]> git.proxmox.com Git - mirror_ubuntu-zesty-kernel.git/blob - drivers/iommu/amd_iommu.c
projects / mirror_ubuntu-zesty-kernel.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
iommu: Add a new type field in iommu_resv_region
[mirror_ubuntu-zesty-kernel.git] / drivers / iommu / amd_iommu.c
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/acpi.h>
23 #include <linux/amba/bus.h>
24 #include <linux/platform_device.h>
25 #include <linux/pci-ats.h>
26 #include <linux/bitmap.h>
27 #include <linux/slab.h>
28 #include <linux/debugfs.h>
29 #include <linux/scatterlist.h>
30 #include <linux/dma-mapping.h>
31 #include <linux/iommu-helper.h>
32 #include <linux/iommu.h>
33 #include <linux/delay.h>
34 #include <linux/amd-iommu.h>
35 #include <linux/notifier.h>
36 #include <linux/export.h>
37 #include <linux/irq.h>
38 #include <linux/msi.h>
39 #include <linux/dma-contiguous.h>
40 #include <linux/irqdomain.h>
41 #include <linux/percpu.h>
42 #include <linux/iova.h>
43 #include <asm/irq_remapping.h>
44 #include <asm/io_apic.h>
45 #include <asm/apic.h>
46 #include <asm/hw_irq.h>
47 #include <asm/msidef.h>
48 #include <asm/proto.h>
49 #include <asm/iommu.h>
50 #include <asm/gart.h>
51 #include <asm/dma.h>
52
53 #include "amd_iommu_proto.h"
54 #include "amd_iommu_types.h"
55 #include "irq_remapping.h"
56
57 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
58
59 #define LOOP_TIMEOUT 100000
60
61 /* IO virtual address start page frame number */
62 #define IOVA_START_PFN (1)
63 #define IOVA_PFN(addr) ((addr) >> PAGE_SHIFT)
64 #define DMA_32BIT_PFN IOVA_PFN(DMA_BIT_MASK(32))
65
66 /* Reserved IOVA ranges */
67 #define MSI_RANGE_START (0xfee00000)
68 #define MSI_RANGE_END (0xfeefffff)
69 #define HT_RANGE_START (0xfd00000000ULL)
70 #define HT_RANGE_END (0xffffffffffULL)
71
72 /*
73 * This bitmap is used to advertise the page sizes our hardware support
74 * to the IOMMU core, which will then use this information to split
75 * physically contiguous memory regions it is mapping into page sizes
76 * that we support.
77 *
78 * 512GB Pages are not supported due to a hardware bug
79 */
80 #define AMD_IOMMU_PGSIZES ((~0xFFFUL) & ~(2ULL << 38))
81
82 static DEFINE_RWLOCK(amd_iommu_devtable_lock);
83
84 /* List of all available dev_data structures */
85 static LIST_HEAD(dev_data_list);
86 static DEFINE_SPINLOCK(dev_data_list_lock);
87
88 LIST_HEAD(ioapic_map);
89 LIST_HEAD(hpet_map);
90 LIST_HEAD(acpihid_map);
91
92 #define FLUSH_QUEUE_SIZE 256
93
94 struct flush_queue_entry {
95 unsigned long iova_pfn;
96 unsigned long pages;
97 struct dma_ops_domain *dma_dom;
98 };
99
100 struct flush_queue {
101 spinlock_t lock;
102 unsigned next;
103 struct flush_queue_entry *entries;
104 };
105
106 static DEFINE_PER_CPU(struct flush_queue, flush_queue);
107
108 static atomic_t queue_timer_on;
109 static struct timer_list queue_timer;
110
111 /*
112 * Domain for untranslated devices - only allocated
113 * if iommu=pt passed on kernel cmd line.
114 */
115 static const struct iommu_ops amd_iommu_ops;
116
117 static ATOMIC_NOTIFIER_HEAD(ppr_notifier);
118 int amd_iommu_max_glx_val = -1;
119
120 static struct dma_map_ops amd_iommu_dma_ops;
121
122 /*
123 * This struct contains device specific data for the IOMMU
124 */
125 struct iommu_dev_data {
126 struct list_head list; /* For domain->dev_list */
127 struct list_head dev_data_list; /* For global dev_data_list */
128 struct protection_domain *domain; /* Domain the device is bound to */
129 u16 devid; /* PCI Device ID */
130 u16 alias; /* Alias Device ID */
131 bool iommu_v2; /* Device can make use of IOMMUv2 */
132 bool passthrough; /* Device is identity mapped */
133 struct {
134 bool enabled;
135 int qdep;
136 } ats; /* ATS state */
137 bool pri_tlp; /* PASID TLB required for
138 PPR completions */
139 u32 errata; /* Bitmap for errata to apply */
140 bool use_vapic; /* Enable device to use vapic mode */
141 };
142
143 /*
144 * general struct to manage commands send to an IOMMU
145 */
146 struct iommu_cmd {
147 u32 data[4];
148 };
149
150 struct kmem_cache *amd_iommu_irq_cache;
151
152 static void update_domain(struct protection_domain *domain);
153 static int protection_domain_init(struct protection_domain *domain);
154 static void detach_device(struct device *dev);
155
156 /*
157 * Data container for a dma_ops specific protection domain
158 */
159 struct dma_ops_domain {
160 /* generic protection domain information */
161 struct protection_domain domain;
162
163 /* IOVA RB-Tree */
164 struct iova_domain iovad;
165 };
166
167 static struct iova_domain reserved_iova_ranges;
168 static struct lock_class_key reserved_rbtree_key;
169
170 /****************************************************************************
171 *
172 * Helper functions
173 *
174 ****************************************************************************/
175
176 static inline int match_hid_uid(struct device *dev,
177 struct acpihid_map_entry *entry)
178 {
179 const char *hid, *uid;
180
181 hid = acpi_device_hid(ACPI_COMPANION(dev));
182 uid = acpi_device_uid(ACPI_COMPANION(dev));
183
184 if (!hid || !(*hid))
185 return -ENODEV;
186
187 if (!uid || !(*uid))
188 return strcmp(hid, entry->hid);
189
190 if (!(*entry->uid))
191 return strcmp(hid, entry->hid);
192
193 return (strcmp(hid, entry->hid) || strcmp(uid, entry->uid));
194 }
195
196 static inline u16 get_pci_device_id(struct device *dev)
197 {
198 struct pci_dev *pdev = to_pci_dev(dev);
199
200 return PCI_DEVID(pdev->bus->number, pdev->devfn);
201 }
202
203 static inline int get_acpihid_device_id(struct device *dev,
204 struct acpihid_map_entry **entry)
205 {
206 struct acpihid_map_entry *p;
207
208 list_for_each_entry(p, &acpihid_map, list) {
209 if (!match_hid_uid(dev, p)) {
210 if (entry)
211 *entry = p;
212 return p->devid;
213 }
214 }
215 return -EINVAL;
216 }
217
218 static inline int get_device_id(struct device *dev)
219 {
220 int devid;
221
222 if (dev_is_pci(dev))
223 devid = get_pci_device_id(dev);
224 else
225 devid = get_acpihid_device_id(dev, NULL);
226
227 return devid;
228 }
229
230 static struct protection_domain *to_pdomain(struct iommu_domain *dom)
231 {
232 return container_of(dom, struct protection_domain, domain);
233 }
234
235 static struct dma_ops_domain* to_dma_ops_domain(struct protection_domain *domain)
236 {
237 BUG_ON(domain->flags != PD_DMA_OPS_MASK);
238 return container_of(domain, struct dma_ops_domain, domain);
239 }
240
241 static struct iommu_dev_data *alloc_dev_data(u16 devid)
242 {
243 struct iommu_dev_data *dev_data;
244 unsigned long flags;
245
246 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
247 if (!dev_data)
248 return NULL;
249
250 dev_data->devid = devid;
251
252 spin_lock_irqsave(&dev_data_list_lock, flags);
253 list_add_tail(&dev_data->dev_data_list, &dev_data_list);
254 spin_unlock_irqrestore(&dev_data_list_lock, flags);
255
256 return dev_data;
257 }
258
259 static struct iommu_dev_data *search_dev_data(u16 devid)
260 {
261 struct iommu_dev_data *dev_data;
262 unsigned long flags;
263
264 spin_lock_irqsave(&dev_data_list_lock, flags);
265 list_for_each_entry(dev_data, &dev_data_list, dev_data_list) {
266 if (dev_data->devid == devid)
267 goto out_unlock;
268 }
269
270 dev_data = NULL;
271
272 out_unlock:
273 spin_unlock_irqrestore(&dev_data_list_lock, flags);
274
275 return dev_data;
276 }
277
278 static int __last_alias(struct pci_dev *pdev, u16 alias, void *data)
279 {
280 *(u16 *)data = alias;
281 return 0;
282 }
283
284 static u16 get_alias(struct device *dev)
285 {
286 struct pci_dev *pdev = to_pci_dev(dev);
287 u16 devid, ivrs_alias, pci_alias;
288
289 /* The callers make sure that get_device_id() does not fail here */
290 devid = get_device_id(dev);
291 ivrs_alias = amd_iommu_alias_table[devid];
292 pci_for_each_dma_alias(pdev, __last_alias, &pci_alias);
293
294 if (ivrs_alias == pci_alias)
295 return ivrs_alias;
296
297 /*
298 * DMA alias showdown
299 *
300 * The IVRS is fairly reliable in telling us about aliases, but it
301 * can't know about every screwy device. If we don't have an IVRS
302 * reported alias, use the PCI reported alias. In that case we may
303 * still need to initialize the rlookup and dev_table entries if the
304 * alias is to a non-existent device.
305 */
306 if (ivrs_alias == devid) {
307 if (!amd_iommu_rlookup_table[pci_alias]) {
308 amd_iommu_rlookup_table[pci_alias] =
309 amd_iommu_rlookup_table[devid];
310 memcpy(amd_iommu_dev_table[pci_alias].data,
311 amd_iommu_dev_table[devid].data,
312 sizeof(amd_iommu_dev_table[pci_alias].data));
313 }
314
315 return pci_alias;
316 }
317
318 pr_info("AMD-Vi: Using IVRS reported alias %02x:%02x.%d "
319 "for device %s[%04x:%04x], kernel reported alias "
320 "%02x:%02x.%d\n", PCI_BUS_NUM(ivrs_alias), PCI_SLOT(ivrs_alias),
321 PCI_FUNC(ivrs_alias), dev_name(dev), pdev->vendor, pdev->device,
322 PCI_BUS_NUM(pci_alias), PCI_SLOT(pci_alias),
323 PCI_FUNC(pci_alias));
324
325 /*
326 * If we don't have a PCI DMA alias and the IVRS alias is on the same
327 * bus, then the IVRS table may know about a quirk that we don't.
328 */
329 if (pci_alias == devid &&
330 PCI_BUS_NUM(ivrs_alias) == pdev->bus->number) {
331 pci_add_dma_alias(pdev, ivrs_alias & 0xff);
332 pr_info("AMD-Vi: Added PCI DMA alias %02x.%d for %s\n",
333 PCI_SLOT(ivrs_alias), PCI_FUNC(ivrs_alias),
334 dev_name(dev));
335 }
336
337 return ivrs_alias;
338 }
339
340 static struct iommu_dev_data *find_dev_data(u16 devid)
341 {
342 struct iommu_dev_data *dev_data;
343
344 dev_data = search_dev_data(devid);
345
346 if (dev_data == NULL)
347 dev_data = alloc_dev_data(devid);
348
349 return dev_data;
350 }
351
352 static struct iommu_dev_data *get_dev_data(struct device *dev)
353 {
354 return dev->archdata.iommu;
355 }
356
357 /*
358 * Find or create an IOMMU group for a acpihid device.
359 */
360 static struct iommu_group *acpihid_device_group(struct device *dev)
361 {
362 struct acpihid_map_entry *p, *entry = NULL;
363 int devid;
364
365 devid = get_acpihid_device_id(dev, &entry);
366 if (devid < 0)
367 return ERR_PTR(devid);
368
369 list_for_each_entry(p, &acpihid_map, list) {
370 if ((devid == p->devid) && p->group)
371 entry->group = p->group;
372 }
373
374 if (!entry->group)
375 entry->group = generic_device_group(dev);
376 else
377 iommu_group_ref_get(entry->group);
378
379 return entry->group;
380 }
381
382 static bool pci_iommuv2_capable(struct pci_dev *pdev)
383 {
384 static const int caps[] = {
385 PCI_EXT_CAP_ID_ATS,
386 PCI_EXT_CAP_ID_PRI,
387 PCI_EXT_CAP_ID_PASID,
388 };
389 int i, pos;
390
391 for (i = 0; i < 3; ++i) {
392 pos = pci_find_ext_capability(pdev, caps[i]);
393 if (pos == 0)
394 return false;
395 }
396
397 return true;
398 }
399
400 static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
401 {
402 struct iommu_dev_data *dev_data;
403
404 dev_data = get_dev_data(&pdev->dev);
405
406 return dev_data->errata & (1 << erratum) ? true : false;
407 }
408
409 /*
410 * This function checks if the driver got a valid device from the caller to
411 * avoid dereferencing invalid pointers.
412 */
413 static bool check_device(struct device *dev)
414 {
415 int devid;
416
417 if (!dev || !dev->dma_mask)
418 return false;
419
420 devid = get_device_id(dev);
421 if (devid < 0)
422 return false;
423
424 /* Out of our scope? */
425 if (devid > amd_iommu_last_bdf)
426 return false;
427
428 if (amd_iommu_rlookup_table[devid] == NULL)
429 return false;
430
431 return true;
432 }
433
434 static void init_iommu_group(struct device *dev)
435 {
436 struct iommu_group *group;
437
438 group = iommu_group_get_for_dev(dev);
439 if (IS_ERR(group))
440 return;
441
442 iommu_group_put(group);
443 }
444
445 static int iommu_init_device(struct device *dev)
446 {
447 struct iommu_dev_data *dev_data;
448 int devid;
449
450 if (dev->archdata.iommu)
451 return 0;
452
453 devid = get_device_id(dev);
454 if (devid < 0)
455 return devid;
456
457 dev_data = find_dev_data(devid);
458 if (!dev_data)
459 return -ENOMEM;
460
461 dev_data->alias = get_alias(dev);
462
463 if (dev_is_pci(dev) && pci_iommuv2_capable(to_pci_dev(dev))) {
464 struct amd_iommu *iommu;
465
466 iommu = amd_iommu_rlookup_table[dev_data->devid];
467 dev_data->iommu_v2 = iommu->is_iommu_v2;
468 }
469
470 dev->archdata.iommu = dev_data;
471
472 iommu_device_link(amd_iommu_rlookup_table[dev_data->devid]->iommu_dev,
473 dev);
474
475 return 0;
476 }
477
478 static void iommu_ignore_device(struct device *dev)
479 {
480 u16 alias;
481 int devid;
482
483 devid = get_device_id(dev);
484 if (devid < 0)
485 return;
486
487 alias = get_alias(dev);
488
489 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
490 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
491
492 amd_iommu_rlookup_table[devid] = NULL;
493 amd_iommu_rlookup_table[alias] = NULL;
494 }
495
496 static void iommu_uninit_device(struct device *dev)
497 {
498 int devid;
499 struct iommu_dev_data *dev_data;
500
501 devid = get_device_id(dev);
502 if (devid < 0)
503 return;
504
505 dev_data = search_dev_data(devid);
506 if (!dev_data)
507 return;
508
509 if (dev_data->domain)
510 detach_device(dev);
511
512 iommu_device_unlink(amd_iommu_rlookup_table[dev_data->devid]->iommu_dev,
513 dev);
514
515 iommu_group_remove_device(dev);
516
517 /* Remove dma-ops */
518 dev->archdata.dma_ops = NULL;
519
520 /*
521 * We keep dev_data around for unplugged devices and reuse it when the
522 * device is re-plugged - not doing so would introduce a ton of races.
