• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (C) 2007-2010 Advanced Micro Devices, Inc.
3  * Author: Joerg Roedel <jroedel@suse.de>
4  *         Leo Duran <leo.duran@amd.com>
5  *
6  * This program is free software; you can redistribute it and/or modify it
7  * under the terms of the GNU General Public License version 2 as published
8  * by the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it will be useful,
11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
13  * GNU General Public License for more details.
14  *
15  * You should have received a copy of the GNU General Public License
16  * along with this program; if not, write to the Free Software
17  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307 USA
18  */
19 
20 #include <linux/ratelimit.h>
21 #include <linux/pci.h>
22 #include <linux/pci-ats.h>
23 #include <linux/bitmap.h>
24 #include <linux/slab.h>
25 #include <linux/debugfs.h>
26 #include <linux/scatterlist.h>
27 #include <linux/dma-mapping.h>
28 #include <linux/iommu-helper.h>
29 #include <linux/iommu.h>
30 #include <linux/delay.h>
31 #include <linux/amd-iommu.h>
32 #include <linux/notifier.h>
33 #include <linux/export.h>
34 #include <linux/irq.h>
35 #include <linux/msi.h>
36 #include <linux/dma-contiguous.h>
37 #include <linux/irqdomain.h>
38 #include <asm/irq_remapping.h>
39 #include <asm/io_apic.h>
40 #include <asm/apic.h>
41 #include <asm/hw_irq.h>
42 #include <asm/msidef.h>
43 #include <asm/proto.h>
44 #include <asm/iommu.h>
45 #include <asm/gart.h>
46 #include <asm/dma.h>
47 
48 #include "amd_iommu_proto.h"
49 #include "amd_iommu_types.h"
50 #include "irq_remapping.h"
51 
52 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
53 
54 #define LOOP_TIMEOUT	100000
55 
56 /*
57  * This bitmap is used to advertise the page sizes our hardware support
58  * to the IOMMU core, which will then use this information to split
59  * physically contiguous memory regions it is mapping into page sizes
60  * that we support.
61  *
62  * 512GB Pages are not supported due to a hardware bug
63  */
64 #define AMD_IOMMU_PGSIZES	((~0xFFFUL) & ~(2ULL << 38))
65 
66 static DEFINE_RWLOCK(amd_iommu_devtable_lock);
67 
68 /* List of all available dev_data structures */
69 static LIST_HEAD(dev_data_list);
70 static DEFINE_SPINLOCK(dev_data_list_lock);
71 
72 LIST_HEAD(ioapic_map);
73 LIST_HEAD(hpet_map);
74 
75 /*
76  * Domain for untranslated devices - only allocated
77  * if iommu=pt passed on kernel cmd line.
78  */
79 static const struct iommu_ops amd_iommu_ops;
80 
81 static ATOMIC_NOTIFIER_HEAD(ppr_notifier);
82 int amd_iommu_max_glx_val = -1;
83 
84 static struct dma_map_ops amd_iommu_dma_ops;
85 
86 /*
87  * This struct contains device specific data for the IOMMU
88  */
89 struct iommu_dev_data {
90 	struct list_head list;		  /* For domain->dev_list */
91 	struct list_head dev_data_list;	  /* For global dev_data_list */
92 	struct protection_domain *domain; /* Domain the device is bound to */
93 	u16 devid;			  /* PCI Device ID */
94 	u16 alias;			  /* Alias Device ID */
95 	bool iommu_v2;			  /* Device can make use of IOMMUv2 */
96 	bool passthrough;		  /* Device is identity mapped */
97 	struct {
98 		bool enabled;
99 		int qdep;
100 	} ats;				  /* ATS state */
101 	bool pri_tlp;			  /* PASID TLB required for
102 					     PPR completions */
103 	u32 errata;			  /* Bitmap for errata to apply */
104 };
105 
106 /*
107  * general struct to manage commands send to an IOMMU
108  */
109 struct iommu_cmd {
110 	u32 data[4];
111 };
112 
113 struct kmem_cache *amd_iommu_irq_cache;
114 
115 static void update_domain(struct protection_domain *domain);
116 static int protection_domain_init(struct protection_domain *domain);
117 
118 /****************************************************************************
119  *
120  * Helper functions
121  *
122  ****************************************************************************/
123 
to_pdomain(struct iommu_domain * dom)124 static struct protection_domain *to_pdomain(struct iommu_domain *dom)
125 {
126 	return container_of(dom, struct protection_domain, domain);
127 }
128 
get_device_id(struct device * dev)129 static inline u16 get_device_id(struct device *dev)
130 {
131 	struct pci_dev *pdev = to_pci_dev(dev);
132 
133 	return PCI_DEVID(pdev->bus->number, pdev->devfn);
134 }
135 
alloc_dev_data(u16 devid)136 static struct iommu_dev_data *alloc_dev_data(u16 devid)
137 {
138 	struct iommu_dev_data *dev_data;
139 	unsigned long flags;
140 
141 	dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
142 	if (!dev_data)
143 		return NULL;
144 
145 	dev_data->devid = devid;
146 
147 	spin_lock_irqsave(&dev_data_list_lock, flags);
148 	list_add_tail(&dev_data->dev_data_list, &dev_data_list);
149 	spin_unlock_irqrestore(&dev_data_list_lock, flags);
150 
151 	return dev_data;
152 }
153 
search_dev_data(u16 devid)154 static struct iommu_dev_data *search_dev_data(u16 devid)
155 {
156 	struct iommu_dev_data *dev_data;
157 	unsigned long flags;
158 
159 	spin_lock_irqsave(&dev_data_list_lock, flags);
160 	list_for_each_entry(dev_data, &dev_data_list, dev_data_list) {
161 		if (dev_data->devid == devid)
162 			goto out_unlock;
163 	}
164 
165 	dev_data = NULL;
166 
167 out_unlock:
168 	spin_unlock_irqrestore(&dev_data_list_lock, flags);
169 
170 	return dev_data;
171 }
172 
__last_alias(struct pci_dev * pdev,u16 alias,void * data)173 static int __last_alias(struct pci_dev *pdev, u16 alias, void *data)
174 {
175 	*(u16 *)data = alias;
176 	return 0;
177 }
178 
get_alias(struct device * dev)179 static u16 get_alias(struct device *dev)
180 {
181 	struct pci_dev *pdev = to_pci_dev(dev);
182 	u16 devid, ivrs_alias, pci_alias;
183 
184 	devid = get_device_id(dev);
185 	ivrs_alias = amd_iommu_alias_table[devid];
186 	pci_for_each_dma_alias(pdev, __last_alias, &pci_alias);
187 
188 	if (ivrs_alias == pci_alias)
189 		return ivrs_alias;
190 
191 	/*
192 	 * DMA alias showdown
193 	 *
194 	 * The IVRS is fairly reliable in telling us about aliases, but it
195 	 * can't know about every screwy device.  If we don't have an IVRS
196 	 * reported alias, use the PCI reported alias.  In that case we may
197 	 * still need to initialize the rlookup and dev_table entries if the
198 	 * alias is to a non-existent device.
199 	 */
200 	if (ivrs_alias == devid) {
201 		if (!amd_iommu_rlookup_table[pci_alias]) {
202 			amd_iommu_rlookup_table[pci_alias] =
203 				amd_iommu_rlookup_table[devid];
204 			memcpy(amd_iommu_dev_table[pci_alias].data,
205 			       amd_iommu_dev_table[devid].data,
206 			       sizeof(amd_iommu_dev_table[pci_alias].data));
207 		}
208 
209 		return pci_alias;
210 	}
211 
212 	pr_info("AMD-Vi: Using IVRS reported alias %02x:%02x.%d "
213 		"for device %s[%04x:%04x], kernel reported alias "
214 		"%02x:%02x.%d\n", PCI_BUS_NUM(ivrs_alias), PCI_SLOT(ivrs_alias),
215 		PCI_FUNC(ivrs_alias), dev_name(dev), pdev->vendor, pdev->device,
216 		PCI_BUS_NUM(pci_alias), PCI_SLOT(pci_alias),
217 		PCI_FUNC(pci_alias));
218 
219 	/*
220 	 * If we don't have a PCI DMA alias and the IVRS alias is on the same
221 	 * bus, then the IVRS table may know about a quirk that we don't.
222 	 */
223 	if (pci_alias == devid &&
224 	    PCI_BUS_NUM(ivrs_alias) == pdev->bus->number) {
225 		pdev->dev_flags |= PCI_DEV_FLAGS_DMA_ALIAS_DEVFN;
226 		pdev->dma_alias_devfn = ivrs_alias & 0xff;
227 		pr_info("AMD-Vi: Added PCI DMA alias %02x.%d for %s\n",
228 			PCI_SLOT(ivrs_alias), PCI_FUNC(ivrs_alias),
229 			dev_name(dev));
230 	}
231 
232 	return ivrs_alias;
233 }
234 
find_dev_data(u16 devid)235 static struct iommu_dev_data *find_dev_data(u16 devid)
236 {
237 	struct iommu_dev_data *dev_data;
238 
239 	dev_data = search_dev_data(devid);
240 
241 	if (dev_data == NULL)
242 		dev_data = alloc_dev_data(devid);
243 
244 	return dev_data;
245 }
246 
get_dev_data(struct device * dev)247 static struct iommu_dev_data *get_dev_data(struct device *dev)
248 {
249 	return dev->archdata.iommu;
250 }
251 
pci_iommuv2_capable(struct pci_dev * pdev)252 static bool pci_iommuv2_capable(struct pci_dev *pdev)
253 {
254 	static const int caps[] = {
255 		PCI_EXT_CAP_ID_ATS,
256 		PCI_EXT_CAP_ID_PRI,
257 		PCI_EXT_CAP_ID_PASID,
258 	};
259 	int i, pos;
260 
261 	for (i = 0; i < 3; ++i) {
262 		pos = pci_find_ext_capability(pdev, caps[i]);
263 		if (pos == 0)
264 			return false;
265 	}
266 
267 	return true;
268 }
269 
pdev_pri_erratum(struct pci_dev * pdev,u32 erratum)270 static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
271 {
272 	struct iommu_dev_data *dev_data;
273 
274 	dev_data = get_dev_data(&pdev->dev);
275 
276 	return dev_data->errata & (1 << erratum) ? true : false;
277 }
278 
279 /*
280  * This function actually applies the mapping to the page table of the
281  * dma_ops domain.
282  */
alloc_unity_mapping(struct dma_ops_domain * dma_dom,struct unity_map_entry * e)283 static void alloc_unity_mapping(struct dma_ops_domain *dma_dom,
284 				struct unity_map_entry *e)
285 {
286 	u64 addr;
287 
288 	for (addr = e->address_start; addr < e->address_end;
289 	     addr += PAGE_SIZE) {
290 		if (addr < dma_dom->aperture_size)
291 			__set_bit(addr >> PAGE_SHIFT,
292 				  dma_dom->aperture[0]->bitmap);
293 	}
294 }
295 
296 /*
297  * Inits the unity mappings required for a specific device
298  */
init_unity_mappings_for_device(struct device * dev,struct dma_ops_domain * dma_dom)299 static void init_unity_mappings_for_device(struct device *dev,
300 					   struct dma_ops_domain *dma_dom)
301 {
302 	struct unity_map_entry *e;
303 	u16 devid;
304 
305 	devid = get_device_id(dev);
306 
307 	list_for_each_entry(e, &amd_iommu_unity_map, list) {
308 		if (!(devid >= e->devid_start && devid <= e->devid_end))
309 			continue;
310 		alloc_unity_mapping(dma_dom, e);
311 	}
312 }
313 
314 /*
315  * This function checks if the driver got a valid device from the caller to
316  * avoid dereferencing invalid pointers.
317  */
check_device(struct device * dev)318 static bool check_device(struct device *dev)
319 {
320 	u16 devid;
321 
322 	if (!dev || !dev->dma_mask)
323 		return false;
324 
325 	/* No PCI device */
326 	if (!dev_is_pci(dev))
327 		return false;
328 
329 	devid = get_device_id(dev);
330 
331 	/* Out of our scope? */
332 	if (devid > amd_iommu_last_bdf)
333 		return false;
334 
335 	if (amd_iommu_rlookup_table[devid] == NULL)
336 		return false;
337 
338 	return true;
339 }
340 
init_iommu_group(struct device * dev)341 static void init_iommu_group(struct device *dev)
342 {
343 	struct dma_ops_domain *dma_domain;
344 	struct iommu_domain *domain;
345 	struct iommu_group *group;
346 
347 	group = iommu_group_get_for_dev(dev);
348 	if (IS_ERR(group))
349 		return;
350 
351 	domain = iommu_group_default_domain(group);
352 	if (!domain)
353 		goto out;
354 
355 	if (to_pdomain(domain)->flags == PD_DMA_OPS_MASK) {
356 		dma_domain = to_pdomain(domain)->priv;
357 		init_unity_mappings_for_device(dev, dma_domain);
358 	}
359 
360 out:
361 	iommu_group_put(group);
362 }
363 
iommu_init_device(struct device * dev)364 static int iommu_init_device(struct device *dev)
365 {
366 	struct pci_dev *pdev = to_pci_dev(dev);
367 	struct iommu_dev_data *dev_data;
368 
369 	if (dev->archdata.iommu)
370 		return 0;
371 
372 	dev_data = find_dev_data(get_device_id(dev));
373 	if (!dev_data)
374 		return -ENOMEM;
375 
376 	dev_data->alias = get_alias(dev);
377 
378 	if (pci_iommuv2_capable(pdev)) {
379 		struct amd_iommu *iommu;
380 
381 		iommu              = amd_iommu_rlookup_table[dev_data->devid];
382 		dev_data->iommu_v2 = iommu->is_iommu_v2;
383 	}
384 
385 	dev->archdata.iommu = dev_data;
386 
387 	iommu_device_link(amd_iommu_rlookup_table[dev_data->devid]->iommu_dev,
388 			  dev);
389 
390 	return 0;
391 }
392 
iommu_ignore_device(struct device * dev)393 static void iommu_ignore_device(struct device *dev)
394 {
395 	u16 devid, alias;
396 
397 	devid = get_device_id(dev);
398 	alias = get_alias(dev);
399 
400 	memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
401 	memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
402 
403 	amd_iommu_rlookup_table[devid] = NULL;
404 	amd_iommu_rlookup_table[alias] = NULL;
405 }
406 
iommu_uninit_device(struct device * dev)407 static void iommu_uninit_device(struct device *dev)
408 {
409 	struct iommu_dev_data *dev_data = search_dev_data(get_device_id(dev));
410 
411 	if (!dev_data)
412 		return;
413 
414 	iommu_device_unlink(amd_iommu_rlookup_table[dev_data->devid]->iommu_dev,
415 			    dev);
416 
417 	iommu_group_remove_device(dev);
418 
419 	/* Remove dma-ops */
420 	dev->archdata.dma_ops = NULL;
421 
422 	/*
423 	 * We keep dev_data around for unplugged devices and reuse it when the
424 	 * device is re-plugged - not doing so would introduce a ton of races.
