projects / deliverable / linux.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
mm, page_alloc: distinguish between being unable to sleep, unwilling to sleep and...
[deliverable/linux.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/pci-ats.h>
23 #include <linux/bitmap.h>
24 #include <linux/slab.h>
25 #include <linux/debugfs.h>
26 #include <linux/scatterlist.h>
27 #include <linux/dma-mapping.h>
28 #include <linux/iommu-helper.h>
29 #include <linux/iommu.h>
30 #include <linux/delay.h>
31 #include <linux/amd-iommu.h>
32 #include <linux/notifier.h>
33 #include <linux/export.h>
34 #include <linux/irq.h>
35 #include <linux/msi.h>
36 #include <linux/dma-contiguous.h>
37 #include <linux/irqdomain.h>
38 #include <asm/irq_remapping.h>
39 #include <asm/io_apic.h>
40 #include <asm/apic.h>
41 #include <asm/hw_irq.h>
42 #include <asm/msidef.h>
43 #include <asm/proto.h>
44 #include <asm/iommu.h>
45 #include <asm/gart.h>
46 #include <asm/dma.h>
47
48 #include "amd_iommu_proto.h"
49 #include "amd_iommu_types.h"
50 #include "irq_remapping.h"
51
52 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
53
54 #define LOOP_TIMEOUT 100000
55
56 /*
57 * This bitmap is used to advertise the page sizes our hardware support
58 * to the IOMMU core, which will then use this information to split
59 * physically contiguous memory regions it is mapping into page sizes
60 * that we support.
61 *
62 * 512GB Pages are not supported due to a hardware bug
63 */
64 #define AMD_IOMMU_PGSIZES ((~0xFFFUL) & ~(2ULL << 38))
65
66 static DEFINE_RWLOCK(amd_iommu_devtable_lock);
67
68 /* List of all available dev_data structures */
69 static LIST_HEAD(dev_data_list);
70 static DEFINE_SPINLOCK(dev_data_list_lock);
71
72 LIST_HEAD(ioapic_map);
73 LIST_HEAD(hpet_map);
74
75 /*
76 * Domain for untranslated devices - only allocated
77 * if iommu=pt passed on kernel cmd line.
78 */
79 static const struct iommu_ops amd_iommu_ops;
80
81 static ATOMIC_NOTIFIER_HEAD(ppr_notifier);
82 int amd_iommu_max_glx_val = -1;
83
84 static struct dma_map_ops amd_iommu_dma_ops;
85
86 /*
87 * This struct contains device specific data for the IOMMU
88 */
89 struct iommu_dev_data {
90 struct list_head list; /* For domain->dev_list */
91 struct list_head dev_data_list; /* For global dev_data_list */
92 struct protection_domain *domain; /* Domain the device is bound to */
93 u16 devid; /* PCI Device ID */
94 bool iommu_v2; /* Device can make use of IOMMUv2 */
95 bool passthrough; /* Device is identity mapped */
96 struct {
97 bool enabled;
98 int qdep;
99 } ats; /* ATS state */
100 bool pri_tlp; /* PASID TLB required for
101 PPR completions */
102 u32 errata; /* Bitmap for errata to apply */
103 };
104
105 /*
106 * general struct to manage commands send to an IOMMU
107 */
108 struct iommu_cmd {
109 u32 data[4];
110 };
111
112 struct kmem_cache *amd_iommu_irq_cache;
113
114 static void update_domain(struct protection_domain *domain);
115 static int protection_domain_init(struct protection_domain *domain);
116
117 /****************************************************************************
118 *
119 * Helper functions
120 *
121 ****************************************************************************/
122
123 static struct protection_domain *to_pdomain(struct iommu_domain *dom)
124 {
125 return container_of(dom, struct protection_domain, domain);
126 }
127
128 static struct iommu_dev_data *alloc_dev_data(u16 devid)
129 {
130 struct iommu_dev_data *dev_data;
131 unsigned long flags;
132
133 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
134 if (!dev_data)
135 return NULL;
136
137 dev_data->devid = devid;
138
139 spin_lock_irqsave(&dev_data_list_lock, flags);
140 list_add_tail(&dev_data->dev_data_list, &dev_data_list);
141 spin_unlock_irqrestore(&dev_data_list_lock, flags);
142
143 return dev_data;
144 }
145
146 static struct iommu_dev_data *search_dev_data(u16 devid)
147 {
148 struct iommu_dev_data *dev_data;
149 unsigned long flags;
150
151 spin_lock_irqsave(&dev_data_list_lock, flags);
152 list_for_each_entry(dev_data, &dev_data_list, dev_data_list) {
153 if (dev_data->devid == devid)
154 goto out_unlock;
155 }
156
157 dev_data = NULL;
158
159 out_unlock:
160 spin_unlock_irqrestore(&dev_data_list_lock, flags);
161
162 return dev_data;
163 }
164
165 static struct iommu_dev_data *find_dev_data(u16 devid)
166 {
167 struct iommu_dev_data *dev_data;
168
169 dev_data = search_dev_data(devid);
170
171 if (dev_data == NULL)
172 dev_data = alloc_dev_data(devid);
173
174 return dev_data;
175 }
176
177 static inline u16 get_device_id(struct device *dev)
178 {
179 struct pci_dev *pdev = to_pci_dev(dev);
180
181 return PCI_DEVID(pdev->bus->number, pdev->devfn);
182 }
183
184 static struct iommu_dev_data *get_dev_data(struct device *dev)
185 {
186 return dev->archdata.iommu;
187 }
188
189 static bool pci_iommuv2_capable(struct pci_dev *pdev)
190 {
191 static const int caps[] = {
192 PCI_EXT_CAP_ID_ATS,
193 PCI_EXT_CAP_ID_PRI,
194 PCI_EXT_CAP_ID_PASID,
195 };
196 int i, pos;
197
198 for (i = 0; i < 3; ++i) {
199 pos = pci_find_ext_capability(pdev, caps[i]);
200 if (pos == 0)
201 return false;
202 }
203
204 return true;
205 }
206
207 static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
208 {
209 struct iommu_dev_data *dev_data;
210
211 dev_data = get_dev_data(&pdev->dev);
212
213 return dev_data->errata & (1 << erratum) ? true : false;
214 }
215
216 /*
217 * This function actually applies the mapping to the page table of the
218 * dma_ops domain.
219 */
220 static void alloc_unity_mapping(struct dma_ops_domain *dma_dom,
221 struct unity_map_entry *e)
222 {
223 u64 addr;
224
225 for (addr = e->address_start; addr < e->address_end;
226 addr += PAGE_SIZE) {
227 if (addr < dma_dom->aperture_size)
228 __set_bit(addr >> PAGE_SHIFT,
229 dma_dom->aperture[0]->bitmap);
230 }
231 }
232
233 /*
234 * Inits the unity mappings required for a specific device
235 */
236 static void init_unity_mappings_for_device(struct device *dev,
237 struct dma_ops_domain *dma_dom)
238 {
239 struct unity_map_entry *e;
240 u16 devid;
241
242 devid = get_device_id(dev);
243
244 list_for_each_entry(e, &amd_iommu_unity_map, list) {
245 if (!(devid >= e->devid_start && devid <= e->devid_end))
246 continue;
247 alloc_unity_mapping(dma_dom, e);
248 }
249 }
250
251 /*
252 * This function checks if the driver got a valid device from the caller to
253 * avoid dereferencing invalid pointers.
254 */
255 static bool check_device(struct device *dev)
256 {
257 u16 devid;
258
259 if (!dev || !dev->dma_mask)
260 return false;
261
262 /* No PCI device */
263 if (!dev_is_pci(dev))
264 return false;
265
266 devid = get_device_id(dev);
267
268 /* Out of our scope? */
269 if (devid > amd_iommu_last_bdf)
270 return false;
271
272 if (amd_iommu_rlookup_table[devid] == NULL)
273 return false;
274
275 return true;
276 }
277
278 static void init_iommu_group(struct device *dev)
279 {
280 struct dma_ops_domain *dma_domain;
281 struct iommu_domain *domain;
282 struct iommu_group *group;
283
284 group = iommu_group_get_for_dev(dev);
285 if (IS_ERR(group))
286 return;
287
288 domain = iommu_group_default_domain(group);
289 if (!domain)
290 goto out;
291
292 dma_domain = to_pdomain(domain)->priv;
293
294 init_unity_mappings_for_device(dev, dma_domain);
295 out:
296 iommu_group_put(group);
297 }
298
299 static int iommu_init_device(struct device *dev)
300 {
301 struct pci_dev *pdev = to_pci_dev(dev);
302 struct iommu_dev_data *dev_data;
303
304 if (dev->archdata.iommu)
305 return 0;
306
307 dev_data = find_dev_data(get_device_id(dev));
308 if (!dev_data)
309 return -ENOMEM;
310
311 if (pci_iommuv2_capable(pdev)) {
312 struct amd_iommu *iommu;
313
314 iommu = amd_iommu_rlookup_table[dev_data->devid];
315 dev_data->iommu_v2 = iommu->is_iommu_v2;
316 }
317
318 dev->archdata.iommu = dev_data;
319
320 iommu_device_link(amd_iommu_rlookup_table[dev_data->devid]->iommu_dev,
321 dev);
322
323 return 0;
324 }
325
326 static void iommu_ignore_device(struct device *dev)
327 {
328 u16 devid, alias;
329
330 devid = get_device_id(dev);
331 alias = amd_iommu_alias_table[devid];
332
333 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
334 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
335
336 amd_iommu_rlookup_table[devid] = NULL;
337 amd_iommu_rlookup_table[alias] = NULL;
338 }
339
340 static void iommu_uninit_device(struct device *dev)
341 {
342 struct iommu_dev_data *dev_data = search_dev_data(get_device_id(dev));
343
344 if (!dev_data)
345 return;
346
347 iommu_device_unlink(amd_iommu_rlookup_table[dev_data->devid]->iommu_dev,
348 dev);
349
350 iommu_group_remove_device(dev);
351
352 /* Remove dma-ops */
353 dev->archdata.dma_ops = NULL;
354
355 /*
356 * We keep dev_data around for unplugged devices and reuse it when the
357 * device is re-plugged - not doing so would introduce a ton of races.