523 */
524 }
525
526 /****************************************************************************
527 *
528 * Interrupt handling functions
529 *
530 ****************************************************************************/
531
532 static void dump_dte_entry(u16 devid)
533 {
534 int i;
535
536 for (i = 0; i < 4; ++i)
537 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
538 amd_iommu_dev_table[devid].data[i]);
539 }
540
541 static void dump_command(unsigned long phys_addr)
542 {
543 struct iommu_cmd *cmd = phys_to_virt(phys_addr);
544 int i;
545
546 for (i = 0; i < 4; ++i)
547 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
548 }
549
550 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
551 {
552 int type, devid, domid, flags;
553 volatile u32 *event = __evt;
554 int count = 0;
555 u64 address;
556
557 retry:
558 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
559 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
560 domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
561 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
562 address = (u64)(((u64)event[3]) << 32) | event[2];
563
564 if (type == 0) {
565 /* Did we hit the erratum? */
566 if (++count == LOOP_TIMEOUT) {
567 pr_err("AMD-Vi: No event written to event log\n");
568 return;
569 }
570 udelay(1);
571 goto retry;
572 }
573
574 printk(KERN_ERR "AMD-Vi: Event logged [");
575
576 switch (type) {
577 case EVENT_TYPE_ILL_DEV:
578 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
579 "address=0x%016llx flags=0x%04x]\n",
580 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
581 address, flags);
582 dump_dte_entry(devid);
583 break;
584 case EVENT_TYPE_IO_FAULT:
585 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
586 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
587 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
588 domid, address, flags);
589 break;
590 case EVENT_TYPE_DEV_TAB_ERR:
591 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
592 "address=0x%016llx flags=0x%04x]\n",
593 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
594 address, flags);
595 break;
596 case EVENT_TYPE_PAGE_TAB_ERR:
597 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
598 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
599 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
600 domid, address, flags);
601 break;
602 case EVENT_TYPE_ILL_CMD:
603 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
604 dump_command(address);
605 break;
606 case EVENT_TYPE_CMD_HARD_ERR:
607 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
608 "flags=0x%04x]\n", address, flags);
609 break;
610 case EVENT_TYPE_IOTLB_INV_TO:
611 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
612 "address=0x%016llx]\n",
613 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
614 address);
615 break;
616 case EVENT_TYPE_INV_DEV_REQ:
617 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
618 "address=0x%016llx flags=0x%04x]\n",
619 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
620 address, flags);
621 break;
622 default:
623 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
624 }
625
626 memset(__evt, 0, 4 * sizeof(u32));
627 }
628
629 static void iommu_poll_events(struct amd_iommu *iommu)
630 {
631 u32 head, tail;
632
633 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
634 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
635
636 while (head != tail) {
637 iommu_print_event(iommu, iommu->evt_buf + head);
638 head = (head + EVENT_ENTRY_SIZE) % EVT_BUFFER_SIZE;
639 }
640
641 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
642 }
643
644 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
645 {
646 struct amd_iommu_fault fault;
647
648 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
649 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
650 return;
651 }
652
653 fault.address = raw[1];
654 fault.pasid = PPR_PASID(raw[0]);
655 fault.device_id = PPR_DEVID(raw[0]);
656 fault.tag = PPR_TAG(raw[0]);
657 fault.flags = PPR_FLAGS(raw[0]);
658
659 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
660 }
661
662 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
663 {
664 u32 head, tail;
665
666 if (iommu->ppr_log == NULL)
667 return;
668
669 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
670 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
671
672 while (head != tail) {
673 volatile u64 *raw;
674 u64 entry[2];
675 int i;
676
677 raw = (u64 *)(iommu->ppr_log + head);
678
679 /*
680 * Hardware bug: Interrupt may arrive before the entry is
681 * written to memory. If this happens we need to wait for the
682 * entry to arrive.
683 */
684 for (i = 0; i < LOOP_TIMEOUT; ++i) {
685 if (PPR_REQ_TYPE(raw[0]) != 0)
686 break;
687 udelay(1);
688 }
689
690 /* Avoid memcpy function-call overhead */
691 entry[0] = raw[0];
692 entry[1] = raw[1];
693
694 /*
695 * To detect the hardware bug we need to clear the entry
696 * back to zero.
697 */
698 raw[0] = raw[1] = 0UL;
699
700 /* Update head pointer of hardware ring-buffer */
701 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
702 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
703
704 /* Handle PPR entry */
705 iommu_handle_ppr_entry(iommu, entry);
706
707 /* Refresh ring-buffer information */
708 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
709 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
710 }
711 }
712
713 #ifdef CONFIG_IRQ_REMAP
714 static int (*iommu_ga_log_notifier)(u32);
715
716 int amd_iommu_register_ga_log_notifier(int (*notifier)(u32))
717 {
718 iommu_ga_log_notifier = notifier;
719
720 return 0;
721 }
722 EXPORT_SYMBOL(amd_iommu_register_ga_log_notifier);
723
724 static void iommu_poll_ga_log(struct amd_iommu *iommu)
725 {
726 u32 head, tail, cnt = 0;
727
728 if (iommu->ga_log == NULL)
729 return;
730
731 head = readl(iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
732 tail = readl(iommu->mmio_base + MMIO_GA_TAIL_OFFSET);
733
734 while (head != tail) {
735 volatile u64 *raw;
736 u64 log_entry;
737
738 raw = (u64 *)(iommu->ga_log + head);
739 cnt++;
740
741 /* Avoid memcpy function-call overhead */
742 log_entry = *raw;
743
744 /* Update head pointer of hardware ring-buffer */
745 head = (head + GA_ENTRY_SIZE) % GA_LOG_SIZE;
746 writel(head, iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
747
748 /* Handle GA entry */
749 switch (GA_REQ_TYPE(log_entry)) {
750 case GA_GUEST_NR:
751 if (!iommu_ga_log_notifier)
752 break;
753
754 pr_debug("AMD-Vi: %s: devid=%#x, ga_tag=%#x\n",
755 __func__, GA_DEVID(log_entry),
756 GA_TAG(log_entry));
757
758 if (iommu_ga_log_notifier(GA_TAG(log_entry)) != 0)
759 pr_err("AMD-Vi: GA log notifier failed.\n");
760 break;
761 default:
762 break;
763 }
764 }
765 }
766 #endif /* CONFIG_IRQ_REMAP */
767
768 #define AMD_IOMMU_INT_MASK \
769 (MMIO_STATUS_EVT_INT_MASK | \
770 MMIO_STATUS_PPR_INT_MASK | \
771 MMIO_STATUS_GALOG_INT_MASK)
772
773 irqreturn_t amd_iommu_int_thread(int irq, void *data)
774 {
775 struct amd_iommu *iommu = (struct amd_iommu *) data;
776 u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
777
778 while (status & AMD_IOMMU_INT_MASK) {
779 /* Enable EVT and PPR and GA interrupts again */
780 writel(AMD_IOMMU_INT_MASK,
781 iommu->mmio_base + MMIO_STATUS_OFFSET);
782
783 if (status & MMIO_STATUS_EVT_INT_MASK) {
784 pr_devel("AMD-Vi: Processing IOMMU Event Log\n");
785 iommu_poll_events(iommu);
786 }
787
788 if (status & MMIO_STATUS_PPR_INT_MASK) {
789 pr_devel("AMD-Vi: Processing IOMMU PPR Log\n");
790 iommu_poll_ppr_log(iommu);
791 }
792
793 #ifdef CONFIG_IRQ_REMAP
794 if (status & MMIO_STATUS_GALOG_INT_MASK) {
795 pr_devel("AMD-Vi: Processing IOMMU GA Log\n");
796 iommu_poll_ga_log(iommu);
797 }
798 #endif
799
800 /*
801 * Hardware bug: ERBT1312
802 * When re-enabling interrupt (by writing 1
803 * to clear the bit), the hardware might also try to set
804 * the interrupt bit in the event status register.
805 * In this scenario, the bit will be set, and disable
806 * subsequent interrupts.
807 *
808 * Workaround: The IOMMU driver should read back the
809 * status register and check if the interrupt bits are cleared.
810 * If not, driver will need to go through the interrupt handler
811 * again and re-clear the bits
812 */
813 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
814 }
815 return IRQ_HANDLED;
816 }
817
818 irqreturn_t amd_iommu_int_handler(int irq, void *data)
819 {
820 return IRQ_WAKE_THREAD;
821 }
822
823 /****************************************************************************
824 *
825 * IOMMU command queuing functions
826 *
827 ****************************************************************************/
828
829 static int wait_on_sem(volatile u64 *sem)
830 {
831 int i = 0;
832
833 while (*sem == 0 && i < LOOP_TIMEOUT) {
834 udelay(1);
835 i += 1;
836 }
837
838 if (i == LOOP_TIMEOUT) {
839 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
840 return -EIO;
841 }
842
843 return 0;
844 }
845
846 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
847 struct iommu_cmd *cmd,
848 u32 tail)
849 {
850 u8 *target;
851
852 target = iommu->cmd_buf + tail;
853 tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
854
855 /* Copy command to buffer */
856 memcpy(target, cmd, sizeof(*cmd));
857
858 /* Tell the IOMMU about it */
859 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
860 }
861
862 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
863 {
864 WARN_ON(address & 0x7ULL);
865
866 memset(cmd, 0, sizeof(*cmd));
867 cmd->data[0] = lower_32_bits(__pa(address)) | CMD_COMPL_WAIT_STORE_MASK;
868 cmd->data[1] = upper_32_bits(__pa(address));
869 cmd->data[2] = 1;
870 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
871 }
872
873 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
874 {
875 memset(cmd, 0, sizeof(*cmd));
876 cmd->data[0] = devid;
877 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
878 }
879
880 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
881 size_t size, u16 domid, int pde)
882 {
883 u64 pages;
884 bool s;
885
886 pages = iommu_num_pages(address, size, PAGE_SIZE);
887 s = false;
888
889 if (pages > 1) {
890 /*
891 * If we have to flush more than one page, flush all
892 * TLB entries for this domain
893 */
894 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
895 s = true;
896 }
897
898 address &= PAGE_MASK;
899
900 memset(cmd, 0, sizeof(*cmd));
901 cmd->data[1] |= domid;
902 cmd->data[2] = lower_32_bits(address);
903 cmd->data[3] = upper_32_bits(address);
904 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
905 if (s) /* size bit - we flush more than one 4kb page */
906 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
907 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
908 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
909 }
910
911 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
912 u64 address, size_t size)
913 {
914 u64 pages;
915 bool s;
916
917 pages = iommu_num_pages(address, size, PAGE_SIZE);
918 s = false;
919
920 if (pages > 1) {
921 /*
922 * If we have to flush more than one page, flush all
923 * TLB entries for this domain
924 */
925 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
926 s = true;
927 }
928
929 address &= PAGE_MASK;
930
931 memset(cmd, 0, sizeof(*cmd));
932 cmd->data[0] = devid;
933 cmd->data[0] |= (qdep & 0xff) << 24;
934 cmd->data[1] = devid;
935 cmd->data[2] = lower_32_bits(address);
936 cmd->data[3] = upper_32_bits(address);
937 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
938 if (s)
939 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
940 }
941
942 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
943 u64 address, bool size)
944 {
945 memset(cmd, 0, sizeof(*cmd));
946
947 address &= ~(0xfffULL);
948
949 cmd->data[0] = pasid;
950 cmd->data[1] = domid;
951 cmd->data[2] = lower_32_bits(address);
952 cmd->data[3] = upper_32_bits(address);
953 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
954 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
955 if (size)
956 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
957 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
958 }
959
960 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
961 int qdep, u64 address, bool size)
962 {
963 memset(cmd, 0, sizeof(*cmd));
964
965 address &= ~(0xfffULL);
966
967 cmd->data[0] = devid;
968 cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
969 cmd->data[0] |= (qdep & 0xff) << 24;
970 cmd->data[1] = devid;
971 cmd->data[1] |= (pasid & 0xff) << 16;
972 cmd->data[2] = lower_32_bits(address);
973 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
974 cmd->data[3] = upper_32_bits(address);
975 if (size)
976 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
977 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
978 }
979
980 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
981 int status, int tag, bool gn)
982 {
983 memset(cmd, 0, sizeof(*cmd));
984
985 cmd->data[0] = devid;
986 if (gn) {
987 cmd->data[1] = pasid;
988 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
989 }
990 cmd->data[3] = tag & 0x1ff;
991 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
992
993 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
994 }
995
996 static void build_inv_all(struct iommu_cmd *cmd)
997 {
998 memset(cmd, 0, sizeof(*cmd));
999 CMD_SET_TYPE(cmd, CMD_INV_ALL);
1000 }
1001
1002 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
1003 {
1004 memset(cmd, 0, sizeof(*cmd));
1005 cmd->data[0] = devid;
1006 CMD_SET_TYPE(cmd, CMD_INV_IRT);
1007 }
1008
1009 /*
1010 * Writes the command to the IOMMUs command buffer and informs the
1011 * hardware about the new command.
1012 */
1013 static int __iommu_queue_command_sync(struct amd_iommu *iommu,
1014 struct iommu_cmd *cmd,
1015 bool sync)
1016 {
1017 u32 left, tail, head, next_tail;
1018
1019 again:
1020
1021 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
1022 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
1023 next_tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
1024 left = (head - next_tail) % CMD_BUFFER_SIZE;
1025
1026 if (left <= 0x20) {
1027 struct iommu_cmd sync_cmd;
1028 int ret;
1029
1030 iommu->cmd_sem = 0;
1031
1032 build_completion_wait(&sync_cmd, (u64)&iommu->cmd_sem);
1033 copy_cmd_to_buffer(iommu, &sync_cmd, tail);
1034
1035 if ((ret = wait_on_sem(&iommu->cmd_sem)) != 0)
1036 return ret;
1037
1038 goto again;
1039 }
1040
1041 copy_cmd_to_buffer(iommu, cmd, tail);
1042
1043 /* We need to sync now to make sure all commands are processed */
1044 iommu->need_sync = sync;
1045
1046 return 0;
1047 }
1048
1049 static int iommu_queue_command_sync(struct amd_iommu *iommu,
1050 struct iommu_cmd *cmd,
1051 bool sync)
1052 {
1053 unsigned long flags;
1054 int ret;
1055
1056 spin_lock_irqsave(&iommu->lock, flags);
1057 ret = __iommu_queue_command_sync(iommu, cmd, sync);
1058 spin_unlock_irqrestore(&iommu->lock, flags);
1059
1060 return ret;
1061 }
1062
1063 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
1064 {
1065 return iommu_queue_command_sync(iommu, cmd, true);
1066 }
1067
1068 /*
1069 * This function queues a completion wait command into the command
1070 * buffer of an IOMMU
1071 */
1072 static int iommu_completion_wait(struct amd_iommu *iommu)
1073 {
1074 struct iommu_cmd cmd;
1075 unsigned long flags;
1076 int ret;
1077
1078 if (!iommu->need_sync)
1079 return 0;
1080
1081
1082 build_completion_wait(&cmd, (u64)&iommu->cmd_sem);
1083
1084 spin_lock_irqsave(&iommu->lock, flags);
1085
1086 iommu->cmd_sem = 0;
1087
1088 ret = __iommu_queue_command_sync(iommu, &cmd, false);
1089 if (ret)
1090 goto out_unlock;
1091
1092 ret = wait_on_sem(&iommu->cmd_sem);
1093
1094 out_unlock:
1095 spin_unlock_irqrestore(&iommu->lock, flags);
1096
1097 return ret;
1098 }
1099
1100 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1101 {
1102 struct iommu_cmd cmd;
1103
1104 build_inv_dte(&cmd, devid);
1105
1106 return iommu_queue_command(iommu, &cmd);
1107 }
1108
1109 static void iommu_flush_dte_all(struct amd_iommu *iommu)
1110 {
1111 u32 devid;
1112
1113 for (devid = 0; devid <= 0xffff; ++devid)
1114 iommu_flush_dte(iommu, devid);
1115
1116 iommu_completion_wait(iommu);
1117 }
1118
1119 /*
1120 * This function uses heavy locking and may disable irqs for some time. But
1121 * this is no issue because it is only called during resume.
