size_t pgsize, size_t pgcount,
2303 				    struct iommu_iotlb_gather *gather)
2304 {
2305 	struct protection_domain *domain = to_pdomain(dom);
2306 	struct io_pgtable_ops *ops = &domain->iop.iop.ops;
2307 	size_t r;
2308 
2309 	if ((amd_iommu_pgtable == AMD_IOMMU_V1) &&
2310 	    (domain->iop.mode == PAGE_MODE_NONE))
2311 		return 0;
2312 
2313 	r = (ops->unmap_pages) ? ops->unmap_pages(ops, iova, pgsize, pgcount, NULL) : 0;
2314 
2315 	if (r)
2316 		amd_iommu_iotlb_gather_add_page(dom, gather, iova, r);
2317 
2318 	return r;
2319 }
2320 
amd_iommu_iova_to_phys(struct iommu_domain * dom,dma_addr_t iova)2321 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
2322 					  dma_addr_t iova)
2323 {
2324 	struct protection_domain *domain = to_pdomain(dom);
2325 	struct io_pgtable_ops *ops = &domain->iop.iop.ops;
2326 
2327 	return ops->iova_to_phys(ops, iova);
2328 }
2329 
amd_iommu_capable(struct device * dev,enum iommu_cap cap)2330 static bool amd_iommu_capable(struct device *dev, enum iommu_cap cap)
2331 {
2332 	switch (cap) {
2333 	case IOMMU_CAP_CACHE_COHERENCY:
2334 		return true;
2335 	case IOMMU_CAP_NOEXEC:
2336 		return false;
2337 	case IOMMU_CAP_PRE_BOOT_PROTECTION:
2338 		return amdr_ivrs_remap_support;
2339 	case IOMMU_CAP_ENFORCE_CACHE_COHERENCY:
2340 		return true;
2341 	case IOMMU_CAP_DEFERRED_FLUSH:
2342 		return true;
2343 	default:
2344 		break;
2345 	}
2346 
2347 	return false;
2348 }
2349 
amd_iommu_get_resv_regions(struct device * dev,struct list_head * head)2350 static void amd_iommu_get_resv_regions(struct device *dev,
2351 				       struct list_head *head)
2352 {
2353 	struct iommu_resv_region *region;
2354 	struct unity_map_entry *entry;
2355 	struct amd_iommu *iommu;
2356 	struct amd_iommu_pci_seg *pci_seg;
2357 	int devid, sbdf;
2358 
2359 	sbdf = get_device_sbdf_id(dev);
2360 	if (sbdf < 0)
2361 		return;
2362 
2363 	devid = PCI_SBDF_TO_DEVID(sbdf);
2364 	iommu = rlookup_amd_iommu(dev);
2365 	if (!iommu)
2366 		return;
2367 	pci_seg = iommu->pci_seg;
2368 
2369 	list_for_each_entry(entry, &pci_seg->unity_map, list) {
2370 		int type, prot = 0;
2371 		size_t length;
2372 
2373 		if (devid < entry->devid_start || devid > entry->devid_end)
2374 			continue;
2375 
2376 		type   = IOMMU_RESV_DIRECT;
2377 		length = entry->address_end - entry->address_start;
2378 		if (entry->prot & IOMMU_PROT_IR)
2379 			prot |= IOMMU_READ;
2380 		if (entry->prot & IOMMU_PROT_IW)
2381 			prot |= IOMMU_WRITE;
2382 		if (entry->prot & IOMMU_UNITY_MAP_FLAG_EXCL_RANGE)
2383 			/* Exclusion range */
2384 			type = IOMMU_RESV_RESERVED;
2385 
2386 		region = iommu_alloc_resv_region(entry->address_start,
2387 						 length, prot, type,
2388 						 GFP_KERNEL);
2389 		if (!region) {
2390 			dev_err(dev, "Out of memory allocating dm-regions\n");
2391 			return;
2392 		}
2393 		list_add_tail(&region->list, head);
2394 	}
2395 
2396 	region = iommu_alloc_resv_region(MSI_RANGE_START,
2397 					 MSI_RANGE_END - MSI_RANGE_START + 1,
2398 					 0, IOMMU_RESV_MSI, GFP_KERNEL);
2399 	if (!region)
2400 		return;
2401 	list_add_tail(&region->list, head);
2402 
2403 	region = iommu_alloc_resv_region(HT_RANGE_START,
2404 					 HT_RANGE_END - HT_RANGE_START + 1,
2405 					 0, IOMMU_RESV_RESERVED, GFP_KERNEL);
2406 	if (!region)
2407 		return;
2408 	list_add_tail(&region->list, head);
2409 }
2410 
amd_iommu_is_attach_deferred(struct device * dev)2411 bool amd_iommu_is_attach_deferred(struct device *dev)
2412 {
2413 	struct iommu_dev_data *dev_data = dev_iommu_priv_get(dev);
2414 
2415 	return dev_data->defer_attach;
2416 }
2417 EXPORT_SYMBOL_GPL(amd_iommu_is_attach_deferred);
2418 
amd_iommu_flush_iotlb_all(struct iommu_domain * domain)2419 static void amd_iommu_flush_iotlb_all(struct iommu_domain *domain)
2420 {
2421 	struct protection_domain *dom = to_pdomain(domain);
2422 	unsigned long flags;
2423 
2424 	spin_lock_irqsave(&dom->lock, flags);
2425 	amd_iommu_domain_flush_tlb_pde(dom);
2426 	amd_iommu_domain_flush_complete(dom);
2427 	spin_unlock_irqrestore(&dom->lock, flags);
2428 }
2429 
amd_iommu_iotlb_sync(struct iommu_domain * domain,struct iommu_iotlb_gather * gather)2430 static void amd_iommu_iotlb_sync(struct iommu_domain *domain,
2431 				 struct iommu_iotlb_gather *gather)
2432 {
2433 	struct protection_domain *dom = to_pdomain(domain);
2434 	unsigned long flags;
2435 
2436 	spin_lock_irqsave(&dom->lock, flags);
2437 	domain_flush_pages(dom, gather->start, gather->end - gather->start + 1, 1);
2438 	amd_iommu_domain_flush_complete(dom);
2439 	spin_unlock_irqrestore(&dom->lock, flags);
2440 }
2441 
amd_iommu_def_domain_type(struct device * dev)2442 static int amd_iommu_def_domain_type(struct device *dev)
2443 {
2444 	struct iommu_dev_data *dev_data;
2445 
2446 	dev_data = dev_iommu_priv_get(dev);
2447 	if (!dev_data)
2448 		return 0;
2449 
2450 	/*
2451 	 * Do not identity map IOMMUv2 capable devices when:
2452 	 *  - memory encryption is active, because some of those devices
2453 	 *    (AMD GPUs) don't have the encryption bit in their DMA-mask
2454 	 *    and require remapping.
2455 	 *  - SNP is enabled, because it prohibits DTE[Mode]=0.