358 */
359 }
360
361 #ifdef CONFIG_AMD_IOMMU_STATS
362
363 /*
364 * Initialization code for statistics collection
365 */
366
367 DECLARE_STATS_COUNTER(compl_wait);
368 DECLARE_STATS_COUNTER(cnt_map_single);
369 DECLARE_STATS_COUNTER(cnt_unmap_single);
370 DECLARE_STATS_COUNTER(cnt_map_sg);
371 DECLARE_STATS_COUNTER(cnt_unmap_sg);
372 DECLARE_STATS_COUNTER(cnt_alloc_coherent);
373 DECLARE_STATS_COUNTER(cnt_free_coherent);
374 DECLARE_STATS_COUNTER(cross_page);
375 DECLARE_STATS_COUNTER(domain_flush_single);
376 DECLARE_STATS_COUNTER(domain_flush_all);
377 DECLARE_STATS_COUNTER(alloced_io_mem);
378 DECLARE_STATS_COUNTER(total_map_requests);
379 DECLARE_STATS_COUNTER(complete_ppr);
380 DECLARE_STATS_COUNTER(invalidate_iotlb);
381 DECLARE_STATS_COUNTER(invalidate_iotlb_all);
382 DECLARE_STATS_COUNTER(pri_requests);
383
384 static struct dentry *stats_dir;
385 static struct dentry *de_fflush;
386
387 static void amd_iommu_stats_add(struct __iommu_counter *cnt)
388 {
389 if (stats_dir == NULL)
390 return;
391
392 cnt->dent = debugfs_create_u64(cnt->name, 0444, stats_dir,
393 &cnt->value);
394 }
395
396 static void amd_iommu_stats_init(void)
397 {
398 stats_dir = debugfs_create_dir("amd-iommu", NULL);
399 if (stats_dir == NULL)
400 return;
401
402 de_fflush = debugfs_create_bool("fullflush", 0444, stats_dir,
403 &amd_iommu_unmap_flush);
404
405 amd_iommu_stats_add(&compl_wait);
406 amd_iommu_stats_add(&cnt_map_single);
407 amd_iommu_stats_add(&cnt_unmap_single);
408 amd_iommu_stats_add(&cnt_map_sg);
409 amd_iommu_stats_add(&cnt_unmap_sg);
410 amd_iommu_stats_add(&cnt_alloc_coherent);
411 amd_iommu_stats_add(&cnt_free_coherent);
412 amd_iommu_stats_add(&cross_page);
413 amd_iommu_stats_add(&domain_flush_single);
414 amd_iommu_stats_add(&domain_flush_all);
415 amd_iommu_stats_add(&alloced_io_mem);
416 amd_iommu_stats_add(&total_map_requests);
417 amd_iommu_stats_add(&complete_ppr);
418 amd_iommu_stats_add(&invalidate_iotlb);
419 amd_iommu_stats_add(&invalidate_iotlb_all);
420 amd_iommu_stats_add(&pri_requests);
421 }
422
423 #endif
424
425 /****************************************************************************
426 *
427 * Interrupt handling functions
428 *
429 ****************************************************************************/
430
431 static void dump_dte_entry(u16 devid)
432 {
433 int i;
434
435 for (i = 0; i < 4; ++i)
436 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
437 amd_iommu_dev_table[devid].data[i]);
438 }
439
440 static void dump_command(unsigned long phys_addr)
441 {
442 struct iommu_cmd *cmd = phys_to_virt(phys_addr);
443 int i;
444
445 for (i = 0; i < 4; ++i)
446 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
447 }
448
449 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
450 {
451 int type, devid, domid, flags;
452 volatile u32 *event = __evt;
453 int count = 0;
454 u64 address;
455
456 retry:
457 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
458 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
459 domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
460 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
461 address = (u64)(((u64)event[3]) << 32) | event[2];
462
463 if (type == 0) {
464 /* Did we hit the erratum? */
465 if (++count == LOOP_TIMEOUT) {
466 pr_err("AMD-Vi: No event written to event log\n");
467 return;
468 }
469 udelay(1);
470 goto retry;
471 }
472
473 printk(KERN_ERR "AMD-Vi: Event logged [");
474
475 switch (type) {
476 case EVENT_TYPE_ILL_DEV:
477 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
478 "address=0x%016llx flags=0x%04x]\n",
479 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
480 address, flags);
481 dump_dte_entry(devid);
482 break;
483 case EVENT_TYPE_IO_FAULT:
484 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
485 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
486 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
487 domid, address, flags);
488 break;
489 case EVENT_TYPE_DEV_TAB_ERR:
490 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
491 "address=0x%016llx flags=0x%04x]\n",
492 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
493 address, flags);
494 break;
495 case EVENT_TYPE_PAGE_TAB_ERR:
496 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
497 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
498 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
499 domid, address, flags);
500 break;
501 case EVENT_TYPE_ILL_CMD:
502 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
503 dump_command(address);
504 break;
505 case EVENT_TYPE_CMD_HARD_ERR:
506 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
507 "flags=0x%04x]\n", address, flags);
508 break;
509 case EVENT_TYPE_IOTLB_INV_TO:
510 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
511 "address=0x%016llx]\n",
512 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
513 address);
514 break;
515 case EVENT_TYPE_INV_DEV_REQ:
516 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
517 "address=0x%016llx flags=0x%04x]\n",
518 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
519 address, flags);
520 break;
521 default:
522 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
523 }
524
525 memset(__evt, 0, 4 * sizeof(u32));
526 }
527
528 static void iommu_poll_events(struct amd_iommu *iommu)
529 {
530 u32 head, tail;
531
532 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
533 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
534
535 while (head != tail) {
536 iommu_print_event(iommu, iommu->evt_buf + head);
537 head = (head + EVENT_ENTRY_SIZE) % EVT_BUFFER_SIZE;
538 }
539
540 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
541 }
542
543 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
544 {
545 struct amd_iommu_fault fault;
546
547 INC_STATS_COUNTER(pri_requests);
548
549 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
550 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
551 return;
552 }
553
554 fault.address = raw[1];
555 fault.pasid = PPR_PASID(raw[0]);
556 fault.device_id = PPR_DEVID(raw[0]);
557 fault.tag = PPR_TAG(raw[0]);
558 fault.flags = PPR_FLAGS(raw[0]);
559
560 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
561 }
562
563 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
564 {
565 u32 head, tail;
566
567 if (iommu->ppr_log == NULL)
568 return;
569
570 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
571 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
572
573 while (head != tail) {
574 volatile u64 *raw;
575 u64 entry[2];
576 int i;
577
578 raw = (u64 *)(iommu->ppr_log + head);
579
580 /*
581 * Hardware bug: Interrupt may arrive before the entry is
582 * written to memory. If this happens we need to wait for the
583 * entry to arrive.
584 */
585 for (i = 0; i < LOOP_TIMEOUT; ++i) {
586 if (PPR_REQ_TYPE(raw[0]) != 0)
587 break;
588 udelay(1);
589 }
590
591 /* Avoid memcpy function-call overhead */
592 entry[0] = raw[0];
593 entry[1] = raw[1];
594
595 /*
596 * To detect the hardware bug we need to clear the entry
597 * back to zero.
598 */
599 raw[0] = raw[1] = 0UL;
600
601 /* Update head pointer of hardware ring-buffer */
602 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
603 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
604
605 /* Handle PPR entry */
606 iommu_handle_ppr_entry(iommu, entry);
607
608 /* Refresh ring-buffer information */
609 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
610 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
611 }
612 }
613
614 irqreturn_t amd_iommu_int_thread(int irq, void *data)
615 {
616 struct amd_iommu *iommu = (struct amd_iommu *) data;
617 u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
618
619 while (status & (MMIO_STATUS_EVT_INT_MASK | MMIO_STATUS_PPR_INT_MASK)) {
620 /* Enable EVT and PPR interrupts again */
621 writel((MMIO_STATUS_EVT_INT_MASK | MMIO_STATUS_PPR_INT_MASK),
622 iommu->mmio_base + MMIO_STATUS_OFFSET);
623
624 if (status & MMIO_STATUS_EVT_INT_MASK) {
625 pr_devel("AMD-Vi: Processing IOMMU Event Log\n");
626 iommu_poll_events(iommu);
627 }
628
629 if (status & MMIO_STATUS_PPR_INT_MASK) {
630 pr_devel("AMD-Vi: Processing IOMMU PPR Log\n");
631 iommu_poll_ppr_log(iommu);
632 }
633
634 /*
635 * Hardware bug: ERBT1312
636 * When re-enabling interrupt (by writing 1
637 * to clear the bit), the hardware might also try to set
638 * the interrupt bit in the event status register.
639 * In this scenario, the bit will be set, and disable
640 * subsequent interrupts.
641 *
642 * Workaround: The IOMMU driver should read back the
643 * status register and check if the interrupt bits are cleared.
644 * If not, driver will need to go through the interrupt handler
645 * again and re-clear the bits
646 */
647 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
648 }
649 return IRQ_HANDLED;
650 }
651
652 irqreturn_t amd_iommu_int_handler(int irq, void *data)
653 {
654 return IRQ_WAKE_THREAD;
655 }
656
657 /****************************************************************************
658 *
659 * IOMMU command queuing functions
660 *
661 ****************************************************************************/
662
663 static int wait_on_sem(volatile u64 *sem)
664 {
665 int i = 0;
666
667 while (*sem == 0 && i < LOOP_TIMEOUT) {
668 udelay(1);
669 i += 1;
670 }
671
672 if (i == LOOP_TIMEOUT) {
673 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
674 return -EIO;
675 }
676
677 return 0;
678 }
679
680 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
681 struct iommu_cmd *cmd,
682 u32 tail)
683 {
684 u8 *target;
685
686 target = iommu->cmd_buf + tail;
687 tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
688
689 /* Copy command to buffer */
690 memcpy(target, cmd, sizeof(*cmd));
691
692 /* Tell the IOMMU about it */
693 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
694 }
695
696 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
697 {
698 WARN_ON(address & 0x7ULL);
699
700 memset(cmd, 0, sizeof(*cmd));
701 cmd->data[0] = lower_32_bits(__pa(address)) | CMD_COMPL_WAIT_STORE_MASK;
702 cmd->data[1] = upper_32_bits(__pa(address));
703 cmd->data[2] = 1;
704 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
705 }
706
707 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
708 {
709 memset(cmd, 0, sizeof(*cmd));
710 cmd->data[0] = devid;
711 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
712 }
713
714 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
715 size_t size, u16 domid, int pde)
716 {
717 u64 pages;
718 bool s;
719
720 pages = iommu_num_pages(address, size, PAGE_SIZE);
721 s = false;
722
723 if (pages > 1) {
724 /*
725 * If we have to flush more than one page, flush all
726 * TLB entries for this domain
727 */
728 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
729 s = true;
730 }
731
732 address &= PAGE_MASK;
733
734 memset(cmd, 0, sizeof(*cmd));
735 cmd->data[1] |= domid;
736 cmd->data[2] = lower_32_bits(address);
737 cmd->data[3] = upper_32_bits(address);
738 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
739 if (s) /* size bit - we flush more than one 4kb page */
740 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
741 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
742 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
743 }
744
745 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
746 u64 address, size_t size)
747 {
748 u64 pages;
749 bool s;
750
751 pages = iommu_num_pages(address, size, PAGE_SIZE);
752 s = false;
753
754 if (pages > 1) {
755 /*
756 * If we have to flush more than one page, flush all
757 * TLB entries for this domain
758 */
759 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
760 s = true;
761 }
762
763 address &= PAGE_MASK;
764
765 memset(cmd, 0, sizeof(*cmd));
766 cmd->data[0] = devid;
767 cmd->data[0] |= (qdep & 0xff) << 24;
768 cmd->data[1] = devid;
769 cmd->data[2] = lower_32_bits(address);
770 cmd->data[3] = upper_32_bits(address);
771 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
772 if (s)
773 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
774 }
775
776 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
777 u64 address, bool size)
778 {
779 memset(cmd, 0, sizeof(*cmd));
780
781 address &= ~(0xfffULL);
782
783 cmd->data[0] = pasid;
784 cmd->data[1] = domid;
785 cmd->data[2] = lower_32_bits(address);
786 cmd->data[3] = upper_32_bits(address);
787 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
788 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
789 if (size)
790 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
791 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
792 }
793
794 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
795 int qdep, u64 address, bool size)
796 {
797 memset(cmd, 0, sizeof(*cmd));
798
799 address &= ~(0xfffULL);
800
801 cmd->data[0] = devid;
802 cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
803 cmd->data[0] |= (qdep & 0xff) << 24;
804 cmd->data[1] = devid;
805 cmd->data[1] |= (pasid & 0xff) << 16;
806 cmd->data[2] = lower_32_bits(address);
807 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
808 cmd->data[3] = upper_32_bits(address);
809 if (size)
810 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
811 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
812 }
813
814 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
815 int status, int tag, bool gn)
816 {
817 memset(cmd, 0, sizeof(*cmd));
818
819 cmd->data[0] = devid;
820 if (gn) {
821 cmd->data[1] = pasid;
822 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
823 }
824 cmd->data[3] = tag & 0x1ff;
825 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
826
827 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
828 }
829
830 static void build_inv_all(struct iommu_cmd *cmd)
831 {
832 memset(cmd, 0, sizeof(*cmd));
833 CMD_SET_TYPE(cmd, CMD_INV_ALL);
834 }
835
836 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
837 {
838 memset(cmd, 0, sizeof(*cmd));
839 cmd->data[0] = devid;
840 CMD_SET_TYPE(cmd, CMD_INV_IRT);
841 }
842
843 /*
844 * Writes the command to the IOMMUs command buffer and informs the
845 * hardware about the new command.