1122 */
1123 static void iommu_flush_tlb_all(struct amd_iommu *iommu)
1124 {
1125 u32 dom_id;
1126
1127 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
1128 struct iommu_cmd cmd;
1129 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1130 dom_id, 1);
1131 iommu_queue_command(iommu, &cmd);
1132 }
1133
1134 iommu_completion_wait(iommu);
1135 }
1136
1137 static void iommu_flush_all(struct amd_iommu *iommu)
1138 {
1139 struct iommu_cmd cmd;
1140
1141 build_inv_all(&cmd);
1142
1143 iommu_queue_command(iommu, &cmd);
1144 iommu_completion_wait(iommu);
1145 }
1146
1147 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1148 {
1149 struct iommu_cmd cmd;
1150
1151 build_inv_irt(&cmd, devid);
1152
1153 iommu_queue_command(iommu, &cmd);
1154 }
1155
1156 static void iommu_flush_irt_all(struct amd_iommu *iommu)
1157 {
1158 u32 devid;
1159
1160 for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
1161 iommu_flush_irt(iommu, devid);
1162
1163 iommu_completion_wait(iommu);
1164 }
1165
1166 void iommu_flush_all_caches(struct amd_iommu *iommu)
1167 {
1168 if (iommu_feature(iommu, FEATURE_IA)) {
1169 iommu_flush_all(iommu);
1170 } else {
1171 iommu_flush_dte_all(iommu);
1172 iommu_flush_irt_all(iommu);
1173 iommu_flush_tlb_all(iommu);
1174 }
1175 }
1176
1177 /*
1178 * Command send function for flushing on-device TLB
1179 */
1180 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1181 u64 address, size_t size)
1182 {
1183 struct amd_iommu *iommu;
1184 struct iommu_cmd cmd;
1185 int qdep;
1186
1187 qdep = dev_data->ats.qdep;
1188 iommu = amd_iommu_rlookup_table[dev_data->devid];
1189
1190 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1191
1192 return iommu_queue_command(iommu, &cmd);
1193 }
1194
1195 /*
1196 * Command send function for invalidating a device table entry
1197 */
1198 static int device_flush_dte(struct iommu_dev_data *dev_data)
1199 {
1200 struct amd_iommu *iommu;
1201 u16 alias;
1202 int ret;
1203
1204 iommu = amd_iommu_rlookup_table[dev_data->devid];
1205 alias = dev_data->alias;
1206
1207 ret = iommu_flush_dte(iommu, dev_data->devid);
1208 if (!ret && alias != dev_data->devid)
1209 ret = iommu_flush_dte(iommu, alias);
1210 if (ret)
1211 return ret;
1212
1213 if (dev_data->ats.enabled)
1214 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1215
1216 return ret;
1217 }
1218
1219 /*
1220 * TLB invalidation function which is called from the mapping functions.
1221 * It invalidates a single PTE if the range to flush is within a single
1222 * page. Otherwise it flushes the whole TLB of the IOMMU.
1223 */
1224 static void __domain_flush_pages(struct protection_domain *domain,
1225 u64 address, size_t size, int pde)
1226 {
1227 struct iommu_dev_data *dev_data;
1228 struct iommu_cmd cmd;
1229 int ret = 0, i;
1230
1231 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1232
1233 for (i = 0; i < amd_iommus_present; ++i) {
1234 if (!domain->dev_iommu[i])
1235 continue;
1236
1237 /*
1238 * Devices of this domain are behind this IOMMU
1239 * We need a TLB flush
1240 */
1241 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1242 }
1243
1244 list_for_each_entry(dev_data, &domain->dev_list, list) {
1245
1246 if (!dev_data->ats.enabled)
1247 continue;
1248
1249 ret |= device_flush_iotlb(dev_data, address, size);
1250 }
1251
1252 WARN_ON(ret);
1253 }
1254
1255 static void domain_flush_pages(struct protection_domain *domain,
1256 u64 address, size_t size)
1257 {
1258 __domain_flush_pages(domain, address, size, 0);
1259 }
1260
1261 /* Flush the whole IO/TLB for a given protection domain */
1262 static void domain_flush_tlb(struct protection_domain *domain)
1263 {
1264 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1265 }
1266
1267 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1268 static void domain_flush_tlb_pde(struct protection_domain *domain)
1269 {
1270 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1271 }
1272
1273 static void domain_flush_complete(struct protection_domain *domain)
1274 {
1275 int i;
1276
1277 for (i = 0; i < amd_iommus_present; ++i) {
1278 if (domain && !domain->dev_iommu[i])
1279 continue;
1280
1281 /*
1282 * Devices of this domain are behind this IOMMU
1283 * We need to wait for completion of all commands.
1284 */
1285 iommu_completion_wait(amd_iommus[i]);
1286 }
1287 }
1288
1289
1290 /*
1291 * This function flushes the DTEs for all devices in domain
1292 */
1293 static void domain_flush_devices(struct protection_domain *domain)
1294 {
1295 struct iommu_dev_data *dev_data;
1296
1297 list_for_each_entry(dev_data, &domain->dev_list, list)
1298 device_flush_dte(dev_data);
1299 }
1300
1301 /****************************************************************************
1302 *
1303 * The functions below are used the create the page table mappings for
1304 * unity mapped regions.
1305 *
1306 ****************************************************************************/
1307
1308 /*
1309 * This function is used to add another level to an IO page table. Adding
1310 * another level increases the size of the address space by 9 bits to a size up
1311 * to 64 bits.
1312 */
1313 static bool increase_address_space(struct protection_domain *domain,
1314 gfp_t gfp)
1315 {
1316 u64 *pte;
1317
1318 if (domain->mode == PAGE_MODE_6_LEVEL)
1319 /* address space already 64 bit large */
1320 return false;
1321
1322 pte = (void *)get_zeroed_page(gfp);
1323 if (!pte)
1324 return false;
1325
1326 *pte = PM_LEVEL_PDE(domain->mode,
1327 virt_to_phys(domain->pt_root));
1328 domain->pt_root = pte;
1329 domain->mode += 1;
1330 domain->updated = true;
1331
1332 return true;
1333 }
1334
1335 static u64 *alloc_pte(struct protection_domain *domain,
1336 unsigned long address,
1337 unsigned long page_size,
1338 u64 **pte_page,
1339 gfp_t gfp)
1340 {
1341 int level, end_lvl;
1342 u64 *pte, *page;
1343
1344 BUG_ON(!is_power_of_2(page_size));
1345
1346 while (address > PM_LEVEL_SIZE(domain->mode))
1347 increase_address_space(domain, gfp);
1348
1349 level = domain->mode - 1;
1350 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1351 address = PAGE_SIZE_ALIGN(address, page_size);
1352 end_lvl = PAGE_SIZE_LEVEL(page_size);
1353
1354 while (level > end_lvl) {
1355 u64 __pte, __npte;
1356
1357 __pte = *pte;
1358
1359 if (!IOMMU_PTE_PRESENT(__pte)) {
1360 page = (u64 *)get_zeroed_page(gfp);
1361 if (!page)
1362 return NULL;
1363
1364 __npte = PM_LEVEL_PDE(level, virt_to_phys(page));
1365
1366 /* pte could have been changed somewhere. */
1367 if (cmpxchg64(pte, __pte, __npte) != __pte) {
1368 free_page((unsigned long)page);
1369 continue;
1370 }
1371 }
1372
1373 /* No level skipping support yet */
1374 if (PM_PTE_LEVEL(*pte) != level)
1375 return NULL;
1376
1377 level -= 1;
1378
1379 pte = IOMMU_PTE_PAGE(*pte);
1380
1381 if (pte_page && level == end_lvl)
1382 *pte_page = pte;
1383
1384 pte = &pte[PM_LEVEL_INDEX(level, address)];
1385 }
1386
1387 return pte;
1388 }
1389
1390 /*
1391 * This function checks if there is a PTE for a given dma address. If
1392 * there is one, it returns the pointer to it.
1393 */
1394 static u64 *fetch_pte(struct protection_domain *domain,
1395 unsigned long address,
1396 unsigned long *page_size)
1397 {
1398 int level;
1399 u64 *pte;
1400
1401 if (address > PM_LEVEL_SIZE(domain->mode))
1402 return NULL;
1403
1404 level = domain->mode - 1;
1405 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1406 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1407
1408 while (level > 0) {
1409
1410 /* Not Present */
1411 if (!IOMMU_PTE_PRESENT(*pte))
1412 return NULL;
1413
1414 /* Large PTE */
1415 if (PM_PTE_LEVEL(*pte) == 7 ||
1416 PM_PTE_LEVEL(*pte) == 0)
1417 break;
1418
1419 /* No level skipping support yet */
1420 if (PM_PTE_LEVEL(*pte) != level)
1421 return NULL;
1422
1423 level -= 1;
1424
1425 /* Walk to the next level */
1426 pte = IOMMU_PTE_PAGE(*pte);
1427 pte = &pte[PM_LEVEL_INDEX(level, address)];
1428 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1429 }
1430
1431 if (PM_PTE_LEVEL(*pte) == 0x07) {
1432 unsigned long pte_mask;
1433
1434 /*
1435 * If we have a series of large PTEs, make
1436 * sure to return a pointer to the first one.
1437 */
1438 *page_size = pte_mask = PTE_PAGE_SIZE(*pte);
1439 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1440 pte = (u64 *)(((unsigned long)pte) & pte_mask);
1441 }
1442
1443 return pte;
1444 }
1445
1446 /*
1447 * Generic mapping functions. It maps a physical address into a DMA
1448 * address space. It allocates the page table pages if necessary.
1449 * In the future it can be extended to a generic mapping function
1450 * supporting all features of AMD IOMMU page tables like level skipping
1451 * and full 64 bit address spaces.
1452 */
1453 static int iommu_map_page(struct protection_domain *dom,
1454 unsigned long bus_addr,
1455 unsigned long phys_addr,
1456 unsigned long page_size,
1457 int prot,
1458 gfp_t gfp)
1459 {
1460 u64 __pte, *pte;
1461 int i, count;
1462
1463 BUG_ON(!IS_ALIGNED(bus_addr, page_size));
1464 BUG_ON(!IS_ALIGNED(phys_addr, page_size));
1465
1466 if (!(prot & IOMMU_PROT_MASK))
1467 return -EINVAL;
1468
1469 count = PAGE_SIZE_PTE_COUNT(page_size);
1470 pte = alloc_pte(dom, bus_addr, page_size, NULL, gfp);
1471
1472 if (!pte)
1473 return -ENOMEM;
1474
1475 for (i = 0; i < count; ++i)
1476 if (IOMMU_PTE_PRESENT(pte[i]))
1477 return -EBUSY;
1478
1479 if (count > 1) {
1480 __pte = PAGE_SIZE_PTE(phys_addr, page_size);
1481 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_P | IOMMU_PTE_FC;
1482 } else
1483 __pte = phys_addr | IOMMU_PTE_P | IOMMU_PTE_FC;
1484
1485 if (prot & IOMMU_PROT_IR)
1486 __pte |= IOMMU_PTE_IR;
1487 if (prot & IOMMU_PROT_IW)
1488 __pte |= IOMMU_PTE_IW;
1489
1490 for (i = 0; i < count; ++i)
1491 pte[i] = __pte;
1492
1493 update_domain(dom);
1494
1495 return 0;
1496 }
1497
1498 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1499 unsigned long bus_addr,
1500 unsigned long page_size)
1501 {
1502 unsigned long long unmapped;
1503 unsigned long unmap_size;
1504 u64 *pte;
1505
1506 BUG_ON(!is_power_of_2(page_size));
1507
1508 unmapped = 0;
1509
1510 while (unmapped < page_size) {
1511
1512 pte = fetch_pte(dom, bus_addr, &unmap_size);
1513
1514 if (pte) {
1515 int i, count;
1516
1517 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1518 for (i = 0; i < count; i++)
1519 pte[i] = 0ULL;
1520 }
1521
1522 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1523 unmapped += unmap_size;
1524 }
1525
1526 BUG_ON(unmapped && !is_power_of_2(unmapped));
1527
1528 return unmapped;
1529 }
1530
1531 /****************************************************************************
1532 *
1533 * The next functions belong to the address allocator for the dma_ops
1534 * interface functions.
1535 *
1536 ****************************************************************************/
1537
1538
1539 static unsigned long dma_ops_alloc_iova(struct device *dev,
1540 struct dma_ops_domain *dma_dom,
1541 unsigned int pages, u64 dma_mask)
1542 {
1543 unsigned long pfn = 0;
1544
1545 pages = __roundup_pow_of_two(pages);
1546
1547 if (dma_mask > DMA_BIT_MASK(32))
1548 pfn = alloc_iova_fast(&dma_dom->iovad, pages,
1549 IOVA_PFN(DMA_BIT_MASK(32)));
1550
1551 if (!pfn)
1552 pfn = alloc_iova_fast(&dma_dom->iovad, pages, IOVA_PFN(dma_mask));
1553
1554 return (pfn << PAGE_SHIFT);
1555 }
1556
1557 static void dma_ops_free_iova(struct dma_ops_domain *dma_dom,
1558 unsigned long address,
1559 unsigned int pages)
1560 {
1561 pages = __roundup_pow_of_two(pages);
1562 address >>= PAGE_SHIFT;
1563
1564 free_iova_fast(&dma_dom->iovad, address, pages);
1565 }
1566
1567 /****************************************************************************
1568 *
1569 * The next functions belong to the domain allocation. A domain is
1570 * allocated for every IOMMU as the default domain. If device isolation
1571 * is enabled, every device get its own domain. The most important thing
1572 * about domains is the page table mapping the DMA address space they
1573 * contain.
1574 *
1575 ****************************************************************************/
1576
1577 /*
1578 * This function adds a protection domain to the global protection domain list
1579 */
1580 static void add_domain_to_list(struct protection_domain *domain)
1581 {
1582 unsigned long flags;
1583
1584 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1585 list_add(&domain->list, &amd_iommu_pd_list);
1586 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1587 }
1588
1589 /*
1590 * This function removes a protection domain to the global
1591 * protection domain list
1592 */
1593 static void del_domain_from_list(struct protection_domain *domain)
1594 {
1595 unsigned long flags;
1596
1597 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1598 list_del(&domain->list);
1599 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1600 }
1601
1602 static u16 domain_id_alloc(void)
1603 {
1604 unsigned long flags;
1605 int id;
1606
1607 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1608 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1609 BUG_ON(id == 0);
1610 if (id > 0 && id < MAX_DOMAIN_ID)
1611 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1612 else
1613 id = 0;
1614 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1615
1616 return id;
1617 }
1618
1619 static void domain_id_free(int id)
1620 {
1621 unsigned long flags;
1622
1623 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1624 if (id > 0 && id < MAX_DOMAIN_ID)
1625 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1626 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1627 }
1628
1629 #define DEFINE_FREE_PT_FN(LVL, FN) \
1630 static void free_pt_##LVL (unsigned long __pt) \
1631 { \
1632 unsigned long p; \
1633 u64 *pt; \
1634 int i; \
1635 \
1636 pt = (u64 *)__pt; \
1637 \
1638 for (i = 0; i < 512; ++i) { \
1639 /* PTE present? */ \
1640 if (!IOMMU_PTE_PRESENT(pt[i])) \
1641 continue; \
1642 \
1643 /* Large PTE? */ \
1644 if (PM_PTE_LEVEL(pt[i]) == 0 || \
1645 PM_PTE_LEVEL(pt[i]) == 7) \
1646 continue; \
1647 \
1648 p = (unsigned long)IOMMU_PTE_PAGE(pt[i]); \
1649 FN(p); \
1650 } \
1651 free_page((unsigned long)pt); \
1652 }
1653
1654 DEFINE_FREE_PT_FN(l2, free_page)
1655 DEFINE_FREE_PT_FN(l3, free_pt_l2)
1656 DEFINE_FREE_PT_FN(l4, free_pt_l3)
1657 DEFINE_FREE_PT_FN(l5, free_pt_l4)
1658 DEFINE_FREE_PT_FN(l6, free_pt_l5)
1659
1660 static void free_pagetable(struct protection_domain *domain)
1661 {
1662 unsigned long root = (unsigned long)domain->pt_root;
1663
1664 switch (domain->mode) {
1665 case PAGE_MODE_NONE:
1666 break;
1667 case PAGE_MODE_1_LEVEL:
1668 free_page(root);
1669 break;
1670 case PAGE_MODE_2_LEVEL:
1671 free_pt_l2(root);
1672 break;
1673 case PAGE_MODE_3_LEVEL:
1674 free_pt_l3(root);
1675 break;
1676 case PAGE_MODE_4_LEVEL:
1677 free_pt_l4(root);
1678 break;
1679 case PAGE_MODE_5_LEVEL:
1680 free_pt_l5(root);
1681 break;
1682 case PAGE_MODE_6_LEVEL:
1683 free_pt_l6(root);
1684 break;
1685 default:
1686 BUG();
1687 }
1688 }
1689
1690 static void free_gcr3_tbl_level1(u64 *tbl)
1691 {
1692 u64 *ptr;
1693 int i;
1694
1695 for (i = 0; i < 512; ++i) {
1696 if (!(tbl[i] & GCR3_VALID))
1697 continue;
1698
1699 ptr = __va(tbl[i] & PAGE_MASK);
1700
1701 free_page((unsigned long)ptr);
1702 }
1703 }
1704
1705 static void free_gcr3_tbl_level2(u64 *tbl)
1706 {
1707 u64 *ptr;
1708 int i;
1709
1710 for (i = 0; i < 512; ++i) {
1711 if (!(tbl[i] & GCR3_VALID))
1712 continue;
1713
1714 ptr = __va(tbl[i] & PAGE_MASK);
1715
1716 free_gcr3_tbl_level1(ptr);
1717 }
1718 }
1719
1720 static void free_gcr3_table(struct protection_domain *domain)
1721 {
1722 if (domain->glx == 2)
1723 free_gcr3_tbl_level2(domain->gcr3_tbl);
1724 else if (domain->glx == 1)
1725 free_gcr3_tbl_level1(domain->gcr3_tbl);
1726 else
1727 BUG_ON(domain->glx != 0);
1728
1729 free_page((unsigned long)domain->gcr3_tbl);
1730 }
1731
1732 /*
1733 * Free a domain, only used if something went wrong in the
1734 * allocation path and we need to free an already allocated page table
1735 */
1736 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1737 {
1738 if (!dom)
1739 return;
1740
1741 del_domain_from_list(&dom->domain);
1742
1743 put_iova_domain(&dom->iovad);
1744
1745 free_pagetable(&dom->domain);
1746
1747 if (dom->domain.id)
1748 domain_id_free(dom->domain.id);
1749
1750 kfree(dom);
1751 }
1752
1753 /*
1754 * Allocates a new protection domain usable for the dma_ops functions.