2456 	 */
2457 	if (dev_data->iommu_v2 &&
2458 	    !cc_platform_has(CC_ATTR_MEM_ENCRYPT) &&
2459 	    !amd_iommu_snp_en) {
2460 		return IOMMU_DOMAIN_IDENTITY;
2461 	}
2462 
2463 	return 0;
2464 }
2465 
amd_iommu_enforce_cache_coherency(struct iommu_domain * domain)2466 static bool amd_iommu_enforce_cache_coherency(struct iommu_domain *domain)
2467 {
2468 	/* IOMMU_PTE_FC is always set */
2469 	return true;
2470 }
2471 
2472 const struct iommu_ops amd_iommu_ops = {
2473 	.capable = amd_iommu_capable,
2474 	.domain_alloc = amd_iommu_domain_alloc,
2475 	.probe_device = amd_iommu_probe_device,
2476 	.release_device = amd_iommu_release_device,
2477 	.probe_finalize = amd_iommu_probe_finalize,
2478 	.device_group = amd_iommu_device_group,
2479 	.get_resv_regions = amd_iommu_get_resv_regions,
2480 	.is_attach_deferred = amd_iommu_is_attach_deferred,
2481 	.pgsize_bitmap	= AMD_IOMMU_PGSIZES,
2482 	.def_domain_type = amd_iommu_def_domain_type,
2483 	.default_domain_ops = &(const struct iommu_domain_ops) {
2484 		.attach_dev	= amd_iommu_attach_device,
2485 		.map_pages	= amd_iommu_map_pages,
2486 		.unmap_pages	= amd_iommu_unmap_pages,
2487 		.iotlb_sync_map	= amd_iommu_iotlb_sync_map,
2488 		.iova_to_phys	= amd_iommu_iova_to_phys,
2489 		.flush_iotlb_all = amd_iommu_flush_iotlb_all,
2490 		.iotlb_sync	= amd_iommu_iotlb_sync,
2491 		.free		= amd_iommu_domain_free,
2492 		.enforce_cache_coherency = amd_iommu_enforce_cache_coherency,
2493 	}
2494 };
2495 
2496 /*****************************************************************************
2497  *
2498  * The next functions do a basic initialization of IOMMU for pass through
2499  * mode
2500  *
2501  * In passthrough mode the IOMMU is initialized and enabled but not used for
2502  * DMA-API translation.
2503  *
2504  *****************************************************************************/
2505 
2506 /* IOMMUv2 specific functions */
amd_iommu_register_ppr_notifier(struct notifier_block * nb)2507 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
2508 {
2509 	return atomic_notifier_chain_register(&ppr_notifier, nb);
2510 }
2511 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
2512 
amd_iommu_unregister_ppr_notifier(struct notifier_block * nb)2513 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
2514 {
2515 	return atomic_notifier_chain_unregister(&ppr_notifier, nb);
2516 }
2517 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
2518 
amd_iommu_domain_direct_map(struct iommu_domain * dom)2519 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
2520 {
2521 	struct protection_domain *domain = to_pdomain(dom);
2522 	unsigned long flags;
2523 
2524 	spin_lock_irqsave(&domain->lock, flags);
2525 
2526 	if (domain->iop.pgtbl_cfg.tlb)
2527 		free_io_pgtable_ops(&domain->iop.iop.ops);
2528 
2529 	spin_unlock_irqrestore(&domain->lock, flags);
2530 }
2531 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
2532 
2533 /* Note: This function expects iommu_domain->lock to be held prior calling the function. */
domain_enable_v2(struct protection_domain * domain,int pasids)2534 static int domain_enable_v2(struct protection_domain *domain, int pasids)
2535 {
2536 	int levels;
2537 
2538 	/* Number of GCR3 table levels required */
2539 	for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
2540 		levels += 1;
2541 
2542 	if (levels > amd_iommu_max_glx_val)
2543 		return -EINVAL;
2544 
2545 	domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
2546 	if (domain->gcr3_tbl == NULL)
2547 		return -ENOMEM;
2548 
2549 	domain->glx      = levels;
2550 	domain->flags   |= PD_IOMMUV2_MASK;
2551 
2552 	amd_iommu_domain_update(domain);
2553 
2554 	return 0;
2555 }
2556 
amd_iommu_domain_enable_v2(struct iommu_domain * dom,int pasids)2557 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
2558 {
2559 	struct protection_domain *pdom = to_pdomain(dom);
2560 	unsigned long flags;
2561 	int ret;
2562 
2563 	spin_lock_irqsave(&pdom->lock, flags);
2564 
2565 	/*
2566 	 * Save us all sanity checks whether devices already in the
2567 	 * domain support IOMMUv2. Just force that the domain has no
2568 	 * devices attached when it is switched into IOMMUv2 mode.
2569 	 */
2570 	ret = -EBUSY;
2571 	if (pdom->dev_cnt > 0 || pdom->flags & PD_IOMMUV2_MASK)
2572 		goto out;
2573 
2574 	if (!pdom->gcr3_tbl)
2575 		ret = domain_enable_v2(pdom, pasids);
2576 
2577 out:
2578 	spin_unlock_irqrestore(&pdom->lock, flags);
2579 	return ret;
2580 }
2581 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
2582 
__flush_pasid(struct protection_domain * domain,u32 pasid,u64 address,bool size)2583 static int __flush_pasid(struct protection_domain *domain, u32 pasid,
2584 			 u64 address, bool size)
2585 {
2586 	struct iommu_dev_data *dev_data;
2587 	struct iommu_cmd cmd;
2588 	int i, ret;
2589 
2590 	if (!(domain->flags & PD_IOMMUV2_MASK))
2591 		return -EINVAL;
2592 
2593 	build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
2594 
2595 	/*
2596 	 * IOMMU TLB needs to be flushed before Device TLB to
2597 	 * prevent device TLB refill from IOMMU TLB
2598 	 */
2599 	for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
2600 		if (domain->dev_iommu[i] == 0)
2601 			continue;
2602 
2603 		ret = iommu_queue_command(amd_iommus[i], &cmd);
2604 		if (ret != 0)
2605 			goto out;
2606 	}
2607 
2608 	/* Wait until IOMMU TLB flushes are complete */
2609 	amd_iommu_domain_flush_complete(domain);
2610 
2611 	/* Now flush device TLBs */
2612 	list_for_each_entry(dev_data, &domain->dev_list, list) {
2613 		struct amd_iommu *iommu;
2614 		int qdep;
2615 
2616 		/*
2617 		   There might be non-IOMMUv2 capable devices in an IOMMUv2
2618 		 * domain.