846 */
847 static int iommu_queue_command_sync(struct amd_iommu *iommu,
848 struct iommu_cmd *cmd,
849 bool sync)
850 {
851 u32 left, tail, head, next_tail;
852 unsigned long flags;
853
854 again:
855 spin_lock_irqsave(&iommu->lock, flags);
856
857 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
858 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
859 next_tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
860 left = (head - next_tail) % CMD_BUFFER_SIZE;
861
862 if (left <= 2) {
863 struct iommu_cmd sync_cmd;
864 volatile u64 sem = 0;
865 int ret;
866
867 build_completion_wait(&sync_cmd, (u64)&sem);
868 copy_cmd_to_buffer(iommu, &sync_cmd, tail);
869
870 spin_unlock_irqrestore(&iommu->lock, flags);
871
872 if ((ret = wait_on_sem(&sem)) != 0)
873 return ret;
874
875 goto again;
876 }
877
878 copy_cmd_to_buffer(iommu, cmd, tail);
879
880 /* We need to sync now to make sure all commands are processed */
881 iommu->need_sync = sync;
882
883 spin_unlock_irqrestore(&iommu->lock, flags);
884
885 return 0;
886 }
887
888 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
889 {
890 return iommu_queue_command_sync(iommu, cmd, true);
891 }
892
893 /*
894 * This function queues a completion wait command into the command
895 * buffer of an IOMMU
896 */
897 static int iommu_completion_wait(struct amd_iommu *iommu)
898 {
899 struct iommu_cmd cmd;
900 volatile u64 sem = 0;
901 int ret;
902
903 if (!iommu->need_sync)
904 return 0;
905
906 build_completion_wait(&cmd, (u64)&sem);
907
908 ret = iommu_queue_command_sync(iommu, &cmd, false);
909 if (ret)
910 return ret;
911
912 return wait_on_sem(&sem);
913 }
914
915 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
916 {
917 struct iommu_cmd cmd;
918
919 build_inv_dte(&cmd, devid);
920
921 return iommu_queue_command(iommu, &cmd);
922 }
923
924 static void iommu_flush_dte_all(struct amd_iommu *iommu)
925 {
926 u32 devid;
927
928 for (devid = 0; devid <= 0xffff; ++devid)
929 iommu_flush_dte(iommu, devid);
930
931 iommu_completion_wait(iommu);
932 }
933
934 /*
935 * This function uses heavy locking and may disable irqs for some time. But
936 * this is no issue because it is only called during resume.
937 */
938 static void iommu_flush_tlb_all(struct amd_iommu *iommu)
939 {
940 u32 dom_id;
941
942 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
943 struct iommu_cmd cmd;
944 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
945 dom_id, 1);
946 iommu_queue_command(iommu, &cmd);
947 }
948
949 iommu_completion_wait(iommu);
950 }
951
952 static void iommu_flush_all(struct amd_iommu *iommu)
953 {
954 struct iommu_cmd cmd;
955
956 build_inv_all(&cmd);
957
958 iommu_queue_command(iommu, &cmd);
959 iommu_completion_wait(iommu);
960 }
961
962 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
963 {
964 struct iommu_cmd cmd;
965
966 build_inv_irt(&cmd, devid);
967
968 iommu_queue_command(iommu, &cmd);
969 }
970
971 static void iommu_flush_irt_all(struct amd_iommu *iommu)
972 {
973 u32 devid;
974
975 for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
976 iommu_flush_irt(iommu, devid);
977
978 iommu_completion_wait(iommu);
979 }
980
981 void iommu_flush_all_caches(struct amd_iommu *iommu)
982 {
983 if (iommu_feature(iommu, FEATURE_IA)) {
984 iommu_flush_all(iommu);
985 } else {
986 iommu_flush_dte_all(iommu);
987 iommu_flush_irt_all(iommu);
988 iommu_flush_tlb_all(iommu);
989 }
990 }
991
992 /*
993 * Command send function for flushing on-device TLB
994 */
995 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
996 u64 address, size_t size)
997 {
998 struct amd_iommu *iommu;
999 struct iommu_cmd cmd;
1000 int qdep;
1001
1002 qdep = dev_data->ats.qdep;
1003 iommu = amd_iommu_rlookup_table[dev_data->devid];
1004
1005 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1006
1007 return iommu_queue_command(iommu, &cmd);
1008 }
1009
1010 /*
1011 * Command send function for invalidating a device table entry
1012 */
1013 static int device_flush_dte(struct iommu_dev_data *dev_data)
1014 {
1015 struct amd_iommu *iommu;
1016 u16 alias;
1017 int ret;
1018
1019 iommu = amd_iommu_rlookup_table[dev_data->devid];
1020 alias = amd_iommu_alias_table[dev_data->devid];
1021
1022 ret = iommu_flush_dte(iommu, dev_data->devid);
1023 if (!ret && alias != dev_data->devid)
1024 ret = iommu_flush_dte(iommu, alias);
1025 if (ret)
1026 return ret;
1027
1028 if (dev_data->ats.enabled)
1029 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1030
1031 return ret;
1032 }
1033
1034 /*
1035 * TLB invalidation function which is called from the mapping functions.
1036 * It invalidates a single PTE if the range to flush is within a single
1037 * page. Otherwise it flushes the whole TLB of the IOMMU.
1038 */
1039 static void __domain_flush_pages(struct protection_domain *domain,
1040 u64 address, size_t size, int pde)
1041 {
1042 struct iommu_dev_data *dev_data;
1043 struct iommu_cmd cmd;
1044 int ret = 0, i;
1045
1046 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1047
1048 for (i = 0; i < amd_iommus_present; ++i) {
1049 if (!domain->dev_iommu[i])
1050 continue;
1051
1052 /*
1053 * Devices of this domain are behind this IOMMU
1054 * We need a TLB flush
1055 */
1056 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1057 }
1058
1059 list_for_each_entry(dev_data, &domain->dev_list, list) {
1060
1061 if (!dev_data->ats.enabled)
1062 continue;
1063
1064 ret |= device_flush_iotlb(dev_data, address, size);
1065 }
1066
1067 WARN_ON(ret);
1068 }
1069
1070 static void domain_flush_pages(struct protection_domain *domain,
1071 u64 address, size_t size)
1072 {
1073 __domain_flush_pages(domain, address, size, 0);
1074 }
1075
1076 /* Flush the whole IO/TLB for a given protection domain */
1077 static void domain_flush_tlb(struct protection_domain *domain)
1078 {
1079 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1080 }
1081
1082 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1083 static void domain_flush_tlb_pde(struct protection_domain *domain)
1084 {
1085 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1086 }
1087
1088 static void domain_flush_complete(struct protection_domain *domain)
1089 {
1090 int i;
1091
1092 for (i = 0; i < amd_iommus_present; ++i) {
1093 if (!domain->dev_iommu[i])
1094 continue;
1095
1096 /*
1097 * Devices of this domain are behind this IOMMU
1098 * We need to wait for completion of all commands.
1099 */
1100 iommu_completion_wait(amd_iommus[i]);
1101 }
1102 }
1103
1104
1105 /*
1106 * This function flushes the DTEs for all devices in domain
1107 */
1108 static void domain_flush_devices(struct protection_domain *domain)
1109 {
1110 struct iommu_dev_data *dev_data;
1111
1112 list_for_each_entry(dev_data, &domain->dev_list, list)
1113 device_flush_dte(dev_data);
1114 }
1115
1116 /****************************************************************************
1117 *
1118 * The functions below are used the create the page table mappings for
1119 * unity mapped regions.
1120 *
1121 ****************************************************************************/
1122
1123 /*
1124 * This function is used to add another level to an IO page table. Adding
1125 * another level increases the size of the address space by 9 bits to a size up
1126 * to 64 bits.
1127 */
1128 static bool increase_address_space(struct protection_domain *domain,
1129 gfp_t gfp)
1130 {
1131 u64 *pte;
1132
1133 if (domain->mode == PAGE_MODE_6_LEVEL)
1134 /* address space already 64 bit large */
1135 return false;
1136
1137 pte = (void *)get_zeroed_page(gfp);
1138 if (!pte)
1139 return false;
1140
1141 *pte = PM_LEVEL_PDE(domain->mode,
1142 virt_to_phys(domain->pt_root));
1143 domain->pt_root = pte;
1144 domain->mode += 1;
1145 domain->updated = true;
1146
1147 return true;
1148 }
1149
1150 static u64 *alloc_pte(struct protection_domain *domain,
1151 unsigned long address,
1152 unsigned long page_size,
1153 u64 **pte_page,
1154 gfp_t gfp)
1155 {
1156 int level, end_lvl;
1157 u64 *pte, *page;
1158
1159 BUG_ON(!is_power_of_2(page_size));
1160
1161 while (address > PM_LEVEL_SIZE(domain->mode))
1162 increase_address_space(domain, gfp);
1163
1164 level = domain->mode - 1;
1165 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1166 address = PAGE_SIZE_ALIGN(address, page_size);
1167 end_lvl = PAGE_SIZE_LEVEL(page_size);
1168
1169 while (level > end_lvl) {
1170 if (!IOMMU_PTE_PRESENT(*pte)) {
1171 page = (u64 *)get_zeroed_page(gfp);
1172 if (!page)
1173 return NULL;
1174 *pte = PM_LEVEL_PDE(level, virt_to_phys(page));
1175 }
1176
1177 /* No level skipping support yet */
1178 if (PM_PTE_LEVEL(*pte) != level)
1179 return NULL;
1180
1181 level -= 1;
1182
1183 pte = IOMMU_PTE_PAGE(*pte);
1184
1185 if (pte_page && level == end_lvl)
1186 *pte_page = pte;
1187
1188 pte = &pte[PM_LEVEL_INDEX(level, address)];
1189 }
1190
1191 return pte;
1192 }
1193
1194 /*
1195 * This function checks if there is a PTE for a given dma address. If
1196 * there is one, it returns the pointer to it.
1197 */
1198 static u64 *fetch_pte(struct protection_domain *domain,
1199 unsigned long address,
1200 unsigned long *page_size)
1201 {
1202 int level;
1203 u64 *pte;
1204
1205 if (address > PM_LEVEL_SIZE(domain->mode))
1206 return NULL;
1207
1208 level = domain->mode - 1;
1209 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1210 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1211
1212 while (level > 0) {
1213
1214 /* Not Present */
1215 if (!IOMMU_PTE_PRESENT(*pte))
1216 return NULL;
1217
1218 /* Large PTE */
1219 if (PM_PTE_LEVEL(*pte) == 7 ||
1220 PM_PTE_LEVEL(*pte) == 0)
1221 break;
1222
1223 /* No level skipping support yet */
1224 if (PM_PTE_LEVEL(*pte) != level)
1225 return NULL;
1226
1227 level -= 1;
1228
1229 /* Walk to the next level */
1230 pte = IOMMU_PTE_PAGE(*pte);
1231 pte = &pte[PM_LEVEL_INDEX(level, address)];
1232 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1233 }
1234
1235 if (PM_PTE_LEVEL(*pte) == 0x07) {
1236 unsigned long pte_mask;
1237
1238 /*
1239 * If we have a series of large PTEs, make
1240 * sure to return a pointer to the first one.
1241 */
1242 *page_size = pte_mask = PTE_PAGE_SIZE(*pte);
1243 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1244 pte = (u64 *)(((unsigned long)pte) & pte_mask);
1245 }
1246
1247 return pte;
1248 }
1249
1250 /*
1251 * Generic mapping functions. It maps a physical address into a DMA
1252 * address space. It allocates the page table pages if necessary.
1253 * In the future it can be extended to a generic mapping function
1254 * supporting all features of AMD IOMMU page tables like level skipping
1255 * and full 64 bit address spaces.