1755 * It also initializes the page table and the address allocator data
1756 * structures required for the dma_ops interface
1757 */
1758 static struct dma_ops_domain *dma_ops_domain_alloc(void)
1759 {
1760 struct dma_ops_domain *dma_dom;
1761
1762 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1763 if (!dma_dom)
1764 return NULL;
1765
1766 if (protection_domain_init(&dma_dom->domain))
1767 goto free_dma_dom;
1768
1769 dma_dom->domain.mode = PAGE_MODE_3_LEVEL;
1770 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1771 dma_dom->domain.flags = PD_DMA_OPS_MASK;
1772 if (!dma_dom->domain.pt_root)
1773 goto free_dma_dom;
1774
1775 init_iova_domain(&dma_dom->iovad, PAGE_SIZE,
1776 IOVA_START_PFN, DMA_32BIT_PFN);
1777
1778 /* Initialize reserved ranges */
1779 copy_reserved_iova(&reserved_iova_ranges, &dma_dom->iovad);
1780
1781 add_domain_to_list(&dma_dom->domain);
1782
1783 return dma_dom;
1784
1785 free_dma_dom:
1786 dma_ops_domain_free(dma_dom);
1787
1788 return NULL;
1789 }
1790
1791 /*
1792 * little helper function to check whether a given protection domain is a
1793 * dma_ops domain
1794 */
1795 static bool dma_ops_domain(struct protection_domain *domain)
1796 {
1797 return domain->flags & PD_DMA_OPS_MASK;
1798 }
1799
1800 static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
1801 {
1802 u64 pte_root = 0;
1803 u64 flags = 0;
1804
1805 if (domain->mode != PAGE_MODE_NONE)
1806 pte_root = virt_to_phys(domain->pt_root);
1807
1808 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
1809 << DEV_ENTRY_MODE_SHIFT;
1810 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
1811
1812 flags = amd_iommu_dev_table[devid].data[1];
1813
1814 if (ats)
1815 flags |= DTE_FLAG_IOTLB;
1816
1817 if (domain->flags & PD_IOMMUV2_MASK) {
1818 u64 gcr3 = __pa(domain->gcr3_tbl);
1819 u64 glx = domain->glx;
1820 u64 tmp;
1821
1822 pte_root |= DTE_FLAG_GV;
1823 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1824
1825 /* First mask out possible old values for GCR3 table */
1826 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1827 flags &= ~tmp;
1828
1829 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1830 flags &= ~tmp;
1831
1832 /* Encode GCR3 table into DTE */
1833 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1834 pte_root |= tmp;
1835
1836 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1837 flags |= tmp;
1838
1839 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1840 flags |= tmp;
1841 }
1842
1843 flags &= ~(0xffffUL);
1844 flags |= domain->id;
1845
1846 amd_iommu_dev_table[devid].data[1] = flags;
1847 amd_iommu_dev_table[devid].data[0] = pte_root;
1848 }
1849
1850 static void clear_dte_entry(u16 devid)
1851 {
1852 /* remove entry from the device table seen by the hardware */
1853 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
1854 amd_iommu_dev_table[devid].data[1] &= DTE_FLAG_MASK;
1855
1856 amd_iommu_apply_erratum_63(devid);
1857 }
1858
1859 static void do_attach(struct iommu_dev_data *dev_data,
1860 struct protection_domain *domain)
1861 {
1862 struct amd_iommu *iommu;
1863 u16 alias;
1864 bool ats;
1865
1866 iommu = amd_iommu_rlookup_table[dev_data->devid];
1867 alias = dev_data->alias;
1868 ats = dev_data->ats.enabled;
1869
1870 /* Update data structures */
1871 dev_data->domain = domain;
1872 list_add(&dev_data->list, &domain->dev_list);
1873
1874 /* Do reference counting */
1875 domain->dev_iommu[iommu->index] += 1;
1876 domain->dev_cnt += 1;
1877
1878 /* Update device table */
1879 set_dte_entry(dev_data->devid, domain, ats);
1880 if (alias != dev_data->devid)
1881 set_dte_entry(alias, domain, ats);
1882
1883 device_flush_dte(dev_data);
1884 }
1885
1886 static void do_detach(struct iommu_dev_data *dev_data)
1887 {
1888 struct amd_iommu *iommu;
1889 u16 alias;
1890
1891 /*
1892 * First check if the device is still attached. It might already
1893 * be detached from its domain because the generic
1894 * iommu_detach_group code detached it and we try again here in
1895 * our alias handling.
1896 */
1897 if (!dev_data->domain)
1898 return;
1899
1900 iommu = amd_iommu_rlookup_table[dev_data->devid];
1901 alias = dev_data->alias;
1902
1903 /* decrease reference counters */
1904 dev_data->domain->dev_iommu[iommu->index] -= 1;
1905 dev_data->domain->dev_cnt -= 1;
1906
1907 /* Update data structures */
1908 dev_data->domain = NULL;
1909 list_del(&dev_data->list);
1910 clear_dte_entry(dev_data->devid);
1911 if (alias != dev_data->devid)
1912 clear_dte_entry(alias);
1913
1914 /* Flush the DTE entry */
1915 device_flush_dte(dev_data);
1916 }
1917
1918 /*
1919 * If a device is not yet associated with a domain, this function does
1920 * assigns it visible for the hardware
1921 */
1922 static int __attach_device(struct iommu_dev_data *dev_data,
1923 struct protection_domain *domain)
1924 {
1925 int ret;
1926
1927 /*
1928 * Must be called with IRQs disabled. Warn here to detect early
1929 * when its not.
1930 */
1931 WARN_ON(!irqs_disabled());
1932
1933 /* lock domain */
1934 spin_lock(&domain->lock);
1935
1936 ret = -EBUSY;
1937 if (dev_data->domain != NULL)
1938 goto out_unlock;
1939
1940 /* Attach alias group root */
1941 do_attach(dev_data, domain);
1942
1943 ret = 0;
1944
1945 out_unlock:
1946
1947 /* ready */
1948 spin_unlock(&domain->lock);
1949
1950 return ret;
1951 }
1952
1953
1954 static void pdev_iommuv2_disable(struct pci_dev *pdev)
1955 {
1956 pci_disable_ats(pdev);
1957 pci_disable_pri(pdev);
1958 pci_disable_pasid(pdev);
1959 }
1960
1961 /* FIXME: Change generic reset-function to do the same */
1962 static int pri_reset_while_enabled(struct pci_dev *pdev)
1963 {
1964 u16 control;
1965 int pos;
1966
1967 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
1968 if (!pos)
1969 return -EINVAL;
1970
1971 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
1972 control |= PCI_PRI_CTRL_RESET;
1973 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
1974
1975 return 0;
1976 }
1977
1978 static int pdev_iommuv2_enable(struct pci_dev *pdev)
1979 {
1980 bool reset_enable;
1981 int reqs, ret;
1982
1983 /* FIXME: Hardcode number of outstanding requests for now */
1984 reqs = 32;
1985 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
1986 reqs = 1;
1987 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
1988
1989 /* Only allow access to user-accessible pages */
1990 ret = pci_enable_pasid(pdev, 0);
1991 if (ret)
1992 goto out_err;
1993
1994 /* First reset the PRI state of the device */
1995 ret = pci_reset_pri(pdev);
1996 if (ret)
1997 goto out_err;
1998
1999 /* Enable PRI */
2000 ret = pci_enable_pri(pdev, reqs);
2001 if (ret)
2002 goto out_err;
2003
2004 if (reset_enable) {
2005 ret = pri_reset_while_enabled(pdev);
2006 if (ret)
2007 goto out_err;
2008 }
2009
2010 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2011 if (ret)
2012 goto out_err;
2013
2014 return 0;
2015
2016 out_err:
2017 pci_disable_pri(pdev);
2018 pci_disable_pasid(pdev);
2019
2020 return ret;
2021 }
2022
2023 /* FIXME: Move this to PCI code */
2024 #define PCI_PRI_TLP_OFF (1 << 15)
2025
2026 static bool pci_pri_tlp_required(struct pci_dev *pdev)
2027 {
2028 u16 status;
2029 int pos;
2030
2031 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2032 if (!pos)
2033 return false;
2034
2035 pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
2036
2037 return (status & PCI_PRI_TLP_OFF) ? true : false;
2038 }
2039
2040 /*
2041 * If a device is not yet associated with a domain, this function
2042 * assigns it visible for the hardware
2043 */
2044 static int attach_device(struct device *dev,
2045 struct protection_domain *domain)
2046 {
2047 struct pci_dev *pdev;
2048 struct iommu_dev_data *dev_data;
2049 unsigned long flags;
2050 int ret;
2051
2052 dev_data = get_dev_data(dev);
2053
2054 if (!dev_is_pci(dev))
2055 goto skip_ats_check;
2056
2057 pdev = to_pci_dev(dev);
2058 if (domain->flags & PD_IOMMUV2_MASK) {
2059 if (!dev_data->passthrough)
2060 return -EINVAL;
2061
2062 if (dev_data->iommu_v2) {
2063 if (pdev_iommuv2_enable(pdev) != 0)
2064 return -EINVAL;
2065
2066 dev_data->ats.enabled = true;
2067 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2068 dev_data->pri_tlp = pci_pri_tlp_required(pdev);
2069 }
2070 } else if (amd_iommu_iotlb_sup &&
2071 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2072 dev_data->ats.enabled = true;
2073 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2074 }
2075
2076 skip_ats_check:
2077 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2078 ret = __attach_device(dev_data, domain);
2079 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2080
2081 /*
2082 * We might boot into a crash-kernel here. The crashed kernel
2083 * left the caches in the IOMMU dirty. So we have to flush
2084 * here to evict all dirty stuff.
2085 */
2086 domain_flush_tlb_pde(domain);
2087
2088 return ret;
2089 }
2090
2091 /*
2092 * Removes a device from a protection domain (unlocked)
2093 */
2094 static void __detach_device(struct iommu_dev_data *dev_data)
2095 {
2096 struct protection_domain *domain;
2097
2098 /*
2099 * Must be called with IRQs disabled. Warn here to detect early
2100 * when its not.
2101 */
2102 WARN_ON(!irqs_disabled());
2103
2104 if (WARN_ON(!dev_data->domain))
2105 return;
2106
2107 domain = dev_data->domain;
2108
2109 spin_lock(&domain->lock);
2110
2111 do_detach(dev_data);
2112
2113 spin_unlock(&domain->lock);
2114 }
2115
2116 /*
2117 * Removes a device from a protection domain (with devtable_lock held)
2118 */
2119 static void detach_device(struct device *dev)
2120 {
2121 struct protection_domain *domain;
2122 struct iommu_dev_data *dev_data;
2123 unsigned long flags;
2124
2125 dev_data = get_dev_data(dev);
2126 domain = dev_data->domain;
2127
2128 /* lock device table */
2129 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2130 __detach_device(dev_data);
2131 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2132
2133 if (!dev_is_pci(dev))
2134 return;
2135
2136 if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2)
2137 pdev_iommuv2_disable(to_pci_dev(dev));
2138 else if (dev_data->ats.enabled)
2139 pci_disable_ats(to_pci_dev(dev));
2140
2141 dev_data->ats.enabled = false;
2142 }
2143
2144 static int amd_iommu_add_device(struct device *dev)
2145 {
2146 struct iommu_dev_data *dev_data;
2147 struct iommu_domain *domain;
2148 struct amd_iommu *iommu;
2149 int ret, devid;
2150
2151 if (!check_device(dev) || get_dev_data(dev))
2152 return 0;
2153
2154 devid = get_device_id(dev);
2155 if (devid < 0)
2156 return devid;
2157
2158 iommu = amd_iommu_rlookup_table[devid];
2159
2160 ret = iommu_init_device(dev);
2161 if (ret) {
2162 if (ret != -ENOTSUPP)
2163 pr_err("Failed to initialize device %s - trying to proceed anyway\n",
2164 dev_name(dev));
2165
2166 iommu_ignore_device(dev);
2167 dev->archdata.dma_ops = &nommu_dma_ops;
2168 goto out;
2169 }
2170 init_iommu_group(dev);
2171
2172 dev_data = get_dev_data(dev);
2173
2174 BUG_ON(!dev_data);
2175
2176 if (iommu_pass_through || dev_data->iommu_v2)
2177 iommu_request_dm_for_dev(dev);
2178
2179 /* Domains are initialized for this device - have a look what we ended up with */
2180 domain = iommu_get_domain_for_dev(dev);
2181 if (domain->type == IOMMU_DOMAIN_IDENTITY)
2182 dev_data->passthrough = true;
2183 else
2184 dev->archdata.dma_ops = &amd_iommu_dma_ops;
2185
2186 out:
2187 iommu_completion_wait(iommu);
2188
2189 return 0;
2190 }
2191
2192 static void amd_iommu_remove_device(struct device *dev)
2193 {
2194 struct amd_iommu *iommu;
2195 int devid;
2196
2197 if (!check_device(dev))
2198 return;
2199
2200 devid = get_device_id(dev);
2201 if (devid < 0)
2202 return;
2203
2204 iommu = amd_iommu_rlookup_table[devid];
2205
2206 iommu_uninit_device(dev);
2207 iommu_completion_wait(iommu);
2208 }
2209
2210 static struct iommu_group *amd_iommu_device_group(struct device *dev)
2211 {
2212 if (dev_is_pci(dev))
2213 return pci_device_group(dev);
2214
2215 return acpihid_device_group(dev);
2216 }
2217
2218 /*****************************************************************************
2219 *
2220 * The next functions belong to the dma_ops mapping/unmapping code.
2221 *
2222 *****************************************************************************/
2223
2224 static void __queue_flush(struct flush_queue *queue)
2225 {
2226 struct protection_domain *domain;
2227 unsigned long flags;
2228 int idx;
2229
2230 /* First flush TLB of all known domains */
2231 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
2232 list_for_each_entry(domain, &amd_iommu_pd_list, list)
2233 domain_flush_tlb(domain);
2234 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
2235
2236 /* Wait until flushes have completed */
2237 domain_flush_complete(NULL);
2238
2239 for (idx = 0; idx < queue->next; ++idx) {
2240 struct flush_queue_entry *entry;
2241
2242 entry = queue->entries + idx;
2243
2244 free_iova_fast(&entry->dma_dom->iovad,
2245 entry->iova_pfn,
2246 entry->pages);
2247
2248 /* Not really necessary, just to make sure we catch any bugs */
2249 entry->dma_dom = NULL;
2250 }
2251
2252 queue->next = 0;
2253 }
2254
2255 static void queue_flush_all(void)
2256 {
2257 int cpu;
2258
2259 for_each_possible_cpu(cpu) {
2260 struct flush_queue *queue;
2261 unsigned long flags;
2262
2263 queue = per_cpu_ptr(&flush_queue, cpu);
2264 spin_lock_irqsave(&queue->lock, flags);
2265 if (queue->next > 0)
2266 __queue_flush(queue);
2267 spin_unlock_irqrestore(&queue->lock, flags);
2268 }
2269 }
2270
2271 static void queue_flush_timeout(unsigned long unsused)
2272 {
2273 atomic_set(&queue_timer_on, 0);
2274 queue_flush_all();
2275 }
2276
2277 static void queue_add(struct dma_ops_domain *dma_dom,
2278 unsigned long address, unsigned long pages)
2279 {
2280 struct flush_queue_entry *entry;
2281 struct flush_queue *queue;
2282 unsigned long flags;
2283 int idx;
2284
2285 pages = __roundup_pow_of_two(pages);
2286 address >>= PAGE_SHIFT;
2287
2288 queue = get_cpu_ptr(&flush_queue);
2289 spin_lock_irqsave(&queue->lock, flags);
2290
2291 if (queue->next == FLUSH_QUEUE_SIZE)
2292 __queue_flush(queue);
2293
2294 idx = queue->next++;
2295 entry = queue->entries + idx;
2296
2297 entry->iova_pfn = address;
2298 entry->pages = pages;
2299 entry->dma_dom = dma_dom;
2300
2301 spin_unlock_irqrestore(&queue->lock, flags);
2302
2303 if (atomic_cmpxchg(&queue_timer_on, 0, 1) == 0)
2304 mod_timer(&queue_timer, jiffies + msecs_to_jiffies(10));
2305
2306 put_cpu_ptr(&flush_queue);
2307 }
2308
2309
2310 /*
2311 * In the dma_ops path we only have the struct device. This function
2312 * finds the corresponding IOMMU, the protection domain and the
2313 * requestor id for a given device.