2619 		 */
2620 		if (!dev_data->ats.enabled)
2621 			continue;
2622 
2623 		qdep  = dev_data->ats.qdep;
2624 		iommu = rlookup_amd_iommu(dev_data->dev);
2625 		if (!iommu)
2626 			continue;
2627 		build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
2628 				      qdep, address, size);
2629 
2630 		ret = iommu_queue_command(iommu, &cmd);
2631 		if (ret != 0)
2632 			goto out;
2633 	}
2634 
2635 	/* Wait until all device TLBs are flushed */
2636 	amd_iommu_domain_flush_complete(domain);
2637 
2638 	ret = 0;
2639 
2640 out:
2641 
2642 	return ret;
2643 }
2644 
__amd_iommu_flush_page(struct protection_domain * domain,u32 pasid,u64 address)2645 static int __amd_iommu_flush_page(struct protection_domain *domain, u32 pasid,
2646 				  u64 address)
2647 {
2648 	return __flush_pasid(domain, pasid, address, false);
2649 }
2650 
amd_iommu_flush_page(struct iommu_domain * dom,u32 pasid,u64 address)2651 int amd_iommu_flush_page(struct iommu_domain *dom, u32 pasid,
2652 			 u64 address)
2653 {
2654 	struct protection_domain *domain = to_pdomain(dom);
2655 	unsigned long flags;
2656 	int ret;
2657 
2658 	spin_lock_irqsave(&domain->lock, flags);
2659 	ret = __amd_iommu_flush_page(domain, pasid, address);
2660 	spin_unlock_irqrestore(&domain->lock, flags);
2661 
2662 	return ret;
2663 }
2664 EXPORT_SYMBOL(amd_iommu_flush_page);
2665 
__amd_iommu_flush_tlb(struct protection_domain * domain,u32 pasid)2666 static int __amd_iommu_flush_tlb(struct protection_domain *domain, u32 pasid)
2667 {
2668 	return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
2669 			     true);
2670 }
2671 
amd_iommu_flush_tlb(struct iommu_domain * dom,u32 pasid)2672 int amd_iommu_flush_tlb(struct iommu_domain *dom, u32 pasid)
2673 {
2674 	struct protection_domain *domain = to_pdomain(dom);
2675 	unsigned long flags;
2676 	int ret;
2677 
2678 	spin_lock_irqsave(&domain->lock, flags);
2679 	ret = __amd_iommu_flush_tlb(domain, pasid);
2680 	spin_unlock_irqrestore(&domain->lock, flags);
2681 
2682 	return ret;
2683 }
2684 EXPORT_SYMBOL(amd_iommu_flush_tlb);
2685 
__get_gcr3_pte(u64 * root,int level,u32 pasid,bool alloc)2686 static u64 *__get_gcr3_pte(u64 *root, int level, u32 pasid, bool alloc)
2687 {
2688 	int index;
2689 	u64 *pte;
2690 
2691 	while (true) {
2692 
2693 		index = (pasid >> (9 * level)) & 0x1ff;
2694 		pte   = &root[index];
2695 
2696 		if (level == 0)
2697 			break;
2698 
2699 		if (!(*pte & GCR3_VALID)) {
2700 			if (!alloc)
2701 				return NULL;
2702 
2703 			root = (void *)get_zeroed_page(GFP_ATOMIC);
2704 			if (root == NULL)
2705 				return NULL;
2706 
2707 			*pte = iommu_virt_to_phys(root) | GCR3_VALID;
2708 		}
2709 
2710 		root = iommu_phys_to_virt(*pte & PAGE_MASK);
2711 
2712 		level -= 1;
2713 	}
2714 
2715 	return pte;
2716 }
2717 
__set_gcr3(struct protection_domain * domain,u32 pasid,unsigned long cr3)2718 static int __set_gcr3(struct protection_domain *domain, u32 pasid,
2719 		      unsigned long cr3)
2720 {
2721 	u64 *pte;
2722 
2723 	if (domain->iop.mode != PAGE_MODE_NONE)
2724 		return -EINVAL;
2725 
2726 	pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
2727 	if (pte == NULL)
2728 		return -ENOMEM;
2729 
2730 	*pte = (cr3 & PAGE_MASK) | GCR3_VALID;
2731 
2732 	return __amd_iommu_flush_tlb(domain, pasid);
2733 }
2734 
__clear_gcr3(struct protection_domain * domain,u32 pasid)2735 static int __clear_gcr3(struct protection_domain *domain, u32 pasid)
2736 {
2737 	u64 *pte;
2738 
2739 	if (domain->iop.mode != PAGE_MODE_NONE)
2740 		return -EINVAL;
2741 
2742 	pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
2743 	if (pte == NULL)
2744 		return 0;
2745 
2746 	*pte = 0;
2747 
2748 	return __amd_iommu_flush_tlb(domain, pasid);
2749 }
2750 
amd_iommu_domain_set_gcr3(struct iommu_domain * dom,u32 pasid,unsigned long cr3)2751 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, u32 pasid,
2752 			      unsigned long cr3)
2753 {
2754 	struct protection_domain *domain = to_pdomain(dom);
2755 	unsigned long flags;
2756 	int ret;
2757 
2758 	spin_lock_irqsave(&domain->lock, flags);
2759 	ret = __set_gcr3(domain, pasid, cr3);
2760 	spin_unlock_irqrestore(&domain->lock, flags);
2761 
2762 	return ret;
2763 }
2764 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
2765 
amd_iommu_domain_clear_gcr3(struct iommu_domain * dom,u32 pasid)2766 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, u32 pasid)
2767 {
2768 	struct protection_domain *domain = to_pdomain(dom);
2769 	unsigned long flags;
2770 	int ret;
2771 
2772 	spin_lock_irqsave(&domain->lock, flags);
2773 	ret = __clear_gcr3(domain, pasid);
2774 	spin_unlock_irqrestore(&domain->lock, flags);
2775 
2776 	return ret;
2777 }
2778 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
2779 
amd_iommu_complete_ppr(struct pci_dev * pdev,u32 pasid,int status,int tag)2780 int amd_iommu_complete_ppr(struct pci_dev *pdev, u32 pasid,
2781 			   int status, int tag)
2782 {
2783 	struct iommu_dev_data *dev_data;
2784 	struct amd_iommu *iommu;
2785 	struct iommu_cmd cmd;
2786 
2787 	dev_data = dev_iommu_priv_get(&pdev->dev);
2788 	iommu    = rlookup_amd_iommu(&pdev->dev);
2789 	if (!iommu)
2790 		return -ENODEV;
2791 
2792 	build_complete_ppr(&cmd, dev_data->devid, pasid, status,
2793 			   tag, dev_data->pri_tlp);
2794 
2795 	return iommu_queue_command(iommu, &cmd);
2796 }
2797 EXPORT_SYMBOL(amd_iommu_complete_ppr);
2798 
amd_iommu_device_info(struct pci_dev * pdev,struct amd_iommu_device_info * info)2799 int amd_iommu_device_info(struct pci_dev *pdev,
2800                           struct amd_iommu_device_info *info)
2801 {
2802 	int max_pasids;
2803 	int pos;
2804 
2805 	if (pdev == NULL || info == NULL)
2806 		return -EINVAL;
2807 
2808 	if (!amd_iommu_v2_supported())
2809 		return -EINVAL;
2810 
2811 	memset(info, 0, sizeof(*info));
2812 
2813 	if (pci_ats_supported(pdev))
2814 		info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
2815 
2816 	pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2817 	if (pos)
2818 		info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
2819 
2820 	pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
2821 	if (pos) {
2822 		int features;
2823 
2824 		max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
2825 		max_pasids = min(max_pasids, (1 << 20));
2826 
2827 		info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
2828 		info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
2829 
2830 		features = pci_pasid_features(pdev);
2831 		if (features & PCI_PASID_CAP_EXEC)
2832 			info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
2833 		if (features & PCI_PASID_CAP_PRIV)
2834 			info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
2835 	}
2836 
2837 	return 0;
2838 }
2839 EXPORT_SYMBOL(amd_iommu_device_info);
2840 
2841 #ifdef CONFIG_IRQ_REMAP
2842 
2843 /*****************************************************************************
2844  *
2845  * Interrupt Remapping Implementation
2846  *
2847  *****************************************************************************/
2848 
2849 static struct irq_chip amd_ir_chip;
2850 static DEFINE_SPINLOCK(iommu_table_lock);
2851 
iommu_flush_irt_and_complete(struct amd_iommu * iommu,u16 devid)2852 static void iommu_flush_irt_and_complete(struct amd_iommu *iommu, u16 devid)
2853 {
2854 	int ret;
2855 	u64 data;
2856 	unsigned long flags;
2857 	struct iommu_cmd cmd, cmd2;
2858 