425 	 */
426 }
427 
428 #ifdef CONFIG_AMD_IOMMU_STATS
429 
430 /*
431  * Initialization code for statistics collection
432  */
433 
434 DECLARE_STATS_COUNTER(compl_wait);
435 DECLARE_STATS_COUNTER(cnt_map_single);
436 DECLARE_STATS_COUNTER(cnt_unmap_single);
437 DECLARE_STATS_COUNTER(cnt_map_sg);
438 DECLARE_STATS_COUNTER(cnt_unmap_sg);
439 DECLARE_STATS_COUNTER(cnt_alloc_coherent);
440 DECLARE_STATS_COUNTER(cnt_free_coherent);
441 DECLARE_STATS_COUNTER(cross_page);
442 DECLARE_STATS_COUNTER(domain_flush_single);
443 DECLARE_STATS_COUNTER(domain_flush_all);
444 DECLARE_STATS_COUNTER(alloced_io_mem);
445 DECLARE_STATS_COUNTER(total_map_requests);
446 DECLARE_STATS_COUNTER(complete_ppr);
447 DECLARE_STATS_COUNTER(invalidate_iotlb);
448 DECLARE_STATS_COUNTER(invalidate_iotlb_all);
449 DECLARE_STATS_COUNTER(pri_requests);
450 
451 static struct dentry *stats_dir;
452 static struct dentry *de_fflush;
453 
amd_iommu_stats_add(struct __iommu_counter * cnt)454 static void amd_iommu_stats_add(struct __iommu_counter *cnt)
455 {
456 	if (stats_dir == NULL)
457 		return;
458 
459 	cnt->dent = debugfs_create_u64(cnt->name, 0444, stats_dir,
460 				       &cnt->value);
461 }
462 
amd_iommu_stats_init(void)463 static void amd_iommu_stats_init(void)
464 {
465 	stats_dir = debugfs_create_dir("amd-iommu", NULL);
466 	if (stats_dir == NULL)
467 		return;
468 
469 	de_fflush  = debugfs_create_bool("fullflush", 0444, stats_dir,
470 					 &amd_iommu_unmap_flush);
471 
472 	amd_iommu_stats_add(&compl_wait);
473 	amd_iommu_stats_add(&cnt_map_single);
474 	amd_iommu_stats_add(&cnt_unmap_single);
475 	amd_iommu_stats_add(&cnt_map_sg);
476 	amd_iommu_stats_add(&cnt_unmap_sg);
477 	amd_iommu_stats_add(&cnt_alloc_coherent);
478 	amd_iommu_stats_add(&cnt_free_coherent);
479 	amd_iommu_stats_add(&cross_page);
480 	amd_iommu_stats_add(&domain_flush_single);
481 	amd_iommu_stats_add(&domain_flush_all);
482 	amd_iommu_stats_add(&alloced_io_mem);
483 	amd_iommu_stats_add(&total_map_requests);
484 	amd_iommu_stats_add(&complete_ppr);
485 	amd_iommu_stats_add(&invalidate_iotlb);
486 	amd_iommu_stats_add(&invalidate_iotlb_all);
487 	amd_iommu_stats_add(&pri_requests);
488 }
489 
490 #endif
491 
492 /****************************************************************************
493  *
494  * Interrupt handling functions
495  *
496  ****************************************************************************/
497 
dump_dte_entry(u16 devid)498 static void dump_dte_entry(u16 devid)
499 {
500 	int i;
501 
502 	for (i = 0; i < 4; ++i)
503 		pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
504 			amd_iommu_dev_table[devid].data[i]);
505 }
506 
dump_command(unsigned long phys_addr)507 static void dump_command(unsigned long phys_addr)
508 {
509 	struct iommu_cmd *cmd = phys_to_virt(phys_addr);
510 	int i;
511 
512 	for (i = 0; i < 4; ++i)
513 		pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
514 }
515 
iommu_print_event(struct amd_iommu * iommu,void * __evt)516 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
517 {
518 	int type, devid, domid, flags;
519 	volatile u32 *event = __evt;
520 	int count = 0;
521 	u64 address;
522 
523 retry:
524 	type    = (event[1] >> EVENT_TYPE_SHIFT)  & EVENT_TYPE_MASK;
525 	devid   = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
526 	domid   = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
527 	flags   = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
528 	address = (u64)(((u64)event[3]) << 32) | event[2];
529 
530 	if (type == 0) {
531 		/* Did we hit the erratum? */
532 		if (++count == LOOP_TIMEOUT) {
533 			pr_err("AMD-Vi: No event written to event log\n");
534 			return;
535 		}
536 		udelay(1);
537 		goto retry;
538 	}
539 
540 	printk(KERN_ERR "AMD-Vi: Event logged [");
541 
542 	switch (type) {
543 	case EVENT_TYPE_ILL_DEV:
544 		printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
545 		       "address=0x%016llx flags=0x%04x]\n",
546 		       PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
547 		       address, flags);
548 		dump_dte_entry(devid);
549 		break;
550 	case EVENT_TYPE_IO_FAULT:
551 		printk("IO_PAGE_FAULT device=%02x:%02x.%x "
552 		       "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
553 		       PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
554 		       domid, address, flags);
555 		break;
556 	case EVENT_TYPE_DEV_TAB_ERR:
557 		printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
558 		       "address=0x%016llx flags=0x%04x]\n",
559 		       PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
560 		       address, flags);
561 		break;
562 	case EVENT_TYPE_PAGE_TAB_ERR:
563 		printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
564 		       "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
565 		       PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
566 		       domid, address, flags);
567 		break;
568 	case EVENT_TYPE_ILL_CMD:
569 		printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
570 		dump_command(address);
571 		break;
572 	case EVENT_TYPE_CMD_HARD_ERR:
573 		printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
574 		       "flags=0x%04x]\n", address, flags);
575 		break;
576 	case EVENT_TYPE_IOTLB_INV_TO:
577 		printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
578 		       "address=0x%016llx]\n",
579 		       PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
580 		       address);
581 		break;
582 	case EVENT_TYPE_INV_DEV_REQ:
583 		printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
584 		       "address=0x%016llx flags=0x%04x]\n",
585 		       PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
586 		       address, flags);
587 		break;
588 	default:
589 		printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
590 	}
591 
592 	memset(__evt, 0, 4 * sizeof(u32));
593 }
594 
iommu_poll_events(struct amd_iommu * iommu)595 static void iommu_poll_events(struct amd_iommu *iommu)
596 {
597 	u32 head, tail;
598 
599 	head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
600 	tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
601 
602 	while (head != tail) {
603 		iommu_print_event(iommu, iommu->evt_buf + head);
604 		head = (head + EVENT_ENTRY_SIZE) % EVT_BUFFER_SIZE;
605 	}
606 
607 	writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
608 }
609 
iommu_handle_ppr_entry(struct amd_iommu * iommu,u64 * raw)610 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
611 {
612 	struct amd_iommu_fault fault;
613 
614 	INC_STATS_COUNTER(pri_requests);
615 
616 	if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
617 		pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
618 		return;
619 	}
620 
621 	fault.address   = raw[1];
622 	fault.pasid     = PPR_PASID(raw[0]);
623 	fault.device_id = PPR_DEVID(raw[0]);
624 	fault.tag       = PPR_TAG(raw[0]);
625 	fault.flags     = PPR_FLAGS(raw[0]);
626 
627 	atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
628 }
629 
iommu_poll_ppr_log(struct amd_iommu * iommu)630 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
631 {
632 	u32 head, tail;
633 
634 	if (iommu->ppr_log == NULL)
635 		return;
636 
637 	head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
638 	tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
639 
640 	while (head != tail) {
641 		volatile u64 *raw;
642 		u64 entry[2];
643 		int i;
644 
645 		raw = (u64 *)(iommu->ppr_log + head);
646 
647 		/*
648 		 * Hardware bug: Interrupt may arrive before the entry is
649 		 * written to memory. If this happens we need to wait for the
650 		 * entry to arrive.
651 		 */
652 		for (i = 0; i < LOOP_TIMEOUT; ++i) {
653 			if (PPR_REQ_TYPE(raw[0]) != 0)
654 				break;
655 			udelay(1);
656 		}
657 
658 		/* Avoid memcpy function-call overhead */
659 		entry[0] = raw[0];
660 		entry[1] = raw[1];
661 
662 		/*
663 		 * To detect the hardware bug we need to clear the entry
664 		 * back to zero.
665 		 */
666 		raw[0] = raw[1] = 0UL;
667 
668 		/* Update head pointer of hardware ring-buffer */
669 		head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
670 		writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
671 
672 		/* Handle PPR entry */
673 		iommu_handle_ppr_entry(iommu, entry);
674 
675 		/* Refresh ring-buffer information */
676 		head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
677 		tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
678 	}
679 }
680 
amd_iommu_int_thread(int irq,void * data)681 irqreturn_t amd_iommu_int_thread(int irq, void *data)
682 {
683 	struct amd_iommu *iommu = (struct amd_iommu *) data;
684 	u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
685 
686 	while (status & (MMIO_STATUS_EVT_INT_MASK | MMIO_STATUS_PPR_INT_MASK)) {
687 		/* Enable EVT and PPR interrupts again */
688 		writel((MMIO_STATUS_EVT_INT_MASK | MMIO_STATUS_PPR_INT_MASK),
689 			iommu->mmio_base + MMIO_STATUS_OFFSET);
690 
691 		if (status & MMIO_STATUS_EVT_INT_MASK) {
692 			pr_devel("AMD-Vi: Processing IOMMU Event Log\n");
693 			iommu_poll_events(iommu);
694 		}
695 
696 		if (status & MMIO_STATUS_PPR_INT_MASK) {
697 			pr_devel("AMD-Vi: Processing IOMMU PPR Log\n");
698 			iommu_poll_ppr_log(iommu);
699 		}
700 
701 		/*
702 		 * Hardware bug: ERBT1312
703 		 * When re-enabling interrupt (by writing 1
704 		 * to clear the bit), the hardware might also try to set
705 		 * the interrupt bit in the event status register.
706 		 * In this scenario, the bit will be set, and disable
707 		 * subsequent interrupts.
708 		 *
709 		 * Workaround: The IOMMU driver should read back the
710 		 * status register and check if the interrupt bits are cleared.
711 		 * If not, driver will need to go through the interrupt handler
712 		 * again and re-clear the bits
713 		 */
714 		status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
715 	}
716 	return IRQ_HANDLED;
717 }
718 
amd_iommu_int_handler(int irq,void * data)719 irqreturn_t amd_iommu_int_handler(int irq, void *data)
720 {
721 	return IRQ_WAKE_THREAD;
722 }
723 
724 /****************************************************************************
725  *
726  * IOMMU command queuing functions
727  *
728  ****************************************************************************/
729 
wait_on_sem(volatile u64 * sem)730 static int wait_on_sem(volatile u64 *sem)
731 {
732 	int i = 0;
733 
734 	while (*sem == 0 && i < LOOP_TIMEOUT) {
735 		udelay(1);
736 		i += 1;
737 	}
738 
739 	if (i == LOOP_TIMEOUT) {
740 		pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
741 		return -EIO;
742 	}
743 
744 	return 0;
745 }
746 
copy_cmd_to_buffer(struct amd_iommu * iommu,struct iommu_cmd * cmd,u32 tail)747 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
748 			       struct iommu_cmd *cmd,
749 			       u32 tail)
750 {
751 	u8 *target;
752 
753 	target = iommu->cmd_buf + tail;
754 	tail   = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
755 
756 	/* Copy command to buffer */
757 	memcpy(target, cmd, sizeof(*cmd));
758 
759 	/* Tell the IOMMU about it */
760 	writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
761 }
762 
build_completion_wait(struct iommu_cmd * cmd,u64 address)763 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
764 {
765 	WARN_ON(address & 0x7ULL);
766 
767 	memset(cmd, 0, sizeof(*cmd));
768 	cmd->data[0] = lower_32_bits(__pa(address)) | CMD_COMPL_WAIT_STORE_MASK;
769 	cmd->data[1] = upper_32_bits(__pa(address));
770 	cmd->data[2] = 1;
771 	CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
772 }
773 
build_inv_dte(struct iommu_cmd * cmd,u16 devid)774 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
775 {
776 	memset(cmd, 0, sizeof(*cmd));
777 	cmd->data[0] = devid;
778 	CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
779 }
780 
build_inv_iommu_pages(struct iommu_cmd * cmd,u64 address,size_t size,u16 domid,int pde)781 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
782 				  size_t size, u16 domid, int pde)
783 {
784 	u64 pages;
785 	bool s;
786 
787 	pages = iommu_num_pages(address, size, PAGE_SIZE);
788 	s     = false;
789 
790 	if (pages > 1) {
791 		/*
792 		 * If we have to flush more than one page, flush all
793 		 * TLB entries for this domain
794 		 */
795 		address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
796 		s = true;
797 	}
798 
799 	address &= PAGE_MASK;
800 
801 	memset(cmd, 0, sizeof(*cmd));
802 	cmd->data[1] |= domid;
803 	cmd->data[2]  = lower_32_bits(address);
804 	cmd->data[3]  = upper_32_bits(address);
805 	CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
806 	if (s) /* size bit - we flush more than one 4kb page */
807 		cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
808 	if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
809 		cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
810 }
811 
build_inv_iotlb_pages(struct iommu_cmd * cmd,u16 devid,int qdep,u64 address,size_t size)812 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
813 				  u64 address, size_t size)
814 {
815 	u64 pages;
816 	bool s;
817 
818 	pages = iommu_num_pages(address, size, PAGE_SIZE);
819 	s     = false;
820 
821 	if (pages > 1) {
822 		/*
823 		 * If we have to flush more than one page, flush all
824 		 * TLB entries for this domain
825 		 */
826 		address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
827 		s = true;
828 	}
829 
830 	address &= PAGE_MASK;
831 
832 	memset(cmd, 0, sizeof(*cmd));
833 	cmd->data[0]  = devid;
834 	cmd->data[0] |= (qdep & 0xff) << 24;
835 	cmd->data[1]  = devid;
836 	cmd->data[2]  = lower_32_bits(address);
837 	cmd->data[3]  = upper_32_bits(address);
838 	CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
839 	if (s)
840 		cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
841 }
842 
build_inv_iommu_pasid(struct iommu_cmd * cmd,u16 domid,int pasid,u64 address,bool size)843 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
844 				  u64 address, bool size)
845 {
846 	memset(cmd, 0, sizeof(*cmd));
847 
848 	address &= ~(0xfffULL);
849 
850 	cmd->data[0]  = pasid;
851 	cmd->data[1]  = domid;
852 	cmd->data[2]  = lower_32_bits(address);
853 	cmd->data[3]  = upper_32_bits(address);
854 	cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
855 	cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
856 	if (size)
857 		cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
858 	CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
859 }
860 
build_inv_iotlb_pasid(struct iommu_cmd * cmd,u16 devid,int pasid,int qdep,u64 address,bool size)861 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
862 				  int qdep, u64 address, bool size)
863 {
864 	memset(cmd, 0, sizeof(*cmd));
865 
866 	address &= ~(0xfffULL);
867 
868 	cmd->data[0]  = devid;
869 	cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
870 	cmd->data[0] |= (qdep  & 0xff) << 24;
871 	cmd->data[1]  = devid;
872 	cmd->data[1] |= (pasid & 0xff) << 16;
873 	cmd->data[2]  = lower_32_bits(address);
874 	cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
875 	cmd->data[3]  = upper_32_bits(address);
876 	if (size)
877 		cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
878 	CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
879 }
880 
build_complete_ppr(struct iommu_cmd * cmd,u16 devid,int pasid,int status,int tag,bool gn)881 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
882 			       int status, int tag, bool gn)
883 {
884 	memset(cmd, 0, sizeof(*cmd));
885 
886 	cmd->data[0]  = devid;
887 	if (gn) {
888 		cmd->data[1]  = pasid;
889 		cmd->data[2]  = CMD_INV_IOMMU_PAGES_GN_MASK;
890 	}
891 	cmd->data[3]  = tag & 0x1ff;
892 	cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
893 
894 	CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
895 }
896 
build_inv_all(struct iommu_cmd * cmd)897 static void build_inv_all(struct iommu_cmd *cmd)
898 {
899 	memset(cmd, 0, sizeof(*cmd));
900 	CMD_SET_TYPE(cmd, CMD_INV_ALL);
901 }
902 
build_inv_irt(struct iommu_cmd * cmd,u16 devid)903 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
904 {
905 	memset(cmd, 0, sizeof(*cmd));
906 	cmd->data[0] = devid;
907 	CMD_SET_TYPE(cmd, CMD_INV_IRT);
908 }
909 
910 /*
911  * Writes the command to the IOMMUs command buffer and informs the
912  * hardware about the new command.
913  */
iommu_queue_command_sync(struct amd_iommu * iommu,struct iommu_cmd * cmd,bool sync)914 static int iommu_queue_command_sync(struct amd_iommu *iommu,
915 				    struct iommu_cmd *cmd,
916 				    bool sync)
917 {
918 	u32 left, tail, head, next_tail;
919 	unsigned long flags;
920 
921 again:
922 	spin_lock_irqsave(&iommu->lock, flags);
923 
924 	head      = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
925 	tail      = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
926 	next_tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
927 	left      = (head - next_tail) % CMD_BUFFER_SIZE;
928 
929 	if (left <= 0x20) {
930 		struct iommu_cmd sync_cmd;
931 		volatile u64 sem = 0;
932 		int ret;
933 
934 		build_completion_wait(&sync_cmd, (u64)&sem);
935 		copy_cmd_to_buffer(iommu, &sync_cmd, tail);
936 
937 		spin_unlock_irqrestore(&iommu->lock, flags);
938 
939 		if ((ret = wait_on_sem(&sem)) != 0)
940 			return ret;
941 
942 		goto again;
943 	}
944 
945 	copy_cmd_to_buffer(iommu, cmd, tail);
946 
947 	/* We need to sync now to make sure all commands are processed */
948 	iommu->need_sync = sync;
949 
950 	spin_unlock_irqrestore(&iommu->lock, flags);
951 
952 	return 0;
953 }
954 
iommu_queue_command(struct amd_iommu * iommu,struct iommu_cmd * cmd)955 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
956 {
957 	return iommu_queue_command_sync(iommu, cmd, true);
958 }
959 
960 /*
961  * This function queues a completion wait command into the command
962  * buffer of an IOMMU
963  */
iommu_completion_wait(struct amd_iommu * iommu)964 static int iommu_completion_wait(struct amd_iommu *iommu)
965 {
966 	struct iommu_cmd cmd;
967 	volatile u64 sem = 0;
968 	int ret;
969 
970 	if (!iommu->need_sync)
971 		return 0;
972 
973 	build_completion_wait(&cmd, (u64)&sem);
974 
975 	ret = iommu_queue_command_sync(iommu, &cmd, false);
976 	if (ret)
977 		return ret;
978 
979 	return wait_on_sem(&sem);
980 }
981 
iommu_flush_dte(struct amd_iommu * iommu,u16 devid)982 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
983 {
984 	struct iommu_cmd cmd;
985 
986 	build_inv_dte(&cmd, devid);
987 
988 	return iommu_queue_command(iommu, &cmd);
989 }
990 
iommu_flush_dte_all(struct amd_iommu * iommu)991 static void iommu_flush_dte_all(struct amd_iommu *iommu)
992 {
993 	u32 devid;
994 
995 	for (devid = 0; devid <= 0xffff; ++devid)
996 		iommu_flush_dte(iommu, devid);
997 
998 	iommu_completion_wait(iommu);
999 }
1000 
1001 /*
1002  * This function uses heavy locking and may disable irqs for some time. But
1003  * this is no issue because it is only called during resume.