2314 * If the device is not yet associated with a domain this is also done
2315 * in this function.
2316 */
2317 static struct protection_domain *get_domain(struct device *dev)
2318 {
2319 struct protection_domain *domain;
2320
2321 if (!check_device(dev))
2322 return ERR_PTR(-EINVAL);
2323
2324 domain = get_dev_data(dev)->domain;
2325 if (!dma_ops_domain(domain))
2326 return ERR_PTR(-EBUSY);
2327
2328 return domain;
2329 }
2330
2331 static void update_device_table(struct protection_domain *domain)
2332 {
2333 struct iommu_dev_data *dev_data;
2334
2335 list_for_each_entry(dev_data, &domain->dev_list, list) {
2336 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled);
2337
2338 if (dev_data->devid == dev_data->alias)
2339 continue;
2340
2341 /* There is an alias, update device table entry for it */
2342 set_dte_entry(dev_data->alias, domain, dev_data->ats.enabled);
2343 }
2344 }
2345
2346 static void update_domain(struct protection_domain *domain)
2347 {
2348 if (!domain->updated)
2349 return;
2350
2351 update_device_table(domain);
2352
2353 domain_flush_devices(domain);
2354 domain_flush_tlb_pde(domain);
2355
2356 domain->updated = false;
2357 }
2358
2359 static int dir2prot(enum dma_data_direction direction)
2360 {
2361 if (direction == DMA_TO_DEVICE)
2362 return IOMMU_PROT_IR;
2363 else if (direction == DMA_FROM_DEVICE)
2364 return IOMMU_PROT_IW;
2365 else if (direction == DMA_BIDIRECTIONAL)
2366 return IOMMU_PROT_IW | IOMMU_PROT_IR;
2367 else
2368 return 0;
2369 }
2370 /*
2371 * This function contains common code for mapping of a physically
2372 * contiguous memory region into DMA address space. It is used by all
2373 * mapping functions provided with this IOMMU driver.
2374 * Must be called with the domain lock held.
2375 */
2376 static dma_addr_t __map_single(struct device *dev,
2377 struct dma_ops_domain *dma_dom,
2378 phys_addr_t paddr,
2379 size_t size,
2380 enum dma_data_direction direction,
2381 u64 dma_mask)
2382 {
2383 dma_addr_t offset = paddr & ~PAGE_MASK;
2384 dma_addr_t address, start, ret;
2385 unsigned int pages;
2386 int prot = 0;
2387 int i;
2388
2389 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2390 paddr &= PAGE_MASK;
2391
2392 address = dma_ops_alloc_iova(dev, dma_dom, pages, dma_mask);
2393 if (address == DMA_ERROR_CODE)
2394 goto out;
2395
2396 prot = dir2prot(direction);
2397
2398 start = address;
2399 for (i = 0; i < pages; ++i) {
2400 ret = iommu_map_page(&dma_dom->domain, start, paddr,
2401 PAGE_SIZE, prot, GFP_ATOMIC);
2402 if (ret)
2403 goto out_unmap;
2404
2405 paddr += PAGE_SIZE;
2406 start += PAGE_SIZE;
2407 }
2408 address += offset;
2409
2410 if (unlikely(amd_iommu_np_cache)) {
2411 domain_flush_pages(&dma_dom->domain, address, size);
2412 domain_flush_complete(&dma_dom->domain);
2413 }
2414
2415 out:
2416 return address;
2417
2418 out_unmap:
2419
2420 for (--i; i >= 0; --i) {
2421 start -= PAGE_SIZE;
2422 iommu_unmap_page(&dma_dom->domain, start, PAGE_SIZE);
2423 }
2424
2425 domain_flush_tlb(&dma_dom->domain);
2426 domain_flush_complete(&dma_dom->domain);
2427
2428 dma_ops_free_iova(dma_dom, address, pages);
2429
2430 return DMA_ERROR_CODE;
2431 }
2432
2433 /*
2434 * Does the reverse of the __map_single function. Must be called with
2435 * the domain lock held too
2436 */
2437 static void __unmap_single(struct dma_ops_domain *dma_dom,
2438 dma_addr_t dma_addr,
2439 size_t size,
2440 int dir)
2441 {
2442 dma_addr_t flush_addr;
2443 dma_addr_t i, start;
2444 unsigned int pages;
2445
2446 flush_addr = dma_addr;
2447 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2448 dma_addr &= PAGE_MASK;
2449 start = dma_addr;
2450
2451 for (i = 0; i < pages; ++i) {
2452 iommu_unmap_page(&dma_dom->domain, start, PAGE_SIZE);
2453 start += PAGE_SIZE;
2454 }
2455
2456 if (amd_iommu_unmap_flush) {
2457 dma_ops_free_iova(dma_dom, dma_addr, pages);
2458 domain_flush_tlb(&dma_dom->domain);
2459 domain_flush_complete(&dma_dom->domain);
2460 } else {
2461 queue_add(dma_dom, dma_addr, pages);
2462 }
2463 }
2464
2465 /*
2466 * The exported map_single function for dma_ops.
2467 */
2468 static dma_addr_t map_page(struct device *dev, struct page *page,
2469 unsigned long offset, size_t size,
2470 enum dma_data_direction dir,
2471 unsigned long attrs)
2472 {
2473 phys_addr_t paddr = page_to_phys(page) + offset;
2474 struct protection_domain *domain;
2475 struct dma_ops_domain *dma_dom;
2476 u64 dma_mask;
2477
2478 domain = get_domain(dev);
2479 if (PTR_ERR(domain) == -EINVAL)
2480 return (dma_addr_t)paddr;
2481 else if (IS_ERR(domain))
2482 return DMA_ERROR_CODE;
2483
2484 dma_mask = *dev->dma_mask;
2485 dma_dom = to_dma_ops_domain(domain);
2486
2487 return __map_single(dev, dma_dom, paddr, size, dir, dma_mask);
2488 }
2489
2490 /*
2491 * The exported unmap_single function for dma_ops.
2492 */
2493 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2494 enum dma_data_direction dir, unsigned long attrs)
2495 {
2496 struct protection_domain *domain;
2497 struct dma_ops_domain *dma_dom;
2498
2499 domain = get_domain(dev);
2500 if (IS_ERR(domain))
2501 return;
2502
2503 dma_dom = to_dma_ops_domain(domain);
2504
2505 __unmap_single(dma_dom, dma_addr, size, dir);
2506 }
2507
2508 static int sg_num_pages(struct device *dev,
2509 struct scatterlist *sglist,
2510 int nelems)
2511 {
2512 unsigned long mask, boundary_size;
2513 struct scatterlist *s;
2514 int i, npages = 0;
2515
2516 mask = dma_get_seg_boundary(dev);
2517 boundary_size = mask + 1 ? ALIGN(mask + 1, PAGE_SIZE) >> PAGE_SHIFT :
2518 1UL << (BITS_PER_LONG - PAGE_SHIFT);
2519
2520 for_each_sg(sglist, s, nelems, i) {
2521 int p, n;
2522
2523 s->dma_address = npages << PAGE_SHIFT;
2524 p = npages % boundary_size;
2525 n = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2526 if (p + n > boundary_size)
2527 npages += boundary_size - p;
2528 npages += n;
2529 }
2530
2531 return npages;
2532 }
2533
2534 /*
2535 * The exported map_sg function for dma_ops (handles scatter-gather
2536 * lists).
2537 */
2538 static int map_sg(struct device *dev, struct scatterlist *sglist,
2539 int nelems, enum dma_data_direction direction,
2540 unsigned long attrs)
2541 {
2542 int mapped_pages = 0, npages = 0, prot = 0, i;
2543 struct protection_domain *domain;
2544 struct dma_ops_domain *dma_dom;
2545 struct scatterlist *s;
2546 unsigned long address;
2547 u64 dma_mask;
2548
2549 domain = get_domain(dev);
2550 if (IS_ERR(domain))
2551 return 0;
2552
2553 dma_dom = to_dma_ops_domain(domain);
2554 dma_mask = *dev->dma_mask;
2555
2556 npages = sg_num_pages(dev, sglist, nelems);
2557
2558 address = dma_ops_alloc_iova(dev, dma_dom, npages, dma_mask);
2559 if (address == DMA_ERROR_CODE)
2560 goto out_err;
2561
2562 prot = dir2prot(direction);
2563
2564 /* Map all sg entries */
2565 for_each_sg(sglist, s, nelems, i) {
2566 int j, pages = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2567
2568 for (j = 0; j < pages; ++j) {
2569 unsigned long bus_addr, phys_addr;
2570 int ret;
2571
2572 bus_addr = address + s->dma_address + (j << PAGE_SHIFT);
2573 phys_addr = (sg_phys(s) & PAGE_MASK) + (j << PAGE_SHIFT);
2574 ret = iommu_map_page(domain, bus_addr, phys_addr, PAGE_SIZE, prot, GFP_ATOMIC);
2575 if (ret)
2576 goto out_unmap;
2577
2578 mapped_pages += 1;
2579 }
2580 }
2581
2582 /* Everything is mapped - write the right values into s->dma_address */
2583 for_each_sg(sglist, s, nelems, i) {
2584 s->dma_address += address + s->offset;
2585 s->dma_length = s->length;
2586 }
2587
2588 return nelems;
2589
2590 out_unmap:
2591 pr_err("%s: IOMMU mapping error in map_sg (io-pages: %d)\n",
2592 dev_name(dev), npages);
2593
2594 for_each_sg(sglist, s, nelems, i) {
2595 int j, pages = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2596
2597 for (j = 0; j < pages; ++j) {
2598 unsigned long bus_addr;
2599
2600 bus_addr = address + s->dma_address + (j << PAGE_SHIFT);
2601 iommu_unmap_page(domain, bus_addr, PAGE_SIZE);
2602
2603 if (--mapped_pages)
2604 goto out_free_iova;
2605 }
2606 }
2607
2608 out_free_iova:
2609 free_iova_fast(&dma_dom->iovad, address, npages);
2610
2611 out_err:
2612 return 0;
2613 }
2614
2615 /*
2616 * The exported map_sg function for dma_ops (handles scatter-gather
2617 * lists).
2618 */
2619 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2620 int nelems, enum dma_data_direction dir,
2621 unsigned long attrs)
2622 {
2623 struct protection_domain *domain;
2624 struct dma_ops_domain *dma_dom;
2625 unsigned long startaddr;
2626 int npages = 2;
2627
2628 domain = get_domain(dev);
2629 if (IS_ERR(domain))
2630 return;
2631
2632 startaddr = sg_dma_address(sglist) & PAGE_MASK;
2633 dma_dom = to_dma_ops_domain(domain);
2634 npages = sg_num_pages(dev, sglist, nelems);
2635
2636 __unmap_single(dma_dom, startaddr, npages << PAGE_SHIFT, dir);
2637 }
2638
2639 /*
2640 * The exported alloc_coherent function for dma_ops.
2641 */
2642 static void *alloc_coherent(struct device *dev, size_t size,
2643 dma_addr_t *dma_addr, gfp_t flag,
2644 unsigned long attrs)
2645 {
2646 u64 dma_mask = dev->coherent_dma_mask;
2647 struct protection_domain *domain;
2648 struct dma_ops_domain *dma_dom;
2649 struct page *page;
2650
2651 domain = get_domain(dev);
2652 if (PTR_ERR(domain) == -EINVAL) {
2653 page = alloc_pages(flag, get_order(size));
2654 *dma_addr = page_to_phys(page);
2655 return page_address(page);
2656 } else if (IS_ERR(domain))
2657 return NULL;
2658
2659 dma_dom = to_dma_ops_domain(domain);
2660 size = PAGE_ALIGN(size);
2661 dma_mask = dev->coherent_dma_mask;
2662 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2663 flag |= __GFP_ZERO;
2664
2665 page = alloc_pages(flag | __GFP_NOWARN, get_order(size));
2666 if (!page) {
2667 if (!gfpflags_allow_blocking(flag))
2668 return NULL;
2669
2670 page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT,
2671 get_order(size));
2672 if (!page)
2673 return NULL;
2674 }
2675
2676 if (!dma_mask)
2677 dma_mask = *dev->dma_mask;
2678
2679 *dma_addr = __map_single(dev, dma_dom, page_to_phys(page),
2680 size, DMA_BIDIRECTIONAL, dma_mask);
2681
2682 if (*dma_addr == DMA_ERROR_CODE)
2683 goto out_free;
2684
2685 return page_address(page);
2686
2687 out_free:
2688
2689 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2690 __free_pages(page, get_order(size));
2691
2692 return NULL;
2693 }
2694
2695 /*
2696 * The exported free_coherent function for dma_ops.
2697 */
2698 static void free_coherent(struct device *dev, size_t size,
2699 void *virt_addr, dma_addr_t dma_addr,
2700 unsigned long attrs)
2701 {
2702 struct protection_domain *domain;
2703 struct dma_ops_domain *dma_dom;
2704 struct page *page;
2705
2706 page = virt_to_page(virt_addr);
2707 size = PAGE_ALIGN(size);
2708
2709 domain = get_domain(dev);
2710 if (IS_ERR(domain))
2711 goto free_mem;
2712
2713 dma_dom = to_dma_ops_domain(domain);
2714
2715 __unmap_single(dma_dom, dma_addr, size, DMA_BIDIRECTIONAL);
2716
2717 free_mem:
2718 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2719 __free_pages(page, get_order(size));
2720 }
2721
2722 /*
2723 * This function is called by the DMA layer to find out if we can handle a
2724 * particular device. It is part of the dma_ops.
2725 */
2726 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2727 {
2728 return check_device(dev);
2729 }
2730
2731 static struct dma_map_ops amd_iommu_dma_ops = {
2732 .alloc = alloc_coherent,
2733 .free = free_coherent,
2734 .map_page = map_page,
2735 .unmap_page = unmap_page,
2736 .map_sg = map_sg,
2737 .unmap_sg = unmap_sg,
2738 .dma_supported = amd_iommu_dma_supported,
2739 };
2740
2741 static int init_reserved_iova_ranges(void)
2742 {
2743 struct pci_dev *pdev = NULL;
2744 struct iova *val;
2745
2746 init_iova_domain(&reserved_iova_ranges, PAGE_SIZE,
2747 IOVA_START_PFN, DMA_32BIT_PFN);
2748
2749 lockdep_set_class(&reserved_iova_ranges.iova_rbtree_lock,
2750 &reserved_rbtree_key);
2751
2752 /* MSI memory range */
2753 val = reserve_iova(&reserved_iova_ranges,
2754 IOVA_PFN(MSI_RANGE_START), IOVA_PFN(MSI_RANGE_END));
2755 if (!val) {
2756 pr_err("Reserving MSI range failed\n");
2757 return -ENOMEM;
2758 }
2759
2760 /* HT memory range */
2761 val = reserve_iova(&reserved_iova_ranges,
2762 IOVA_PFN(HT_RANGE_START), IOVA_PFN(HT_RANGE_END));
2763 if (!val) {
2764 pr_err("Reserving HT range failed\n");
2765 return -ENOMEM;
2766 }
2767
2768 /*
2769 * Memory used for PCI resources
2770 * FIXME: Check whether we can reserve the PCI-hole completly
2771 */
2772 for_each_pci_dev(pdev) {
2773 int i;
2774
2775 for (i = 0; i < PCI_NUM_RESOURCES; ++i) {
2776 struct resource *r = &pdev->resource[i];
2777
2778 if (!(r->flags & IORESOURCE_MEM))
2779 continue;
2780
2781 val = reserve_iova(&reserved_iova_ranges,
2782 IOVA_PFN(r->start),
2783 IOVA_PFN(r->end));
2784 if (!val) {
2785 pr_err("Reserve pci-resource range failed\n");
2786 return -ENOMEM;
2787 }
2788 }
2789 }
2790
2791 return 0;
2792 }
2793
2794 int __init amd_iommu_init_api(void)
2795 {
2796 int ret, cpu, err = 0;
2797
2798 ret = iova_cache_get();
2799 if (ret)
2800 return ret;
2801
2802 ret = init_reserved_iova_ranges();
2803 if (ret)
2804 return ret;
2805
2806 for_each_possible_cpu(cpu) {
2807 struct flush_queue *queue = per_cpu_ptr(&flush_queue, cpu);
2808
2809 queue->entries = kzalloc(FLUSH_QUEUE_SIZE *
2810 sizeof(*queue->entries),
2811 GFP_KERNEL);
2812 if (!queue->entries)
2813 goto out_put_iova;
2814
2815 spin_lock_init(&queue->lock);
2816 }
2817
2818 err = bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
2819 if (err)
2820 return err;
2821 #ifdef CONFIG_ARM_AMBA
2822 err = bus_set_iommu(&amba_bustype, &amd_iommu_ops);
2823 if (err)
2824 return err;
2825 #endif
2826 err = bus_set_iommu(&platform_bus_type, &amd_iommu_ops);
2827 if (err)
2828 return err;
2829 return 0;
2830
2831 out_put_iova:
2832 for_each_possible_cpu(cpu) {
2833 struct flush_queue *queue = per_cpu_ptr(&flush_queue, cpu);
2834
2835 kfree(queue->entries);
2836 }
2837
2838 return -ENOMEM;
2839 }
2840
2841 int __init amd_iommu_init_dma_ops(void)
2842 {
2843 setup_timer(&queue_timer, queue_flush_timeout, 0);
2844 atomic_set(&queue_timer_on, 0);
2845
2846 swiotlb = iommu_pass_through ? 1 : 0;
2847 iommu_detected = 1;
2848
2849 /*
2850 * In case we don't initialize SWIOTLB (actually the common case
2851 * when AMD IOMMU is enabled), make sure there are global
2852 * dma_ops set as a fall-back for devices not handled by this
2853 * driver (for example non-PCI devices).