1256 */
1257 static int iommu_map_page(struct protection_domain *dom,
1258 unsigned long bus_addr,
1259 unsigned long phys_addr,
1260 int prot,
1261 unsigned long page_size)
1262 {
1263 u64 __pte, *pte;
1264 int i, count;
1265
1266 BUG_ON(!IS_ALIGNED(bus_addr, page_size));
1267 BUG_ON(!IS_ALIGNED(phys_addr, page_size));
1268
1269 if (!(prot & IOMMU_PROT_MASK))
1270 return -EINVAL;
1271
1272 count = PAGE_SIZE_PTE_COUNT(page_size);
1273 pte = alloc_pte(dom, bus_addr, page_size, NULL, GFP_KERNEL);
1274
1275 if (!pte)
1276 return -ENOMEM;
1277
1278 for (i = 0; i < count; ++i)
1279 if (IOMMU_PTE_PRESENT(pte[i]))
1280 return -EBUSY;
1281
1282 if (count > 1) {
1283 __pte = PAGE_SIZE_PTE(phys_addr, page_size);
1284 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_P | IOMMU_PTE_FC;
1285 } else
1286 __pte = phys_addr | IOMMU_PTE_P | IOMMU_PTE_FC;
1287
1288 if (prot & IOMMU_PROT_IR)
1289 __pte |= IOMMU_PTE_IR;
1290 if (prot & IOMMU_PROT_IW)
1291 __pte |= IOMMU_PTE_IW;
1292
1293 for (i = 0; i < count; ++i)
1294 pte[i] = __pte;
1295
1296 update_domain(dom);
1297
1298 return 0;
1299 }
1300
1301 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1302 unsigned long bus_addr,
1303 unsigned long page_size)
1304 {
1305 unsigned long long unmapped;
1306 unsigned long unmap_size;
1307 u64 *pte;
1308
1309 BUG_ON(!is_power_of_2(page_size));
1310
1311 unmapped = 0;
1312
1313 while (unmapped < page_size) {
1314
1315 pte = fetch_pte(dom, bus_addr, &unmap_size);
1316
1317 if (pte) {
1318 int i, count;
1319
1320 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1321 for (i = 0; i < count; i++)
1322 pte[i] = 0ULL;
1323 }
1324
1325 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1326 unmapped += unmap_size;
1327 }
1328
1329 BUG_ON(unmapped && !is_power_of_2(unmapped));
1330
1331 return unmapped;
1332 }
1333
1334 /****************************************************************************
1335 *
1336 * The next functions belong to the address allocator for the dma_ops
1337 * interface functions. They work like the allocators in the other IOMMU
1338 * drivers. Its basically a bitmap which marks the allocated pages in
1339 * the aperture. Maybe it could be enhanced in the future to a more
1340 * efficient allocator.
1341 *
1342 ****************************************************************************/
1343
1344 /*
1345 * The address allocator core functions.
1346 *
1347 * called with domain->lock held
1348 */
1349
1350 /*
1351 * Used to reserve address ranges in the aperture (e.g. for exclusion
1352 * ranges.
1353 */
1354 static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
1355 unsigned long start_page,
1356 unsigned int pages)
1357 {
1358 unsigned int i, last_page = dom->aperture_size >> PAGE_SHIFT;
1359
1360 if (start_page + pages > last_page)
1361 pages = last_page - start_page;
1362
1363 for (i = start_page; i < start_page + pages; ++i) {
1364 int index = i / APERTURE_RANGE_PAGES;
1365 int page = i % APERTURE_RANGE_PAGES;
1366 __set_bit(page, dom->aperture[index]->bitmap);
1367 }
1368 }
1369
1370 /*
1371 * This function is used to add a new aperture range to an existing
1372 * aperture in case of dma_ops domain allocation or address allocation
1373 * failure.
1374 */
1375 static int alloc_new_range(struct dma_ops_domain *dma_dom,
1376 bool populate, gfp_t gfp)
1377 {
1378 int index = dma_dom->aperture_size >> APERTURE_RANGE_SHIFT;
1379 struct amd_iommu *iommu;
1380 unsigned long i, old_size, pte_pgsize;
1381
1382 #ifdef CONFIG_IOMMU_STRESS
1383 populate = false;
1384 #endif
1385
1386 if (index >= APERTURE_MAX_RANGES)
1387 return -ENOMEM;
1388
1389 dma_dom->aperture[index] = kzalloc(sizeof(struct aperture_range), gfp);
1390 if (!dma_dom->aperture[index])
1391 return -ENOMEM;
1392
1393 dma_dom->aperture[index]->bitmap = (void *)get_zeroed_page(gfp);
1394 if (!dma_dom->aperture[index]->bitmap)
1395 goto out_free;
1396
1397 dma_dom->aperture[index]->offset = dma_dom->aperture_size;
1398
1399 if (populate) {
1400 unsigned long address = dma_dom->aperture_size;
1401 int i, num_ptes = APERTURE_RANGE_PAGES / 512;
1402 u64 *pte, *pte_page;
1403
1404 for (i = 0; i < num_ptes; ++i) {
1405 pte = alloc_pte(&dma_dom->domain, address, PAGE_SIZE,
1406 &pte_page, gfp);
1407 if (!pte)
1408 goto out_free;
1409
1410 dma_dom->aperture[index]->pte_pages[i] = pte_page;
1411
1412 address += APERTURE_RANGE_SIZE / 64;
1413 }
1414 }
1415
1416 old_size = dma_dom->aperture_size;
1417 dma_dom->aperture_size += APERTURE_RANGE_SIZE;
1418
1419 /* Reserve address range used for MSI messages */
1420 if (old_size < MSI_ADDR_BASE_LO &&
1421 dma_dom->aperture_size > MSI_ADDR_BASE_LO) {
1422 unsigned long spage;
1423 int pages;
1424
1425 pages = iommu_num_pages(MSI_ADDR_BASE_LO, 0x10000, PAGE_SIZE);
1426 spage = MSI_ADDR_BASE_LO >> PAGE_SHIFT;
1427
1428 dma_ops_reserve_addresses(dma_dom, spage, pages);
1429 }
1430
1431 /* Initialize the exclusion range if necessary */
1432 for_each_iommu(iommu) {
1433 if (iommu->exclusion_start &&
1434 iommu->exclusion_start >= dma_dom->aperture[index]->offset
1435 && iommu->exclusion_start < dma_dom->aperture_size) {
1436 unsigned long startpage;
1437 int pages = iommu_num_pages(iommu->exclusion_start,
1438 iommu->exclusion_length,
1439 PAGE_SIZE);
1440 startpage = iommu->exclusion_start >> PAGE_SHIFT;
1441 dma_ops_reserve_addresses(dma_dom, startpage, pages);
1442 }
1443 }
1444
1445 /*
1446 * Check for areas already mapped as present in the new aperture
1447 * range and mark those pages as reserved in the allocator. Such
1448 * mappings may already exist as a result of requested unity
1449 * mappings for devices.
1450 */
1451 for (i = dma_dom->aperture[index]->offset;
1452 i < dma_dom->aperture_size;
1453 i += pte_pgsize) {
1454 u64 *pte = fetch_pte(&dma_dom->domain, i, &pte_pgsize);
1455 if (!pte || !IOMMU_PTE_PRESENT(*pte))
1456 continue;
1457
1458 dma_ops_reserve_addresses(dma_dom, i >> PAGE_SHIFT,
1459 pte_pgsize >> 12);
1460 }
1461
1462 update_domain(&dma_dom->domain);
1463
1464 return 0;
1465
1466 out_free:
1467 update_domain(&dma_dom->domain);
1468
1469 free_page((unsigned long)dma_dom->aperture[index]->bitmap);
1470
1471 kfree(dma_dom->aperture[index]);
1472 dma_dom->aperture[index] = NULL;
1473
1474 return -ENOMEM;
1475 }
1476
1477 static unsigned long dma_ops_area_alloc(struct device *dev,
1478 struct dma_ops_domain *dom,
1479 unsigned int pages,
1480 unsigned long align_mask,
1481 u64 dma_mask,
1482 unsigned long start)
1483 {
1484 unsigned long next_bit = dom->next_address % APERTURE_RANGE_SIZE;
1485 int max_index = dom->aperture_size >> APERTURE_RANGE_SHIFT;
1486 int i = start >> APERTURE_RANGE_SHIFT;
1487 unsigned long boundary_size, mask;
1488 unsigned long address = -1;
1489 unsigned long limit;
1490
1491 next_bit >>= PAGE_SHIFT;
1492
1493 mask = dma_get_seg_boundary(dev);
1494
1495 boundary_size = mask + 1 ? ALIGN(mask + 1, PAGE_SIZE) >> PAGE_SHIFT :
1496 1UL << (BITS_PER_LONG - PAGE_SHIFT);
1497
1498 for (;i < max_index; ++i) {
1499 unsigned long offset = dom->aperture[i]->offset >> PAGE_SHIFT;
1500
1501 if (dom->aperture[i]->offset >= dma_mask)
1502 break;
1503
1504 limit = iommu_device_max_index(APERTURE_RANGE_PAGES, offset,
1505 dma_mask >> PAGE_SHIFT);
1506
1507 address = iommu_area_alloc(dom->aperture[i]->bitmap,
1508 limit, next_bit, pages, 0,
1509 boundary_size, align_mask);
1510 if (address != -1) {
1511 address = dom->aperture[i]->offset +
1512 (address << PAGE_SHIFT);
1513 dom->next_address = address + (pages << PAGE_SHIFT);
1514 break;
1515 }
1516
1517 next_bit = 0;
1518 }
1519
1520 return address;
1521 }
1522
1523 static unsigned long dma_ops_alloc_addresses(struct device *dev,
1524 struct dma_ops_domain *dom,
1525 unsigned int pages,
1526 unsigned long align_mask,
1527 u64 dma_mask)
1528 {
1529 unsigned long address;
1530
1531 #ifdef CONFIG_IOMMU_STRESS
1532 dom->next_address = 0;
1533 dom->need_flush = true;
1534 #endif
1535
1536 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1537 dma_mask, dom->next_address);
1538
1539 if (address == -1) {
1540 dom->next_address = 0;
1541 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1542 dma_mask, 0);
1543 dom->need_flush = true;
1544 }
1545
1546 if (unlikely(address == -1))
1547 address = DMA_ERROR_CODE;
1548
1549 WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
1550
1551 return address;
1552 }
1553
1554 /*
1555 * The address free function.
1556 *
1557 * called with domain->lock held
1558 */
1559 static void dma_ops_free_addresses(struct dma_ops_domain *dom,
1560 unsigned long address,
1561 unsigned int pages)
1562 {
1563 unsigned i = address >> APERTURE_RANGE_SHIFT;
1564 struct aperture_range *range = dom->aperture[i];
1565
1566 BUG_ON(i >= APERTURE_MAX_RANGES || range == NULL);
1567
1568 #ifdef CONFIG_IOMMU_STRESS
1569 if (i < 4)
1570 return;
1571 #endif
1572
1573 if (address >= dom->next_address)
1574 dom->need_flush = true;
1575
1576 address = (address % APERTURE_RANGE_SIZE) >> PAGE_SHIFT;
1577
1578 bitmap_clear(range->bitmap, address, pages);
1579
1580 }
1581
1582 /****************************************************************************
1583 *
1584 * The next functions belong to the domain allocation. A domain is
1585 * allocated for every IOMMU as the default domain. If device isolation
1586 * is enabled, every device get its own domain. The most important thing
1587 * about domains is the page table mapping the DMA address space they
1588 * contain.