1004  */
iommu_flush_tlb_all(struct amd_iommu * iommu)1005 static void iommu_flush_tlb_all(struct amd_iommu *iommu)
1006 {
1007 	u32 dom_id;
1008 
1009 	for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
1010 		struct iommu_cmd cmd;
1011 		build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1012 				      dom_id, 1);
1013 		iommu_queue_command(iommu, &cmd);
1014 	}
1015 
1016 	iommu_completion_wait(iommu);
1017 }
1018 
iommu_flush_all(struct amd_iommu * iommu)1019 static void iommu_flush_all(struct amd_iommu *iommu)
1020 {
1021 	struct iommu_cmd cmd;
1022 
1023 	build_inv_all(&cmd);
1024 
1025 	iommu_queue_command(iommu, &cmd);
1026 	iommu_completion_wait(iommu);
1027 }
1028 
iommu_flush_irt(struct amd_iommu * iommu,u16 devid)1029 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1030 {
1031 	struct iommu_cmd cmd;
1032 
1033 	build_inv_irt(&cmd, devid);
1034 
1035 	iommu_queue_command(iommu, &cmd);
1036 }
1037 
iommu_flush_irt_all(struct amd_iommu * iommu)1038 static void iommu_flush_irt_all(struct amd_iommu *iommu)
1039 {
1040 	u32 devid;
1041 
1042 	for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
1043 		iommu_flush_irt(iommu, devid);
1044 
1045 	iommu_completion_wait(iommu);
1046 }
1047 
iommu_flush_all_caches(struct amd_iommu * iommu)1048 void iommu_flush_all_caches(struct amd_iommu *iommu)
1049 {
1050 	if (iommu_feature(iommu, FEATURE_IA)) {
1051 		iommu_flush_all(iommu);
1052 	} else {
1053 		iommu_flush_dte_all(iommu);
1054 		iommu_flush_irt_all(iommu);
1055 		iommu_flush_tlb_all(iommu);
1056 	}
1057 }
1058 
1059 /*
1060  * Command send function for flushing on-device TLB
1061  */
device_flush_iotlb(struct iommu_dev_data * dev_data,u64 address,size_t size)1062 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1063 			      u64 address, size_t size)
1064 {
1065 	struct amd_iommu *iommu;
1066 	struct iommu_cmd cmd;
1067 	int qdep;
1068 
1069 	qdep     = dev_data->ats.qdep;
1070 	iommu    = amd_iommu_rlookup_table[dev_data->devid];
1071 
1072 	build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1073 
1074 	return iommu_queue_command(iommu, &cmd);
1075 }
1076 
1077 /*
1078  * Command send function for invalidating a device table entry
1079  */
device_flush_dte(struct iommu_dev_data * dev_data)1080 static int device_flush_dte(struct iommu_dev_data *dev_data)
1081 {
1082 	struct amd_iommu *iommu;
1083 	u16 alias;
1084 	int ret;
1085 
1086 	iommu = amd_iommu_rlookup_table[dev_data->devid];
1087 	alias = dev_data->alias;
1088 
1089 	ret = iommu_flush_dte(iommu, dev_data->devid);
1090 	if (!ret && alias != dev_data->devid)
1091 		ret = iommu_flush_dte(iommu, alias);
1092 	if (ret)
1093 		return ret;
1094 
1095 	if (dev_data->ats.enabled)
1096 		ret = device_flush_iotlb(dev_data, 0, ~0UL);
1097 
1098 	return ret;
1099 }
1100 
1101 /*
1102  * TLB invalidation function which is called from the mapping functions.
1103  * It invalidates a single PTE if the range to flush is within a single
1104  * page. Otherwise it flushes the whole TLB of the IOMMU.
1105  */
__domain_flush_pages(struct protection_domain * domain,u64 address,size_t size,int pde)1106 static void __domain_flush_pages(struct protection_domain *domain,
1107 				 u64 address, size_t size, int pde)
1108 {
1109 	struct iommu_dev_data *dev_data;
1110 	struct iommu_cmd cmd;
1111 	int ret = 0, i;
1112 
1113 	build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1114 
1115 	for (i = 0; i < amd_iommus_present; ++i) {
1116 		if (!domain->dev_iommu[i])
1117 			continue;
1118 
1119 		/*
1120 		 * Devices of this domain are behind this IOMMU
1121 		 * We need a TLB flush
1122 		 */
1123 		ret |= iommu_queue_command(amd_iommus[i], &cmd);
1124 	}
1125 
1126 	list_for_each_entry(dev_data, &domain->dev_list, list) {
1127 
1128 		if (!dev_data->ats.enabled)
1129 			continue;
1130 
1131 		ret |= device_flush_iotlb(dev_data, address, size);
1132 	}
1133 
1134 	WARN_ON(ret);
1135 }
1136 
domain_flush_pages(struct protection_domain * domain,u64 address,size_t size)1137 static void domain_flush_pages(struct protection_domain *domain,
1138 			       u64 address, size_t size)
1139 {
1140 	__domain_flush_pages(domain, address, size, 0);
1141 }
1142 
1143 /* Flush the whole IO/TLB for a given protection domain */
domain_flush_tlb(struct protection_domain * domain)1144 static void domain_flush_tlb(struct protection_domain *domain)
1145 {
1146 	__domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1147 }
1148 
1149 /* Flush the whole IO/TLB for a given protection domain - including PDE */
domain_flush_tlb_pde(struct protection_domain * domain)1150 static void domain_flush_tlb_pde(struct protection_domain *domain)
1151 {
1152 	__domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1153 }
1154 
domain_flush_complete(struct protection_domain * domain)1155 static void domain_flush_complete(struct protection_domain *domain)
1156 {
1157 	int i;
1158 
1159 	for (i = 0; i < amd_iommus_present; ++i) {
1160 		if (!domain->dev_iommu[i])
1161 			continue;
1162 
1163 		/*
1164 		 * Devices of this domain are behind this IOMMU
1165 		 * We need to wait for completion of all commands.
1166 		 */
1167 		iommu_completion_wait(amd_iommus[i]);
1168 	}
1169 }
1170 
1171 
1172 /*
1173  * This function flushes the DTEs for all devices in domain
1174  */
domain_flush_devices(struct protection_domain * domain)1175 static void domain_flush_devices(struct protection_domain *domain)
1176 {
1177 	struct iommu_dev_data *dev_data;
1178 
1179 	list_for_each_entry(dev_data, &domain->dev_list, list)
1180 		device_flush_dte(dev_data);
1181 }
1182 
1183 /****************************************************************************
1184  *
1185  * The functions below are used the create the page table mappings for
1186  * unity mapped regions.
1187  *
1188  ****************************************************************************/
1189 
1190 /*
1191  * This function is used to add another level to an IO page table. Adding
1192  * another level increases the size of the address space by 9 bits to a size up
1193  * to 64 bits.
1194  */
increase_address_space(struct protection_domain * domain,gfp_t gfp)1195 static bool increase_address_space(struct protection_domain *domain,
1196 				   gfp_t gfp)
1197 {
1198 	u64 *pte;
1199 
1200 	if (domain->mode == PAGE_MODE_6_LEVEL)
1201 		/* address space already 64 bit large */
1202 		return false;
1203 
1204 	pte = (void *)get_zeroed_page(gfp);
1205 	if (!pte)
1206 		return false;
1207 
1208 	*pte             = PM_LEVEL_PDE(domain->mode,
1209 					virt_to_phys(domain->pt_root));
1210 	domain->pt_root  = pte;
1211 	domain->mode    += 1;
1212 	domain->updated  = true;
1213 
1214 	return true;
1215 }
1216 
alloc_pte(struct protection_domain * domain,unsigned long address,unsigned long page_size,u64 ** pte_page,gfp_t gfp)1217 static u64 *alloc_pte(struct protection_domain *domain,
1218 		      unsigned long address,
1219 		      unsigned long page_size,
1220 		      u64 **pte_page,
1221 		      gfp_t gfp)
1222 {
1223 	int level, end_lvl;
1224 	u64 *pte, *page;
1225 
1226 	BUG_ON(!is_power_of_2(page_size));
1227 
1228 	while (address > PM_LEVEL_SIZE(domain->mode))
1229 		increase_address_space(domain, gfp);
1230 
1231 	level   = domain->mode - 1;
1232 	pte     = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1233 	address = PAGE_SIZE_ALIGN(address, page_size);
1234 	end_lvl = PAGE_SIZE_LEVEL(page_size);
1235 
1236 	while (level > end_lvl) {
1237 		if (!IOMMU_PTE_PRESENT(*pte)) {
1238 			page = (u64 *)get_zeroed_page(gfp);
1239 			if (!page)
1240 				return NULL;
1241 			*pte = PM_LEVEL_PDE(level, virt_to_phys(page));
1242 		}
1243 
1244 		/* No level skipping support yet */
1245 		if (PM_PTE_LEVEL(*pte) != level)
1246 			return NULL;
1247 
1248 		level -= 1;
1249 
1250 		pte = IOMMU_PTE_PAGE(*pte);
1251 
1252 		if (pte_page && level == end_lvl)
1253 			*pte_page = pte;
1254 
1255 		pte = &pte[PM_LEVEL_INDEX(level, address)];
1256 	}
1257 
1258 	return pte;
1259 }
1260 
1261 /*
1262  * This function checks if there is a PTE for a given dma address. If
1263  * there is one, it returns the pointer to it.
1264  */
fetch_pte(struct protection_domain * domain,unsigned long address,unsigned long * page_size)1265 static u64 *fetch_pte(struct protection_domain *domain,
1266 		      unsigned long address,
1267 		      unsigned long *page_size)
1268 {
1269 	int level;
1270 	u64 *pte;
1271 
1272 	if (address > PM_LEVEL_SIZE(domain->mode))
1273 		return NULL;
1274 
1275 	level	   =  domain->mode - 1;
1276 	pte	   = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1277 	*page_size =  PTE_LEVEL_PAGE_SIZE(level);
1278 
1279 	while (level > 0) {
1280 
1281 		/* Not Present */
1282 		if (!IOMMU_PTE_PRESENT(*pte))
1283 			return NULL;
1284 
1285 		/* Large PTE */
1286 		if (PM_PTE_LEVEL(*pte) == 7 ||
1287 		    PM_PTE_LEVEL(*pte) == 0)
1288 			break;
1289 
1290 		/* No level skipping support yet */
1291 		if (PM_PTE_LEVEL(*pte) != level)
1292 			return NULL;
1293 
1294 		level -= 1;
1295 
1296 		/* Walk to the next level */
1297 		pte	   = IOMMU_PTE_PAGE(*pte);
1298 		pte	   = &pte[PM_LEVEL_INDEX(level, address)];
1299 		*page_size = PTE_LEVEL_PAGE_SIZE(level);
1300 	}
1301 
1302 	if (PM_PTE_LEVEL(*pte) == 0x07) {
1303 		unsigned long pte_mask;
1304 
1305 		/*
1306 		 * If we have a series of large PTEs, make
1307 		 * sure to return a pointer to the first one.
1308 		 */
1309 		*page_size = pte_mask = PTE_PAGE_SIZE(*pte);
1310 		pte_mask   = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1311 		pte        = (u64 *)(((unsigned long)pte) & pte_mask);
1312 	}
1313 
1314 	return pte;
1315 }
1316 
1317 /*
1318  * Generic mapping functions. It maps a physical address into a DMA
1319  * address space. It allocates the page table pages if necessary.
1320  * In the future it can be extended to a generic mapping function
1321  * supporting all features of AMD IOMMU page tables like level skipping
1322  * and full 64 bit address spaces.
1323  */
iommu_map_page(struct protection_domain * dom,unsigned long bus_addr,unsigned long phys_addr,int prot,unsigned long page_size)1324 static int iommu_map_page(struct protection_domain *dom,
1325 			  unsigned long bus_addr,
1326 			  unsigned long phys_addr,
1327 			  int prot,
1328 			  unsigned long page_size)
1329 {
1330 	u64 __pte, *pte;
1331 	int i, count;
1332 
1333 	BUG_ON(!IS_ALIGNED(bus_addr, page_size));
1334 	BUG_ON(!IS_ALIGNED(phys_addr, page_size));
1335 
1336 	if (!(prot & IOMMU_PROT_MASK))
1337 		return -EINVAL;
1338 
1339 	count = PAGE_SIZE_PTE_COUNT(page_size);
1340 	pte   = alloc_pte(dom, bus_addr, page_size, NULL, GFP_KERNEL);
1341 
1342 	if (!pte)
1343 		return -ENOMEM;
1344 
1345 	for (i = 0; i < count; ++i)
1346 		if (IOMMU_PTE_PRESENT(pte[i]))
1347 			return -EBUSY;
1348 
1349 	if (count > 1) {
1350 		__pte = PAGE_SIZE_PTE(phys_addr, page_size);
1351 		__pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_P | IOMMU_PTE_FC;
1352 	} else
1353 		__pte = phys_addr | IOMMU_PTE_P | IOMMU_PTE_FC;
1354 
1355 	if (prot & IOMMU_PROT_IR)
1356 		__pte |= IOMMU_PTE_IR;
1357 	if (prot & IOMMU_PROT_IW)
1358 		__pte |= IOMMU_PTE_IW;
1359 
1360 	for (i = 0; i < count; ++i)
1361 		pte[i] = __pte;
1362 
1363 	update_domain(dom);
1364 
1365 	return 0;
1366 }
1367 
iommu_unmap_page(struct protection_domain * dom,unsigned long bus_addr,unsigned long page_size)1368 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1369 				      unsigned long bus_addr,
1370 				      unsigned long page_size)
1371 {
1372 	unsigned long long unmapped;
1373 	unsigned long unmap_size;
1374 	u64 *pte;
1375 
1376 	BUG_ON(!is_power_of_2(page_size));
1377 
1378 	unmapped = 0;
1379 
1380 	while (unmapped < page_size) {
1381 
1382 		pte = fetch_pte(dom, bus_addr, &unmap_size);
1383 
1384 		if (pte) {
1385 			int i, count;
1386 
1387 			count = PAGE_SIZE_PTE_COUNT(unmap_size);
1388 			for (i = 0; i < count; i++)
1389 				pte[i] = 0ULL;
1390 		}
1391 
1392 		bus_addr  = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1393 		unmapped += unmap_size;
1394 	}
1395 
1396 	BUG_ON(unmapped && !is_power_of_2(unmapped));
1397 
1398 	return unmapped;
1399 }
1400 
1401 /****************************************************************************
1402  *
1403  * The next functions belong to the address allocator for the dma_ops
1404  * interface functions. They work like the allocators in the other IOMMU
1405  * drivers. Its basically a bitmap which marks the allocated pages in
1406  * the aperture. Maybe it could be enhanced in the future to a more
1407  * efficient allocator.
1408  *
1409  ****************************************************************************/
1410 
1411 /*
1412  * The address allocator core functions.
1413  *
1414  * called with domain->lock held
1415  */
1416 
1417 /*
1418  * Used to reserve address ranges in the aperture (e.g. for exclusion
1419  * ranges.
1420  */
dma_ops_reserve_addresses(struct dma_ops_domain * dom,unsigned long start_page,unsigned int pages)1421 static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
1422 				      unsigned long start_page,
1423 				      unsigned int pages)
1424 {
1425 	unsigned int i, last_page = dom->aperture_size >> PAGE_SHIFT;
1426 
1427 	if (start_page + pages > last_page)
1428 		pages = last_page - start_page;
1429 
1430 	for (i = start_page; i < start_page + pages; ++i) {
1431 		int index = i / APERTURE_RANGE_PAGES;
1432 		int page  = i % APERTURE_RANGE_PAGES;
1433 		__set_bit(page, dom->aperture[index]->bitmap);
1434 	}
1435 }
1436 
1437 /*
1438  * This function is used to add a new aperture range to an existing
1439  * aperture in case of dma_ops domain allocation or address allocation
1440  * failure.
1441  */
alloc_new_range(struct dma_ops_domain * dma_dom,bool populate,gfp_t gfp)1442 static int alloc_new_range(struct dma_ops_domain *dma_dom,
1443 			   bool populate, gfp_t gfp)
1444 {
1445 	int index = dma_dom->aperture_size >> APERTURE_RANGE_SHIFT;
1446 	struct amd_iommu *iommu;
1447 	unsigned long i, old_size, pte_pgsize;
1448 
1449 #ifdef CONFIG_IOMMU_STRESS
1450 	populate = false;
1451 #endif
1452 
1453 	if (index >= APERTURE_MAX_RANGES)
1454 		return -ENOMEM;
1455 
1456 	dma_dom->aperture[index] = kzalloc(sizeof(struct aperture_range), gfp);
1457 	if (!dma_dom->aperture[index])
1458 		return -ENOMEM;
1459 
1460 	dma_dom->aperture[index]->bitmap = (void *)get_zeroed_page(gfp);
1461 	if (!dma_dom->aperture[index]->bitmap)
1462 		goto out_free;
1463 
1464 	dma_dom->aperture[index]->offset = dma_dom->aperture_size;
1465 
1466 	if (populate) {
1467 		unsigned long address = dma_dom->aperture_size;
1468 		int i, num_ptes = APERTURE_RANGE_PAGES / 512;
1469 		u64 *pte, *pte_page;
1470 
1471 		for (i = 0; i < num_ptes; ++i) {
1472 			pte = alloc_pte(&dma_dom->domain, address, PAGE_SIZE,
1473 					&pte_page, gfp);
1474 			if (!pte)
1475 				goto out_free;
1476 
1477 			dma_dom->aperture[index]->pte_pages[i] = pte_page;
1478 
1479 			address += APERTURE_RANGE_SIZE / 64;
1480 		}
1481 	}
1482 
1483 	old_size                = dma_dom->aperture_size;
1484 	dma_dom->aperture_size += APERTURE_RANGE_SIZE;
1485 
1486 	/* Reserve address range used for MSI messages */
1487 	if (old_size < MSI_ADDR_BASE_LO &&
1488 	    dma_dom->aperture_size > MSI_ADDR_BASE_LO) {
1489 		unsigned long spage;
1490 		int pages;
1491 
1492 		pages = iommu_num_pages(MSI_ADDR_BASE_LO, 0x10000, PAGE_SIZE);
1493 		spage = MSI_ADDR_BASE_LO >> PAGE_SHIFT;
1494 
1495 		dma_ops_reserve_addresses(dma_dom, spage, pages);
1496 	}
1497 
1498 	/* Initialize the exclusion range if necessary */
1499 	for_each_iommu(iommu) {
1500 		if (iommu->exclusion_start &&
1501 		    iommu->exclusion_start >= dma_dom->aperture[index]->offset
1502 		    && iommu->exclusion_start < dma_dom->aperture_size) {
1503 			unsigned long startpage;
1504 			int pages = iommu_num_pages(iommu->exclusion_start,
1505 						    iommu->exclusion_length,
1506 						    PAGE_SIZE);
1507 			startpage = iommu->exclusion_start >> PAGE_SHIFT;
1508 			dma_ops_reserve_addresses(dma_dom, startpage, pages);
1509 		}
1510 	}
1511 
1512 	/*
1513 	 * Check for areas already mapped as present in the new aperture
1514 	 * range and mark those pages as reserved in the allocator. Such
1515 	 * mappings may already exist as a result of requested unity
1516 	 * mappings for devices.