2854 */
2855 if (!swiotlb)
2856 dma_ops = &nommu_dma_ops;
2857
2858 if (amd_iommu_unmap_flush)
2859 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
2860 else
2861 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
2862
2863 return 0;
2864
2865 }
2866
2867 /*****************************************************************************
2868 *
2869 * The following functions belong to the exported interface of AMD IOMMU
2870 *
2871 * This interface allows access to lower level functions of the IOMMU
2872 * like protection domain handling and assignement of devices to domains
2873 * which is not possible with the dma_ops interface.
2874 *
2875 *****************************************************************************/
2876
2877 static void cleanup_domain(struct protection_domain *domain)
2878 {
2879 struct iommu_dev_data *entry;
2880 unsigned long flags;
2881
2882 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2883
2884 while (!list_empty(&domain->dev_list)) {
2885 entry = list_first_entry(&domain->dev_list,
2886 struct iommu_dev_data, list);
2887 __detach_device(entry);
2888 }
2889
2890 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2891 }
2892
2893 static void protection_domain_free(struct protection_domain *domain)
2894 {
2895 if (!domain)
2896 return;
2897
2898 del_domain_from_list(domain);
2899
2900 if (domain->id)
2901 domain_id_free(domain->id);
2902
2903 kfree(domain);
2904 }
2905
2906 static int protection_domain_init(struct protection_domain *domain)
2907 {
2908 spin_lock_init(&domain->lock);
2909 mutex_init(&domain->api_lock);
2910 domain->id = domain_id_alloc();
2911 if (!domain->id)
2912 return -ENOMEM;
2913 INIT_LIST_HEAD(&domain->dev_list);
2914
2915 return 0;
2916 }
2917
2918 static struct protection_domain *protection_domain_alloc(void)
2919 {
2920 struct protection_domain *domain;
2921
2922 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2923 if (!domain)
2924 return NULL;
2925
2926 if (protection_domain_init(domain))
2927 goto out_err;
2928
2929 add_domain_to_list(domain);
2930
2931 return domain;
2932
2933 out_err:
2934 kfree(domain);
2935
2936 return NULL;
2937 }
2938
2939 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
2940 {
2941 struct protection_domain *pdomain;
2942 struct dma_ops_domain *dma_domain;
2943
2944 switch (type) {
2945 case IOMMU_DOMAIN_UNMANAGED:
2946 pdomain = protection_domain_alloc();
2947 if (!pdomain)
2948 return NULL;
2949
2950 pdomain->mode = PAGE_MODE_3_LEVEL;
2951 pdomain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2952 if (!pdomain->pt_root) {
2953 protection_domain_free(pdomain);
2954 return NULL;
2955 }
2956
2957 pdomain->domain.geometry.aperture_start = 0;
2958 pdomain->domain.geometry.aperture_end = ~0ULL;
2959 pdomain->domain.geometry.force_aperture = true;
2960
2961 break;
2962 case IOMMU_DOMAIN_DMA:
2963 dma_domain = dma_ops_domain_alloc();
2964 if (!dma_domain) {
2965 pr_err("AMD-Vi: Failed to allocate\n");
2966 return NULL;
2967 }
2968 pdomain = &dma_domain->domain;
2969 break;
2970 case IOMMU_DOMAIN_IDENTITY:
2971 pdomain = protection_domain_alloc();
2972 if (!pdomain)
2973 return NULL;
2974
2975 pdomain->mode = PAGE_MODE_NONE;
2976 break;
2977 default:
2978 return NULL;
2979 }
2980
2981 return &pdomain->domain;
2982 }
2983
2984 static void amd_iommu_domain_free(struct iommu_domain *dom)
2985 {
2986 struct protection_domain *domain;
2987 struct dma_ops_domain *dma_dom;
2988
2989 domain = to_pdomain(dom);
2990
2991 if (domain->dev_cnt > 0)
2992 cleanup_domain(domain);
2993
2994 BUG_ON(domain->dev_cnt != 0);
2995
2996 if (!dom)
2997 return;
2998
2999 switch (dom->type) {
3000 case IOMMU_DOMAIN_DMA:
3001 /*
3002 * First make sure the domain is no longer referenced from the
3003 * flush queue
3004 */
3005 queue_flush_all();
3006
3007 /* Now release the domain */
3008 dma_dom = to_dma_ops_domain(domain);
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
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 int devid;
3029
3030 if (!check_device(dev))
3031 return;
3032
3033 devid = get_device_id(dev);
3034 if (devid < 0)
3035 return;
3036
3037 if (dev_data->domain != NULL)
3038 detach_device(dev);
3039
3040 iommu = amd_iommu_rlookup_table[devid];
3041 if (!iommu)
3042 return;
3043
3044 #ifdef CONFIG_IRQ_REMAP
3045 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) &&
3046 (dom->type == IOMMU_DOMAIN_UNMANAGED))
3047 dev_data->use_vapic = 0;
3048 #endif
3049
3050 iommu_completion_wait(iommu);
3051 }
3052
3053 static int amd_iommu_attach_device(struct iommu_domain *dom,
3054 struct device *dev)
3055 {
3056 struct protection_domain *domain = to_pdomain(dom);
3057 struct iommu_dev_data *dev_data;
3058 struct amd_iommu *iommu;
3059 int ret;
3060
3061 if (!check_device(dev))
3062 return -EINVAL;
3063
3064 dev_data = dev->archdata.iommu;
3065
3066 iommu = amd_iommu_rlookup_table[dev_data->devid];
3067 if (!iommu)
3068 return -EINVAL;
3069
3070 if (dev_data->domain)
3071 detach_device(dev);
3072
3073 ret = attach_device(dev, domain);
3074
3075 #ifdef CONFIG_IRQ_REMAP
3076 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
3077 if (dom->type == IOMMU_DOMAIN_UNMANAGED)
3078 dev_data->use_vapic = 1;
3079 else
3080 dev_data->use_vapic = 0;
3081 }
3082 #endif
3083
3084 iommu_completion_wait(iommu);
3085
3086 return ret;
3087 }
3088
3089 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
3090 phys_addr_t paddr, size_t page_size, int iommu_prot)
3091 {
3092 struct protection_domain *domain = to_pdomain(dom);
3093 int prot = 0;
3094 int ret;
3095
3096 if (domain->mode == PAGE_MODE_NONE)
3097 return -EINVAL;
3098
3099 if (iommu_prot & IOMMU_READ)
3100 prot |= IOMMU_PROT_IR;
3101 if (iommu_prot & IOMMU_WRITE)
3102 prot |= IOMMU_PROT_IW;
3103
3104 mutex_lock(&domain->api_lock);
3105 ret = iommu_map_page(domain, iova, paddr, page_size, prot, GFP_KERNEL);
3106 mutex_unlock(&domain->api_lock);
3107
3108 return ret;
3109 }
3110
3111 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3112 size_t page_size)
3113 {
3114 struct protection_domain *domain = to_pdomain(dom);
3115 size_t unmap_size;
3116
3117 if (domain->mode == PAGE_MODE_NONE)
3118 return -EINVAL;
3119
3120 mutex_lock(&domain->api_lock);
3121 unmap_size = iommu_unmap_page(domain, iova, page_size);
3122 mutex_unlock(&domain->api_lock);
3123
3124 domain_flush_tlb_pde(domain);
3125
3126 return unmap_size;
3127 }
3128
3129 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3130 dma_addr_t iova)
3131 {
3132 struct protection_domain *domain = to_pdomain(dom);
3133 unsigned long offset_mask, pte_pgsize;
3134 u64 *pte, __pte;
3135
3136 if (domain->mode == PAGE_MODE_NONE)
3137 return iova;
3138
3139 pte = fetch_pte(domain, iova, &pte_pgsize);
3140
3141 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3142 return 0;
3143
3144 offset_mask = pte_pgsize - 1;
3145 __pte = *pte & PM_ADDR_MASK;
3146
3147 return (__pte & ~offset_mask) | (iova & offset_mask);
3148 }
3149
3150 static bool amd_iommu_capable(enum iommu_cap cap)
3151 {
3152 switch (cap) {
3153 case IOMMU_CAP_CACHE_COHERENCY:
3154 return true;
3155 case IOMMU_CAP_INTR_REMAP:
3156 return (irq_remapping_enabled == 1);
3157 case IOMMU_CAP_NOEXEC:
3158 return false;
3159 }
3160
3161 return false;
3162 }
3163
3164 static void amd_iommu_get_resv_regions(struct device *dev,
3165 struct list_head *head)
3166 {
3167 struct unity_map_entry *entry;
3168 int devid;
3169
3170 devid = get_device_id(dev);
3171 if (devid < 0)
3172 return;
3173
3174 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
3175 struct iommu_resv_region *region;
3176
3177 if (devid < entry->devid_start || devid > entry->devid_end)
3178 continue;
3179
3180 region = kzalloc(sizeof(*region), GFP_KERNEL);
3181 if (!region) {
3182 pr_err("Out of memory allocating dm-regions for %s\n",
3183 dev_name(dev));
3184 return;
3185 }
3186
3187 region->start = entry->address_start;
3188 region->length = entry->address_end - entry->address_start;
3189 region->type = IOMMU_RESV_DIRECT;
3190 if (entry->prot & IOMMU_PROT_IR)
3191 region->prot |= IOMMU_READ;
3192 if (entry->prot & IOMMU_PROT_IW)
3193 region->prot |= IOMMU_WRITE;
3194
3195 list_add_tail(&region->list, head);
3196 }
3197 }
3198
3199 static void amd_iommu_put_resv_regions(struct device *dev,
3200 struct list_head *head)
3201 {
3202 struct iommu_resv_region *entry, *next;
3203
3204 list_for_each_entry_safe(entry, next, head, list)
3205 kfree(entry);
3206 }
3207
3208 static void amd_iommu_apply_resv_region(struct device *dev,
3209 struct iommu_domain *domain,
3210 struct iommu_resv_region *region)
3211 {
3212 struct dma_ops_domain *dma_dom = to_dma_ops_domain(to_pdomain(domain));
3213 unsigned long start, end;
3214
3215 start = IOVA_PFN(region->start);
3216 end = IOVA_PFN(region->start + region->length);
3217
3218 WARN_ON_ONCE(reserve_iova(&dma_dom->iovad, start, end) == NULL);
3219 }
3220
3221 static const struct iommu_ops amd_iommu_ops = {
3222 .capable = amd_iommu_capable,
3223 .domain_alloc = amd_iommu_domain_alloc,
3224 .domain_free = amd_iommu_domain_free,
3225 .attach_dev = amd_iommu_attach_device,
3226 .detach_dev = amd_iommu_detach_device,
3227 .map = amd_iommu_map,
3228 .unmap = amd_iommu_unmap,
3229 .map_sg = default_iommu_map_sg,
3230 .iova_to_phys = amd_iommu_iova_to_phys,
3231 .add_device = amd_iommu_add_device,
3232 .remove_device = amd_iommu_remove_device,
3233 .device_group = amd_iommu_device_group,
3234 .get_resv_regions = amd_iommu_get_resv_regions,
3235 .put_resv_regions = amd_iommu_put_resv_regions,
3236 .apply_resv_region = amd_iommu_apply_resv_region,
3237 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3238 };
3239
3240 /*****************************************************************************
3241 *
3242 * The next functions do a basic initialization of IOMMU for pass through
3243 * mode
3244 *
3245 * In passthrough mode the IOMMU is initialized and enabled but not used for
3246 * DMA-API translation.
3247 *
3248 *****************************************************************************/
3249
3250 /* IOMMUv2 specific functions */
3251 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3252 {
3253 return atomic_notifier_chain_register(&ppr_notifier, nb);
3254 }
3255 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3256
3257 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3258 {
3259 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3260 }
3261 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3262
3263 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3264 {
3265 struct protection_domain *domain = to_pdomain(dom);
3266 unsigned long flags;
3267
3268 spin_lock_irqsave(&domain->lock, flags);
3269
3270 /* Update data structure */
3271 domain->mode = PAGE_MODE_NONE;
3272 domain->updated = true;
3273
3274 /* Make changes visible to IOMMUs */
3275 update_domain(domain);
3276
3277 /* Page-table is not visible to IOMMU anymore, so free it */
3278 free_pagetable(domain);
3279
3280 spin_unlock_irqrestore(&domain->lock, flags);
3281 }
3282 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3283
3284 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3285 {
3286 struct protection_domain *domain = to_pdomain(dom);
3287 unsigned long flags;
3288 int levels, ret;
3289
3290 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3291 return -EINVAL;
3292
3293 /* Number of GCR3 table levels required */
3294 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3295 levels += 1;
3296
3297 if (levels > amd_iommu_max_glx_val)
3298 return -EINVAL;
3299
3300 spin_lock_irqsave(&domain->lock, flags);
3301
3302 /*
3303 * Save us all sanity checks whether devices already in the
3304 * domain support IOMMUv2. Just force that the domain has no
3305 * devices attached when it is switched into IOMMUv2 mode.
3306 */
3307 ret = -EBUSY;
3308 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3309 goto out;
3310
3311 ret = -ENOMEM;
3312 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3313 if (domain->gcr3_tbl == NULL)
3314 goto out;
3315
3316 domain->glx = levels;
3317 domain->flags |= PD_IOMMUV2_MASK;
3318 domain->updated = true;
3319
3320 update_domain(domain);
3321
3322 ret = 0;
3323
3324 out:
3325 spin_unlock_irqrestore(&domain->lock, flags);
3326
3327 return ret;
3328 }
3329 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3330
3331 static int __flush_pasid(struct protection_domain *domain, int pasid,
3332 u64 address, bool size)
3333 {
3334 struct iommu_dev_data *dev_data;
3335 struct iommu_cmd cmd;
3336 int i, ret;
3337
3338 if (!(domain->flags & PD_IOMMUV2_MASK))
3339 return -EINVAL;
3340
3341 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3342
3343 /*
3344 * IOMMU TLB needs to be flushed before Device TLB to
3345 * prevent device TLB refill from IOMMU TLB
3346 */
3347 for (i = 0; i < amd_iommus_present; ++i) {
3348 if (domain->dev_iommu[i] == 0)
3349 continue;
3350
3351 ret = iommu_queue_command(amd_iommus[i], &cmd);
3352 if (ret != 0)
3353 goto out;
3354 }
3355
3356 /* Wait until IOMMU TLB flushes are complete */
3357 domain_flush_complete(domain);
3358
3359 /* Now flush device TLBs */
3360 list_for_each_entry(dev_data, &domain->dev_list, list) {
3361 struct amd_iommu *iommu;
3362 int qdep;
3363
3364 /*
3365 There might be non-IOMMUv2 capable devices in an IOMMUv2
3366 * domain.