1589 *
1590 ****************************************************************************/
1591
1592 /*
1593 * This function adds a protection domain to the global protection domain list
1594 */
1595 static void add_domain_to_list(struct protection_domain *domain)
1596 {
1597 unsigned long flags;
1598
1599 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1600 list_add(&domain->list, &amd_iommu_pd_list);
1601 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1602 }
1603
1604 /*
1605 * This function removes a protection domain to the global
1606 * protection domain list
1607 */
1608 static void del_domain_from_list(struct protection_domain *domain)
1609 {
1610 unsigned long flags;
1611
1612 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1613 list_del(&domain->list);
1614 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1615 }
1616
1617 static u16 domain_id_alloc(void)
1618 {
1619 unsigned long flags;
1620 int id;
1621
1622 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1623 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1624 BUG_ON(id == 0);
1625 if (id > 0 && id < MAX_DOMAIN_ID)
1626 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1627 else
1628 id = 0;
1629 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1630
1631 return id;
1632 }
1633
1634 static void domain_id_free(int id)
1635 {
1636 unsigned long flags;
1637
1638 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1639 if (id > 0 && id < MAX_DOMAIN_ID)
1640 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1641 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1642 }
1643
1644 #define DEFINE_FREE_PT_FN(LVL, FN) \
1645 static void free_pt_##LVL (unsigned long __pt) \
1646 { \
1647 unsigned long p; \
1648 u64 *pt; \
1649 int i; \
1650 \
1651 pt = (u64 *)__pt; \
1652 \
1653 for (i = 0; i < 512; ++i) { \
1654 /* PTE present? */ \
1655 if (!IOMMU_PTE_PRESENT(pt[i])) \
1656 continue; \
1657 \
1658 /* Large PTE? */ \
1659 if (PM_PTE_LEVEL(pt[i]) == 0 || \
1660 PM_PTE_LEVEL(pt[i]) == 7) \
1661 continue; \
1662 \
1663 p = (unsigned long)IOMMU_PTE_PAGE(pt[i]); \
1664 FN(p); \
1665 } \
1666 free_page((unsigned long)pt); \
1667 }
1668
1669 DEFINE_FREE_PT_FN(l2, free_page)
1670 DEFINE_FREE_PT_FN(l3, free_pt_l2)
1671 DEFINE_FREE_PT_FN(l4, free_pt_l3)
1672 DEFINE_FREE_PT_FN(l5, free_pt_l4)
1673 DEFINE_FREE_PT_FN(l6, free_pt_l5)
1674
1675 static void free_pagetable(struct protection_domain *domain)
1676 {
1677 unsigned long root = (unsigned long)domain->pt_root;
1678
1679 switch (domain->mode) {
1680 case PAGE_MODE_NONE:
1681 break;
1682 case PAGE_MODE_1_LEVEL:
1683 free_page(root);
1684 break;
1685 case PAGE_MODE_2_LEVEL:
1686 free_pt_l2(root);
1687 break;
1688 case PAGE_MODE_3_LEVEL:
1689 free_pt_l3(root);
1690 break;
1691 case PAGE_MODE_4_LEVEL:
1692 free_pt_l4(root);
1693 break;
1694 case PAGE_MODE_5_LEVEL:
1695 free_pt_l5(root);
1696 break;
1697 case PAGE_MODE_6_LEVEL:
1698 free_pt_l6(root);
1699 break;
1700 default:
1701 BUG();
1702 }
1703 }
1704
1705 static void free_gcr3_tbl_level1(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_page((unsigned long)ptr);
1717 }
1718 }
1719
1720 static void free_gcr3_tbl_level2(u64 *tbl)
1721 {
1722 u64 *ptr;
1723 int i;
1724
1725 for (i = 0; i < 512; ++i) {
1726 if (!(tbl[i] & GCR3_VALID))
1727 continue;
1728
1729 ptr = __va(tbl[i] & PAGE_MASK);
1730
1731 free_gcr3_tbl_level1(ptr);
1732 }
1733 }
1734
1735 static void free_gcr3_table(struct protection_domain *domain)
1736 {
1737 if (domain->glx == 2)
1738 free_gcr3_tbl_level2(domain->gcr3_tbl);
1739 else if (domain->glx == 1)
1740 free_gcr3_tbl_level1(domain->gcr3_tbl);
1741 else
1742 BUG_ON(domain->glx != 0);
1743
1744 free_page((unsigned long)domain->gcr3_tbl);
1745 }
1746
1747 /*
1748 * Free a domain, only used if something went wrong in the
1749 * allocation path and we need to free an already allocated page table
1750 */
1751 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1752 {
1753 int i;
1754
1755 if (!dom)
1756 return;
1757
1758 del_domain_from_list(&dom->domain);
1759
1760 free_pagetable(&dom->domain);
1761
1762 for (i = 0; i < APERTURE_MAX_RANGES; ++i) {
1763 if (!dom->aperture[i])
1764 continue;
1765 free_page((unsigned long)dom->aperture[i]->bitmap);
1766 kfree(dom->aperture[i]);
1767 }
1768
1769 kfree(dom);
1770 }
1771
1772 /*
1773 * Allocates a new protection domain usable for the dma_ops functions.
1774 * It also initializes the page table and the address allocator data
1775 * structures required for the dma_ops interface
1776 */
1777 static struct dma_ops_domain *dma_ops_domain_alloc(void)
1778 {
1779 struct dma_ops_domain *dma_dom;
1780
1781 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1782 if (!dma_dom)
1783 return NULL;
1784
1785 if (protection_domain_init(&dma_dom->domain))
1786 goto free_dma_dom;
1787
1788 dma_dom->domain.mode = PAGE_MODE_2_LEVEL;
1789 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1790 dma_dom->domain.flags = PD_DMA_OPS_MASK;
1791 dma_dom->domain.priv = dma_dom;
1792 if (!dma_dom->domain.pt_root)
1793 goto free_dma_dom;
1794
1795 dma_dom->need_flush = false;
1796
1797 add_domain_to_list(&dma_dom->domain);
1798
1799 if (alloc_new_range(dma_dom, true, GFP_KERNEL))
1800 goto free_dma_dom;
1801
1802 /*
1803 * mark the first page as allocated so we never return 0 as
1804 * a valid dma-address. So we can use 0 as error value
1805 */
1806 dma_dom->aperture[0]->bitmap[0] = 1;
1807 dma_dom->next_address = 0;
1808
1809
1810 return dma_dom;
1811
1812 free_dma_dom:
1813 dma_ops_domain_free(dma_dom);
1814
1815 return NULL;
1816 }
1817
1818 /*
1819 * little helper function to check whether a given protection domain is a
1820 * dma_ops domain
1821 */
1822 static bool dma_ops_domain(struct protection_domain *domain)
1823 {
1824 return domain->flags & PD_DMA_OPS_MASK;
1825 }
1826
1827 static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
1828 {
1829 u64 pte_root = 0;
1830 u64 flags = 0;
1831
1832 if (domain->mode != PAGE_MODE_NONE)
1833 pte_root = virt_to_phys(domain->pt_root);
1834
1835 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
1836 << DEV_ENTRY_MODE_SHIFT;
1837 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
1838
1839 flags = amd_iommu_dev_table[devid].data[1];
1840
1841 if (ats)
1842 flags |= DTE_FLAG_IOTLB;
1843
1844 if (domain->flags & PD_IOMMUV2_MASK) {
1845 u64 gcr3 = __pa(domain->gcr3_tbl);
1846 u64 glx = domain->glx;
1847 u64 tmp;
1848
1849 pte_root |= DTE_FLAG_GV;
1850 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1851
1852 /* First mask out possible old values for GCR3 table */
1853 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1854 flags &= ~tmp;
1855
1856 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1857 flags &= ~tmp;
1858
1859 /* Encode GCR3 table into DTE */
1860 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1861 pte_root |= tmp;
1862
1863 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1864 flags |= tmp;
1865
1866 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1867 flags |= tmp;
1868 }
1869
1870 flags &= ~(0xffffUL);
1871 flags |= domain->id;
1872
1873 amd_iommu_dev_table[devid].data[1] = flags;
1874 amd_iommu_dev_table[devid].data[0] = pte_root;
1875 }
1876
1877 static void clear_dte_entry(u16 devid)
1878 {
1879 /* remove entry from the device table seen by the hardware */
1880 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
1881 amd_iommu_dev_table[devid].data[1] &= DTE_FLAG_MASK;
1882
1883 amd_iommu_apply_erratum_63(devid);
1884 }
1885
1886 static void do_attach(struct iommu_dev_data *dev_data,
1887 struct protection_domain *domain)
1888 {
1889 struct amd_iommu *iommu;
1890 u16 alias;
1891 bool ats;
1892
1893 iommu = amd_iommu_rlookup_table[dev_data->devid];
1894 alias = amd_iommu_alias_table[dev_data->devid];
1895 ats = dev_data->ats.enabled;
1896
1897 /* Update data structures */
1898 dev_data->domain = domain;
1899 list_add(&dev_data->list, &domain->dev_list);
1900
1901 /* Do reference counting */
1902 domain->dev_iommu[iommu->index] += 1;
1903 domain->dev_cnt += 1;
1904
1905 /* Update device table */
1906 set_dte_entry(dev_data->devid, domain, ats);
1907 if (alias != dev_data->devid)
1908 set_dte_entry(dev_data->devid, domain, ats);
1909
1910 device_flush_dte(dev_data);
1911 }
1912
1913 static void do_detach(struct iommu_dev_data *dev_data)
1914 {
1915 struct amd_iommu *iommu;
1916 u16 alias;
1917
1918 /*
1919 * First check if the device is still attached. It might already
1920 * be detached from its domain because the generic
1921 * iommu_detach_group code detached it and we try again here in
1922 * our alias handling.
1923 */
1924 if (!dev_data->domain)
1925 return;
1926
1927 iommu = amd_iommu_rlookup_table[dev_data->devid];
1928 alias = amd_iommu_alias_table[dev_data->devid];
1929
1930 /* decrease reference counters */
1931 dev_data->domain->dev_iommu[iommu->index] -= 1;
1932 dev_data->domain->dev_cnt -= 1;
1933
1934 /* Update data structures */
1935 dev_data->domain = NULL;
1936 list_del(&dev_data->list);
1937 clear_dte_entry(dev_data->devid);
1938 if (alias != dev_data->devid)
1939 clear_dte_entry(alias);
1940
1941 /* Flush the DTE entry */
1942 device_flush_dte(dev_data);
1943 }
1944
1945 /*
1946 * If a device is not yet associated with a domain, this function does
1947 * assigns it visible for the hardware
1948 */
1949 static int __attach_device(struct iommu_dev_data *dev_data,
1950 struct protection_domain *domain)
1951 {
1952 int ret;
1953
1954 /*
1955 * Must be called with IRQs disabled. Warn here to detect early
1956 * when its not.
1957 */
1958 WARN_ON(!irqs_disabled());
1959
1960 /* lock domain */
1961 spin_lock(&domain->lock);
1962
1963 ret = -EBUSY;
1964 if (dev_data->domain != NULL)
1965 goto out_unlock;
1966
1967 /* Attach alias group root */
1968 do_attach(dev_data, domain);
1969
1970 ret = 0;
1971
1972 out_unlock:
1973
1974 /* ready */
1975 spin_unlock(&domain->lock);
1976
1977 return ret;
1978 }
1979
1980
1981 static void pdev_iommuv2_disable(struct pci_dev *pdev)
1982 {
1983 pci_disable_ats(pdev);
1984 pci_disable_pri(pdev);
1985 pci_disable_pasid(pdev);
1986 }
1987
1988 /* FIXME: Change generic reset-function to do the same */
1989 static int pri_reset_while_enabled(struct pci_dev *pdev)
1990 {
1991 u16 control;
1992 int pos;
1993
1994 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
1995 if (!pos)
1996 return -EINVAL;
1997
1998 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
1999 control |= PCI_PRI_CTRL_RESET;
2000 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
2001
2002 return 0;
2003 }
2004
2005 static int pdev_iommuv2_enable(struct pci_dev *pdev)
2006 {
2007 bool reset_enable;
2008 int reqs, ret;
2009
2010 /* FIXME: Hardcode number of outstanding requests for now */
2011 reqs = 32;
2012 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2013 reqs = 1;
2014 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2015
2016 /* Only allow access to user-accessible pages */
2017 ret = pci_enable_pasid(pdev, 0);
2018 if (ret)
2019 goto out_err;
2020
2021 /* First reset the PRI state of the device */
2022 ret = pci_reset_pri(pdev);
2023 if (ret)
2024 goto out_err;
2025
2026 /* Enable PRI */
2027 ret = pci_enable_pri(pdev, reqs);
2028 if (ret)
2029 goto out_err;
2030
2031 if (reset_enable) {
2032 ret = pri_reset_while_enabled(pdev);
2033 if (ret)
2034 goto out_err;
2035 }
2036
2037 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2038 if (ret)
2039 goto out_err;
2040
2041 return 0;
2042
2043 out_err:
2044 pci_disable_pri(pdev);
2045 pci_disable_pasid(pdev);
2046
2047 return ret;
2048 }
2049
2050 /* FIXME: Move this to PCI code */
2051 #define PCI_PRI_TLP_OFF (1 << 15)
2052
2053 static bool pci_pri_tlp_required(struct pci_dev *pdev)
2054 {
2055 u16 status;
2056 int pos;
2057
2058 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2059 if (!pos)
2060 return false;
2061
2062 pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
2063
2064 return (status & PCI_PRI_TLP_OFF) ? true : false;
2065 }
2066
2067 /*
2068 * If a device is not yet associated with a domain, this function
2069 * assigns it visible for the hardware
2070 */
2071 static int attach_device(struct device *dev,
2072 struct protection_domain *domain)
2073 {
2074 struct pci_dev *pdev = to_pci_dev(dev);
2075 struct iommu_dev_data *dev_data;
2076 unsigned long flags;
2077 int ret;
2078
2079 dev_data = get_dev_data(dev);
2080
2081 if (domain->flags & PD_IOMMUV2_MASK) {
2082 if (!dev_data->passthrough)
2083 return -EINVAL;
2084
2085 if (dev_data->iommu_v2) {
2086 if (pdev_iommuv2_enable(pdev) != 0)
2087 return -EINVAL;
2088
2089 dev_data->ats.enabled = true;
2090 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2091 dev_data->pri_tlp = pci_pri_tlp_required(pdev);
2092 }
2093 } else if (amd_iommu_iotlb_sup &&
2094 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2095 dev_data->ats.enabled = true;
2096 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2097 }
2098
2099 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2100 ret = __attach_device(dev_data, domain);
2101 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2102
2103 /*
2104 * We might boot into a crash-kernel here. The crashed kernel
2105 * left the caches in the IOMMU dirty. So we have to flush
2106 * here to evict all dirty stuff.