1517 	 */
1518 	for (i = dma_dom->aperture[index]->offset;
1519 	     i < dma_dom->aperture_size;
1520 	     i += pte_pgsize) {
1521 		u64 *pte = fetch_pte(&dma_dom->domain, i, &pte_pgsize);
1522 		if (!pte || !IOMMU_PTE_PRESENT(*pte))
1523 			continue;
1524 
1525 		dma_ops_reserve_addresses(dma_dom, i >> PAGE_SHIFT,
1526 					  pte_pgsize >> 12);
1527 	}
1528 
1529 	update_domain(&dma_dom->domain);
1530 
1531 	return 0;
1532 
1533 out_free:
1534 	update_domain(&dma_dom->domain);
1535 
1536 	free_page((unsigned long)dma_dom->aperture[index]->bitmap);
1537 
1538 	kfree(dma_dom->aperture[index]);
1539 	dma_dom->aperture[index] = NULL;
1540 
1541 	return -ENOMEM;
1542 }
1543 
dma_ops_area_alloc(struct device * dev,struct dma_ops_domain * dom,unsigned int pages,unsigned long align_mask,u64 dma_mask,unsigned long start)1544 static unsigned long dma_ops_area_alloc(struct device *dev,
1545 					struct dma_ops_domain *dom,
1546 					unsigned int pages,
1547 					unsigned long align_mask,
1548 					u64 dma_mask,
1549 					unsigned long start)
1550 {
1551 	unsigned long next_bit = dom->next_address % APERTURE_RANGE_SIZE;
1552 	int max_index = dom->aperture_size >> APERTURE_RANGE_SHIFT;
1553 	int i = start >> APERTURE_RANGE_SHIFT;
1554 	unsigned long boundary_size, mask;
1555 	unsigned long address = -1;
1556 	unsigned long limit;
1557 
1558 	next_bit >>= PAGE_SHIFT;
1559 
1560 	mask = dma_get_seg_boundary(dev);
1561 
1562 	boundary_size = mask + 1 ? ALIGN(mask + 1, PAGE_SIZE) >> PAGE_SHIFT :
1563 				   1UL << (BITS_PER_LONG - PAGE_SHIFT);
1564 
1565 	for (;i < max_index; ++i) {
1566 		unsigned long offset = dom->aperture[i]->offset >> PAGE_SHIFT;
1567 
1568 		if (dom->aperture[i]->offset >= dma_mask)
1569 			break;
1570 
1571 		limit = iommu_device_max_index(APERTURE_RANGE_PAGES, offset,
1572 					       dma_mask >> PAGE_SHIFT);
1573 
1574 		address = iommu_area_alloc(dom->aperture[i]->bitmap,
1575 					   limit, next_bit, pages, 0,
1576 					    boundary_size, align_mask);
1577 		if (address != -1) {
1578 			address = dom->aperture[i]->offset +
1579 				  (address << PAGE_SHIFT);
1580 			dom->next_address = address + (pages << PAGE_SHIFT);
1581 			break;
1582 		}
1583 
1584 		next_bit = 0;
1585 	}
1586 
1587 	return address;
1588 }
1589 
dma_ops_alloc_addresses(struct device * dev,struct dma_ops_domain * dom,unsigned int pages,unsigned long align_mask,u64 dma_mask)1590 static unsigned long dma_ops_alloc_addresses(struct device *dev,
1591 					     struct dma_ops_domain *dom,
1592 					     unsigned int pages,
1593 					     unsigned long align_mask,
1594 					     u64 dma_mask)
1595 {
1596 	unsigned long address;
1597 
1598 #ifdef CONFIG_IOMMU_STRESS
1599 	dom->next_address = 0;
1600 	dom->need_flush = true;
1601 #endif
1602 
1603 	address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1604 				     dma_mask, dom->next_address);
1605 
1606 	if (address == -1) {
1607 		dom->next_address = 0;
1608 		address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1609 					     dma_mask, 0);
1610 		dom->need_flush = true;
1611 	}
1612 
1613 	if (unlikely(address == -1))
1614 		address = DMA_ERROR_CODE;
1615 
1616 	WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
1617 
1618 	return address;
1619 }
1620 
1621 /*
1622  * The address free function.
1623  *
1624  * called with domain->lock held
1625  */
dma_ops_free_addresses(struct dma_ops_domain * dom,unsigned long address,unsigned int pages)1626 static void dma_ops_free_addresses(struct dma_ops_domain *dom,
1627 				   unsigned long address,
1628 				   unsigned int pages)
1629 {
1630 	unsigned i = address >> APERTURE_RANGE_SHIFT;
1631 	struct aperture_range *range = dom->aperture[i];
1632 
1633 	BUG_ON(i >= APERTURE_MAX_RANGES || range == NULL);
1634 
1635 #ifdef CONFIG_IOMMU_STRESS
1636 	if (i < 4)
1637 		return;
1638 #endif
1639 
1640 	if (address >= dom->next_address)
1641 		dom->need_flush = true;
1642 
1643 	address = (address % APERTURE_RANGE_SIZE) >> PAGE_SHIFT;
1644 
1645 	bitmap_clear(range->bitmap, address, pages);
1646 
1647 }
1648 
1649 /****************************************************************************
1650  *
1651  * The next functions belong to the domain allocation. A domain is
1652  * allocated for every IOMMU as the default domain. If device isolation
1653  * is enabled, every device get its own domain. The most important thing
1654  * about domains is the page table mapping the DMA address space they
1655  * contain.
1656  *
1657  ****************************************************************************/
1658 
1659 /*
1660  * This function adds a protection domain to the global protection domain list
1661  */
add_domain_to_list(struct protection_domain * domain)1662 static void add_domain_to_list(struct protection_domain *domain)
1663 {
1664 	unsigned long flags;
1665 
1666 	spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1667 	list_add(&domain->list, &amd_iommu_pd_list);
1668 	spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1669 }
1670 
1671 /*
1672  * This function removes a protection domain to the global
1673  * protection domain list
1674  */
del_domain_from_list(struct protection_domain * domain)1675 static void del_domain_from_list(struct protection_domain *domain)
1676 {
1677 	unsigned long flags;
1678 
1679 	spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1680 	list_del(&domain->list);
1681 	spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1682 }
1683 
domain_id_alloc(void)1684 static u16 domain_id_alloc(void)
1685 {
1686 	unsigned long flags;
1687 	int id;
1688 
1689 	write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1690 	id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1691 	BUG_ON(id == 0);
1692 	if (id > 0 && id < MAX_DOMAIN_ID)
1693 		__set_bit(id, amd_iommu_pd_alloc_bitmap);
1694 	else
1695 		id = 0;
1696 	write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1697 
1698 	return id;
1699 }
1700 
domain_id_free(int id)1701 static void domain_id_free(int id)
1702 {
1703 	unsigned long flags;
1704 
1705 	write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1706 	if (id > 0 && id < MAX_DOMAIN_ID)
1707 		__clear_bit(id, amd_iommu_pd_alloc_bitmap);
1708 	write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1709 }
1710 
1711 #define DEFINE_FREE_PT_FN(LVL, FN)				\
1712 static void free_pt_##LVL (unsigned long __pt)			\
1713 {								\
1714 	unsigned long p;					\
1715 	u64 *pt;						\
1716 	int i;							\
1717 								\
1718 	pt = (u64 *)__pt;					\
1719 								\
1720 	for (i = 0; i < 512; ++i) {				\
1721 		/* PTE present? */				\
1722 		if (!IOMMU_PTE_PRESENT(pt[i]))			\
1723 			continue;				\
1724 								\
1725 		/* Large PTE? */				\
1726 		if (PM_PTE_LEVEL(pt[i]) == 0 ||			\
1727 		    PM_PTE_LEVEL(pt[i]) == 7)			\
1728 			continue;				\
1729 								\
1730 		p = (unsigned long)IOMMU_PTE_PAGE(pt[i]);	\
1731 		FN(p);						\
1732 	}							\
1733 	free_page((unsigned long)pt);				\
1734 }
1735 
DEFINE_FREE_PT_FN(l2,free_page)1736 DEFINE_FREE_PT_FN(l2, free_page)
1737 DEFINE_FREE_PT_FN(l3, free_pt_l2)
1738 DEFINE_FREE_PT_FN(l4, free_pt_l3)
1739 DEFINE_FREE_PT_FN(l5, free_pt_l4)
1740 DEFINE_FREE_PT_FN(l6, free_pt_l5)
1741 
1742 static void free_pagetable(struct protection_domain *domain)
1743 {
1744 	unsigned long root = (unsigned long)domain->pt_root;
1745 
1746 	switch (domain->mode) {
1747 	case PAGE_MODE_NONE:
1748 		break;
1749 	case PAGE_MODE_1_LEVEL:
1750 		free_page(root);
1751 		break;
1752 	case PAGE_MODE_2_LEVEL:
1753 		free_pt_l2(root);
1754 		break;
1755 	case PAGE_MODE_3_LEVEL:
1756 		free_pt_l3(root);
1757 		break;
1758 	case PAGE_MODE_4_LEVEL:
1759 		free_pt_l4(root);
1760 		break;
1761 	case PAGE_MODE_5_LEVEL:
1762 		free_pt_l5(root);
1763 		break;
1764 	case PAGE_MODE_6_LEVEL:
1765 		free_pt_l6(root);
1766 		break;
1767 	default:
1768 		BUG();
1769 	}
1770 }
1771 
free_gcr3_tbl_level1(u64 * tbl)1772 static void free_gcr3_tbl_level1(u64 *tbl)
1773 {
1774 	u64 *ptr;
1775 	int i;
1776 
1777 	for (i = 0; i < 512; ++i) {
1778 		if (!(tbl[i] & GCR3_VALID))
1779 			continue;
1780 
1781 		ptr = __va(tbl[i] & PAGE_MASK);
1782 
1783 		free_page((unsigned long)ptr);
1784 	}
1785 }
1786 
free_gcr3_tbl_level2(u64 * tbl)1787 static void free_gcr3_tbl_level2(u64 *tbl)
1788 {
1789 	u64 *ptr;
1790 	int i;
1791 
1792 	for (i = 0; i < 512; ++i) {
1793 		if (!(tbl[i] & GCR3_VALID))
1794 			continue;
1795 
1796 		ptr = __va(tbl[i] & PAGE_MASK);
1797 
1798 		free_gcr3_tbl_level1(ptr);
1799 	}
1800 }
1801 
free_gcr3_table(struct protection_domain * domain)1802 static void free_gcr3_table(struct protection_domain *domain)
1803 {
1804 	if (domain->glx == 2)
1805 		free_gcr3_tbl_level2(domain->gcr3_tbl);
1806 	else if (domain->glx == 1)
1807 		free_gcr3_tbl_level1(domain->gcr3_tbl);
1808 	else
1809 		BUG_ON(domain->glx != 0);
1810 
1811 	free_page((unsigned long)domain->gcr3_tbl);
1812 }
1813 
1814 /*
1815  * Free a domain, only used if something went wrong in the
1816  * allocation path and we need to free an already allocated page table
1817  */
dma_ops_domain_free(struct dma_ops_domain * dom)1818 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1819 {
1820 	int i;
1821 
1822 	if (!dom)
1823 		return;
1824 
1825 	del_domain_from_list(&dom->domain);
1826 
1827 	free_pagetable(&dom->domain);
1828 
1829 	for (i = 0; i < APERTURE_MAX_RANGES; ++i) {
1830 		if (!dom->aperture[i])
1831 			continue;
1832 		free_page((unsigned long)dom->aperture[i]->bitmap);
1833 		kfree(dom->aperture[i]);
1834 	}
1835 
1836 	if (dom->domain.id)
1837 		domain_id_free(dom->domain.id);
1838 
1839 	kfree(dom);
1840 }
1841 
1842 /*
1843  * Allocates a new protection domain usable for the dma_ops functions.
1844  * It also initializes the page table and the address allocator data
1845  * structures required for the dma_ops interface
1846  */
dma_ops_domain_alloc(void)1847 static struct dma_ops_domain *dma_ops_domain_alloc(void)
1848 {
1849 	struct dma_ops_domain *dma_dom;
1850 
1851 	dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1852 	if (!dma_dom)
1853 		return NULL;
1854 
1855 	if (protection_domain_init(&dma_dom->domain))
1856 		goto free_dma_dom;
1857 
1858 	dma_dom->domain.mode = PAGE_MODE_2_LEVEL;
1859 	dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1860 	dma_dom->domain.flags = PD_DMA_OPS_MASK;
1861 	dma_dom->domain.priv = dma_dom;
1862 	if (!dma_dom->domain.pt_root)
1863 		goto free_dma_dom;
1864 
1865 	dma_dom->need_flush = false;
1866 
1867 	add_domain_to_list(&dma_dom->domain);
1868 
1869 	if (alloc_new_range(dma_dom, true, GFP_KERNEL))
1870 		goto free_dma_dom;
1871 
1872 	/*
1873 	 * mark the first page as allocated so we never return 0 as
1874 	 * a valid dma-address. So we can use 0 as error value
1875 	 */
1876 	dma_dom->aperture[0]->bitmap[0] = 1;
1877 	dma_dom->next_address = 0;
1878 
1879 
1880 	return dma_dom;
1881 
1882 free_dma_dom:
1883 	dma_ops_domain_free(dma_dom);
1884 
1885 	return NULL;
1886 }
1887 
1888 /*
1889  * little helper function to check whether a given protection domain is a
1890  * dma_ops domain
1891  */
dma_ops_domain(struct protection_domain * domain)1892 static bool dma_ops_domain(struct protection_domain *domain)
1893 {
1894 	return domain->flags & PD_DMA_OPS_MASK;
1895 }
1896 
set_dte_entry(u16 devid,struct protection_domain * domain,bool ats)1897 static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
1898 {
1899 	u64 pte_root = 0;
1900 	u64 flags = 0;
1901 
1902 	if (domain->mode != PAGE_MODE_NONE)
1903 		pte_root = virt_to_phys(domain->pt_root);
1904 
1905 	pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
1906 		    << DEV_ENTRY_MODE_SHIFT;
1907 	pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
1908 
1909 	flags = amd_iommu_dev_table[devid].data[1];
1910 
1911 	if (ats)
1912 		flags |= DTE_FLAG_IOTLB;
1913 
1914 	if (domain->flags & PD_IOMMUV2_MASK) {
1915 		u64 gcr3 = __pa(domain->gcr3_tbl);
1916 		u64 glx  = domain->glx;
1917 		u64 tmp;
1918 
1919 		pte_root |= DTE_FLAG_GV;
1920 		pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1921 
1922 		/* First mask out possible old values for GCR3 table */
1923 		tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1924 		flags    &= ~tmp;
1925 
1926 		tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1927 		flags    &= ~tmp;
1928 
1929 		/* Encode GCR3 table into DTE */
1930 		tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1931 		pte_root |= tmp;
1932 
1933 		tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1934 		flags    |= tmp;
1935 
1936 		tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1937 		flags    |= tmp;
1938 	}
1939 
1940 	flags &= ~(0xffffUL);
1941 	flags |= domain->id;
1942 
1943 	amd_iommu_dev_table[devid].data[1]  = flags;
1944 	amd_iommu_dev_table[devid].data[0]  = pte_root;
1945 }
1946 
clear_dte_entry(u16 devid)1947 static void clear_dte_entry(u16 devid)
1948 {
1949 	/* remove entry from the device table seen by the hardware */
1950 	amd_iommu_dev_table[devid].data[0]  = IOMMU_PTE_P | IOMMU_PTE_TV;
1951 	amd_iommu_dev_table[devid].data[1] &= DTE_FLAG_MASK;
1952 
1953 	amd_iommu_apply_erratum_63(devid);
1954 }
1955 
do_attach(struct iommu_dev_data * dev_data,struct protection_domain * domain)1956 static void do_attach(struct iommu_dev_data *dev_data,
1957 		      struct protection_domain *domain)
1958 {
1959 	struct amd_iommu *iommu;
1960 	u16 alias;
1961 	bool ats;
1962 
1963 	iommu = amd_iommu_rlookup_table[dev_data->devid];
1964 	alias = dev_data->alias;
1965 	ats   = dev_data->ats.enabled;
1966 
1967 	/* Update data structures */
1968 	dev_data->domain = domain;
1969 	list_add(&dev_data->list, &domain->dev_list);
1970 
1971 	/* Do reference counting */
1972 	domain->dev_iommu[iommu->index] += 1;
1973 	domain->dev_cnt                 += 1;
1974 
1975 	/* Update device table */
1976 	set_dte_entry(dev_data->devid, domain, ats);
1977 	if (alias != dev_data->devid)
1978 		set_dte_entry(alias, domain, ats);
1979 
1980 	device_flush_dte(dev_data);
1981 }
1982 
do_detach(struct iommu_dev_data * dev_data)1983 static void do_detach(struct iommu_dev_data *dev_data)
1984 {
1985 	struct protection_domain *domain = dev_data->domain;
1986 	struct amd_iommu *iommu;
1987 	u16 alias;
1988 
1989 	/*
1990 	 * First check if the device is still attached. It might already
1991 	 * be detached from its domain because the generic
1992 	 * iommu_detach_group code detached it and we try again here in
1993 	 * our alias handling.
1994 	 */
1995 	if (!dev_data->domain)
1996 		return;
1997 
1998 	iommu = amd_iommu_rlookup_table[dev_data->devid];
1999 	alias = dev_data->alias;
2000 
2001 	/* Update data structures */
2002 	dev_data->domain = NULL;
2003 	list_del(&dev_data->list);
2004 	clear_dte_entry(dev_data->devid);
2005 	if (alias != dev_data->devid)
2006 		clear_dte_entry(alias);
2007 
2008 	/* Flush the DTE entry */
2009 	device_flush_dte(dev_data);
2010 
2011 	/* Flush IOTLB */
2012 	domain_flush_tlb_pde(domain);
2013 
2014 	/* Wait for the flushes to finish */
2015 	domain_flush_complete(domain);
2016 
2017 	/* decrease reference counters - needs to happen after the flushes */
2018 	domain->dev_iommu[iommu->index] -= 1;
2019 	domain->dev_cnt                 -= 1;
2020 }
2021 
2022 /*
2023  * If a device is not yet associated with a domain, this function does
2024  * assigns it visible for the hardware
2025  */
__attach_device(struct iommu_dev_data * dev_data,struct protection_domain * domain)2026 static int __attach_device(struct iommu_dev_data *dev_data,
2027 			   struct protection_domain *domain)
2028 {
2029 	int ret;
2030 
2031 	/*
2032 	 * Must be called with IRQs disabled. Warn here to detect early
2033 	 * when its not.