3367 */
3368 if (!dev_data->ats.enabled)
3369 continue;
3370
3371 qdep = dev_data->ats.qdep;
3372 iommu = amd_iommu_rlookup_table[dev_data->devid];
3373
3374 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3375 qdep, address, size);
3376
3377 ret = iommu_queue_command(iommu, &cmd);
3378 if (ret != 0)
3379 goto out;
3380 }
3381
3382 /* Wait until all device TLBs are flushed */
3383 domain_flush_complete(domain);
3384
3385 ret = 0;
3386
3387 out:
3388
3389 return ret;
3390 }
3391
3392 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3393 u64 address)
3394 {
3395 return __flush_pasid(domain, pasid, address, false);
3396 }
3397
3398 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3399 u64 address)
3400 {
3401 struct protection_domain *domain = to_pdomain(dom);
3402 unsigned long flags;
3403 int ret;
3404
3405 spin_lock_irqsave(&domain->lock, flags);
3406 ret = __amd_iommu_flush_page(domain, pasid, address);
3407 spin_unlock_irqrestore(&domain->lock, flags);
3408
3409 return ret;
3410 }
3411 EXPORT_SYMBOL(amd_iommu_flush_page);
3412
3413 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3414 {
3415 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3416 true);
3417 }
3418
3419 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3420 {
3421 struct protection_domain *domain = to_pdomain(dom);
3422 unsigned long flags;
3423 int ret;
3424
3425 spin_lock_irqsave(&domain->lock, flags);
3426 ret = __amd_iommu_flush_tlb(domain, pasid);
3427 spin_unlock_irqrestore(&domain->lock, flags);
3428
3429 return ret;
3430 }
3431 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3432
3433 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3434 {
3435 int index;
3436 u64 *pte;
3437
3438 while (true) {
3439
3440 index = (pasid >> (9 * level)) & 0x1ff;
3441 pte = &root[index];
3442
3443 if (level == 0)
3444 break;
3445
3446 if (!(*pte & GCR3_VALID)) {
3447 if (!alloc)
3448 return NULL;
3449
3450 root = (void *)get_zeroed_page(GFP_ATOMIC);
3451 if (root == NULL)
3452 return NULL;
3453
3454 *pte = __pa(root) | GCR3_VALID;
3455 }
3456
3457 root = __va(*pte & PAGE_MASK);
3458
3459 level -= 1;
3460 }
3461
3462 return pte;
3463 }
3464
3465 static int __set_gcr3(struct protection_domain *domain, int pasid,
3466 unsigned long cr3)
3467 {
3468 u64 *pte;
3469
3470 if (domain->mode != PAGE_MODE_NONE)
3471 return -EINVAL;
3472
3473 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3474 if (pte == NULL)
3475 return -ENOMEM;
3476
3477 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3478
3479 return __amd_iommu_flush_tlb(domain, pasid);
3480 }
3481
3482 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3483 {
3484 u64 *pte;
3485
3486 if (domain->mode != PAGE_MODE_NONE)
3487 return -EINVAL;
3488
3489 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3490 if (pte == NULL)
3491 return 0;
3492
3493 *pte = 0;
3494
3495 return __amd_iommu_flush_tlb(domain, pasid);
3496 }
3497
3498 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3499 unsigned long cr3)
3500 {
3501 struct protection_domain *domain = to_pdomain(dom);
3502 unsigned long flags;
3503 int ret;
3504
3505 spin_lock_irqsave(&domain->lock, flags);
3506 ret = __set_gcr3(domain, pasid, cr3);
3507 spin_unlock_irqrestore(&domain->lock, flags);
3508
3509 return ret;
3510 }
3511 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3512
3513 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3514 {
3515 struct protection_domain *domain = to_pdomain(dom);
3516 unsigned long flags;
3517 int ret;
3518
3519 spin_lock_irqsave(&domain->lock, flags);
3520 ret = __clear_gcr3(domain, pasid);
3521 spin_unlock_irqrestore(&domain->lock, flags);
3522
3523 return ret;
3524 }
3525 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3526
3527 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3528 int status, int tag)
3529 {
3530 struct iommu_dev_data *dev_data;
3531 struct amd_iommu *iommu;
3532 struct iommu_cmd cmd;
3533
3534 dev_data = get_dev_data(&pdev->dev);
3535 iommu = amd_iommu_rlookup_table[dev_data->devid];
3536
3537 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3538 tag, dev_data->pri_tlp);
3539
3540 return iommu_queue_command(iommu, &cmd);
3541 }
3542 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3543
3544 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3545 {
3546 struct protection_domain *pdomain;
3547
3548 pdomain = get_domain(&pdev->dev);
3549 if (IS_ERR(pdomain))
3550 return NULL;
3551
3552 /* Only return IOMMUv2 domains */
3553 if (!(pdomain->flags & PD_IOMMUV2_MASK))
3554 return NULL;
3555
3556 return &pdomain->domain;
3557 }
3558 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3559
3560 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3561 {
3562 struct iommu_dev_data *dev_data;
3563
3564 if (!amd_iommu_v2_supported())
3565 return;
3566
3567 dev_data = get_dev_data(&pdev->dev);
3568 dev_data->errata |= (1 << erratum);
3569 }
3570 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3571
3572 int amd_iommu_device_info(struct pci_dev *pdev,
3573 struct amd_iommu_device_info *info)
3574 {
3575 int max_pasids;
3576 int pos;
3577
3578 if (pdev == NULL || info == NULL)
3579 return -EINVAL;
3580
3581 if (!amd_iommu_v2_supported())
3582 return -EINVAL;
3583
3584 memset(info, 0, sizeof(*info));
3585
3586 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3587 if (pos)
3588 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3589
3590 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3591 if (pos)
3592 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3593
3594 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3595 if (pos) {
3596 int features;
3597
3598 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3599 max_pasids = min(max_pasids, (1 << 20));
3600
3601 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3602 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3603
3604 features = pci_pasid_features(pdev);
3605 if (features & PCI_PASID_CAP_EXEC)
3606 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3607 if (features & PCI_PASID_CAP_PRIV)
3608 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3609 }
3610
3611 return 0;
3612 }
3613 EXPORT_SYMBOL(amd_iommu_device_info);
3614
3615 #ifdef CONFIG_IRQ_REMAP
3616
3617 /*****************************************************************************
3618 *
3619 * Interrupt Remapping Implementation
3620 *
3621 *****************************************************************************/
3622
3623 static struct irq_chip amd_ir_chip;
3624
3625 #define DTE_IRQ_PHYS_ADDR_MASK (((1ULL << 45)-1) << 6)
3626 #define DTE_IRQ_REMAP_INTCTL (2ULL << 60)
3627 #define DTE_IRQ_TABLE_LEN (8ULL << 1)
3628 #define DTE_IRQ_REMAP_ENABLE 1ULL
3629
3630 static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3631 {
3632 u64 dte;
3633
3634 dte = amd_iommu_dev_table[devid].data[2];
3635 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
3636 dte |= virt_to_phys(table->table);
3637 dte |= DTE_IRQ_REMAP_INTCTL;
3638 dte |= DTE_IRQ_TABLE_LEN;
3639 dte |= DTE_IRQ_REMAP_ENABLE;
3640
3641 amd_iommu_dev_table[devid].data[2] = dte;
3642 }
3643
3644 static struct irq_remap_table *get_irq_table(u16 devid, bool ioapic)
3645 {
3646 struct irq_remap_table *table = NULL;
3647 struct amd_iommu *iommu;
3648 unsigned long flags;
3649 u16 alias;
3650
3651 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3652
3653 iommu = amd_iommu_rlookup_table[devid];
3654 if (!iommu)
3655 goto out_unlock;
3656
3657 table = irq_lookup_table[devid];
3658 if (table)
3659 goto out_unlock;
3660
3661 alias = amd_iommu_alias_table[devid];
3662 table = irq_lookup_table[alias];
3663 if (table) {
3664 irq_lookup_table[devid] = table;
3665 set_dte_irq_entry(devid, table);
3666 iommu_flush_dte(iommu, devid);
3667 goto out;
3668 }
3669
3670 /* Nothing there yet, allocate new irq remapping table */
3671 table = kzalloc(sizeof(*table), GFP_ATOMIC);
3672 if (!table)
3673 goto out_unlock;
3674
3675 /* Initialize table spin-lock */
3676 spin_lock_init(&table->lock);
3677
3678 if (ioapic)
3679 /* Keep the first 32 indexes free for IOAPIC interrupts */
3680 table->min_index = 32;
3681
3682 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_ATOMIC);
3683 if (!table->table) {
3684 kfree(table);
3685 table = NULL;
3686 goto out_unlock;
3687 }
3688
3689 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
3690 memset(table->table, 0,
3691 MAX_IRQS_PER_TABLE * sizeof(u32));
3692 else
3693 memset(table->table, 0,
3694 (MAX_IRQS_PER_TABLE * (sizeof(u64) * 2)));
3695
3696 if (ioapic) {
3697 int i;
3698
3699 for (i = 0; i < 32; ++i)
3700 iommu->irte_ops->set_allocated(table, i);
3701 }
3702
3703 irq_lookup_table[devid] = table;
3704 set_dte_irq_entry(devid, table);
3705 iommu_flush_dte(iommu, devid);
3706 if (devid != alias) {
3707 irq_lookup_table[alias] = table;
3708 set_dte_irq_entry(alias, table);
3709 iommu_flush_dte(iommu, alias);
3710 }
3711
3712 out:
3713 iommu_completion_wait(iommu);
3714
3715 out_unlock:
3716 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3717
3718 return table;
3719 }
3720
3721 static int alloc_irq_index(u16 devid, int count)
3722 {
3723 struct irq_remap_table *table;
3724 unsigned long flags;
3725 int index, c;
3726 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
3727
3728 if (!iommu)
3729 return -ENODEV;
3730
3731 table = get_irq_table(devid, false);
3732 if (!table)
3733 return -ENODEV;
3734
3735 spin_lock_irqsave(&table->lock, flags);
3736
3737 /* Scan table for free entries */
3738 for (c = 0, index = table->min_index;
3739 index < MAX_IRQS_PER_TABLE;
3740 ++index) {
3741 if (!iommu->irte_ops->is_allocated(table, index))
3742 c += 1;
3743 else
3744 c = 0;
3745
3746 if (c == count) {
3747 for (; c != 0; --c)
3748 iommu->irte_ops->set_allocated(table, index - c + 1);
3749
3750 index -= count - 1;
3751 goto out;
3752 }
3753 }
3754
3755 index = -ENOSPC;
3756
3757 out:
3758 spin_unlock_irqrestore(&table->lock, flags);
3759
3760 return index;
3761 }
3762
3763 static int modify_irte_ga(u16 devid, int index, struct irte_ga *irte,
3764 struct amd_ir_data *data)
3765 {
3766 struct irq_remap_table *table;
3767 struct amd_iommu *iommu;
3768 unsigned long flags;
3769 struct irte_ga *entry;
3770
3771 iommu = amd_iommu_rlookup_table[devid];
3772 if (iommu == NULL)
3773 return -EINVAL;
3774
3775 table = get_irq_table(devid, false);
3776 if (!table)
3777 return -ENOMEM;
3778
3779 spin_lock_irqsave(&table->lock, flags);
3780
3781 entry = (struct irte_ga *)table->table;
3782 entry = &entry[index];
3783 entry->lo.fields_remap.valid = 0;
3784 entry->hi.val = irte->hi.val;
3785 entry->lo.val = irte->lo.val;
3786 entry->lo.fields_remap.valid = 1;
3787 if (data)
3788 data->ref = entry;
3789
3790 spin_unlock_irqrestore(&table->lock, flags);
3791
3792 iommu_flush_irt(iommu, devid);
3793 iommu_completion_wait(iommu);
3794
3795 return 0;
3796 }
3797
3798 static int modify_irte(u16 devid, int index, union irte *irte)
3799 {
3800 struct irq_remap_table *table;
3801 struct amd_iommu *iommu;
3802 unsigned long flags;
3803
3804 iommu = amd_iommu_rlookup_table[devid];
3805 if (iommu == NULL)
3806 return -EINVAL;
3807
3808 table = get_irq_table(devid, false);
3809 if (!table)
3810 return -ENOMEM;
3811
3812 spin_lock_irqsave(&table->lock, flags);
3813 table->table[index] = irte->val;
3814 spin_unlock_irqrestore(&table->lock, flags);
3815
3816 iommu_flush_irt(iommu, devid);
3817 iommu_completion_wait(iommu);
3818
3819 return 0;
3820 }
3821
3822 static void free_irte(u16 devid, int index)
3823 {
3824 struct irq_remap_table *table;
3825 struct amd_iommu *iommu;
3826 unsigned long flags;
3827
3828 iommu = amd_iommu_rlookup_table[devid];
3829 if (iommu == NULL)
3830 return;
3831
3832 table = get_irq_table(devid, false);
3833 if (!table)
3834 return;
3835
3836 spin_lock_irqsave(&table->lock, flags);
3837 iommu->irte_ops->clear_allocated(table, index);
3838 spin_unlock_irqrestore(&table->lock, flags);
3839
3840 iommu_flush_irt(iommu, devid);
3841 iommu_completion_wait(iommu);
3842 }
3843
3844 static void irte_prepare(void *entry,
3845 u32 delivery_mode, u32 dest_mode,
3846 u8 vector, u32 dest_apicid, int devid)
3847 {
3848 union irte *irte = (union irte *) entry;
3849
3850 irte->val = 0;
3851 irte->fields.vector = vector;
3852 irte->fields.int_type = delivery_mode;
3853 irte->fields.destination = dest_apicid;
3854 irte->fields.dm = dest_mode;
3855 irte->fields.valid = 1;
3856 }
3857
3858 static void irte_ga_prepare(void *entry,
3859 u32 delivery_mode, u32 dest_mode,
3860 u8 vector, u32 dest_apicid, int devid)
3861 {
3862 struct irte_ga *irte = (struct irte_ga *) entry;
3863 struct iommu_dev_data *dev_data = search_dev_data(devid);
3864
3865 irte->lo.val = 0;
3866 irte->hi.val = 0;
3867 irte->lo.fields_remap.guest_mode = dev_data ? dev_data->use_vapic : 0;
3868 irte->lo.fields_remap.int_type = delivery_mode;
3869 irte->lo.fields_remap.dm = dest_mode;
3870 irte->hi.fields.vector = vector;
3871 irte->lo.fields_remap.destination = dest_apicid;
3872 irte->lo.fields_remap.valid = 1;
3873 }
3874
3875 static void irte_activate(void *entry, u16 devid, u16 index)
3876 {
3877 union irte *irte = (union irte *) entry;
3878
3879 irte->fields.valid = 1;
3880 modify_irte(devid, index, irte);
3881 }
3882
3883 static void irte_ga_activate(void *entry, u16 devid, u16 index)
3884 {
3885 struct irte_ga *irte = (struct irte_ga *) entry;
3886
3887 irte->lo.fields_remap.valid = 1;
3888 modify_irte_ga(devid, index, irte, NULL);
3889 }
3890
3891 static void irte_deactivate(void *entry, u16 devid, u16 index)
3892 {
3893 union irte *irte = (union irte *) entry;
3894
3895 irte->fields.valid = 0;
3896 modify_irte(devid, index, irte);
3897 }
3898
3899 static void irte_ga_deactivate(void *entry, u16 devid, u16 index)
3900 {
3901 struct irte_ga *irte = (struct irte_ga *) entry;
3902
3903 irte->lo.fields_remap.valid = 0;
3904 modify_irte_ga(devid, index, irte, NULL);
3905 }
3906
3907 static void irte_set_affinity(void *entry, u16 devid, u16 index,
3908 u8 vector, u32 dest_apicid)
3909 {
3910 union irte *irte = (union irte *) entry;
3911
3912 irte->fields.vector = vector;
3913 irte->fields.destination = dest_apicid;
3914 modify_irte(devid, index, irte);
3915 }
3916
3917 static void irte_ga_set_affinity(void *entry, u16 devid, u16 index,
3918 u8 vector, u32 dest_apicid)
3919 {
3920 struct irte_ga *irte = (struct irte_ga *) entry;
3921 struct iommu_dev_data *dev_data = search_dev_data(devid);
3922
3923 if (!dev_data || !dev_data->use_vapic) {
3924 irte->hi.fields.vector = vector;
3925 irte->lo.fields_remap.destination = dest_apicid;