2107 */
2108 domain_flush_tlb_pde(domain);
2109
2110 return ret;
2111 }
2112
2113 /*
2114 * Removes a device from a protection domain (unlocked)
2115 */
2116 static void __detach_device(struct iommu_dev_data *dev_data)
2117 {
2118 struct protection_domain *domain;
2119
2120 /*
2121 * Must be called with IRQs disabled. Warn here to detect early
2122 * when its not.
2123 */
2124 WARN_ON(!irqs_disabled());
2125
2126 if (WARN_ON(!dev_data->domain))
2127 return;
2128
2129 domain = dev_data->domain;
2130
2131 spin_lock(&domain->lock);
2132
2133 do_detach(dev_data);
2134
2135 spin_unlock(&domain->lock);
2136 }
2137
2138 /*
2139 * Removes a device from a protection domain (with devtable_lock held)
2140 */
2141 static void detach_device(struct device *dev)
2142 {
2143 struct protection_domain *domain;
2144 struct iommu_dev_data *dev_data;
2145 unsigned long flags;
2146
2147 dev_data = get_dev_data(dev);
2148 domain = dev_data->domain;
2149
2150 /* lock device table */
2151 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2152 __detach_device(dev_data);
2153 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2154
2155 if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2)
2156 pdev_iommuv2_disable(to_pci_dev(dev));
2157 else if (dev_data->ats.enabled)
2158 pci_disable_ats(to_pci_dev(dev));
2159
2160 dev_data->ats.enabled = false;
2161 }
2162
2163 static int amd_iommu_add_device(struct device *dev)
2164 {
2165 struct iommu_dev_data *dev_data;
2166 struct iommu_domain *domain;
2167 struct amd_iommu *iommu;
2168 u16 devid;
2169 int ret;
2170
2171 if (!check_device(dev) || get_dev_data(dev))
2172 return 0;
2173
2174 devid = get_device_id(dev);
2175 iommu = amd_iommu_rlookup_table[devid];
2176
2177 ret = iommu_init_device(dev);
2178 if (ret) {
2179 if (ret != -ENOTSUPP)
2180 pr_err("Failed to initialize device %s - trying to proceed anyway\n",
2181 dev_name(dev));
2182
2183 iommu_ignore_device(dev);
2184 dev->archdata.dma_ops = &nommu_dma_ops;
2185 goto out;
2186 }
2187 init_iommu_group(dev);
2188
2189 dev_data = get_dev_data(dev);
2190
2191 BUG_ON(!dev_data);
2192
2193 if (iommu_pass_through || dev_data->iommu_v2)
2194 iommu_request_dm_for_dev(dev);
2195
2196 /* Domains are initialized for this device - have a look what we ended up with */
2197 domain = iommu_get_domain_for_dev(dev);
2198 if (domain->type == IOMMU_DOMAIN_IDENTITY)
2199 dev_data->passthrough = true;
2200 else
2201 dev->archdata.dma_ops = &amd_iommu_dma_ops;
2202
2203 out:
2204 iommu_completion_wait(iommu);
2205
2206 return 0;
2207 }
2208
2209 static void amd_iommu_remove_device(struct device *dev)
2210 {
2211 struct amd_iommu *iommu;
2212 u16 devid;
2213
2214 if (!check_device(dev))
2215 return;
2216
2217 devid = get_device_id(dev);
2218 iommu = amd_iommu_rlookup_table[devid];
2219
2220 iommu_uninit_device(dev);
2221 iommu_completion_wait(iommu);
2222 }
2223
2224 /*****************************************************************************
2225 *
2226 * The next functions belong to the dma_ops mapping/unmapping code.
2227 *
2228 *****************************************************************************/
2229
2230 /*
2231 * In the dma_ops path we only have the struct device. This function
2232 * finds the corresponding IOMMU, the protection domain and the
2233 * requestor id for a given device.
2234 * If the device is not yet associated with a domain this is also done
2235 * in this function.
2236 */
2237 static struct protection_domain *get_domain(struct device *dev)
2238 {
2239 struct protection_domain *domain;
2240 struct iommu_domain *io_domain;
2241
2242 if (!check_device(dev))
2243 return ERR_PTR(-EINVAL);
2244
2245 io_domain = iommu_get_domain_for_dev(dev);
2246 if (!io_domain)
2247 return NULL;
2248
2249 domain = to_pdomain(io_domain);
2250 if (!dma_ops_domain(domain))
2251 return ERR_PTR(-EBUSY);
2252
2253 return domain;
2254 }
2255
2256 static void update_device_table(struct protection_domain *domain)
2257 {
2258 struct iommu_dev_data *dev_data;
2259
2260 list_for_each_entry(dev_data, &domain->dev_list, list)
2261 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled);
2262 }
2263
2264 static void update_domain(struct protection_domain *domain)
2265 {
2266 if (!domain->updated)
2267 return;
2268
2269 update_device_table(domain);
2270
2271 domain_flush_devices(domain);
2272 domain_flush_tlb_pde(domain);
2273
2274 domain->updated = false;
2275 }
2276
2277 /*
2278 * This function fetches the PTE for a given address in the aperture
2279 */
2280 static u64* dma_ops_get_pte(struct dma_ops_domain *dom,
2281 unsigned long address)
2282 {
2283 struct aperture_range *aperture;
2284 u64 *pte, *pte_page;
2285
2286 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2287 if (!aperture)
2288 return NULL;
2289
2290 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2291 if (!pte) {
2292 pte = alloc_pte(&dom->domain, address, PAGE_SIZE, &pte_page,
2293 GFP_ATOMIC);
2294 aperture->pte_pages[APERTURE_PAGE_INDEX(address)] = pte_page;
2295 } else
2296 pte += PM_LEVEL_INDEX(0, address);
2297
2298 update_domain(&dom->domain);
2299
2300 return pte;
2301 }
2302
2303 /*
2304 * This is the generic map function. It maps one 4kb page at paddr to
2305 * the given address in the DMA address space for the domain.
2306 */
2307 static dma_addr_t dma_ops_domain_map(struct dma_ops_domain *dom,
2308 unsigned long address,
2309 phys_addr_t paddr,
2310 int direction)
2311 {
2312 u64 *pte, __pte;
2313
2314 WARN_ON(address > dom->aperture_size);
2315
2316 paddr &= PAGE_MASK;
2317
2318 pte = dma_ops_get_pte(dom, address);
2319 if (!pte)
2320 return DMA_ERROR_CODE;
2321
2322 __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
2323
2324 if (direction == DMA_TO_DEVICE)
2325 __pte |= IOMMU_PTE_IR;
2326 else if (direction == DMA_FROM_DEVICE)
2327 __pte |= IOMMU_PTE_IW;
2328 else if (direction == DMA_BIDIRECTIONAL)
2329 __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
2330
2331 WARN_ON(*pte);
2332
2333 *pte = __pte;
2334
2335 return (dma_addr_t)address;
2336 }
2337
2338 /*
2339 * The generic unmapping function for on page in the DMA address space.
2340 */
2341 static void dma_ops_domain_unmap(struct dma_ops_domain *dom,
2342 unsigned long address)
2343 {
2344 struct aperture_range *aperture;
2345 u64 *pte;
2346
2347 if (address >= dom->aperture_size)
2348 return;
2349
2350 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2351 if (!aperture)
2352 return;
2353
2354 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2355 if (!pte)
2356 return;
2357
2358 pte += PM_LEVEL_INDEX(0, address);
2359
2360 WARN_ON(!*pte);
2361
2362 *pte = 0ULL;
2363 }
2364
2365 /*
2366 * This function contains common code for mapping of a physically
2367 * contiguous memory region into DMA address space. It is used by all
2368 * mapping functions provided with this IOMMU driver.
2369 * Must be called with the domain lock held.
2370 */
2371 static dma_addr_t __map_single(struct device *dev,
2372 struct dma_ops_domain *dma_dom,
2373 phys_addr_t paddr,
2374 size_t size,
2375 int dir,
2376 bool align,
2377 u64 dma_mask)
2378 {
2379 dma_addr_t offset = paddr & ~PAGE_MASK;
2380 dma_addr_t address, start, ret;
2381 unsigned int pages;
2382 unsigned long align_mask = 0;
2383 int i;
2384
2385 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2386 paddr &= PAGE_MASK;
2387
2388 INC_STATS_COUNTER(total_map_requests);
2389
2390 if (pages > 1)
2391 INC_STATS_COUNTER(cross_page);
2392
2393 if (align)
2394 align_mask = (1UL << get_order(size)) - 1;
2395
2396 retry:
2397 address = dma_ops_alloc_addresses(dev, dma_dom, pages, align_mask,
2398 dma_mask);
2399 if (unlikely(address == DMA_ERROR_CODE)) {
2400 /*
2401 * setting next_address here will let the address
2402 * allocator only scan the new allocated range in the
2403 * first run. This is a small optimization.
2404 */
2405 dma_dom->next_address = dma_dom->aperture_size;
2406
2407 if (alloc_new_range(dma_dom, false, GFP_ATOMIC))
2408 goto out;
2409
2410 /*
2411 * aperture was successfully enlarged by 128 MB, try
2412 * allocation again
2413 */
2414 goto retry;
2415 }
2416
2417 start = address;
2418 for (i = 0; i < pages; ++i) {
2419 ret = dma_ops_domain_map(dma_dom, start, paddr, dir);
2420 if (ret == DMA_ERROR_CODE)
2421 goto out_unmap;
2422
2423 paddr += PAGE_SIZE;
2424 start += PAGE_SIZE;
2425 }
2426 address += offset;
2427
2428 ADD_STATS_COUNTER(alloced_io_mem, size);
2429
2430 if (unlikely(dma_dom->need_flush && !amd_iommu_unmap_flush)) {
2431 domain_flush_tlb(&dma_dom->domain);
2432 dma_dom->need_flush = false;
2433 } else if (unlikely(amd_iommu_np_cache))
2434 domain_flush_pages(&dma_dom->domain, address, size);
2435
2436 out:
2437 return address;
2438
2439 out_unmap:
2440
2441 for (--i; i >= 0; --i) {
2442 start -= PAGE_SIZE;
2443 dma_ops_domain_unmap(dma_dom, start);
2444 }
2445
2446 dma_ops_free_addresses(dma_dom, address, pages);
2447
2448 return DMA_ERROR_CODE;
2449 }
2450
2451 /*
2452 * Does the reverse of the __map_single function. Must be called with
2453 * the domain lock held too
2454 */
2455 static void __unmap_single(struct dma_ops_domain *dma_dom,
2456 dma_addr_t dma_addr,
2457 size_t size,
2458 int dir)
2459 {
2460 dma_addr_t flush_addr;
2461 dma_addr_t i, start;
2462 unsigned int pages;
2463
2464 if ((dma_addr == DMA_ERROR_CODE) ||
2465 (dma_addr + size > dma_dom->aperture_size))
2466 return;
2467
2468 flush_addr = dma_addr;
2469 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2470 dma_addr &= PAGE_MASK;
2471 start = dma_addr;
2472
2473 for (i = 0; i < pages; ++i) {
2474 dma_ops_domain_unmap(dma_dom, start);
2475 start += PAGE_SIZE;
2476 }
2477
2478 SUB_STATS_COUNTER(alloced_io_mem, size);
2479
2480 dma_ops_free_addresses(dma_dom, dma_addr, pages);
2481
2482 if (amd_iommu_unmap_flush || dma_dom->need_flush) {
2483 domain_flush_pages(&dma_dom->domain, flush_addr, size);
2484 dma_dom->need_flush = false;
2485 }
2486 }
2487
2488 /*
2489 * The exported map_single function for dma_ops.