2034 	 */
2035 	WARN_ON(!irqs_disabled());
2036 
2037 	/* lock domain */
2038 	spin_lock(&domain->lock);
2039 
2040 	ret = -EBUSY;
2041 	if (dev_data->domain != NULL)
2042 		goto out_unlock;
2043 
2044 	/* Attach alias group root */
2045 	do_attach(dev_data, domain);
2046 
2047 	ret = 0;
2048 
2049 out_unlock:
2050 
2051 	/* ready */
2052 	spin_unlock(&domain->lock);
2053 
2054 	return ret;
2055 }
2056 
2057 
pdev_iommuv2_disable(struct pci_dev * pdev)2058 static void pdev_iommuv2_disable(struct pci_dev *pdev)
2059 {
2060 	pci_disable_ats(pdev);
2061 	pci_disable_pri(pdev);
2062 	pci_disable_pasid(pdev);
2063 }
2064 
2065 /* FIXME: Change generic reset-function to do the same */
pri_reset_while_enabled(struct pci_dev * pdev)2066 static int pri_reset_while_enabled(struct pci_dev *pdev)
2067 {
2068 	u16 control;
2069 	int pos;
2070 
2071 	pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2072 	if (!pos)
2073 		return -EINVAL;
2074 
2075 	pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
2076 	control |= PCI_PRI_CTRL_RESET;
2077 	pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
2078 
2079 	return 0;
2080 }
2081 
pdev_iommuv2_enable(struct pci_dev * pdev)2082 static int pdev_iommuv2_enable(struct pci_dev *pdev)
2083 {
2084 	bool reset_enable;
2085 	int reqs, ret;
2086 
2087 	/* FIXME: Hardcode number of outstanding requests for now */
2088 	reqs = 32;
2089 	if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2090 		reqs = 1;
2091 	reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2092 
2093 	/* Only allow access to user-accessible pages */
2094 	ret = pci_enable_pasid(pdev, 0);
2095 	if (ret)
2096 		goto out_err;
2097 
2098 	/* First reset the PRI state of the device */
2099 	ret = pci_reset_pri(pdev);
2100 	if (ret)
2101 		goto out_err;
2102 
2103 	/* Enable PRI */
2104 	ret = pci_enable_pri(pdev, reqs);
2105 	if (ret)
2106 		goto out_err;
2107 
2108 	if (reset_enable) {
2109 		ret = pri_reset_while_enabled(pdev);
2110 		if (ret)
2111 			goto out_err;
2112 	}
2113 
2114 	ret = pci_enable_ats(pdev, PAGE_SHIFT);
2115 	if (ret)
2116 		goto out_err;
2117 
2118 	return 0;
2119 
2120 out_err:
2121 	pci_disable_pri(pdev);
2122 	pci_disable_pasid(pdev);
2123 
2124 	return ret;
2125 }
2126 
2127 /* FIXME: Move this to PCI code */
2128 #define PCI_PRI_TLP_OFF		(1 << 15)
2129 
pci_pri_tlp_required(struct pci_dev * pdev)2130 static bool pci_pri_tlp_required(struct pci_dev *pdev)
2131 {
2132 	u16 status;
2133 	int pos;
2134 
2135 	pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2136 	if (!pos)
2137 		return false;
2138 
2139 	pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
2140 
2141 	return (status & PCI_PRI_TLP_OFF) ? true : false;
2142 }
2143 
2144 /*
2145  * If a device is not yet associated with a domain, this function
2146  * assigns it visible for the hardware
2147  */
attach_device(struct device * dev,struct protection_domain * domain)2148 static int attach_device(struct device *dev,
2149 			 struct protection_domain *domain)
2150 {
2151 	struct pci_dev *pdev = to_pci_dev(dev);
2152 	struct iommu_dev_data *dev_data;
2153 	unsigned long flags;
2154 	int ret;
2155 
2156 	dev_data = get_dev_data(dev);
2157 
2158 	if (domain->flags & PD_IOMMUV2_MASK) {
2159 		if (!dev_data->passthrough)
2160 			return -EINVAL;
2161 
2162 		if (dev_data->iommu_v2) {
2163 			if (pdev_iommuv2_enable(pdev) != 0)
2164 				return -EINVAL;
2165 
2166 			dev_data->ats.enabled = true;
2167 			dev_data->ats.qdep    = pci_ats_queue_depth(pdev);
2168 			dev_data->pri_tlp     = pci_pri_tlp_required(pdev);
2169 		}
2170 	} else if (amd_iommu_iotlb_sup &&
2171 		   pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2172 		dev_data->ats.enabled = true;
2173 		dev_data->ats.qdep    = pci_ats_queue_depth(pdev);
2174 	}
2175 
2176 	write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2177 	ret = __attach_device(dev_data, domain);
2178 	write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2179 
2180 	/*
2181 	 * We might boot into a crash-kernel here. The crashed kernel
2182 	 * left the caches in the IOMMU dirty. So we have to flush
2183 	 * here to evict all dirty stuff.
2184 	 */
2185 	domain_flush_tlb_pde(domain);
2186 
2187 	domain_flush_complete(domain);
2188 
2189 	return ret;
2190 }
2191 
2192 /*
2193  * Removes a device from a protection domain (unlocked)
2194  */
__detach_device(struct iommu_dev_data * dev_data)2195 static void __detach_device(struct iommu_dev_data *dev_data)
2196 {
2197 	struct protection_domain *domain;
2198 
2199 	/*
2200 	 * Must be called with IRQs disabled. Warn here to detect early
2201 	 * when its not.
2202 	 */
2203 	WARN_ON(!irqs_disabled());
2204 
2205 	if (WARN_ON(!dev_data->domain))
2206 		return;
2207 
2208 	domain = dev_data->domain;
2209 
2210 	spin_lock(&domain->lock);
2211 
2212 	do_detach(dev_data);
2213 
2214 	spin_unlock(&domain->lock);
2215 }
2216 
2217 /*
2218  * Removes a device from a protection domain (with devtable_lock held)
2219  */
detach_device(struct device * dev)2220 static void detach_device(struct device *dev)
2221 {
2222 	struct protection_domain *domain;
2223 	struct iommu_dev_data *dev_data;
2224 	unsigned long flags;
2225 
2226 	dev_data = get_dev_data(dev);
2227 	domain   = dev_data->domain;
2228 
2229 	/* lock device table */
2230 	write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2231 	__detach_device(dev_data);
2232 	write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2233 
2234 	if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2)
2235 		pdev_iommuv2_disable(to_pci_dev(dev));
2236 	else if (dev_data->ats.enabled)
2237 		pci_disable_ats(to_pci_dev(dev));
2238 
2239 	dev_data->ats.enabled = false;
2240 }
2241 
amd_iommu_add_device(struct device * dev)2242 static int amd_iommu_add_device(struct device *dev)
2243 {
2244 	struct iommu_dev_data *dev_data;
2245 	struct iommu_domain *domain;
2246 	struct amd_iommu *iommu;
2247 	u16 devid;
2248 	int ret;
2249 
2250 	if (!check_device(dev) || get_dev_data(dev))
2251 		return 0;
2252 
2253 	devid = get_device_id(dev);
2254 	iommu = amd_iommu_rlookup_table[devid];
2255 
2256 	ret = iommu_init_device(dev);
2257 	if (ret) {
2258 		if (ret != -ENOTSUPP)
2259 			pr_err("Failed to initialize device %s - trying to proceed anyway\n",
2260 				dev_name(dev));
2261 
2262 		iommu_ignore_device(dev);
2263 		dev->archdata.dma_ops = &nommu_dma_ops;
2264 		goto out;
2265 	}
2266 	init_iommu_group(dev);
2267 
2268 	dev_data = get_dev_data(dev);
2269 
2270 	BUG_ON(!dev_data);
2271 
2272 	if (iommu_pass_through || dev_data->iommu_v2)
2273 		iommu_request_dm_for_dev(dev);
2274 
2275 	/* Domains are initialized for this device - have a look what we ended up with */
2276 	domain = iommu_get_domain_for_dev(dev);
2277 	if (domain->type == IOMMU_DOMAIN_IDENTITY)
2278 		dev_data->passthrough = true;
2279 	else
2280 		dev->archdata.dma_ops = &amd_iommu_dma_ops;
2281 
2282 out:
2283 	iommu_completion_wait(iommu);
2284 
2285 	return 0;
2286 }
2287 
amd_iommu_remove_device(struct device * dev)2288 static void amd_iommu_remove_device(struct device *dev)
2289 {
2290 	struct amd_iommu *iommu;
2291 	u16 devid;
2292 
2293 	if (!check_device(dev))
2294 		return;
2295 
2296 	devid = get_device_id(dev);
2297 	iommu = amd_iommu_rlookup_table[devid];
2298 
2299 	iommu_uninit_device(dev);
2300 	iommu_completion_wait(iommu);
2301 }
2302 
2303 /*****************************************************************************
2304  *
2305  * The next functions belong to the dma_ops mapping/unmapping code.
2306  *
2307  *****************************************************************************/
2308 
2309 /*
2310  * In the dma_ops path we only have the struct device. This function
2311  * finds the corresponding IOMMU, the protection domain and the
2312  * requestor id for a given device.
2313  * If the device is not yet associated with a domain this is also done
2314  * in this function.
2315  */
get_domain(struct device * dev)2316 static struct protection_domain *get_domain(struct device *dev)
2317 {
2318 	struct protection_domain *domain;
2319 	struct iommu_domain *io_domain;
2320 
2321 	if (!check_device(dev))
2322 		return ERR_PTR(-EINVAL);
2323 
2324 	io_domain = iommu_get_domain_for_dev(dev);
2325 	if (!io_domain)
2326 		return NULL;
2327 
2328 	domain = to_pdomain(io_domain);
2329 	if (!dma_ops_domain(domain))
2330 		return ERR_PTR(-EBUSY);
2331 
2332 	return domain;
2333 }
2334 
update_device_table(struct protection_domain * domain)2335 static void update_device_table(struct protection_domain *domain)
2336 {
2337 	struct iommu_dev_data *dev_data;
2338 
2339 	list_for_each_entry(dev_data, &domain->dev_list, list) {
2340 		set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled);
2341 
2342 		if (dev_data->devid == dev_data->alias)
2343 			continue;
2344 
2345 		/* There is an alias, update device table entry for it */
2346 		set_dte_entry(dev_data->alias, domain, dev_data->ats.enabled);
2347 	}
2348 }
2349 
update_domain(struct protection_domain * domain)2350 static void update_domain(struct protection_domain *domain)
2351 {
2352 	if (!domain->updated)
2353 		return;
2354 
2355 	update_device_table(domain);
2356 
2357 	domain_flush_devices(domain);
2358 	domain_flush_tlb_pde(domain);
2359 
2360 	domain->updated = false;
2361 }
2362 
2363 /*
2364  * This function fetches the PTE for a given address in the aperture
2365  */
dma_ops_get_pte(struct dma_ops_domain * dom,unsigned long address)2366 static u64* dma_ops_get_pte(struct dma_ops_domain *dom,
2367 			    unsigned long address)
2368 {
2369 	struct aperture_range *aperture;
2370 	u64 *pte, *pte_page;
2371 
2372 	aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2373 	if (!aperture)
2374 		return NULL;
2375 
2376 	pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2377 	if (!pte) {
2378 		pte = alloc_pte(&dom->domain, address, PAGE_SIZE, &pte_page,
2379 				GFP_ATOMIC);
2380 		aperture->pte_pages[APERTURE_PAGE_INDEX(address)] = pte_page;
2381 	} else
2382 		pte += PM_LEVEL_INDEX(0, address);
2383 
2384 	update_domain(&dom->domain);
2385 
2386 	return pte;
2387 }
2388 
2389 /*
2390  * This is the generic map function. It maps one 4kb page at paddr to
2391  * the given address in the DMA address space for the domain.
2392  */
dma_ops_domain_map(struct dma_ops_domain * dom,unsigned long address,phys_addr_t paddr,int direction)2393 static dma_addr_t dma_ops_domain_map(struct dma_ops_domain *dom,
2394 				     unsigned long address,
2395 				     phys_addr_t paddr,
2396 				     int direction)
2397 {
2398 	u64 *pte, __pte;
2399 
2400 	WARN_ON(address > dom->aperture_size);
2401 
2402 	paddr &= PAGE_MASK;
2403 
2404 	pte  = dma_ops_get_pte(dom, address);
2405 	if (!pte)
2406 		return DMA_ERROR_CODE;
2407 
2408 	__pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
2409 
2410 	if (direction == DMA_TO_DEVICE)
2411 		__pte |= IOMMU_PTE_IR;
2412 	else if (direction == DMA_FROM_DEVICE)
2413 		__pte |= IOMMU_PTE_IW;
2414 	else if (direction == DMA_BIDIRECTIONAL)
2415 		__pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
2416 
2417 	WARN_ON(*pte);
2418 
2419 	*pte = __pte;
2420 
2421 	return (dma_addr_t)address;
2422 }
2423 
2424 /*
2425  * The generic unmapping function for on page in the DMA address space.
2426  */
dma_ops_domain_unmap(struct dma_ops_domain * dom,unsigned long address)2427 static void dma_ops_domain_unmap(struct dma_ops_domain *dom,
2428 				 unsigned long address)
2429 {
2430 	struct aperture_range *aperture;
2431 	u64 *pte;
2432 
2433 	if (address >= dom->aperture_size)
2434 		return;
2435 
2436 	aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2437 	if (!aperture)
2438 		return;
2439 
2440 	pte  = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2441 	if (!pte)
2442 		return;
2443 
2444 	pte += PM_LEVEL_INDEX(0, address);
2445 
2446 	WARN_ON(!*pte);
2447 
2448 	*pte = 0ULL;
2449 }
2450 
2451 /*
2452  * This function contains common code for mapping of a physically
2453  * contiguous memory region into DMA address space. It is used by all
2454  * mapping functions provided with this IOMMU driver.
2455  * Must be called with the domain lock held.
2456  */
__map_single(struct device * dev,struct dma_ops_domain * dma_dom,phys_addr_t paddr,size_t size,int dir,bool align,u64 dma_mask)2457 static dma_addr_t __map_single(struct device *dev,
2458 			       struct dma_ops_domain *dma_dom,
2459 			       phys_addr_t paddr,
2460 			       size_t size,
2461 			       int dir,
2462 			       bool align,
2463 			       u64 dma_mask)
2464 {
2465 	dma_addr_t offset = paddr & ~PAGE_MASK;
2466 	dma_addr_t address, start, ret;
2467 	unsigned int pages;
2468 	unsigned long align_mask = 0;
2469 	int i;
2470 
2471 	pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2472 	paddr &= PAGE_MASK;
2473 
2474 	INC_STATS_COUNTER(total_map_requests);
2475 
2476 	if (pages > 1)
2477 		INC_STATS_COUNTER(cross_page);
2478 
2479 	if (align)
2480 		align_mask = (1UL << get_order(size)) - 1;
2481 
2482 retry:
2483 	address = dma_ops_alloc_addresses(dev, dma_dom, pages, align_mask,
2484 					  dma_mask);
2485 	if (unlikely(address == DMA_ERROR_CODE)) {
2486 		/*
2487 		 * setting next_address here will let the address
2488 		 * allocator only scan the new allocated range in the
2489 		 * first run. This is a small optimization.
2490 		 */
2491 		dma_dom->next_address = dma_dom->aperture_size;
2492 
2493 		if (alloc_new_range(dma_dom, false, GFP_ATOMIC))
2494 			goto out;
2495 
2496 		/*
2497 		 * aperture was successfully enlarged by 128 MB, try
2498 		 * allocation again
2499 		 */
2500 		goto retry;
2501 	}
2502 
2503 	start = address;
2504 	for (i = 0; i < pages; ++i) {
2505 		ret = dma_ops_domain_map(dma_dom, start, paddr, dir);
2506 		if (ret == DMA_ERROR_CODE)
2507 			goto out_unmap;
2508 
2509 		paddr += PAGE_SIZE;
2510 		start += PAGE_SIZE;
2511 	}
2512 	address += offset;
2513 
2514 	ADD_STATS_COUNTER(alloced_io_mem, size);
2515 
2516 	if (unlikely(dma_dom->need_flush && !amd_iommu_unmap_flush)) {
2517 		domain_flush_tlb(&dma_dom->domain);
2518 		dma_dom->need_flush = false;
2519 	} else if (unlikely(amd_iommu_np_cache))
2520 		domain_flush_pages(&dma_dom->domain, address, size);
2521 
2522 out:
2523 	return address;
2524 
2525 out_unmap:
2526 
2527 	for (--i; i >= 0; --i) {
2528 		start -= PAGE_SIZE;
2529 		dma_ops_domain_unmap(dma_dom, start);
2530 	}
2531 
2532 	dma_ops_free_addresses(dma_dom, address, pages);
2533 
2534 	return DMA_ERROR_CODE;
2535 }
2536 
2537 /*
2538  * Does the reverse of the __map_single function. Must be called with
2539  * the domain lock held too
2540  */
__unmap_single(struct dma_ops_domain * dma_dom,dma_addr_t dma_addr,size_t size,int dir)2541 static void __unmap_single(struct dma_ops_domain *dma_dom,
2542 			   dma_addr_t dma_addr,
2543 			   size_t size,
2544 			   int dir)
2545 {
2546 	dma_addr_t flush_addr;
2547 	dma_addr_t i, start;
2548 	unsigned int pages;
2549 
2550 	if ((dma_addr == DMA_ERROR_CODE) ||
2551 	    (dma_addr + size > dma_dom->aperture_size))
2552 		return;
2553 
2554 	flush_addr = dma_addr;
2555 	pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2556 	dma_addr &= PAGE_MASK;
2557 	start = dma_addr;
2558 
2559 	for (i = 0; i < pages; ++i) {
2560 		dma_ops_domain_unmap(dma_dom, start);
2561 		start += PAGE_SIZE;
2562 	}
2563 
2564 	SUB_STATS_COUNTER(alloced_io_mem, size);
2565 
2566 	dma_ops_free_addresses(dma_dom, dma_addr, pages);
2567 
2568 	if (amd_iommu_unmap_flush || dma_dom->need_flush) {
2569 		domain_flush_pages(&dma_dom->domain, flush_addr, size);
2570 		dma_dom->need_flush = false;
2571 	}
2572 }
2573 
2574 /*
2575  * The exported map_single function for dma_ops.