2859 	if (iommu->irtcachedis_enabled)
2860 		return;
2861 
2862 	build_inv_irt(&cmd, devid);
2863 	data = atomic64_add_return(1, &iommu->cmd_sem_val);
2864 	build_completion_wait(&cmd2, iommu, data);
2865 
2866 	raw_spin_lock_irqsave(&iommu->lock, flags);
2867 	ret = __iommu_queue_command_sync(iommu, &cmd, true);
2868 	if (ret)
2869 		goto out;
2870 	ret = __iommu_queue_command_sync(iommu, &cmd2, false);
2871 	if (ret)
2872 		goto out;
2873 	wait_on_sem(iommu, data);
2874 out:
2875 	raw_spin_unlock_irqrestore(&iommu->lock, flags);
2876 }
2877 
set_dte_irq_entry(struct amd_iommu * iommu,u16 devid,struct irq_remap_table * table)2878 static void set_dte_irq_entry(struct amd_iommu *iommu, u16 devid,
2879 			      struct irq_remap_table *table)
2880 {
2881 	u64 dte;
2882 	struct dev_table_entry *dev_table = get_dev_table(iommu);
2883 
2884 	dte	= dev_table[devid].data[2];
2885 	dte	&= ~DTE_IRQ_PHYS_ADDR_MASK;
2886 	dte	|= iommu_virt_to_phys(table->table);
2887 	dte	|= DTE_IRQ_REMAP_INTCTL;
2888 	dte	|= DTE_INTTABLEN;
2889 	dte	|= DTE_IRQ_REMAP_ENABLE;
2890 
2891 	dev_table[devid].data[2] = dte;
2892 }
2893 
get_irq_table(struct amd_iommu * iommu,u16 devid)2894 static struct irq_remap_table *get_irq_table(struct amd_iommu *iommu, u16 devid)
2895 {
2896 	struct irq_remap_table *table;
2897 	struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
2898 
2899 	if (WARN_ONCE(!pci_seg->rlookup_table[devid],
2900 		      "%s: no iommu for devid %x:%x\n",
2901 		      __func__, pci_seg->id, devid))
2902 		return NULL;
2903 
2904 	table = pci_seg->irq_lookup_table[devid];
2905 	if (WARN_ONCE(!table, "%s: no table for devid %x:%x\n",
2906 		      __func__, pci_seg->id, devid))
2907 		return NULL;
2908 
2909 	return table;
2910 }
2911 
__alloc_irq_table(void)2912 static struct irq_remap_table *__alloc_irq_table(void)
2913 {
2914 	struct irq_remap_table *table;
2915 
2916 	table = kzalloc(sizeof(*table), GFP_KERNEL);
2917 	if (!table)
2918 		return NULL;
2919 
2920 	table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_KERNEL);
2921 	if (!table->table) {
2922 		kfree(table);
2923 		return NULL;
2924 	}
2925 	raw_spin_lock_init(&table->lock);
2926 
2927 	if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
2928 		memset(table->table, 0,
2929 		       MAX_IRQS_PER_TABLE * sizeof(u32));
2930 	else
2931 		memset(table->table, 0,
2932 		       (MAX_IRQS_PER_TABLE * (sizeof(u64) * 2)));
2933 	return table;
2934 }
2935 
set_remap_table_entry(struct amd_iommu * iommu,u16 devid,struct irq_remap_table * table)2936 static void set_remap_table_entry(struct amd_iommu *iommu, u16 devid,
2937 				  struct irq_remap_table *table)
2938 {
2939 	struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
2940 
2941 	pci_seg->irq_lookup_table[devid] = table;
2942 	set_dte_irq_entry(iommu, devid, table);
2943 	iommu_flush_dte(iommu, devid);
2944 }
2945 
set_remap_table_entry_alias(struct pci_dev * pdev,u16 alias,void * data)2946 static int set_remap_table_entry_alias(struct pci_dev *pdev, u16 alias,
2947 				       void *data)
2948 {
2949 	struct irq_remap_table *table = data;
2950 	struct amd_iommu_pci_seg *pci_seg;
2951 	struct amd_iommu *iommu = rlookup_amd_iommu(&pdev->dev);
2952 
2953 	if (!iommu)
2954 		return -EINVAL;
2955 
2956 	pci_seg = iommu->pci_seg;
2957 	pci_seg->irq_lookup_table[alias] = table;
2958 	set_dte_irq_entry(iommu, alias, table);
2959 	iommu_flush_dte(pci_seg->rlookup_table[alias], alias);
2960 
2961 	return 0;
2962 }
2963 
alloc_irq_table(struct amd_iommu * iommu,u16 devid,struct pci_dev * pdev)2964 static struct irq_remap_table *alloc_irq_table(struct amd_iommu *iommu,
2965 					       u16 devid, struct pci_dev *pdev)
2966 {
2967 	struct irq_remap_table *table = NULL;
2968 	struct irq_remap_table *new_table = NULL;
2969 	struct amd_iommu_pci_seg *pci_seg;
2970 	unsigned long flags;
2971 	u16 alias;
2972 
2973 	spin_lock_irqsave(&iommu_table_lock, flags);
2974 
2975 	pci_seg = iommu->pci_seg;
2976 	table = pci_seg->irq_lookup_table[devid];
2977 	if (table)
2978 		goto out_unlock;
2979 
2980 	alias = pci_seg->alias_table[devid];
2981 	table = pci_seg->irq_lookup_table[alias];
2982 	if (table) {
2983 		set_remap_table_entry(iommu, devid, table);
2984 		goto out_wait;
2985 	}
2986 	spin_unlock_irqrestore(&iommu_table_lock, flags);
2987 
2988 	/* Nothing there yet, allocate new irq remapping table */
2989 	new_table = __alloc_irq_table();
2990 	if (!new_table)
2991 		return NULL;
2992 
2993 	spin_lock_irqsave(&iommu_table_lock, flags);
2994 
2995 	table = pci_seg->irq_lookup_table[devid];
2996 	if (table)
2997 		goto out_unlock;
2998 
2999 	table = pci_seg->irq_lookup_table[alias];
3000 	if (table) {
3001 		set_remap_table_entry(iommu, devid, table);
3002 		goto out_wait;
3003 	}
3004 
3005 	table = new_table;
3006 	new_table = NULL;
3007 
3008 	if (pdev)
3009 		pci_for_each_dma_alias(pdev, set_remap_table_entry_alias,
3010 				       table);
3011 	else
3012 		set_remap_table_entry(iommu, devid, table);
3013 
3014 	if (devid != alias)
3015 		set_remap_table_entry(iommu, alias, table);
3016 
3017 out_wait:
3018 	iommu_completion_wait(iommu);
3019 
3020 out_unlock:
3021 	spin_unlock_irqrestore(&iommu_table_lock, flags);
3022 
3023 	if (new_table) {
3024 		kmem_cache_free(amd_iommu_irq_cache, new_table->table);
3025 		kfree(new_table);
3026 	}
3027 	return table;
3028 }
3029 
alloc_irq_index(struct amd_iommu * iommu,u16 devid,int count,bool align,struct pci_dev * pdev)3030 static int alloc_irq_index(struct amd_iommu *iommu, u16 devid, int count,
3031 			   bool align, struct pci_dev *pdev)
3032 {
3033 	struct irq_remap_table *table;
3034 	int index, c, alignment = 1;
3035 	unsigned long flags;
3036 
3037 	table = alloc_irq_table(iommu, devid, pdev);
3038 	if (!table)
3039 		return -ENODEV;
3040 
3041 	if (align)
3042 		alignment = roundup_pow_of_two(count);
3043 
3044 	raw_spin_lock_irqsave(&table->lock, flags);
3045 
3046 	/* Scan table for free entries */
3047 	for (index = ALIGN(table->min_index, alignment), c = 0;
3048 	     index < MAX_IRQS_PER_TABLE;) {
3049 		if (!iommu->irte_ops->is_allocated(table, index)) {
3050 			c += 1;
3051 		} else {
3052 			c     = 0;
3053 			index = ALIGN(index + 1, alignment);
3054 			continue;
3055 		}
3056 
3057 		if (c == count)	{
3058 			for (; c != 0; --c)
3059 				iommu->irte_ops->set_allocated(table, index - c + 1);
3060 
3061 			index -= count - 1;
3062 			goto out;
3063 		}
3064 
3065 		index++;
3066 	}
3067 
3068 	index = -ENOSPC;
3069 
3070 out:
3071 	raw_spin_unlock_irqrestore(&table->lock, flags);
3072 
3073 	return index;
3074 }
3075 
modify_irte_ga(struct amd_iommu * iommu,u16 devid,int index,struct irte_ga * irte)3076 static int modify_irte_ga(struct amd_iommu *iommu, u16 devid, int index,
3077 			  struct irte_ga *irte)
3078 {
3079 	struct irq_remap_table *table;
3080 	struct irte_ga *entry;
3081 	unsigned long flags;
3082 	u128 old;
3083 
3084 	table = get_irq_table(iommu, devid);
3085 	if (!table)
3086 		return -ENOMEM;
3087 
3088 	raw_spin_lock_irqsave(&table->lock, flags);
3089 
3090 	entry = (struct irte_ga *)table->table;
3091 	entry = &entry[index];
3092 
3093 	/*
3094 	 * We use cmpxchg16 to atomically update the 128-bit IRTE,
3095 	 * and it cannot be updated by the hardware or other processors
3096 	 * behind us, so the return value of cmpxchg16 should be the
3097 	 * same as the old value.