2490 */
2491 static dma_addr_t map_page(struct device *dev, struct page *page,
2492 unsigned long offset, size_t size,
2493 enum dma_data_direction dir,
2494 struct dma_attrs *attrs)
2495 {
2496 unsigned long flags;
2497 struct protection_domain *domain;
2498 dma_addr_t addr;
2499 u64 dma_mask;
2500 phys_addr_t paddr = page_to_phys(page) + offset;
2501
2502 INC_STATS_COUNTER(cnt_map_single);
2503
2504 domain = get_domain(dev);
2505 if (PTR_ERR(domain) == -EINVAL)
2506 return (dma_addr_t)paddr;
2507 else if (IS_ERR(domain))
2508 return DMA_ERROR_CODE;
2509
2510 dma_mask = *dev->dma_mask;
2511
2512 spin_lock_irqsave(&domain->lock, flags);
2513
2514 addr = __map_single(dev, domain->priv, paddr, size, dir, false,
2515 dma_mask);
2516 if (addr == DMA_ERROR_CODE)
2517 goto out;
2518
2519 domain_flush_complete(domain);
2520
2521 out:
2522 spin_unlock_irqrestore(&domain->lock, flags);
2523
2524 return addr;
2525 }
2526
2527 /*
2528 * The exported unmap_single function for dma_ops.
2529 */
2530 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2531 enum dma_data_direction dir, struct dma_attrs *attrs)
2532 {
2533 unsigned long flags;
2534 struct protection_domain *domain;
2535
2536 INC_STATS_COUNTER(cnt_unmap_single);
2537
2538 domain = get_domain(dev);
2539 if (IS_ERR(domain))
2540 return;
2541
2542 spin_lock_irqsave(&domain->lock, flags);
2543
2544 __unmap_single(domain->priv, dma_addr, size, dir);
2545
2546 domain_flush_complete(domain);
2547
2548 spin_unlock_irqrestore(&domain->lock, flags);
2549 }
2550
2551 /*
2552 * The exported map_sg function for dma_ops (handles scatter-gather
2553 * lists).
2554 */
2555 static int map_sg(struct device *dev, struct scatterlist *sglist,
2556 int nelems, enum dma_data_direction dir,
2557 struct dma_attrs *attrs)
2558 {
2559 unsigned long flags;
2560 struct protection_domain *domain;
2561 int i;
2562 struct scatterlist *s;
2563 phys_addr_t paddr;
2564 int mapped_elems = 0;
2565 u64 dma_mask;
2566
2567 INC_STATS_COUNTER(cnt_map_sg);
2568
2569 domain = get_domain(dev);
2570 if (IS_ERR(domain))
2571 return 0;
2572
2573 dma_mask = *dev->dma_mask;
2574
2575 spin_lock_irqsave(&domain->lock, flags);
2576
2577 for_each_sg(sglist, s, nelems, i) {
2578 paddr = sg_phys(s);
2579
2580 s->dma_address = __map_single(dev, domain->priv,
2581 paddr, s->length, dir, false,
2582 dma_mask);
2583
2584 if (s->dma_address) {
2585 s->dma_length = s->length;
2586 mapped_elems++;
2587 } else
2588 goto unmap;
2589 }
2590
2591 domain_flush_complete(domain);
2592
2593 out:
2594 spin_unlock_irqrestore(&domain->lock, flags);
2595
2596 return mapped_elems;
2597 unmap:
2598 for_each_sg(sglist, s, mapped_elems, i) {
2599 if (s->dma_address)
2600 __unmap_single(domain->priv, s->dma_address,
2601 s->dma_length, dir);
2602 s->dma_address = s->dma_length = 0;
2603 }
2604
2605 mapped_elems = 0;
2606
2607 goto out;
2608 }
2609
2610 /*
2611 * The exported map_sg function for dma_ops (handles scatter-gather
2612 * lists).
2613 */
2614 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2615 int nelems, enum dma_data_direction dir,
2616 struct dma_attrs *attrs)
2617 {
2618 unsigned long flags;
2619 struct protection_domain *domain;
2620 struct scatterlist *s;
2621 int i;
2622
2623 INC_STATS_COUNTER(cnt_unmap_sg);
2624
2625 domain = get_domain(dev);
2626 if (IS_ERR(domain))
2627 return;
2628
2629 spin_lock_irqsave(&domain->lock, flags);
2630
2631 for_each_sg(sglist, s, nelems, i) {
2632 __unmap_single(domain->priv, s->dma_address,
2633 s->dma_length, dir);
2634 s->dma_address = s->dma_length = 0;
2635 }
2636
2637 domain_flush_complete(domain);
2638
2639 spin_unlock_irqrestore(&domain->lock, flags);
2640 }
2641
2642 /*
2643 * The exported alloc_coherent function for dma_ops.
2644 */
2645 static void *alloc_coherent(struct device *dev, size_t size,
2646 dma_addr_t *dma_addr, gfp_t flag,
2647 struct dma_attrs *attrs)
2648 {
2649 u64 dma_mask = dev->coherent_dma_mask;
2650 struct protection_domain *domain;
2651 unsigned long flags;
2652 struct page *page;
2653
2654 INC_STATS_COUNTER(cnt_alloc_coherent);
2655
2656 domain = get_domain(dev);
2657 if (PTR_ERR(domain) == -EINVAL) {
2658 page = alloc_pages(flag, get_order(size));
2659 *dma_addr = page_to_phys(page);
2660 return page_address(page);
2661 } else if (IS_ERR(domain))
2662 return NULL;
2663
2664 size = PAGE_ALIGN(size);
2665 dma_mask = dev->coherent_dma_mask;
2666 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2667 flag |= __GFP_ZERO;
2668
2669 page = alloc_pages(flag | __GFP_NOWARN, get_order(size));
2670 if (!page) {
2671 if (!gfpflags_allow_blocking(flag))
2672 return NULL;
2673
2674 page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT,
2675 get_order(size));
2676 if (!page)
2677 return NULL;
2678 }
2679
2680 if (!dma_mask)
2681 dma_mask = *dev->dma_mask;
2682
2683 spin_lock_irqsave(&domain->lock, flags);
2684
2685 *dma_addr = __map_single(dev, domain->priv, page_to_phys(page),
2686 size, DMA_BIDIRECTIONAL, true, dma_mask);
2687
2688 if (*dma_addr == DMA_ERROR_CODE) {
2689 spin_unlock_irqrestore(&domain->lock, flags);
2690 goto out_free;
2691 }
2692
2693 domain_flush_complete(domain);
2694
2695 spin_unlock_irqrestore(&domain->lock, flags);
2696
2697 return page_address(page);
2698
2699 out_free:
2700
2701 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2702 __free_pages(page, get_order(size));
2703
2704 return NULL;
2705 }
2706
2707 /*
2708 * The exported free_coherent function for dma_ops.
2709 */
2710 static void free_coherent(struct device *dev, size_t size,
2711 void *virt_addr, dma_addr_t dma_addr,
2712 struct dma_attrs *attrs)
2713 {
2714 struct protection_domain *domain;
2715 unsigned long flags;
2716 struct page *page;
2717
2718 INC_STATS_COUNTER(cnt_free_coherent);
2719
2720 page = virt_to_page(virt_addr);
2721 size = PAGE_ALIGN(size);
2722
2723 domain = get_domain(dev);
2724 if (IS_ERR(domain))
2725 goto free_mem;
2726
2727 spin_lock_irqsave(&domain->lock, flags);
2728
2729 __unmap_single(domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
2730
2731 domain_flush_complete(domain);
2732
2733 spin_unlock_irqrestore(&domain->lock, flags);
2734
2735 free_mem:
2736 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2737 __free_pages(page, get_order(size));
2738 }
2739
2740 /*
2741 * This function is called by the DMA layer to find out if we can handle a
2742 * particular device. It is part of the dma_ops.
2743 */
2744 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2745 {
2746 return check_device(dev);
2747 }
2748
2749 static struct dma_map_ops amd_iommu_dma_ops = {
2750 .alloc = alloc_coherent,
2751 .free = free_coherent,
2752 .map_page = map_page,
2753 .unmap_page = unmap_page,
2754 .map_sg = map_sg,
2755 .unmap_sg = unmap_sg,
2756 .dma_supported = amd_iommu_dma_supported,
2757 };
2758
2759 int __init amd_iommu_init_api(void)
2760 {
2761 return bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
2762 }
2763
2764 int __init amd_iommu_init_dma_ops(void)
2765 {
2766 swiotlb = iommu_pass_through ? 1 : 0;
2767 iommu_detected = 1;
2768
2769 /*
2770 * In case we don't initialize SWIOTLB (actually the common case
2771 * when AMD IOMMU is enabled), make sure there are global
2772 * dma_ops set as a fall-back for devices not handled by this
2773 * driver (for example non-PCI devices).
2774 */
2775 if (!swiotlb)
2776 dma_ops = &nommu_dma_ops;
2777
2778 amd_iommu_stats_init();
2779
2780 if (amd_iommu_unmap_flush)
2781 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
2782 else
2783 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
2784
2785 return 0;
2786 }
2787
2788 /*****************************************************************************
2789 *
2790 * The following functions belong to the exported interface of AMD IOMMU
2791 *
2792 * This interface allows access to lower level functions of the IOMMU
2793 * like protection domain handling and assignement of devices to domains
2794 * which is not possible with the dma_ops interface.