2576  */
map_page(struct device * dev,struct page * page,unsigned long offset,size_t size,enum dma_data_direction dir,struct dma_attrs * attrs)2577 static dma_addr_t map_page(struct device *dev, struct page *page,
2578 			   unsigned long offset, size_t size,
2579 			   enum dma_data_direction dir,
2580 			   struct dma_attrs *attrs)
2581 {
2582 	unsigned long flags;
2583 	struct protection_domain *domain;
2584 	dma_addr_t addr;
2585 	u64 dma_mask;
2586 	phys_addr_t paddr = page_to_phys(page) + offset;
2587 
2588 	INC_STATS_COUNTER(cnt_map_single);
2589 
2590 	domain = get_domain(dev);
2591 	if (PTR_ERR(domain) == -EINVAL)
2592 		return (dma_addr_t)paddr;
2593 	else if (IS_ERR(domain))
2594 		return DMA_ERROR_CODE;
2595 
2596 	dma_mask = *dev->dma_mask;
2597 
2598 	spin_lock_irqsave(&domain->lock, flags);
2599 
2600 	addr = __map_single(dev, domain->priv, paddr, size, dir, false,
2601 			    dma_mask);
2602 	if (addr == DMA_ERROR_CODE)
2603 		goto out;
2604 
2605 	domain_flush_complete(domain);
2606 
2607 out:
2608 	spin_unlock_irqrestore(&domain->lock, flags);
2609 
2610 	return addr;
2611 }
2612 
2613 /*
2614  * The exported unmap_single function for dma_ops.
2615  */
unmap_page(struct device * dev,dma_addr_t dma_addr,size_t size,enum dma_data_direction dir,struct dma_attrs * attrs)2616 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2617 		       enum dma_data_direction dir, struct dma_attrs *attrs)
2618 {
2619 	unsigned long flags;
2620 	struct protection_domain *domain;
2621 
2622 	INC_STATS_COUNTER(cnt_unmap_single);
2623 
2624 	domain = get_domain(dev);
2625 	if (IS_ERR(domain))
2626 		return;
2627 
2628 	spin_lock_irqsave(&domain->lock, flags);
2629 
2630 	__unmap_single(domain->priv, dma_addr, size, dir);
2631 
2632 	domain_flush_complete(domain);
2633 
2634 	spin_unlock_irqrestore(&domain->lock, flags);
2635 }
2636 
2637 /*
2638  * The exported map_sg function for dma_ops (handles scatter-gather
2639  * lists).
2640  */
map_sg(struct device * dev,struct scatterlist * sglist,int nelems,enum dma_data_direction dir,struct dma_attrs * attrs)2641 static int map_sg(struct device *dev, struct scatterlist *sglist,
2642 		  int nelems, enum dma_data_direction dir,
2643 		  struct dma_attrs *attrs)
2644 {
2645 	unsigned long flags;
2646 	struct protection_domain *domain;
2647 	int i;
2648 	struct scatterlist *s;
2649 	phys_addr_t paddr;
2650 	int mapped_elems = 0;
2651 	u64 dma_mask;
2652 
2653 	INC_STATS_COUNTER(cnt_map_sg);
2654 
2655 	domain = get_domain(dev);
2656 	if (IS_ERR(domain))
2657 		return 0;
2658 
2659 	dma_mask = *dev->dma_mask;
2660 
2661 	spin_lock_irqsave(&domain->lock, flags);
2662 
2663 	for_each_sg(sglist, s, nelems, i) {
2664 		paddr = sg_phys(s);
2665 
2666 		s->dma_address = __map_single(dev, domain->priv,
2667 					      paddr, s->length, dir, false,
2668 					      dma_mask);
2669 
2670 		if (s->dma_address) {
2671 			s->dma_length = s->length;
2672 			mapped_elems++;
2673 		} else
2674 			goto unmap;
2675 	}
2676 
2677 	domain_flush_complete(domain);
2678 
2679 out:
2680 	spin_unlock_irqrestore(&domain->lock, flags);
2681 
2682 	return mapped_elems;
2683 unmap:
2684 	for_each_sg(sglist, s, mapped_elems, i) {
2685 		if (s->dma_address)
2686 			__unmap_single(domain->priv, s->dma_address,
2687 				       s->dma_length, dir);
2688 		s->dma_address = s->dma_length = 0;
2689 	}
2690 
2691 	mapped_elems = 0;
2692 
2693 	goto out;
2694 }
2695 
2696 /*
2697  * The exported map_sg function for dma_ops (handles scatter-gather
2698  * lists).
2699  */
unmap_sg(struct device * dev,struct scatterlist * sglist,int nelems,enum dma_data_direction dir,struct dma_attrs * attrs)2700 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2701 		     int nelems, enum dma_data_direction dir,
2702 		     struct dma_attrs *attrs)
2703 {
2704 	unsigned long flags;
2705 	struct protection_domain *domain;
2706 	struct scatterlist *s;
2707 	int i;
2708 
2709 	INC_STATS_COUNTER(cnt_unmap_sg);
2710 
2711 	domain = get_domain(dev);
2712 	if (IS_ERR(domain))
2713 		return;
2714 
2715 	spin_lock_irqsave(&domain->lock, flags);
2716 
2717 	for_each_sg(sglist, s, nelems, i) {
2718 		__unmap_single(domain->priv, s->dma_address,
2719 			       s->dma_length, dir);
2720 		s->dma_address = s->dma_length = 0;
2721 	}
2722 
2723 	domain_flush_complete(domain);
2724 
2725 	spin_unlock_irqrestore(&domain->lock, flags);
2726 }
2727 
2728 /*
2729  * The exported alloc_coherent function for dma_ops.
2730  */
alloc_coherent(struct device * dev,size_t size,dma_addr_t * dma_addr,gfp_t flag,struct dma_attrs * attrs)2731 static void *alloc_coherent(struct device *dev, size_t size,
2732 			    dma_addr_t *dma_addr, gfp_t flag,
2733 			    struct dma_attrs *attrs)
2734 {
2735 	u64 dma_mask = dev->coherent_dma_mask;
2736 	struct protection_domain *domain;
2737 	unsigned long flags;
2738 	struct page *page;
2739 
2740 	INC_STATS_COUNTER(cnt_alloc_coherent);
2741 
2742 	domain = get_domain(dev);
2743 	if (PTR_ERR(domain) == -EINVAL) {
2744 		page = alloc_pages(flag, get_order(size));
2745 		*dma_addr = page_to_phys(page);
2746 		return page_address(page);
2747 	} else if (IS_ERR(domain))
2748 		return NULL;
2749 
2750 	size	  = PAGE_ALIGN(size);
2751 	dma_mask  = dev->coherent_dma_mask;
2752 	flag     &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2753 	flag     |= __GFP_ZERO;
2754 
2755 	page = alloc_pages(flag | __GFP_NOWARN,  get_order(size));
2756 	if (!page) {
2757 		if (!gfpflags_allow_blocking(flag))
2758 			return NULL;
2759 
2760 		page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT,
2761 						 get_order(size));
2762 		if (!page)
2763 			return NULL;
2764 	}
2765 
2766 	if (!dma_mask)
2767 		dma_mask = *dev->dma_mask;
2768 
2769 	spin_lock_irqsave(&domain->lock, flags);
2770 
2771 	*dma_addr = __map_single(dev, domain->priv, page_to_phys(page),
2772 				 size, DMA_BIDIRECTIONAL, true, dma_mask);
2773 
2774 	if (*dma_addr == DMA_ERROR_CODE) {
2775 		spin_unlock_irqrestore(&domain->lock, flags);
2776 		goto out_free;
2777 	}
2778 
2779 	domain_flush_complete(domain);
2780 
2781 	spin_unlock_irqrestore(&domain->lock, flags);
2782 
2783 	return page_address(page);
2784 
2785 out_free:
2786 
2787 	if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2788 		__free_pages(page, get_order(size));
2789 
2790 	return NULL;
2791 }
2792 
2793 /*
2794  * The exported free_coherent function for dma_ops.
2795  */
free_coherent(struct device * dev,size_t size,void * virt_addr,dma_addr_t dma_addr,struct dma_attrs * attrs)2796 static void free_coherent(struct device *dev, size_t size,
2797 			  void *virt_addr, dma_addr_t dma_addr,
2798 			  struct dma_attrs *attrs)
2799 {
2800 	struct protection_domain *domain;
2801 	unsigned long flags;
2802 	struct page *page;
2803 
2804 	INC_STATS_COUNTER(cnt_free_coherent);
2805 
2806 	page = virt_to_page(virt_addr);
2807 	size = PAGE_ALIGN(size);
2808 
2809 	domain = get_domain(dev);
2810 	if (IS_ERR(domain))
2811 		goto free_mem;
2812 
2813 	spin_lock_irqsave(&domain->lock, flags);
2814 
2815 	__unmap_single(domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
2816 
2817 	domain_flush_complete(domain);
2818 
2819 	spin_unlock_irqrestore(&domain->lock, flags);
2820 
2821 free_mem:
2822 	if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2823 		__free_pages(page, get_order(size));
2824 }
2825 
2826 /*
2827  * This function is called by the DMA layer to find out if we can handle a
2828  * particular device. It is part of the dma_ops.
2829  */
amd_iommu_dma_supported(struct device * dev,u64 mask)2830 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2831 {
2832 	return check_device(dev);
2833 }
2834 
2835 static struct dma_map_ops amd_iommu_dma_ops = {
2836 	.alloc = alloc_coherent,
2837 	.free = free_coherent,
2838 	.map_page = map_page,
2839 	.unmap_page = unmap_page,
2840 	.map_sg = map_sg,
2841 	.unmap_sg = unmap_sg,
2842 	.dma_supported = amd_iommu_dma_supported,
2843 };
2844 
amd_iommu_init_api(void)2845 int __init amd_iommu_init_api(void)
2846 {
2847 	return bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
2848 }
2849 
amd_iommu_init_dma_ops(void)2850 int __init amd_iommu_init_dma_ops(void)
2851 {
2852 	swiotlb        = iommu_pass_through ? 1 : 0;
2853 	iommu_detected = 1;
2854 
2855 	/*
2856 	 * In case we don't initialize SWIOTLB (actually the common case
2857 	 * when AMD IOMMU is enabled), make sure there are global
2858 	 * dma_ops set as a fall-back for devices not handled by this
2859 	 * driver (for example non-PCI devices).
2860 	 */
2861 	if (!swiotlb)
2862 		dma_ops = &nommu_dma_ops;
2863 
2864 	amd_iommu_stats_init();
2865 
2866 	if (amd_iommu_unmap_flush)
2867 		pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
2868 	else
2869 		pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
2870 
2871 	return 0;
2872 }
2873 
2874 /*****************************************************************************
2875  *
2876  * The following functions belong to the exported interface of AMD IOMMU
2877  *
2878  * This interface allows access to lower level functions of the IOMMU
2879  * like protection domain handling and assignement of devices to domains
2880  * which is not possible with the dma_ops interface.
2881  *
2882  *****************************************************************************/
2883 
cleanup_domain(struct protection_domain * domain)2884 static void cleanup_domain(struct protection_domain *domain)
2885 {
2886 	struct iommu_dev_data *entry;
2887 	unsigned long flags;
2888 
2889 	write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2890 
2891 	while (!list_empty(&domain->dev_list)) {
2892 		entry = list_first_entry(&domain->dev_list,
2893 					 struct iommu_dev_data, list);
2894 		__detach_device(entry);
2895 	}
2896 
2897 	write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2898 }
2899 
protection_domain_free(struct protection_domain * domain)2900 static void protection_domain_free(struct protection_domain *domain)
2901 {
2902 	if (!domain)
2903 		return;
2904 
2905 	del_domain_from_list(domain);
2906 
2907 	if (domain->id)
2908 		domain_id_free(domain->id);
2909 
2910 	kfree(domain);
2911 }
2912 
protection_domain_init(struct protection_domain * domain)2913 static int protection_domain_init(struct protection_domain *domain)
2914 {
2915 	spin_lock_init(&domain->lock);
2916 	mutex_init(&domain->api_lock);
2917 	domain->id = domain_id_alloc();
2918 	if (!domain->id)
2919 		return -ENOMEM;
2920 	INIT_LIST_HEAD(&domain->dev_list);
2921 
2922 	return 0;
2923 }
2924 
protection_domain_alloc(void)2925 static struct protection_domain *protection_domain_alloc(void)
2926 {
2927 	struct protection_domain *domain;
2928 
2929 	domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2930 	if (!domain)
2931 		return NULL;
2932 
2933 	if (protection_domain_init(domain))
2934 		goto out_err;
2935 
2936 	add_domain_to_list(domain);
2937 
2938 	return domain;
2939 
2940 out_err:
2941 	kfree(domain);
2942 
2943 	return NULL;
2944 }
2945 
amd_iommu_domain_alloc(unsigned type)2946 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
2947 {
2948 	struct protection_domain *pdomain;
2949 	struct dma_ops_domain *dma_domain;
2950 
2951 	switch (type) {
2952 	case IOMMU_DOMAIN_UNMANAGED:
2953 		pdomain = protection_domain_alloc();
2954 		if (!pdomain)
2955 			return NULL;
2956 
2957 		pdomain->mode    = PAGE_MODE_3_LEVEL;
2958 		pdomain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2959 		if (!pdomain->pt_root) {
2960 			protection_domain_free(pdomain);
2961 			return NULL;
2962 		}
2963 
2964 		pdomain->domain.geometry.aperture_start = 0;
2965 		pdomain->domain.geometry.aperture_end   = ~0ULL;
2966 		pdomain->domain.geometry.force_aperture = true;
2967 
2968 		break;
2969 	case IOMMU_DOMAIN_DMA:
2970 		dma_domain = dma_ops_domain_alloc();
2971 		if (!dma_domain) {
2972 			pr_err("AMD-Vi: Failed to allocate\n");
2973 			return NULL;
2974 		}
2975 		pdomain = &dma_domain->domain;
2976 		break;
2977 	case IOMMU_DOMAIN_IDENTITY:
2978 		pdomain = protection_domain_alloc();
2979 		if (!pdomain)
2980 			return NULL;
2981 
2982 		pdomain->mode = PAGE_MODE_NONE;
2983 		break;
2984 	default:
2985 		return NULL;
2986 	}
2987 
2988 	return &pdomain->domain;
2989 }
2990 
amd_iommu_domain_free(struct iommu_domain * dom)2991 static void amd_iommu_domain_free(struct iommu_domain *dom)
2992 {
2993 	struct protection_domain *domain;
2994 	struct dma_ops_domain *dma_dom;
2995 
2996 	domain = to_pdomain(dom);
2997 
2998 	if (domain->dev_cnt > 0)
2999 		cleanup_domain(domain);
3000 
3001 	BUG_ON(domain->dev_cnt != 0);
3002 
3003 	if (!dom)
3004 		return;
3005 
3006 	switch (dom->type) {
3007 	case IOMMU_DOMAIN_DMA:
3008 		dma_dom = domain->priv;
3009 		dma_ops_domain_free(dma_dom);
3010 		break;
3011 	default:
3012 		if (domain->mode != PAGE_MODE_NONE)
3013 			free_pagetable(domain);
3014 
3015 		if (domain->flags & PD_IOMMUV2_MASK)
3016 			free_gcr3_table(domain);
3017 
3018 		protection_domain_free(domain);
3019 		break;
3020 	}
3021 }
3022 
amd_iommu_detach_device(struct iommu_domain * dom,struct device * dev)3023 static void amd_iommu_detach_device(struct iommu_domain *dom,
3024 				    struct device *dev)
3025 {
3026 	struct iommu_dev_data *dev_data = dev->archdata.iommu;
3027 	struct amd_iommu *iommu;
3028 	u16 devid;
3029 
3030 	if (!check_device(dev))
3031 		return;
3032 
3033 	devid = get_device_id(dev);
3034 
3035 	if (dev_data->domain != NULL)
3036 		detach_device(dev);
3037 
3038 	iommu = amd_iommu_rlookup_table[devid];
3039 	if (!iommu)
3040 		return;
3041 
3042 	iommu_completion_wait(iommu);
3043 }
3044 
amd_iommu_attach_device(struct iommu_domain * dom,struct device * dev)3045 static int amd_iommu_attach_device(struct iommu_domain *dom,
3046 				   struct device *dev)
3047 {
3048 	struct protection_domain *domain = to_pdomain(dom);
3049 	struct iommu_dev_data *dev_data;
3050 	struct amd_iommu *iommu;
3051 	int ret;
3052 
3053 	if (!check_device(dev))
3054 		return -EINVAL;
3055 
3056 	dev_data = dev->archdata.iommu;
3057 
3058 	iommu = amd_iommu_rlookup_table[dev_data->devid];
3059 	if (!iommu)
3060 		return -EINVAL;
3061 
3062 	if (dev_data->domain)
3063 		detach_device(dev);
3064 
3065 	ret = attach_device(dev, domain);
3066 
3067 	iommu_completion_wait(iommu);
3068 
3069 	return ret;
3070 }
3071 
amd_iommu_map(struct iommu_domain * dom,unsigned long iova,phys_addr_t paddr,size_t page_size,int iommu_prot)3072 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
3073 			 phys_addr_t paddr, size_t page_size, int iommu_prot)
3074 {
3075 	struct protection_domain *domain = to_pdomain(dom);
3076 	int prot = 0;
3077 	int ret;
3078 
3079 	if (domain->mode == PAGE_MODE_NONE)
3080 		return -EINVAL;
3081 
3082 	if (iommu_prot & IOMMU_READ)
3083 		prot |= IOMMU_PROT_IR;
3084 	if (iommu_prot & IOMMU_WRITE)
3085 		prot |= IOMMU_PROT_IW;
3086 
3087 	mutex_lock(&domain->api_lock);
3088 	ret = iommu_map_page(domain, iova, paddr, prot, page_size);
3089 	mutex_unlock(&domain->api_lock);
3090 
3091 	return ret;
3092 }
3093 
amd_iommu_unmap(struct iommu_domain * dom,unsigned long iova,size_t page_size)3094 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3095 			   size_t page_size)
3096 {
3097 	struct protection_domain *domain = to_pdomain(dom);
3098 	size_t unmap_size;
3099 
3100 	if (domain->mode == PAGE_MODE_NONE)
3101 		return -EINVAL;
3102 
3103 	mutex_lock(&domain->api_lock);
3104 	unmap_size = iommu_unmap_page(domain, iova, page_size);
3105 	mutex_unlock(&domain->api_lock);
3106 
3107 	domain_flush_tlb_pde(domain);
3108 	domain_flush_complete(domain);
3109 
3110 	return unmap_size;
3111 }
3112 
amd_iommu_iova_to_phys(struct iommu_domain * dom,dma_addr_t iova)3113 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3114 					  dma_addr_t iova)
3115 {
3116 	struct protection_domain *domain = to_pdomain(dom);
3117 	unsigned long offset_mask, pte_pgsize;
3118 	u64 *pte, __pte;
3119 
3120 	if (domain->mode == PAGE_MODE_NONE)
3121 		return iova;
3122 
3123 	pte = fetch_pte(domain, iova, &pte_pgsize);
3124 
3125 	if (!pte || !IOMMU_PTE_PRESENT(*pte))
3126 		return 0;
3127 
3128 	offset_mask = pte_pgsize - 1;
3129 	__pte	    = *pte & PM_ADDR_MASK;
3130 
3131 	return (__pte & ~offset_mask) | (iova & offset_mask);
3132 }
3133 
amd_iommu_capable(enum iommu_cap cap)3134 static bool amd_iommu_capable(enum iommu_cap cap)
3135 {
3136 	switch (cap) {
3137 	case IOMMU_CAP_CACHE_COHERENCY:
3138 		return true;
3139 	case IOMMU_CAP_INTR_REMAP:
3140 		return (irq_remapping_enabled == 1);
3141 	case IOMMU_CAP_NOEXEC:
3142 		return false;
3143 	}
3144 
3145 	return false;
3146 }
3147 
amd_iommu_get_dm_regions(struct device * dev,struct list_head * head)3148 static void amd_iommu_get_dm_regions(struct device *dev,
3149 				     struct list_head *head)
3150 {
3151 	struct unity_map_entry *entry;
3152 	u16 devid;
3153 
3154 	devid = get_device_id(dev);
3155 
3156 	list_for_each_entry(entry, &amd_iommu_unity_map, list) {
3157 		struct iommu_dm_region *region;
3158 
3159 		if (devid < entry->devid_start || devid > entry->devid_end)
3160 			continue;
3161 
3162 		region = kzalloc(sizeof(*region), GFP_KERNEL);
3163 		if (!region) {
3164 			pr_err("Out of memory allocating dm-regions for %s\n",
3165 				dev_name(dev));
3166 			return;
3167 		}
3168 
3169 		region->start = entry->address_start;
3170 		region->length = entry->address_end - entry->address_start;
3171 		if (entry->prot & IOMMU_PROT_IR)
3172 			region->prot |= IOMMU_READ;
3173 		if (entry->prot & IOMMU_PROT_IW)
3174 			region->prot |= IOMMU_WRITE;
3175 
3176 		list_add_tail(&region->list, head);
3177 	}
3178 }
3179 
amd_iommu_put_dm_regions(struct device * dev,struct list_head * head)3180 static void amd_iommu_put_dm_regions(struct device *dev,
3181 				     struct list_head *head)
3182 {
3183 	struct iommu_dm_region *entry, *next;
3184 
3185 	list_for_each_entry_safe(entry, next, head, list)
3186 		kfree(entry);
3187 }
3188 
3189 static const struct iommu_ops amd_iommu_ops = {
3190 	.capable = amd_iommu_capable,
3191 	.domain_alloc = amd_iommu_domain_alloc,
3192 	.domain_free  = amd_iommu_domain_free,
3193 	.attach_dev = amd_iommu_attach_device,
3194 	.detach_dev = amd_iommu_detach_device,
3195 	.map = amd_iommu_map,
3196 	.unmap = amd_iommu_unmap,
3197 	.map_sg = default_iommu_map_sg,
3198 	.iova_to_phys = amd_iommu_iova_to_phys,
3199 	.add_device = amd_iommu_add_device,
3200 	.remove_device = amd_iommu_remove_device,
3201 	.device_group = pci_device_group,
3202 	.get_dm_regions = amd_iommu_get_dm_regions,
3203 	.put_dm_regions = amd_iommu_put_dm_regions,
3204 	.pgsize_bitmap	= AMD_IOMMU_PGSIZES,
3205 };
3206 
3207 /*****************************************************************************
3208  *
3209  * The next functions do a basic initialization of IOMMU for pass through
3210  * mode
3211  *
3212  * In passthrough mode the IOMMU is initialized and enabled but not used for
3213  * DMA-API translation.