3098 	 */
3099 	old = entry->irte;
3100 	WARN_ON(!try_cmpxchg128(&entry->irte, &old, irte->irte));
3101 
3102 	raw_spin_unlock_irqrestore(&table->lock, flags);
3103 
3104 	iommu_flush_irt_and_complete(iommu, devid);
3105 
3106 	return 0;
3107 }
3108 
modify_irte(struct amd_iommu * iommu,u16 devid,int index,union irte * irte)3109 static int modify_irte(struct amd_iommu *iommu,
3110 		       u16 devid, int index, union irte *irte)
3111 {
3112 	struct irq_remap_table *table;
3113 	unsigned long flags;
3114 
3115 	table = get_irq_table(iommu, devid);
3116 	if (!table)
3117 		return -ENOMEM;
3118 
3119 	raw_spin_lock_irqsave(&table->lock, flags);
3120 	table->table[index] = irte->val;
3121 	raw_spin_unlock_irqrestore(&table->lock, flags);
3122 
3123 	iommu_flush_irt_and_complete(iommu, devid);
3124 
3125 	return 0;
3126 }
3127 
free_irte(struct amd_iommu * iommu,u16 devid,int index)3128 static void free_irte(struct amd_iommu *iommu, u16 devid, int index)
3129 {
3130 	struct irq_remap_table *table;
3131 	unsigned long flags;
3132 
3133 	table = get_irq_table(iommu, devid);
3134 	if (!table)
3135 		return;
3136 
3137 	raw_spin_lock_irqsave(&table->lock, flags);
3138 	iommu->irte_ops->clear_allocated(table, index);
3139 	raw_spin_unlock_irqrestore(&table->lock, flags);
3140 
3141 	iommu_flush_irt_and_complete(iommu, devid);
3142 }
3143 
irte_prepare(void * entry,u32 delivery_mode,bool dest_mode,u8 vector,u32 dest_apicid,int devid)3144 static void irte_prepare(void *entry,
3145 			 u32 delivery_mode, bool dest_mode,
3146 			 u8 vector, u32 dest_apicid, int devid)
3147 {
3148 	union irte *irte = (union irte *) entry;
3149 
3150 	irte->val                = 0;
3151 	irte->fields.vector      = vector;
3152 	irte->fields.int_type    = delivery_mode;
3153 	irte->fields.destination = dest_apicid;
3154 	irte->fields.dm          = dest_mode;
3155 	irte->fields.valid       = 1;
3156 }
3157 
irte_ga_prepare(void * entry,u32 delivery_mode,bool dest_mode,u8 vector,u32 dest_apicid,int devid)3158 static void irte_ga_prepare(void *entry,
3159 			    u32 delivery_mode, bool dest_mode,
3160 			    u8 vector, u32 dest_apicid, int devid)
3161 {
3162 	struct irte_ga *irte = (struct irte_ga *) entry;
3163 
3164 	irte->lo.val                      = 0;
3165 	irte->hi.val                      = 0;
3166 	irte->lo.fields_remap.int_type    = delivery_mode;
3167 	irte->lo.fields_remap.dm          = dest_mode;
3168 	irte->hi.fields.vector            = vector;
3169 	irte->lo.fields_remap.destination = APICID_TO_IRTE_DEST_LO(dest_apicid);
3170 	irte->hi.fields.destination       = APICID_TO_IRTE_DEST_HI(dest_apicid);
3171 	irte->lo.fields_remap.valid       = 1;
3172 }
3173 
irte_activate(struct amd_iommu * iommu,void * entry,u16 devid,u16 index)3174 static void irte_activate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index)
3175 {
3176 	union irte *irte = (union irte *) entry;
3177 
3178 	irte->fields.valid = 1;
3179 	modify_irte(iommu, devid, index, irte);
3180 }
3181 
irte_ga_activate(struct amd_iommu * iommu,void * entry,u16 devid,u16 index)3182 static void irte_ga_activate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index)
3183 {
3184 	struct irte_ga *irte = (struct irte_ga *) entry;
3185 
3186 	irte->lo.fields_remap.valid = 1;
3187 	modify_irte_ga(iommu, devid, index, irte);
3188 }
3189 
irte_deactivate(struct amd_iommu * iommu,void * entry,u16 devid,u16 index)3190 static void irte_deactivate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index)
3191 {
3192 	union irte *irte = (union irte *) entry;
3193 
3194 	irte->fields.valid = 0;
3195 	modify_irte(iommu, devid, index, irte);
3196 }
3197 
irte_ga_deactivate(struct amd_iommu * iommu,void * entry,u16 devid,u16 index)3198 static void irte_ga_deactivate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index)
3199 {
3200 	struct irte_ga *irte = (struct irte_ga *) entry;
3201 
3202 	irte->lo.fields_remap.valid = 0;
3203 	modify_irte_ga(iommu, devid, index, irte);
3204 }
3205 
irte_set_affinity(struct amd_iommu * iommu,void * entry,u16 devid,u16 index,u8 vector,u32 dest_apicid)3206 static void irte_set_affinity(struct amd_iommu *iommu, void *entry, u16 devid, u16 index,
3207 			      u8 vector, u32 dest_apicid)
3208 {
3209 	union irte *irte = (union irte *) entry;
3210 
3211 	irte->fields.vector = vector;
3212 	irte->fields.destination = dest_apicid;
3213 	modify_irte(iommu, devid, index, irte);
3214 }
3215 
irte_ga_set_affinity(struct amd_iommu * iommu,void * entry,u16 devid,u16 index,u8 vector,u32 dest_apicid)3216 static void irte_ga_set_affinity(struct amd_iommu *iommu, void *entry, u16 devid, u16 index,
3217 				 u8 vector, u32 dest_apicid)
3218 {
3219 	struct irte_ga *irte = (struct irte_ga *) entry;
3220 
3221 	if (!irte->lo.fields_remap.guest_mode) {
3222 		irte->hi.fields.vector = vector;
3223 		irte->lo.fields_remap.destination =
3224 					APICID_TO_IRTE_DEST_LO(dest_apicid);
3225 		irte->hi.fields.destination =
3226 					APICID_TO_IRTE_DEST_HI(dest_apicid);
3227 		modify_irte_ga(iommu, devid, index, irte);
3228 	}
3229 }
3230 
3231 #define IRTE_ALLOCATED (~1U)
irte_set_allocated(struct irq_remap_table * table,int index)3232 static void irte_set_allocated(struct irq_remap_table *table, int index)
3233 {
3234 	table->table[index] = IRTE_ALLOCATED;
3235 }
3236 
irte_ga_set_allocated(struct irq_remap_table * table,int index)3237 static void irte_ga_set_allocated(struct irq_remap_table *table, int index)