3926 irte->lo.fields_remap.guest_mode = 0;
3927 modify_irte_ga(devid, index, irte, NULL);
3928 }
3929 }
3930
3931 #define IRTE_ALLOCATED (~1U)
3932 static void irte_set_allocated(struct irq_remap_table *table, int index)
3933 {
3934 table->table[index] = IRTE_ALLOCATED;
3935 }
3936
3937 static void irte_ga_set_allocated(struct irq_remap_table *table, int index)
3938 {
3939 struct irte_ga *ptr = (struct irte_ga *)table->table;
3940 struct irte_ga *irte = &ptr[index];
3941
3942 memset(&irte->lo.val, 0, sizeof(u64));
3943 memset(&irte->hi.val, 0, sizeof(u64));
3944 irte->hi.fields.vector = 0xff;
3945 }
3946
3947 static bool irte_is_allocated(struct irq_remap_table *table, int index)
3948 {
3949 union irte *ptr = (union irte *)table->table;
3950 union irte *irte = &ptr[index];
3951
3952 return irte->val != 0;
3953 }
3954
3955 static bool irte_ga_is_allocated(struct irq_remap_table *table, int index)
3956 {
3957 struct irte_ga *ptr = (struct irte_ga *)table->table;
3958 struct irte_ga *irte = &ptr[index];
3959
3960 return irte->hi.fields.vector != 0;
3961 }
3962
3963 static void irte_clear_allocated(struct irq_remap_table *table, int index)
3964 {
3965 table->table[index] = 0;
3966 }
3967
3968 static void irte_ga_clear_allocated(struct irq_remap_table *table, int index)
3969 {
3970 struct irte_ga *ptr = (struct irte_ga *)table->table;
3971 struct irte_ga *irte = &ptr[index];
3972
3973 memset(&irte->lo.val, 0, sizeof(u64));
3974 memset(&irte->hi.val, 0, sizeof(u64));
3975 }
3976
3977 static int get_devid(struct irq_alloc_info *info)
3978 {
3979 int devid = -1;
3980
3981 switch (info->type) {
3982 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3983 devid = get_ioapic_devid(info->ioapic_id);
3984 break;
3985 case X86_IRQ_ALLOC_TYPE_HPET:
3986 devid = get_hpet_devid(info->hpet_id);
3987 break;
3988 case X86_IRQ_ALLOC_TYPE_MSI:
3989 case X86_IRQ_ALLOC_TYPE_MSIX:
3990 devid = get_device_id(&info->msi_dev->dev);
3991 break;
3992 default:
3993 BUG_ON(1);
3994 break;
3995 }
3996
3997 return devid;
3998 }
3999
4000 static struct irq_domain *get_ir_irq_domain(struct irq_alloc_info *info)
4001 {
4002 struct amd_iommu *iommu;
4003 int devid;
4004
4005 if (!info)
4006 return NULL;
4007
4008 devid = get_devid(info);
4009 if (devid >= 0) {
4010 iommu = amd_iommu_rlookup_table[devid];
4011 if (iommu)
4012 return iommu->ir_domain;
4013 }
4014
4015 return NULL;
4016 }
4017
4018 static struct irq_domain *get_irq_domain(struct irq_alloc_info *info)
4019 {
4020 struct amd_iommu *iommu;
4021 int devid;
4022
4023 if (!info)
4024 return NULL;
4025
4026 switch (info->type) {
4027 case X86_IRQ_ALLOC_TYPE_MSI:
4028 case X86_IRQ_ALLOC_TYPE_MSIX:
4029 devid = get_device_id(&info->msi_dev->dev);
4030 if (devid < 0)
4031 return NULL;
4032
4033 iommu = amd_iommu_rlookup_table[devid];
4034 if (iommu)
4035 return iommu->msi_domain;
4036 break;
4037 default:
4038 break;
4039 }
4040
4041 return NULL;
4042 }
4043
4044 struct irq_remap_ops amd_iommu_irq_ops = {
4045 .prepare = amd_iommu_prepare,
4046 .enable = amd_iommu_enable,
4047 .disable = amd_iommu_disable,
4048 .reenable = amd_iommu_reenable,
4049 .enable_faulting = amd_iommu_enable_faulting,
4050 .get_ir_irq_domain = get_ir_irq_domain,
4051 .get_irq_domain = get_irq_domain,
4052 };
4053
4054 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
4055 struct irq_cfg *irq_cfg,
4056 struct irq_alloc_info *info,
4057 int devid, int index, int sub_handle)
4058 {
4059 struct irq_2_irte *irte_info = &data->irq_2_irte;
4060 struct msi_msg *msg = &data->msi_entry;
4061 struct IO_APIC_route_entry *entry;
4062 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
4063
4064 if (!iommu)
4065 return;
4066
4067 data->irq_2_irte.devid = devid;
4068 data->irq_2_irte.index = index + sub_handle;
4069 iommu->irte_ops->prepare(data->entry, apic->irq_delivery_mode,
4070 apic->irq_dest_mode, irq_cfg->vector,
4071 irq_cfg->dest_apicid, devid);
4072
4073 switch (info->type) {
4074 case X86_IRQ_ALLOC_TYPE_IOAPIC:
4075 /* Setup IOAPIC entry */
4076 entry = info->ioapic_entry;
4077 info->ioapic_entry = NULL;
4078 memset(entry, 0, sizeof(*entry));
4079 entry->vector = index;
4080 entry->mask = 0;
4081 entry->trigger = info->ioapic_trigger;
4082 entry->polarity = info->ioapic_polarity;
4083 /* Mask level triggered irqs. */
4084 if (info->ioapic_trigger)
4085 entry->mask = 1;
4086 break;
4087
4088 case X86_IRQ_ALLOC_TYPE_HPET:
4089 case X86_IRQ_ALLOC_TYPE_MSI:
4090 case X86_IRQ_ALLOC_TYPE_MSIX:
4091 msg->address_hi = MSI_ADDR_BASE_HI;
4092 msg->address_lo = MSI_ADDR_BASE_LO;
4093 msg->data = irte_info->index;
4094 break;
4095
4096 default:
4097 BUG_ON(1);
4098 break;
4099 }
4100 }
4101
4102 struct amd_irte_ops irte_32_ops = {
4103 .prepare = irte_prepare,
4104 .activate = irte_activate,
4105 .deactivate = irte_deactivate,
4106 .set_affinity = irte_set_affinity,
4107 .set_allocated = irte_set_allocated,
4108 .is_allocated = irte_is_allocated,
4109 .clear_allocated = irte_clear_allocated,
4110 };
4111
4112 struct amd_irte_ops irte_128_ops = {
4113 .prepare = irte_ga_prepare,
4114 .activate = irte_ga_activate,
4115 .deactivate = irte_ga_deactivate,
4116 .set_affinity = irte_ga_set_affinity,
4117 .set_allocated = irte_ga_set_allocated,
4118 .is_allocated = irte_ga_is_allocated,
4119 .clear_allocated = irte_ga_clear_allocated,
4120 };
4121
4122 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
4123 unsigned int nr_irqs, void *arg)
4124 {
4125 struct irq_alloc_info *info = arg;
4126 struct irq_data *irq_data;
4127 struct amd_ir_data *data = NULL;
4128 struct irq_cfg *cfg;
4129 int i, ret, devid;
4130 int index = -1;
4131
4132 if (!info)
4133 return -EINVAL;
4134 if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_MSI &&
4135 info->type != X86_IRQ_ALLOC_TYPE_MSIX)
4136 return -EINVAL;
4137
4138 /*
4139 * With IRQ remapping enabled, don't need contiguous CPU vectors
4140 * to support multiple MSI interrupts.
4141 */
4142 if (info->type == X86_IRQ_ALLOC_TYPE_MSI)
4143 info->flags &= ~X86_IRQ_ALLOC_CONTIGUOUS_VECTORS;
4144
4145 devid = get_devid(info);
4146 if (devid < 0)
4147 return -EINVAL;
4148
4149 ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
4150 if (ret < 0)
4151 return ret;
4152
4153 if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
4154 if (get_irq_table(devid, true))
4155 index = info->ioapic_pin;
4156 else
4157 ret = -ENOMEM;
4158 } else {
4159 index = alloc_irq_index(devid, nr_irqs);
4160 }
4161 if (index < 0) {
4162 pr_warn("Failed to allocate IRTE\n");
4163 ret = index;
4164 goto out_free_parent;
4165 }
4166
4167 for (i = 0; i < nr_irqs; i++) {
4168 irq_data = irq_domain_get_irq_data(domain, virq + i);
4169 cfg = irqd_cfg(irq_data);
4170 if (!irq_data || !cfg) {
4171 ret = -EINVAL;
4172 goto out_free_data;
4173 }
4174
4175 ret = -ENOMEM;
4176 data = kzalloc(sizeof(*data), GFP_KERNEL);
4177 if (!data)
4178 goto out_free_data;
4179
4180 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
4181 data->entry = kzalloc(sizeof(union irte), GFP_KERNEL);
4182 else
4183 data->entry = kzalloc(sizeof(struct irte_ga),
4184 GFP_KERNEL);
4185 if (!data->entry) {
4186 kfree(data);
4187 goto out_free_data;
4188 }
4189
4190 irq_data->hwirq = (devid << 16) + i;
4191 irq_data->chip_data = data;
4192 irq_data->chip = &amd_ir_chip;
4193 irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
4194 irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
4195 }
4196
4197 return 0;
4198
4199 out_free_data:
4200 for (i--; i >= 0; i--) {
4201 irq_data = irq_domain_get_irq_data(domain, virq + i);
4202 if (irq_data)
4203 kfree(irq_data->chip_data);
4204 }
4205 for (i = 0; i < nr_irqs; i++)
4206 free_irte(devid, index + i);
4207 out_free_parent:
4208 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4209 return ret;
4210 }
4211
4212 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
4213 unsigned int nr_irqs)
4214 {
4215 struct irq_2_irte *irte_info;
4216 struct irq_data *irq_data;
4217 struct amd_ir_data *data;
4218 int i;
4219
4220 for (i = 0; i < nr_irqs; i++) {
4221 irq_data = irq_domain_get_irq_data(domain, virq + i);
4222 if (irq_data && irq_data->chip_data) {
4223 data = irq_data->chip_data;
4224 irte_info = &data->irq_2_irte;
4225 free_irte(irte_info->devid, irte_info->index);
4226 kfree(data->entry);
4227 kfree(data);
4228 }
4229 }
4230 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4231 }
4232
4233 static void irq_remapping_activate(struct irq_domain *domain,
4234 struct irq_data *irq_data)
4235 {
4236 struct amd_ir_data *data = irq_data->chip_data;
4237 struct irq_2_irte *irte_info = &data->irq_2_irte;
4238 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4239
4240 if (iommu)
4241 iommu->irte_ops->activate(data->entry, irte_info->devid,
4242 irte_info->index);
4243 }
4244
4245 static void irq_remapping_deactivate(struct irq_domain *domain,
4246 struct irq_data *irq_data)
4247 {
4248 struct amd_ir_data *data = irq_data->chip_data;
4249 struct irq_2_irte *irte_info = &data->irq_2_irte;
4250 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4251
4252 if (iommu)
4253 iommu->irte_ops->deactivate(data->entry, irte_info->devid,
4254 irte_info->index);
4255 }
4256
4257 static struct irq_domain_ops amd_ir_domain_ops = {
4258 .alloc = irq_remapping_alloc,
4259 .free = irq_remapping_free,
4260 .activate = irq_remapping_activate,
4261 .deactivate = irq_remapping_deactivate,
4262 };
4263
4264 static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *vcpu_info)
4265 {
4266 struct amd_iommu *iommu;
4267 struct amd_iommu_pi_data *pi_data = vcpu_info;
4268 struct vcpu_data *vcpu_pi_info = pi_data->vcpu_data;
4269 struct amd_ir_data *ir_data = data->chip_data;
4270 struct irte_ga *irte = (struct irte_ga *) ir_data->entry;
4271 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4272 struct iommu_dev_data *dev_data = search_dev_data(irte_info->devid);
4273
4274 /* Note:
4275 * This device has never been set up for guest mode.
4276 * we should not modify the IRTE
4277 */
4278 if (!dev_data || !dev_data->use_vapic)
4279 return 0;
4280
4281 pi_data->ir_data = ir_data;
4282
4283 /* Note:
4284 * SVM tries to set up for VAPIC mode, but we are in
4285 * legacy mode. So, we force legacy mode instead.
4286 */
4287 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
4288 pr_debug("AMD-Vi: %s: Fall back to using intr legacy remap\n",
4289 __func__);
4290 pi_data->is_guest_mode = false;
4291 }
4292
4293 iommu = amd_iommu_rlookup_table[irte_info->devid];
4294 if (iommu == NULL)
4295 return -EINVAL;
4296
4297 pi_data->prev_ga_tag = ir_data->cached_ga_tag;
4298 if (pi_data->is_guest_mode) {
4299 /* Setting */
4300 irte->hi.fields.ga_root_ptr = (pi_data->base >> 12);
4301 irte->hi.fields.vector = vcpu_pi_info->vector;
4302 irte->lo.fields_vapic.guest_mode = 1;
4303 irte->lo.fields_vapic.ga_tag = pi_data->ga_tag;
4304
4305 ir_data->cached_ga_tag = pi_data->ga_tag;
4306 } else {
4307 /* Un-Setting */
4308 struct irq_cfg *cfg = irqd_cfg(data);
4309
4310 irte->hi.val = 0;
4311 irte->lo.val = 0;
4312 irte->hi.fields.vector = cfg->vector;
4313 irte->lo.fields_remap.guest_mode = 0;
4314 irte->lo.fields_remap.destination = cfg->dest_apicid;
4315 irte->lo.fields_remap.int_type = apic->irq_delivery_mode;
4316 irte->lo.fields_remap.dm = apic->irq_dest_mode;
4317
4318 /*
4319 * This communicates the ga_tag back to the caller
4320 * so that it can do all the necessary clean up.
4321 */
4322 ir_data->cached_ga_tag = 0;
4323 }
4324
4325 return modify_irte_ga(irte_info->devid, irte_info->index, irte, ir_data);
4326 }
4327
4328 static int amd_ir_set_affinity(struct irq_data *data,
4329 const struct cpumask *mask, bool force)
4330 {
4331 struct amd_ir_data *ir_data = data->chip_data;
4332 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4333 struct irq_cfg *cfg = irqd_cfg(data);
4334 struct irq_data *parent = data->parent_data;
4335 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4336 int ret;
4337
4338 if (!iommu)
4339 return -ENODEV;
4340
4341 ret = parent->chip->irq_set_affinity(parent, mask, force);
4342 if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
4343 return ret;
4344
4345 /*
4346 * Atomically updates the IRTE with the new destination, vector
4347 * and flushes the interrupt entry cache.
4348 */
4349 iommu->irte_ops->set_affinity(ir_data->entry, irte_info->devid,
4350 irte_info->index, cfg->vector, cfg->dest_apicid);
4351
4352 /*
4353 * After this point, all the interrupts will start arriving
4354 * at the new destination. So, time to cleanup the previous
4355 * vector allocation.
4356 */
4357 send_cleanup_vector(cfg);
4358
4359 return IRQ_SET_MASK_OK_DONE;
4360 }
4361
4362 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
4363 {
4364 struct amd_ir_data *ir_data = irq_data->chip_data;
4365
4366 *msg = ir_data->msi_entry;
4367 }
4368
4369 static struct irq_chip amd_ir_chip = {
4370 .irq_ack = ir_ack_apic_edge,
4371 .irq_set_affinity = amd_ir_set_affinity,
4372 .irq_set_vcpu_affinity = amd_ir_set_vcpu_affinity,
4373 .irq_compose_msi_msg = ir_compose_msi_msg,
4374 };
4375
4376 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
4377 {
4378 iommu->ir_domain = irq_domain_add_tree(NULL, &amd_ir_domain_ops, iommu);
4379 if (!iommu->ir_domain)
4380 return -ENOMEM;
4381
4382 iommu->ir_domain->parent = arch_get_ir_parent_domain();
4383 iommu->msi_domain = arch_create_msi_irq_domain(iommu->ir_domain);
4384
4385 return 0;
4386 }
4387
4388 int amd_iommu_update_ga(int cpu, bool is_run, void *data)
4389 {
4390 unsigned long flags;
4391 struct amd_iommu *iommu;
4392 struct irq_remap_table *irt;
4393 struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
4394 int devid = ir_data->irq_2_irte.devid;
4395 struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
4396 struct irte_ga *ref = (struct irte_ga *) ir_data->ref;
4397
4398 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
4399 !ref || !entry || !entry->lo.fields_vapic.guest_mode)
4400 return 0;
4401
4402 iommu = amd_iommu_rlookup_table[devid];
4403 if (!iommu)
4404 return -ENODEV;
4405
4406 irt = get_irq_table(devid, false);
4407 if (!irt)
4408 return -ENODEV;
4409
4410 spin_lock_irqsave(&irt->lock, flags);
4411
4412 if (ref->lo.fields_vapic.guest_mode) {
4413 if (cpu >= 0)
4414 ref->lo.fields_vapic.destination = cpu;
4415 ref->lo.fields_vapic.is_run = is_run;
4416 barrier();
4417 }
4418
4419 spin_unlock_irqrestore(&irt->lock, flags);
4420
4421 iommu_flush_irt(iommu, devid);
4422 iommu_completion_wait(iommu);
4423 return 0;
4424 }
4425 EXPORT_SYMBOL(amd_iommu_update_ga);
4426 #endif
ubuntu-zesty-kernel mirror