3214  *
3215  *****************************************************************************/
3216 
3217 /* IOMMUv2 specific functions */
amd_iommu_register_ppr_notifier(struct notifier_block * nb)3218 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3219 {
3220 	return atomic_notifier_chain_register(&ppr_notifier, nb);
3221 }
3222 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3223 
amd_iommu_unregister_ppr_notifier(struct notifier_block * nb)3224 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3225 {
3226 	return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3227 }
3228 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3229 
amd_iommu_domain_direct_map(struct iommu_domain * dom)3230 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3231 {
3232 	struct protection_domain *domain = to_pdomain(dom);
3233 	unsigned long flags;
3234 
3235 	spin_lock_irqsave(&domain->lock, flags);
3236 
3237 	/* Update data structure */
3238 	domain->mode    = PAGE_MODE_NONE;
3239 	domain->updated = true;
3240 
3241 	/* Make changes visible to IOMMUs */
3242 	update_domain(domain);
3243 
3244 	/* Page-table is not visible to IOMMU anymore, so free it */
3245 	free_pagetable(domain);
3246 
3247 	spin_unlock_irqrestore(&domain->lock, flags);
3248 }
3249 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3250 
amd_iommu_domain_enable_v2(struct iommu_domain * dom,int pasids)3251 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3252 {
3253 	struct protection_domain *domain = to_pdomain(dom);
3254 	unsigned long flags;
3255 	int levels, ret;
3256 
3257 	if (pasids <= 0 || pasids > (PASID_MASK + 1))
3258 		return -EINVAL;
3259 
3260 	/* Number of GCR3 table levels required */
3261 	for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3262 		levels += 1;
3263 
3264 	if (levels > amd_iommu_max_glx_val)
3265 		return -EINVAL;
3266 
3267 	spin_lock_irqsave(&domain->lock, flags);
3268 
3269 	/*
3270 	 * Save us all sanity checks whether devices already in the
3271 	 * domain support IOMMUv2. Just force that the domain has no
3272 	 * devices attached when it is switched into IOMMUv2 mode.
3273 	 */
3274 	ret = -EBUSY;
3275 	if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3276 		goto out;
3277 
3278 	ret = -ENOMEM;
3279 	domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3280 	if (domain->gcr3_tbl == NULL)
3281 		goto out;
3282 
3283 	domain->glx      = levels;
3284 	domain->flags   |= PD_IOMMUV2_MASK;
3285 	domain->updated  = true;
3286 
3287 	update_domain(domain);
3288 
3289 	ret = 0;
3290 
3291 out:
3292 	spin_unlock_irqrestore(&domain->lock, flags);
3293 
3294 	return ret;
3295 }
3296 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3297 
__flush_pasid(struct protection_domain * domain,int pasid,u64 address,bool size)3298 static int __flush_pasid(struct protection_domain *domain, int pasid,
3299 			 u64 address, bool size)
3300 {
3301 	struct iommu_dev_data *dev_data;
3302 	struct iommu_cmd cmd;
3303 	int i, ret;
3304 
3305 	if (!(domain->flags & PD_IOMMUV2_MASK))
3306 		return -EINVAL;
3307 
3308 	build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3309 
3310 	/*
3311 	 * IOMMU TLB needs to be flushed before Device TLB to
3312 	 * prevent device TLB refill from IOMMU TLB
3313 	 */
3314 	for (i = 0; i < amd_iommus_present; ++i) {
3315 		if (domain->dev_iommu[i] == 0)
3316 			continue;
3317 
3318 		ret = iommu_queue_command(amd_iommus[i], &cmd);
3319 		if (ret != 0)
3320 			goto out;
3321 	}
3322 
3323 	/* Wait until IOMMU TLB flushes are complete */
3324 	domain_flush_complete(domain);
3325 
3326 	/* Now flush device TLBs */
3327 	list_for_each_entry(dev_data, &domain->dev_list, list) {
3328 		struct amd_iommu *iommu;
3329 		int qdep;
3330 
3331 		/*
3332 		   There might be non-IOMMUv2 capable devices in an IOMMUv2
3333 		 * domain.
3334 		 */
3335 		if (!dev_data->ats.enabled)
3336 			continue;
3337 
3338 		qdep  = dev_data->ats.qdep;
3339 		iommu = amd_iommu_rlookup_table[dev_data->devid];
3340 
3341 		build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3342 				      qdep, address, size);
3343 
3344 		ret = iommu_queue_command(iommu, &cmd);
3345 		if (ret != 0)
3346 			goto out;
3347 	}
3348 
3349 	/* Wait until all device TLBs are flushed */
3350 	domain_flush_complete(domain);
3351 
3352 	ret = 0;
3353 
3354 out:
3355 
3356 	return ret;
3357 }
3358 
__amd_iommu_flush_page(struct protection_domain * domain,int pasid,u64 address)3359 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3360 				  u64 address)
3361 {
3362 	INC_STATS_COUNTER(invalidate_iotlb);
3363 
3364 	return __flush_pasid(domain, pasid, address, false);
3365 }
3366 
amd_iommu_flush_page(struct iommu_domain * dom,int pasid,u64 address)3367 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3368 			 u64 address)
3369 {
3370 	struct protection_domain *domain = to_pdomain(dom);
3371 	unsigned long flags;
3372 	int ret;
3373 
3374 	spin_lock_irqsave(&domain->lock, flags);
3375 	ret = __amd_iommu_flush_page(domain, pasid, address);
3376 	spin_unlock_irqrestore(&domain->lock, flags);
3377 
3378 	return ret;
3379 }
3380 EXPORT_SYMBOL(amd_iommu_flush_page);
3381 
__amd_iommu_flush_tlb(struct protection_domain * domain,int pasid)3382 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3383 {
3384 	INC_STATS_COUNTER(invalidate_iotlb_all);
3385 
3386 	return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3387 			     true);
3388 }
3389 
amd_iommu_flush_tlb(struct iommu_domain * dom,int pasid)3390 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3391 {
3392 	struct protection_domain *domain = to_pdomain(dom);
3393 	unsigned long flags;
3394 	int ret;
3395 
3396 	spin_lock_irqsave(&domain->lock, flags);
3397 	ret = __amd_iommu_flush_tlb(domain, pasid);
3398 	spin_unlock_irqrestore(&domain->lock, flags);
3399 
3400 	return ret;
3401 }
3402 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3403 
__get_gcr3_pte(u64 * root,int level,int pasid,bool alloc)3404 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3405 {
3406 	int index;
3407 	u64 *pte;
3408 
3409 	while (true) {
3410 
3411 		index = (pasid >> (9 * level)) & 0x1ff;
3412 		pte   = &root[index];
3413 
3414 		if (level == 0)
3415 			break;
3416 
3417 		if (!(*pte & GCR3_VALID)) {
3418 			if (!alloc)
3419 				return NULL;
3420 
3421 			root = (void *)get_zeroed_page(GFP_ATOMIC);
3422 			if (root == NULL)
3423 				return NULL;
3424 
3425 			*pte = __pa(root) | GCR3_VALID;
3426 		}
3427 
3428 		root = __va(*pte & PAGE_MASK);
3429 
3430 		level -= 1;
3431 	}
3432 
3433 	return pte;
3434 }
3435 
__set_gcr3(struct protection_domain * domain,int pasid,unsigned long cr3)3436 static int __set_gcr3(struct protection_domain *domain, int pasid,
3437 		      unsigned long cr3)
3438 {
3439 	u64 *pte;
3440 
3441 	if (domain->mode != PAGE_MODE_NONE)
3442 		return -EINVAL;
3443 
3444 	pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3445 	if (pte == NULL)
3446 		return -ENOMEM;
3447 
3448 	*pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3449 
3450 	return __amd_iommu_flush_tlb(domain, pasid);
3451 }
3452 
__clear_gcr3(struct protection_domain * domain,int pasid)3453 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3454 {
3455 	u64 *pte;
3456 
3457 	if (domain->mode != PAGE_MODE_NONE)
3458 		return -EINVAL;
3459 
3460 	pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3461 	if (pte == NULL)
3462 		return 0;
3463 
3464 	*pte = 0;
3465 
3466 	return __amd_iommu_flush_tlb(domain, pasid);
3467 }
3468 
amd_iommu_domain_set_gcr3(struct iommu_domain * dom,int pasid,unsigned long cr3)3469 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3470 			      unsigned long cr3)
3471 {
3472 	struct protection_domain *domain = to_pdomain(dom);
3473 	unsigned long flags;
3474 	int ret;
3475 
3476 	spin_lock_irqsave(&domain->lock, flags);
3477 	ret = __set_gcr3(domain, pasid, cr3);
3478 	spin_unlock_irqrestore(&domain->lock, flags);
3479 
3480 	return ret;
3481 }
3482 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3483 
amd_iommu_domain_clear_gcr3(struct iommu_domain * dom,int pasid)3484 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3485 {
3486 	struct protection_domain *domain = to_pdomain(dom);
3487 	unsigned long flags;
3488 	int ret;
3489 
3490 	spin_lock_irqsave(&domain->lock, flags);
3491 	ret = __clear_gcr3(domain, pasid);
3492 	spin_unlock_irqrestore(&domain->lock, flags);
3493 
3494 	return ret;
3495 }
3496 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3497 
amd_iommu_complete_ppr(struct pci_dev * pdev,int pasid,int status,int tag)3498 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3499 			   int status, int tag)
3500 {
3501 	struct iommu_dev_data *dev_data;
3502 	struct amd_iommu *iommu;
3503 	struct iommu_cmd cmd;
3504 
3505 	INC_STATS_COUNTER(complete_ppr);
3506 
3507 	dev_data = get_dev_data(&pdev->dev);
3508 	iommu    = amd_iommu_rlookup_table[dev_data->devid];
3509 
3510 	build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3511 			   tag, dev_data->pri_tlp);
3512 
3513 	return iommu_queue_command(iommu, &cmd);
3514 }
3515 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3516 
amd_iommu_get_v2_domain(struct pci_dev * pdev)3517 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3518 {
3519 	struct protection_domain *pdomain;
3520 
3521 	pdomain = get_domain(&pdev->dev);
3522 	if (IS_ERR(pdomain))
3523 		return NULL;
3524 
3525 	/* Only return IOMMUv2 domains */
3526 	if (!(pdomain->flags & PD_IOMMUV2_MASK))
3527 		return NULL;
3528 
3529 	return &pdomain->domain;
3530 }
3531 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3532 
amd_iommu_enable_device_erratum(struct pci_dev * pdev,u32 erratum)3533 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3534 {
3535 	struct iommu_dev_data *dev_data;
3536 
3537 	if (!amd_iommu_v2_supported())
3538 		return;
3539 
3540 	dev_data = get_dev_data(&pdev->dev);
3541 	dev_data->errata |= (1 << erratum);
3542 }
3543 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3544 
amd_iommu_device_info(struct pci_dev * pdev,struct amd_iommu_device_info * info)3545 int amd_iommu_device_info(struct pci_dev *pdev,
3546                           struct amd_iommu_device_info *info)
3547 {
3548 	int max_pasids;
3549 	int pos;
3550 
3551 	if (pdev == NULL || info == NULL)
3552 		return -EINVAL;
3553 
3554 	if (!amd_iommu_v2_supported())
3555 		return -EINVAL;
3556 
3557 	memset(info, 0, sizeof(*info));
3558 
3559 	pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3560 	if (pos)
3561 		info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3562 
3563 	pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3564 	if (pos)
3565 		info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3566 
3567 	pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3568 	if (pos) {
3569 		int features;
3570 
3571 		max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3572 		max_pasids = min(max_pasids, (1 << 20));
3573 
3574 		info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3575 		info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3576 
3577 		features = pci_pasid_features(pdev);
3578 		if (features & PCI_PASID_CAP_EXEC)
3579 			info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3580 		if (features & PCI_PASID_CAP_PRIV)
3581 			info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3582 	}
3583 
3584 	return 0;
3585 }
3586 EXPORT_SYMBOL(amd_iommu_device_info);
3587 
3588 #ifdef CONFIG_IRQ_REMAP
3589 
3590 /*****************************************************************************
3591  *
3592  * Interrupt Remapping Implementation
3593  *
3594  *****************************************************************************/
3595 
3596 union irte {
3597 	u32 val;
3598 	struct {
3599 		u32 valid	: 1,
3600 		    no_fault	: 1,
3601 		    int_type	: 3,
3602 		    rq_eoi	: 1,
3603 		    dm		: 1,
3604 		    rsvd_1	: 1,
3605 		    destination	: 8,
3606 		    vector	: 8,
3607 		    rsvd_2	: 8;
3608 	} fields;
3609 };
3610 
3611 struct irq_2_irte {
3612 	u16 devid; /* Device ID for IRTE table */
3613 	u16 index; /* Index into IRTE table*/
3614 };
3615 
3616 struct amd_ir_data {
3617 	struct irq_2_irte			irq_2_irte;
3618 	union irte				irte_entry;
3619 	union {
3620 		struct msi_msg			msi_entry;
3621 	};
3622 };
3623 
3624 static struct irq_chip amd_ir_chip;
3625 
3626 #define DTE_IRQ_PHYS_ADDR_MASK	(((1ULL << 45)-1) << 6)
3627 #define DTE_IRQ_REMAP_INTCTL    (2ULL << 60)
3628 #define DTE_IRQ_TABLE_LEN       (9ULL << 1)
3629 #define DTE_IRQ_REMAP_ENABLE    1ULL
3630 
set_dte_irq_entry(u16 devid,struct irq_remap_table * table)3631 static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3632 {
3633 	u64 dte;
3634 
3635 	dte	= amd_iommu_dev_table[devid].data[2];
3636 	dte	&= ~DTE_IRQ_PHYS_ADDR_MASK;
3637 	dte	|= virt_to_phys(table->table);
3638 	dte	|= DTE_IRQ_REMAP_INTCTL;
3639 	dte	|= DTE_IRQ_TABLE_LEN;
3640 	dte	|= DTE_IRQ_REMAP_ENABLE;
3641 
3642 	amd_iommu_dev_table[devid].data[2] = dte;
3643 }
3644 
3645 #define IRTE_ALLOCATED (~1U)
3646 
get_irq_table(u16 devid,bool ioapic)3647 static struct irq_remap_table *get_irq_table(u16 devid, bool ioapic)
3648 {
3649 	struct irq_remap_table *table = NULL;
3650 	struct amd_iommu *iommu;
3651 	unsigned long flags;
3652 	u16 alias;
3653 
3654 	write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3655 
3656 	iommu = amd_iommu_rlookup_table[devid];
3657 	if (!iommu)
3658 		goto out_unlock;
3659 