3238 {
3239 	struct irte_ga *ptr = (struct irte_ga *)table->table;
3240 	struct irte_ga *irte = &ptr[index];
3241 
3242 	memset(&irte->lo.val, 0, sizeof(u64));
3243 	memset(&irte->hi.val, 0, sizeof(u64));
3244 	irte->hi.fields.vector = 0xff;
3245 }
3246 
irte_is_allocated(struct irq_remap_table * table,int index)3247 static bool irte_is_allocated(struct irq_remap_table *table, int index)
3248 {
3249 	union irte *ptr = (union irte *)table->table;
3250 	union irte *irte = &ptr[index];
3251 
3252 	return irte->val != 0;
3253 }
3254 
irte_ga_is_allocated(struct irq_remap_table * table,int index)3255 static bool irte_ga_is_allocated(struct irq_remap_table *table, int index)
3256 {
3257 	struct irte_ga *ptr = (struct irte_ga *)table->table;
3258 	struct irte_ga *irte = &ptr[index];
3259 
3260 	return irte->hi.fields.vector != 0;
3261 }
3262 
irte_clear_allocated(struct irq_remap_table * table,int index)3263 static void irte_clear_allocated(struct irq_remap_table *table, int index)
3264 {
3265 	table->table[index] = 0;
3266 }
3267 
irte_ga_clear_allocated(struct irq_remap_table * table,int index)3268 static void irte_ga_clear_allocated(struct irq_remap_table *table, int index)
3269 {
3270 	struct irte_ga *ptr = (struct irte_ga *)table->table;
3271 	struct irte_ga *irte = &ptr[index];
3272 
3273 	memset(&irte->lo.val, 0, sizeof(u64));
3274 	memset(&irte->hi.val, 0, sizeof(u64));
3275 }
3276 
get_devid(struct irq_alloc_info * info)3277 static int get_devid(struct irq_alloc_info *info)
3278 {
3279 	switch (info->type) {
3280 	case X86_IRQ_ALLOC_TYPE_IOAPIC:
3281 		return get_ioapic_devid(info->devid);
3282 	case X86_IRQ_ALLOC_TYPE_HPET:
3283 		return get_hpet_devid(info->devid);
3284 	case X86_IRQ_ALLOC_TYPE_PCI_MSI:
3285 	case X86_IRQ_ALLOC_TYPE_PCI_MSIX:
3286 		return get_device_sbdf_id(msi_desc_to_dev(info->desc));
3287 	default:
3288 		WARN_ON_ONCE(1);
3289 		return -1;
3290 	}
3291 }
3292 
3293 struct irq_remap_ops amd_iommu_irq_ops = {
3294 	.prepare		= amd_iommu_prepare,
3295 	.enable			= amd_iommu_enable,
3296 	.disable		= amd_iommu_disable,
3297 	.reenable		= amd_iommu_reenable,
3298 	.enable_faulting	= amd_iommu_enable_faulting,
3299 };
3300 
fill_msi_msg(struct msi_msg * msg,u32 index)3301 static void fill_msi_msg(struct msi_msg *msg, u32 index)
3302 {
3303 	msg->data = index;
3304 	msg->address_lo = 0;
3305 	msg->arch_addr_lo.base_address = X86_MSI_BASE_ADDRESS_LOW;
3306 	msg->address_hi = X86_MSI_BASE_ADDRESS_HIGH;
3307 }
3308 
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)3309 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
3310 				       struct irq_cfg *irq_cfg,
3311 				       struct irq_alloc_info *info,
3312 				       int devid, int index, int sub_handle)
3313 {
3314 	struct irq_2_irte *irte_info = &data->irq_2_irte;
3315 	struct amd_iommu *iommu = data->iommu;
3316 
3317 	if (!iommu)
3318 		return;
3319 
3320 	data->irq_2_irte.devid = devid;
3321 	data->irq_2_irte.index = index + sub_handle;
3322 	iommu->irte_ops->prepare(data->entry, apic->delivery_mode,
3323 				 apic->dest_mode_logical, irq_cfg->vector,
3324 				 irq_cfg->dest_apicid, devid);
3325 
3326 	switch (info->type) {
3327 	case X86_IRQ_ALLOC_TYPE_IOAPIC:
3328 	case X86_IRQ_ALLOC_TYPE_HPET:
3329 	case X86_IRQ_ALLOC_TYPE_PCI_MSI:
3330 	case X86_IRQ_ALLOC_TYPE_PCI_MSIX:
3331 		fill_msi_msg(&data->msi_entry, irte_info->index);
3332 		break;
3333 
3334 	default:
3335 		BUG_ON(1);
3336 		break;
3337 	}
3338 }
3339 
3340 struct amd_irte_ops irte_32_ops = {
3341 	.prepare = irte_prepare,
3342 	.activate = irte_activate,
3343 	.deactivate = irte_deactivate,
3344 	.set_affinity = irte_set_affinity,
3345 	.set_allocated = irte_set_allocated,
3346 	.is_allocated = irte_is_allocated,
3347 	.clear_allocated = irte_clear_allocated,
3348 };
3349 
3350 struct amd_irte_ops irte_128_ops = {
3351 	.prepare = irte_ga_prepare,
3352 	.activate = irte_ga_activate,
3353 	.deactivate = irte_ga_deactivate,
3354 	.set_affinity = irte_ga_set_affinity,
3355 	.set_allocated = irte_ga_set_allocated,
3356 	.is_allocated = irte_ga_is_allocated,
3357 	.clear_allocated = irte_ga_clear_allocated,
3358 };
3359 
irq_remapping_alloc(struct irq_domain * domain,unsigned int virq,unsigned int nr_irqs,void * arg)3360 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
3361 			       unsigned int nr_irqs, void *arg)
3362 {
3363 	struct irq_alloc_info *info = arg;
3364 	struct irq_data *irq_data;
3365 	struct amd_ir_data *data = NULL;
3366 	struct amd_iommu *iommu;
3367 	struct irq_cfg *cfg;
3368 	int i, ret, devid, seg, sbdf;
3369 	int index;
3370 
3371 	if (!info)
3372 		return -EINVAL;
3373 	if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_PCI_MSI)
3374 		return -EINVAL;
3375 
3376 	sbdf = get_devid(info);
3377 	if (sbdf < 0)
3378 		return -EINVAL;
3379 
3380 	seg = PCI_SBDF_TO_SEGID(sbdf);
3381 	devid = PCI_SBDF_TO_DEVID(sbdf);
3382 	iommu = __rlookup_amd_iommu(seg, devid);
3383 	if (!iommu)
3384 		return -EINVAL;
3385 
3386 	ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
3387 	if (ret < 0)
3388 		return ret;
3389 
3390 	if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
3391 		struct irq_remap_table *table;
3392 
3393 		table = alloc_irq_table(iommu, devid, NULL);
3394 		if (table) {
3395 			if (!table->min_index) {
3396 				/*
3397 				 * Keep the first 32 indexes free for IOAPIC
3398 				 * interrupts.
3399 				 */
3400 				table->min_index = 32;
3401 				for (i = 0; i < 32; ++i)
3402 					iommu->irte_ops->set_allocated(table, i);
3403 			}
3404 			WARN_ON(table->min_index != 32);
3405 			index = info->ioapic.pin;
3406 		} else {
3407 			index = -ENOMEM;
3408 		}
3409 	} else if (info->type == X86_IRQ_ALLOC_TYPE_PCI_MSI ||
3410 		   info->type == X86_IRQ_ALLOC_TYPE_PCI_MSIX) {
3411 		bool align = (info->type == X86_IRQ_ALLOC_TYPE_PCI_MSI);
3412 
3413 		index = alloc_irq_index(iommu, devid, nr_irqs, align,
3414 					msi_desc_to_pci_dev(info->desc));
3415 	} else {
3416 		index = alloc_irq_index(iommu, devid, nr_irqs, false, NULL);
3417 	}
3418 
3419 	if (index < 0) {
3420 		pr_warn("Failed to allocate IRTE\n");
3421 		ret = index;
3422 		goto out_free_parent;
3423 	}
3424 
3425 	for (i = 0; i < nr_irqs; i++) {
3426 		irq_data = irq_domain_get_irq_data(domain, virq + i);
3427 		cfg = irq_data ? irqd_cfg(irq_data) : NULL;
3428 		if (!cfg) {
3429 			ret = -EINVAL;
3430 			goto out_free_data;
3431 		}
3432 
3433 		ret = -ENOMEM;
3434 		data = kzalloc(sizeof(*data), GFP_KERNEL);
3435 		if (!data)
3436 			goto out_free_data;
3437 
3438 		if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
3439 			data->entry = kzalloc(sizeof(union irte), GFP_KERNEL);
3440 		else
3441 			data->entry = kzalloc(sizeof(struct irte_ga),
3442 						     GFP_KERNEL);
3443 		if (!data->entry) {
3444 			kfree(data);
3445 			goto out_free_data;
3446 		}
3447 
3448 		data->iommu = iommu;
3449 		irq_data->hwirq = (devid << 16) + i;
3450 		irq_data->chip_data = data;
3451 		irq_data->chip = &amd_ir_chip;
3452 		irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
3453 		irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
3454 	}
3455 
3456 	return 0;
3457 
3458 out_free_data:
3459 	for (i--; i >= 0; i--) {
3460 		irq_data = irq_domain_get_irq_data(domain, virq + i);
3461 		if (irq_data)
3462 			kfree(irq_data->chip_data);
3463 	}
3464 	for (i = 0; i < nr_irqs; i++)
3465 		free_irte(iommu, devid, index + i);
3466 out_free_parent:
3467 	irq_domain_free_irqs_common(domain, virq, nr_irqs);
3468 	return ret;
3469 }
3470 
irq_remapping_free(struct irq_domain * domain,unsigned int virq,unsigned int nr_irqs)3471 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
3472 			       unsigned int nr_irqs)
3473 {
3474 	struct irq_2_irte *irte_info;
3475 	struct irq_data *irq_data;
3476 	struct amd_ir_data *data;
3477 	int i;
3478 
3479 	for (i = 0; i < nr_irqs; i++) {
3480 		irq_data = irq_domain_get_irq_data(domain, virq  + i);
3481 		if (irq_data && irq_data->chip_data) {
3482 			data = irq_data->chip_data;
3483 			irte_info = &data->irq_2_irte;
3484 			free_irte(data->iommu, irte_info->devid, irte_info->index);
3485 			kfree(data->entry);
3486 			kfree(data);
3487 		}
3488 	}
3489 	irq_domain_free_irqs_common(domain, virq, nr_irqs);
3490 }
3491 
3492 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
3493 			       struct amd_ir_data *ir_data,
3494 			       struct irq_2_irte *irte_info,
3495 			       struct irq_cfg *cfg);
3496 
irq_remapping_activate(struct irq_domain * domain,struct irq_data * irq_data,bool reserve)3497 static int irq_remapping_activate(struct irq_domain *domain,
3498 				  struct irq_data *irq_data, bool reserve)
3499 {
3500 	struct amd_ir_data *data = irq_data->chip_data;
3501 	struct irq_2_irte *irte_info = &data->irq_2_irte;
3502 	struct amd_iommu *iommu = data->iommu;
3503 	struct irq_cfg *cfg = irqd_cfg(irq_data);
3504 
3505 	if (!iommu)
3506 		return 0;
3507 
3508 	iommu->irte_ops->activate(iommu, data->entry, irte_info->devid,
3509 				  irte_info->index);
3510 	amd_ir_update_irte(irq_data, iommu, data, irte_info, cfg);
3511 	return 0;
3512 }
3513 
irq_remapping_deactivate(struct irq_domain * domain,struct irq_data * irq_data)3514 static void irq_remapping_deactivate(struct irq_domain *domain,
3515 				     struct irq_data *irq_data)
3516 {
3517 	struct amd_ir_data *data = irq_data->chip_data;
3518 	struct irq_2_irte *irte_info = &data->irq_2_irte;
3519 	struct amd_iommu *iommu = data->iommu;
3520 
3521 	if (iommu)
3522 		iommu->irte_ops->deactivate(iommu, data->entry, irte_info->devid,
3523 					    irte_info->index);
3524 }
3525 
irq_remapping_select(struct irq_domain * d,struct irq_fwspec * fwspec,enum irq_domain_bus_token bus_token)3526 static int irq_remapping_select(struct irq_domain *d, struct irq_fwspec *fwspec,
3527 				enum irq_domain_bus_token bus_token)
3528 {
3529 	struct amd_iommu *iommu;
3530 	int devid = -1;
3531 
3532 	if (!amd_iommu_irq_remap)
3533 		return 0;
3534 
3535 	if (x86_fwspec_is_ioapic(fwspec))
3536 		devid = get_ioapic_devid(fwspec->param[0]);
3537 	else if (x86_fwspec_is_hpet(fwspec))
3538 		devid = get_hpet_devid(fwspec->param[0]);
3539 
3540 	if (devid < 0)
3541 		return 0;
3542 	iommu = __rlookup_amd_iommu((devid >> 16), (devid & 0xffff));
3543 
3544 	return iommu && iommu->ir_domain == d;
3545 }
3546 
3547 static const struct irq_domain_ops amd_ir_domain_ops = {
3548 	.select = irq_remapping_select,
3549 	.alloc = irq_remapping_alloc,
3550 	.free = irq_remapping_free,
3551 	.activate = irq_remapping_activate,
3552 	.deactivate = irq_remapping_deactivate,
3553 };
3554 
amd_iommu_activate_guest_mode(void * data)3555 int amd_iommu_activate_guest_mode(void *data)
3556 {
3557 	struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
3558 	struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
3559 	u64 valid;
3560 
3561 	if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) || !entry)
3562 		return 0;
3563 
3564 	valid = entry->lo.fields_vapic.valid;
3565 
3566 	entry->lo.val = 0;
3567 	entry->hi.val = 0;
3568 
3569 	entry->lo.fields_vapic.valid       = valid;
3570 	entry->lo.fields_vapic.guest_mode  = 1;
3571 	entry->lo.fields_vapic.ga_log_intr = 1;
3572 	entry->hi.fields.ga_root_ptr       = ir_data->ga_root_ptr;
3573 	entry->hi.fields.vector            = ir_data->ga_vector;
3574 	entry->lo.fields_vapic.ga_tag      = ir_data->ga_tag;
3575 
3576 	return modify_irte_ga(ir_data->iommu, ir_data->irq_2_irte.devid,
3577 			      ir_data->irq_2_irte.index, entry);
3578 }
3579 EXPORT_SYMBOL(amd_iommu_activate_guest_mode);
3580 
amd_iommu_deactivate_guest_mode(void * data)3581 int amd_iommu_deactivate_guest_mode(void *data)
3582 {
3583 	struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
3584 	struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
3585 	struct irq_cfg *cfg = ir_data->cfg;
3586 	u64 valid;
3587 
3588 	if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
3589 	    !entry || !entry->lo.fields_vapic.guest_mode)
3590 		return 0;
3591 
3592 	valid = entry->lo.fields_remap.valid;
3593 
3594 	entry->lo.val = 0;
3595 	entry->hi.val = 0;
3596 
3597 	entry->lo.fields_remap.valid       = valid;
3598 	entry->lo.fields_remap.dm          = apic->dest_mode_logical;
3599 	entry->lo.fields_remap.int_type    = apic->delivery_mode;
3600 	entry->hi.fields.vector            = cfg->vector;
3601 	entry->lo.fields_remap.destination =
3602 				APICID_TO_IRTE_DEST_LO(cfg->dest_apicid);
3603 	entry->hi.fields.destination =
3604 				APICID_TO_IRTE_DEST_HI(cfg->dest_apicid);
3605 
3606 	return modify_irte_ga(ir_data->iommu, ir_data->irq_2_irte.devid,
3607 			      ir_data->irq_2_irte.index, entry);
3608 }
3609 EXPORT_SYMBOL(amd_iommu_deactivate_guest_mode);
3610 
amd_ir_set_vcpu_affinity(struct irq_data * data,void * vcpu_info)3611 static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *vcpu_info)
3612 {
3613 	int ret;
3614 	struct amd_iommu_pi_data *pi_data = vcpu_info;
3615 	struct vcpu_data *vcpu_pi_info = pi_data->vcpu_data;
3616 	struct amd_ir_data *ir_data = data->chip_data;
3617 	struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
3618 	struct iommu_dev_data *dev_data;
3619 
3620 	if (ir_data->iommu == NULL)
3621 		return -EINVAL;
3622 
3623 	dev_data = search_dev_data(ir_data->iommu, irte_info->devid);
3624 
3625 	/* Note:
3626 	 * This device has never been set up for guest mode.