3660 	table = irq_lookup_table[devid];
3661 	if (table)
3662 		goto out;
3663 
3664 	alias = amd_iommu_alias_table[devid];
3665 	table = irq_lookup_table[alias];
3666 	if (table) {
3667 		irq_lookup_table[devid] = table;
3668 		set_dte_irq_entry(devid, table);
3669 		iommu_flush_dte(iommu, devid);
3670 		goto out;
3671 	}
3672 
3673 	/* Nothing there yet, allocate new irq remapping table */
3674 	table = kzalloc(sizeof(*table), GFP_ATOMIC);
3675 	if (!table)
3676 		goto out;
3677 
3678 	/* Initialize table spin-lock */
3679 	spin_lock_init(&table->lock);
3680 
3681 	if (ioapic)
3682 		/* Keep the first 32 indexes free for IOAPIC interrupts */
3683 		table->min_index = 32;
3684 
3685 	table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_ATOMIC);
3686 	if (!table->table) {
3687 		kfree(table);
3688 		table = NULL;
3689 		goto out;
3690 	}
3691 
3692 	memset(table->table, 0, MAX_IRQS_PER_TABLE * sizeof(u32));
3693 
3694 	if (ioapic) {
3695 		int i;
3696 
3697 		for (i = 0; i < 32; ++i)
3698 			table->table[i] = IRTE_ALLOCATED;
3699 	}
3700 
3701 	irq_lookup_table[devid] = table;
3702 	set_dte_irq_entry(devid, table);
3703 	iommu_flush_dte(iommu, devid);
3704 	if (devid != alias) {
3705 		irq_lookup_table[alias] = table;
3706 		set_dte_irq_entry(alias, table);
3707 		iommu_flush_dte(iommu, alias);
3708 	}
3709 
3710 out:
3711 	iommu_completion_wait(iommu);
3712 
3713 out_unlock:
3714 	write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3715 
3716 	return table;
3717 }
3718 
alloc_irq_index(u16 devid,int count)3719 static int alloc_irq_index(u16 devid, int count)
3720 {
3721 	struct irq_remap_table *table;
3722 	unsigned long flags;
3723 	int index, c;
3724 
3725 	table = get_irq_table(devid, false);
3726 	if (!table)
3727 		return -ENODEV;
3728 
3729 	spin_lock_irqsave(&table->lock, flags);
3730 
3731 	/* Scan table for free entries */
3732 	for (c = 0, index = table->min_index;
3733 	     index < MAX_IRQS_PER_TABLE;
3734 	     ++index) {
3735 		if (table->table[index] == 0)
3736 			c += 1;
3737 		else
3738 			c = 0;
3739 
3740 		if (c == count)	{
3741 			for (; c != 0; --c)
3742 				table->table[index - c + 1] = IRTE_ALLOCATED;
3743 
3744 			index -= count - 1;
3745 			goto out;
3746 		}
3747 	}
3748 
3749 	index = -ENOSPC;
3750 
3751 out:
3752 	spin_unlock_irqrestore(&table->lock, flags);
3753 
3754 	return index;
3755 }
3756 
modify_irte(u16 devid,int index,union irte irte)3757 static int modify_irte(u16 devid, int index, union irte irte)
3758 {
3759 	struct irq_remap_table *table;
3760 	struct amd_iommu *iommu;
3761 	unsigned long flags;
3762 
3763 	iommu = amd_iommu_rlookup_table[devid];
3764 	if (iommu == NULL)
3765 		return -EINVAL;
3766 
3767 	table = get_irq_table(devid, false);
3768 	if (!table)
3769 		return -ENOMEM;
3770 
3771 	spin_lock_irqsave(&table->lock, flags);
3772 	table->table[index] = irte.val;
3773 	spin_unlock_irqrestore(&table->lock, flags);
3774 
3775 	iommu_flush_irt(iommu, devid);
3776 	iommu_completion_wait(iommu);
3777 
3778 	return 0;
3779 }
3780 
free_irte(u16 devid,int index)3781 static void free_irte(u16 devid, int index)
3782 {
3783 	struct irq_remap_table *table;
3784 	struct amd_iommu *iommu;
3785 	unsigned long flags;
3786 
3787 	iommu = amd_iommu_rlookup_table[devid];
3788 	if (iommu == NULL)
3789 		return;
3790 
3791 	table = get_irq_table(devid, false);
3792 	if (!table)
3793 		return;
3794 
3795 	spin_lock_irqsave(&table->lock, flags);
3796 	table->table[index] = 0;
3797 	spin_unlock_irqrestore(&table->lock, flags);
3798 
3799 	iommu_flush_irt(iommu, devid);
3800 	iommu_completion_wait(iommu);
3801 }
3802 
get_devid(struct irq_alloc_info * info)3803 static int get_devid(struct irq_alloc_info *info)
3804 {
3805 	int devid = -1;
3806 
3807 	switch (info->type) {
3808 	case X86_IRQ_ALLOC_TYPE_IOAPIC:
3809 		devid     = get_ioapic_devid(info->ioapic_id);
3810 		break;
3811 	case X86_IRQ_ALLOC_TYPE_HPET:
3812 		devid     = get_hpet_devid(info->hpet_id);
3813 		break;
3814 	case X86_IRQ_ALLOC_TYPE_MSI:
3815 	case X86_IRQ_ALLOC_TYPE_MSIX:
3816 		devid = get_device_id(&info->msi_dev->dev);
3817 		break;
3818 	default:
3819 		BUG_ON(1);
3820 		break;
3821 	}
3822 
3823 	return devid;
3824 }
3825 
get_ir_irq_domain(struct irq_alloc_info * info)3826 static struct irq_domain *get_ir_irq_domain(struct irq_alloc_info *info)
3827 {
3828 	struct amd_iommu *iommu;
3829 	int devid;
3830 
3831 	if (!info)
3832 		return NULL;
3833 
3834 	devid = get_devid(info);
3835 	if (devid >= 0) {
3836 		iommu = amd_iommu_rlookup_table[devid];
3837 		if (iommu)
3838 			return iommu->ir_domain;
3839 	}
3840 
3841 	return NULL;
3842 }
3843 
get_irq_domain(struct irq_alloc_info * info)3844 static struct irq_domain *get_irq_domain(struct irq_alloc_info *info)
3845 {
3846 	struct amd_iommu *iommu;
3847 	int devid;
3848 
3849 	if (!info)
3850 		return NULL;
3851 
3852 	switch (info->type) {
3853 	case X86_IRQ_ALLOC_TYPE_MSI:
3854 	case X86_IRQ_ALLOC_TYPE_MSIX:
3855 		devid = get_device_id(&info->msi_dev->dev);
3856 		if (devid >= 0) {
3857 			iommu = amd_iommu_rlookup_table[devid];
3858 			if (iommu)
3859 				return iommu->msi_domain;
3860 		}
3861 		break;
3862 	default:
3863 		break;
3864 	}
3865 
3866 	return NULL;
3867 }
3868 
3869 struct irq_remap_ops amd_iommu_irq_ops = {
3870 	.prepare		= amd_iommu_prepare,
3871 	.enable			= amd_iommu_enable,
3872 	.disable		= amd_iommu_disable,
3873 	.reenable		= amd_iommu_reenable,
3874 	.enable_faulting	= amd_iommu_enable_faulting,
3875 	.get_ir_irq_domain	= get_ir_irq_domain,
3876 	.get_irq_domain		= get_irq_domain,
3877 };
3878 
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)3879 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
3880 				       struct irq_cfg *irq_cfg,
3881 				       struct irq_alloc_info *info,
3882 				       int devid, int index, int sub_handle)
3883 {
3884 	struct irq_2_irte *irte_info = &data->irq_2_irte;
3885 	struct msi_msg *msg = &data->msi_entry;
3886 	union irte *irte = &data->irte_entry;
3887 	struct IO_APIC_route_entry *entry;
3888 
3889 	data->irq_2_irte.devid = devid;
3890 	data->irq_2_irte.index = index + sub_handle;
3891 
3892 	/* Setup IRTE for IOMMU */
3893 	irte->val = 0;
3894 	irte->fields.vector      = irq_cfg->vector;
3895 	irte->fields.int_type    = apic->irq_delivery_mode;
3896 	irte->fields.destination = irq_cfg->dest_apicid;
3897 	irte->fields.dm          = apic->irq_dest_mode;
3898 	irte->fields.valid       = 1;
3899 
3900 	switch (info->type) {
3901 	case X86_IRQ_ALLOC_TYPE_IOAPIC:
3902 		/* Setup IOAPIC entry */
3903 		entry = info->ioapic_entry;
3904 		info->ioapic_entry = NULL;
3905 		memset(entry, 0, sizeof(*entry));
3906 		entry->vector        = index;
3907 		entry->mask          = 0;
3908 		entry->trigger       = info->ioapic_trigger;
3909 		entry->polarity      = info->ioapic_polarity;
3910 		/* Mask level triggered irqs. */
3911 		if (info->ioapic_trigger)
3912 			entry->mask = 1;
3913 		break;
3914 
3915 	case X86_IRQ_ALLOC_TYPE_HPET:
3916 	case X86_IRQ_ALLOC_TYPE_MSI:
3917 	case X86_IRQ_ALLOC_TYPE_MSIX:
3918 		msg->address_hi = MSI_ADDR_BASE_HI;
3919 		msg->address_lo = MSI_ADDR_BASE_LO;
3920 		msg->data = irte_info->index;
3921 		break;
3922 
3923 	default:
3924 		BUG_ON(1);
3925 		break;
3926 	}
3927 }
3928 
irq_remapping_alloc(struct irq_domain * domain,unsigned int virq,unsigned int nr_irqs,void * arg)3929 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
3930 			       unsigned int nr_irqs, void *arg)
3931 {
3932 	struct irq_alloc_info *info = arg;
3933 	struct irq_data *irq_data;
3934 	struct amd_ir_data *data;
3935 	struct irq_cfg *cfg;
3936 	int i, ret, devid;
3937 	int index = -1;
3938 
3939 	if (!info)
3940 		return -EINVAL;
3941 	if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_MSI &&
3942 	    info->type != X86_IRQ_ALLOC_TYPE_MSIX)
3943 		return -EINVAL;
3944 
3945 	/*
3946 	 * With IRQ remapping enabled, don't need contiguous CPU vectors
3947 	 * to support multiple MSI interrupts.
3948 	 */
3949 	if (info->type == X86_IRQ_ALLOC_TYPE_MSI)
3950 		info->flags &= ~X86_IRQ_ALLOC_CONTIGUOUS_VECTORS;
3951 
3952 	devid = get_devid(info);
3953 	if (devid < 0)
3954 		return -EINVAL;
3955 
3956 	ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
3957 	if (ret < 0)
3958 		return ret;
3959 
3960 	if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
3961 		if (get_irq_table(devid, true))
3962 			index = info->ioapic_pin;
3963 		else
3964 			ret = -ENOMEM;
3965 	} else {
3966 		index = alloc_irq_index(devid, nr_irqs);
3967 	}
3968 	if (index < 0) {
3969 		pr_warn("Failed to allocate IRTE\n");
3970 		goto out_free_parent;
3971 	}
3972 
3973 	for (i = 0; i < nr_irqs; i++) {
3974 		irq_data = irq_domain_get_irq_data(domain, virq + i);
3975 		cfg = irqd_cfg(irq_data);
3976 		if (!irq_data || !cfg) {
3977 			ret = -EINVAL;
3978 			goto out_free_data;
3979 		}
3980 
3981 		ret = -ENOMEM;
3982 		data = kzalloc(sizeof(*data), GFP_KERNEL);
3983 		if (!data)
3984 			goto out_free_data;
3985 
3986 		irq_data->hwirq = (devid << 16) + i;
3987 		irq_data->chip_data = data;
3988 		irq_data->chip = &amd_ir_chip;
3989 		irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
3990 		irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
3991 	}
3992 
3993 	return 0;
3994 
3995 out_free_data:
3996 	for (i--; i >= 0; i--) {
3997 		irq_data = irq_domain_get_irq_data(domain, virq + i);
3998 		if (irq_data)
3999 			kfree(irq_data->chip_data);
4000 	}
4001 	for (i = 0; i < nr_irqs; i++)
4002 		free_irte(devid, index + i);
4003 out_free_parent:
4004 	irq_domain_free_irqs_common(domain, virq, nr_irqs);
4005 	return ret;
4006 }
4007 
irq_remapping_free(struct irq_domain * domain,unsigned int virq,unsigned int nr_irqs)4008 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
4009 			       unsigned int nr_irqs)
4010 {
4011 	struct irq_2_irte *irte_info;
4012 	struct irq_data *irq_data;
4013 	struct amd_ir_data *data;
4014 	int i;
4015 
4016 	for (i = 0; i < nr_irqs; i++) {
4017 		irq_data = irq_domain_get_irq_data(domain, virq  + i);
4018 		if (irq_data && irq_data->chip_data) {
4019 			data = irq_data->chip_data;
4020 			irte_info = &data->irq_2_irte;
4021 			free_irte(irte_info->devid, irte_info->index);
4022 			kfree(data);
4023 		}
4024 	}
4025 	irq_domain_free_irqs_common(domain, virq, nr_irqs);
4026 }
4027 
irq_remapping_activate(struct irq_domain * domain,struct irq_data * irq_data)4028 static void irq_remapping_activate(struct irq_domain *domain,
4029 				   struct irq_data *irq_data)
4030 {
4031 	struct amd_ir_data *data = irq_data->chip_data;
4032 	struct irq_2_irte *irte_info = &data->irq_2_irte;
4033 
4034 	modify_irte(irte_info->devid, irte_info->index, data->irte_entry);
4035 }
4036 
irq_remapping_deactivate(struct irq_domain * domain,struct irq_data * irq_data)4037 static void irq_remapping_deactivate(struct irq_domain *domain,
4038 				     struct irq_data *irq_data)
4039 {
4040 	struct amd_ir_data *data = irq_data->chip_data;
4041 	struct irq_2_irte *irte_info = &data->irq_2_irte;
4042 	union irte entry;
4043 
4044 	entry.val = 0;
4045 	modify_irte(irte_info->devid, irte_info->index, data->irte_entry);
4046 }
4047 
4048 static struct irq_domain_ops amd_ir_domain_ops = {
4049 	.alloc = irq_remapping_alloc,
4050 	.free = irq_remapping_free,
4051 	.activate = irq_remapping_activate,
4052 	.deactivate = irq_remapping_deactivate,
4053 };
4054 
amd_ir_set_affinity(struct irq_data * data,const struct cpumask * mask,bool force)4055 static int amd_ir_set_affinity(struct irq_data *data,
4056 			       const struct cpumask *mask, bool force)
4057 {
4058 	struct amd_ir_data *ir_data = data->chip_data;
4059 	struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4060 	struct irq_cfg *cfg = irqd_cfg(data);
4061 	struct irq_data *parent = data->parent_data;
4062 	int ret;
4063 
4064 	ret = parent->chip->irq_set_affinity(parent, mask, force);
4065 	if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
4066 		return ret;
4067 
4068 	/*
4069 	 * Atomically updates the IRTE with the new destination, vector
4070 	 * and flushes the interrupt entry cache.
4071 	 */
4072 	ir_data->irte_entry.fields.vector = cfg->vector;
4073 	ir_data->irte_entry.fields.destination = cfg->dest_apicid;
4074 	modify_irte(irte_info->devid, irte_info->index, ir_data->irte_entry);
4075 
4076 	/*
4077 	 * After this point, all the interrupts will start arriving
4078 	 * at the new destination. So, time to cleanup the previous
4079 	 * vector allocation.
4080 	 */
4081 	send_cleanup_vector(cfg);
4082 
4083 	return IRQ_SET_MASK_OK_DONE;
4084 }
4085 
ir_compose_msi_msg(struct irq_data * irq_data,struct msi_msg * msg)4086 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
4087 {
4088 	struct amd_ir_data *ir_data = irq_data->chip_data;
4089 
4090 	*msg = ir_data->msi_entry;
4091 }
4092 
4093 static struct irq_chip amd_ir_chip = {
4094 	.irq_ack = ir_ack_apic_edge,
4095 	.irq_set_affinity = amd_ir_set_affinity,
4096 	.irq_compose_msi_msg = ir_compose_msi_msg,
4097 };
4098 
amd_iommu_create_irq_domain(struct amd_iommu * iommu)4099 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
4100 {
4101 	iommu->ir_domain = irq_domain_add_tree(NULL, &amd_ir_domain_ops, iommu);
4102 	if (!iommu->ir_domain)
4103 		return -ENOMEM;
4104 
4105 	iommu->ir_domain->parent = arch_get_ir_parent_domain();
4106 	iommu->msi_domain = arch_create_msi_irq_domain(iommu->ir_domain);
4107 
4108 	return 0;
4109 }
4110 #endif
4111