3627 	 * we should not modify the IRTE
3628 	 */
3629 	if (!dev_data || !dev_data->use_vapic)
3630 		return -EINVAL;
3631 
3632 	ir_data->cfg = irqd_cfg(data);
3633 	pi_data->ir_data = ir_data;
3634 
3635 	/* Note:
3636 	 * SVM tries to set up for VAPIC mode, but we are in
3637 	 * legacy mode. So, we force legacy mode instead.
3638 	 */
3639 	if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
3640 		pr_debug("%s: Fall back to using intr legacy remap\n",
3641 			 __func__);
3642 		pi_data->is_guest_mode = false;
3643 	}
3644 
3645 	pi_data->prev_ga_tag = ir_data->cached_ga_tag;
3646 	if (pi_data->is_guest_mode) {
3647 		ir_data->ga_root_ptr = (pi_data->base >> 12);
3648 		ir_data->ga_vector = vcpu_pi_info->vector;
3649 		ir_data->ga_tag = pi_data->ga_tag;
3650 		ret = amd_iommu_activate_guest_mode(ir_data);
3651 		if (!ret)
3652 			ir_data->cached_ga_tag = pi_data->ga_tag;
3653 	} else {
3654 		ret = amd_iommu_deactivate_guest_mode(ir_data);
3655 
3656 		/*
3657 		 * This communicates the ga_tag back to the caller
3658 		 * so that it can do all the necessary clean up.
3659 		 */
3660 		if (!ret)
3661 			ir_data->cached_ga_tag = 0;
3662 	}
3663 
3664 	return ret;
3665 }
3666 
3667 
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)3668 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
3669 			       struct amd_ir_data *ir_data,
3670 			       struct irq_2_irte *irte_info,
3671 			       struct irq_cfg *cfg)
3672 {
3673 
3674 	/*
3675 	 * Atomically updates the IRTE with the new destination, vector
3676 	 * and flushes the interrupt entry cache.
3677 	 */
3678 	iommu->irte_ops->set_affinity(iommu, ir_data->entry, irte_info->devid,
3679 				      irte_info->index, cfg->vector,
3680 				      cfg->dest_apicid);
3681 }
3682 
amd_ir_set_affinity(struct irq_data * data,const struct cpumask * mask,bool force)3683 static int amd_ir_set_affinity(struct irq_data *data,
3684 			       const struct cpumask *mask, bool force)
3685 {
3686 	struct amd_ir_data *ir_data = data->chip_data;
3687 	struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
3688 	struct irq_cfg *cfg = irqd_cfg(data);
3689 	struct irq_data *parent = data->parent_data;
3690 	struct amd_iommu *iommu = ir_data->iommu;
3691 	int ret;
3692 
3693 	if (!iommu)
3694 		return -ENODEV;
3695 
3696 	ret = parent->chip->irq_set_affinity(parent, mask, force);
3697 	if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
3698 		return ret;
3699 
3700 	amd_ir_update_irte(data, iommu, ir_data, irte_info, cfg);
3701 	/*
3702 	 * After this point, all the interrupts will start arriving
3703 	 * at the new destination. So, time to cleanup the previous
3704 	 * vector allocation.
3705 	 */
3706 	vector_schedule_cleanup(cfg);
3707 
3708 	return IRQ_SET_MASK_OK_DONE;
3709 }
3710 
ir_compose_msi_msg(struct irq_data * irq_data,struct msi_msg * msg)3711 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
3712 {
3713 	struct amd_ir_data *ir_data = irq_data->chip_data;
3714 
3715 	*msg = ir_data->msi_entry;
3716 }
3717 
3718 static struct irq_chip amd_ir_chip = {
3719 	.name			= "AMD-IR",
3720 	.irq_ack		= apic_ack_irq,
3721 	.irq_set_affinity	= amd_ir_set_affinity,
3722 	.irq_set_vcpu_affinity	= amd_ir_set_vcpu_affinity,
3723 	.irq_compose_msi_msg	= ir_compose_msi_msg,
3724 };
3725 
3726 static const struct msi_parent_ops amdvi_msi_parent_ops = {
3727 	.supported_flags	= X86_VECTOR_MSI_FLAGS_SUPPORTED |
3728 				  MSI_FLAG_MULTI_PCI_MSI |
3729 				  MSI_FLAG_PCI_IMS,
3730 	.prefix			= "IR-",
3731 	.init_dev_msi_info	= msi_parent_init_dev_msi_info,
3732 };
3733 
3734 static const struct msi_parent_ops virt_amdvi_msi_parent_ops = {
3735 	.supported_flags	= X86_VECTOR_MSI_FLAGS_SUPPORTED |
3736 				  MSI_FLAG_MULTI_PCI_MSI,
3737 	.prefix			= "vIR-",
3738 	.init_dev_msi_info	= msi_parent_init_dev_msi_info,
3739 };
3740 
amd_iommu_create_irq_domain(struct amd_iommu * iommu)3741 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
3742 {
3743 	struct fwnode_handle *fn;
3744 
3745 	fn = irq_domain_alloc_named_id_fwnode("AMD-IR", iommu->index);
3746 	if (!fn)
3747 		return -ENOMEM;
3748 	iommu->ir_domain = irq_domain_create_hierarchy(arch_get_ir_parent_domain(), 0, 0,
3749 						       fn, &amd_ir_domain_ops, iommu);
3750 	if (!iommu->ir_domain) {
3751 		irq_domain_free_fwnode(fn);
3752 		return -ENOMEM;
3753 	}
3754 
3755 	irq_domain_update_bus_token(iommu->ir_domain,  DOMAIN_BUS_AMDVI);
3756 	iommu->ir_domain->flags |= IRQ_DOMAIN_FLAG_MSI_PARENT |
3757 				   IRQ_DOMAIN_FLAG_ISOLATED_MSI;
3758 
3759 	if (amd_iommu_np_cache)
3760 		iommu->ir_domain->msi_parent_ops = &virt_amdvi_msi_parent_ops;
3761 	else
3762 		iommu->ir_domain->msi_parent_ops = &amdvi_msi_parent_ops;
3763 
3764 	return 0;
3765 }
3766 
amd_iommu_update_ga(int cpu,bool is_run,void * data)3767 int amd_iommu_update_ga(int cpu, bool is_run, void *data)
3768 {
3769 	struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
3770 	struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
3771 
3772 	if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
3773 	    !entry || !entry->lo.fields_vapic.guest_mode)
3774 		return 0;
3775 
3776 	if (!ir_data->iommu)
3777 		return -ENODEV;
3778 
3779 	if (cpu >= 0) {
3780 		entry->lo.fields_vapic.destination =
3781 					APICID_TO_IRTE_DEST_LO(cpu);
3782 		entry->hi.fields.destination =
3783 					APICID_TO_IRTE_DEST_HI(cpu);
3784 	}
3785 	entry->lo.fields_vapic.is_run = is_run;
3786 
3787 	return modify_irte_ga(ir_data->iommu, ir_data->irq_2_irte.devid,
3788 			      ir_data->irq_2_irte.index, entry);
3789 }
3790 EXPORT_SYMBOL(amd_iommu_update_ga);
3791 #endif
3792