2795 *
2796 *****************************************************************************/
2797
2798 static void cleanup_domain(struct protection_domain *domain)
2799 {
2800 struct iommu_dev_data *entry;
2801 unsigned long flags;
2802
2803 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2804
2805 while (!list_empty(&domain->dev_list)) {
2806 entry = list_first_entry(&domain->dev_list,
2807 struct iommu_dev_data, list);
2808 __detach_device(entry);
2809 }
2810
2811 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2812 }
2813
2814 static void protection_domain_free(struct protection_domain *domain)
2815 {
2816 if (!domain)
2817 return;
2818
2819 del_domain_from_list(domain);
2820
2821 if (domain->id)
2822 domain_id_free(domain->id);
2823
2824 kfree(domain);
2825 }
2826
2827 static int protection_domain_init(struct protection_domain *domain)
2828 {
2829 spin_lock_init(&domain->lock);
2830 mutex_init(&domain->api_lock);
2831 domain->id = domain_id_alloc();
2832 if (!domain->id)
2833 return -ENOMEM;
2834 INIT_LIST_HEAD(&domain->dev_list);
2835
2836 return 0;
2837 }
2838
2839 static struct protection_domain *protection_domain_alloc(void)
2840 {
2841 struct protection_domain *domain;
2842
2843 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2844 if (!domain)
2845 return NULL;
2846
2847 if (protection_domain_init(domain))
2848 goto out_err;
2849
2850 add_domain_to_list(domain);
2851
2852 return domain;
2853
2854 out_err:
2855 kfree(domain);
2856
2857 return NULL;
2858 }
2859
2860 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
2861 {
2862 struct protection_domain *pdomain;
2863 struct dma_ops_domain *dma_domain;
2864
2865 switch (type) {
2866 case IOMMU_DOMAIN_UNMANAGED:
2867 pdomain = protection_domain_alloc();
2868 if (!pdomain)
2869 return NULL;
2870
2871 pdomain->mode = PAGE_MODE_3_LEVEL;
2872 pdomain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2873 if (!pdomain->pt_root) {
2874 protection_domain_free(pdomain);
2875 return NULL;
2876 }
2877
2878 pdomain->domain.geometry.aperture_start = 0;
2879 pdomain->domain.geometry.aperture_end = ~0ULL;
2880 pdomain->domain.geometry.force_aperture = true;
2881
2882 break;
2883 case IOMMU_DOMAIN_DMA:
2884 dma_domain = dma_ops_domain_alloc();
2885 if (!dma_domain) {
2886 pr_err("AMD-Vi: Failed to allocate\n");
2887 return NULL;
2888 }
2889 pdomain = &dma_domain->domain;
2890 break;
2891 case IOMMU_DOMAIN_IDENTITY:
2892 pdomain = protection_domain_alloc();
2893 if (!pdomain)
2894 return NULL;
2895
2896 pdomain->mode = PAGE_MODE_NONE;
2897 break;
2898 default:
2899 return NULL;
2900 }
2901
2902 return &pdomain->domain;
2903 }
2904
2905 static void amd_iommu_domain_free(struct iommu_domain *dom)
2906 {
2907 struct protection_domain *domain;
2908
2909 if (!dom)
2910 return;
2911
2912 domain = to_pdomain(dom);
2913
2914 if (domain->dev_cnt > 0)
2915 cleanup_domain(domain);
2916
2917 BUG_ON(domain->dev_cnt != 0);
2918
2919 if (domain->mode != PAGE_MODE_NONE)
2920 free_pagetable(domain);
2921
2922 if (domain->flags & PD_IOMMUV2_MASK)
2923 free_gcr3_table(domain);
2924
2925 protection_domain_free(domain);
2926 }
2927
2928 static void amd_iommu_detach_device(struct iommu_domain *dom,
2929 struct device *dev)
2930 {
2931 struct iommu_dev_data *dev_data = dev->archdata.iommu;
2932 struct amd_iommu *iommu;
2933 u16 devid;
2934
2935 if (!check_device(dev))
2936 return;
2937
2938 devid = get_device_id(dev);
2939
2940 if (dev_data->domain != NULL)
2941 detach_device(dev);
2942
2943 iommu = amd_iommu_rlookup_table[devid];
2944 if (!iommu)
2945 return;
2946
2947 iommu_completion_wait(iommu);
2948 }
2949
2950 static int amd_iommu_attach_device(struct iommu_domain *dom,
2951 struct device *dev)
2952 {
2953 struct protection_domain *domain = to_pdomain(dom);
2954 struct iommu_dev_data *dev_data;
2955 struct amd_iommu *iommu;
2956 int ret;
2957
2958 if (!check_device(dev))
2959 return -EINVAL;
2960
2961 dev_data = dev->archdata.iommu;
2962
2963 iommu = amd_iommu_rlookup_table[dev_data->devid];
2964 if (!iommu)
2965 return -EINVAL;
2966
2967 if (dev_data->domain)
2968 detach_device(dev);
2969
2970 ret = attach_device(dev, domain);
2971
2972 iommu_completion_wait(iommu);
2973
2974 return ret;
2975 }
2976
2977 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
2978 phys_addr_t paddr, size_t page_size, int iommu_prot)
2979 {
2980 struct protection_domain *domain = to_pdomain(dom);
2981 int prot = 0;
2982 int ret;
2983
2984 if (domain->mode == PAGE_MODE_NONE)
2985 return -EINVAL;
2986
2987 if (iommu_prot & IOMMU_READ)
2988 prot |= IOMMU_PROT_IR;
2989 if (iommu_prot & IOMMU_WRITE)
2990 prot |= IOMMU_PROT_IW;
2991
2992 mutex_lock(&domain->api_lock);
2993 ret = iommu_map_page(domain, iova, paddr, prot, page_size);
2994 mutex_unlock(&domain->api_lock);
2995
2996 return ret;
2997 }
2998
2999 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3000 size_t page_size)
3001 {
3002 struct protection_domain *domain = to_pdomain(dom);
3003 size_t unmap_size;
3004
3005 if (domain->mode == PAGE_MODE_NONE)
3006 return -EINVAL;
3007
3008 mutex_lock(&domain->api_lock);
3009 unmap_size = iommu_unmap_page(domain, iova, page_size);
3010 mutex_unlock(&domain->api_lock);
3011
3012 domain_flush_tlb_pde(domain);
3013
3014 return unmap_size;
3015 }
3016
3017 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3018 dma_addr_t iova)
3019 {
3020 struct protection_domain *domain = to_pdomain(dom);
3021 unsigned long offset_mask, pte_pgsize;
3022 u64 *pte, __pte;
3023
3024 if (domain->mode == PAGE_MODE_NONE)
3025 return iova;
3026
3027 pte = fetch_pte(domain, iova, &pte_pgsize);
3028
3029 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3030 return 0;
3031
3032 offset_mask = pte_pgsize - 1;
3033 __pte = *pte & PM_ADDR_MASK;
3034
3035 return (__pte & ~offset_mask) | (iova & offset_mask);
3036 }
3037
3038 static bool amd_iommu_capable(enum iommu_cap cap)
3039 {
3040 switch (cap) {
3041 case IOMMU_CAP_CACHE_COHERENCY:
3042 return true;
3043 case IOMMU_CAP_INTR_REMAP:
3044 return (irq_remapping_enabled == 1);
3045 case IOMMU_CAP_NOEXEC:
3046 return false;
3047 }
3048
3049 return false;
3050 }
3051
3052 static void amd_iommu_get_dm_regions(struct device *dev,
3053 struct list_head *head)
3054 {
3055 struct unity_map_entry *entry;
3056 u16 devid;
3057
3058 devid = get_device_id(dev);
3059
3060 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
3061 struct iommu_dm_region *region;
3062
3063 if (devid < entry->devid_start || devid > entry->devid_end)
3064 continue;
3065
3066 region = kzalloc(sizeof(*region), GFP_KERNEL);
3067 if (!region) {
3068 pr_err("Out of memory allocating dm-regions for %s\n",
3069 dev_name(dev));
3070 return;
3071 }
3072
3073 region->start = entry->address_start;
3074 region->length = entry->address_end - entry->address_start;
3075 if (entry->prot & IOMMU_PROT_IR)
3076 region->prot |= IOMMU_READ;
3077 if (entry->prot & IOMMU_PROT_IW)
3078 region->prot |= IOMMU_WRITE;
3079
3080 list_add_tail(&region->list, head);
3081 }
3082 }
3083
3084 static void amd_iommu_put_dm_regions(struct device *dev,
3085 struct list_head *head)
3086 {
3087 struct iommu_dm_region *entry, *next;
3088
3089 list_for_each_entry_safe(entry, next, head, list)
3090 kfree(entry);
3091 }
3092
3093 static const struct iommu_ops amd_iommu_ops = {
3094 .capable = amd_iommu_capable,
3095 .domain_alloc = amd_iommu_domain_alloc,
3096 .domain_free = amd_iommu_domain_free,
3097 .attach_dev = amd_iommu_attach_device,
3098 .detach_dev = amd_iommu_detach_device,
3099 .map = amd_iommu_map,
3100 .unmap = amd_iommu_unmap,
3101 .map_sg = default_iommu_map_sg,
3102 .iova_to_phys = amd_iommu_iova_to_phys,
3103 .add_device = amd_iommu_add_device,
3104 .remove_device = amd_iommu_remove_device,
3105 .device_group = pci_device_group,
3106 .get_dm_regions = amd_iommu_get_dm_regions,
3107 .put_dm_regions = amd_iommu_put_dm_regions,
3108 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3109 };
3110
3111 /*****************************************************************************
3112 *
3113 * The next functions do a basic initialization of IOMMU for pass through
3114 * mode
3115 *
3116 * In passthrough mode the IOMMU is initialized and enabled but not used for
3117 * DMA-API translation.
3118 *
3119 *****************************************************************************/
3120
3121 /* IOMMUv2 specific functions */
3122 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3123 {
3124 return atomic_notifier_chain_register(&ppr_notifier, nb);
3125 }
3126 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3127
3128 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3129 {
3130 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3131 }
3132 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3133
3134 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3135 {
3136 struct protection_domain *domain = to_pdomain(dom);
3137 unsigned long flags;
3138
3139 spin_lock_irqsave(&domain->lock, flags);
3140
3141 /* Update data structure */
3142 domain->mode = PAGE_MODE_NONE;
3143 domain->updated = true;
3144
3145 /* Make changes visible to IOMMUs */
3146 update_domain(domain);
3147
3148 /* Page-table is not visible to IOMMU anymore, so free it */
3149 free_pagetable(domain);
3150
3151 spin_unlock_irqrestore(&domain->lock, flags);
3152 }
3153 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3154
3155 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3156 {
3157 struct protection_domain *domain = to_pdomain(dom);
3158 unsigned long flags;
3159 int levels, ret;
3160
3161 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3162 return -EINVAL;
3163
3164 /* Number of GCR3 table levels required */
3165 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3166 levels += 1;
3167
3168 if (levels > amd_iommu_max_glx_val)
3169 return -EINVAL;
3170
3171 spin_lock_irqsave(&domain->lock, flags);
3172
3173 /*
3174 * Save us all sanity checks whether devices already in the
3175 * domain support IOMMUv2. Just force that the domain has no
3176 * devices attached when it is switched into IOMMUv2 mode.
3177 */
3178 ret = -EBUSY;
3179 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3180 goto out;
3181
3182 ret = -ENOMEM;
3183 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3184 if (domain->gcr3_tbl == NULL)
3185 goto out;
3186
3187 domain->glx = levels;
3188 domain->flags |= PD_IOMMUV2_MASK;
3189 domain->updated = true;
3190
3191 update_domain(domain);
3192
3193 ret = 0;
3194
3195 out:
3196 spin_unlock_irqrestore(&domain->lock, flags);
3197
3198 return ret;
3199 }
3200 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3201
3202 static int __flush_pasid(struct protection_domain *domain, int pasid,
3203 u64 address, bool size)
3204 {
3205 struct iommu_dev_data *dev_data;
3206 struct iommu_cmd cmd;
3207 int i, ret;
3208
3209 if (!(domain->flags & PD_IOMMUV2_MASK))
3210 return -EINVAL;
3211
3212 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3213
3214 /*
3215 * IOMMU TLB needs to be flushed before Device TLB to
3216 * prevent device TLB refill from IOMMU TLB
3217 */
3218 for (i = 0; i < amd_iommus_present; ++i) {
3219 if (domain->dev_iommu[i] == 0)
3220 continue;
3221
3222 ret = iommu_queue_command(amd_iommus[i], &cmd);
3223 if (ret != 0)
3224 goto out;
3225 }
3226
3227 /* Wait until IOMMU TLB flushes are complete */
3228 domain_flush_complete(domain);
3229
3230 /* Now flush device TLBs */
3231 list_for_each_entry(dev_data, &domain->dev_list, list) {
3232 struct amd_iommu *iommu;
3233 int qdep;
3234
3235 /*
3236 There might be non-IOMMUv2 capable devices in an IOMMUv2
3237 * domain.
3238 */
3239 if (!dev_data->ats.enabled)
3240 continue;
3241
3242 qdep = dev_data->ats.qdep;
3243 iommu = amd_iommu_rlookup_table[dev_data->devid];
3244
3245 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3246 qdep, address, size);
3247
3248 ret = iommu_queue_command(iommu, &cmd);
3249 if (ret != 0)
3250 goto out;
3251 }
3252
3253 /* Wait until all device TLBs are flushed */
3254 domain_flush_complete(domain);
3255
3256 ret = 0;
3257
3258 out:
3259
3260 return ret;
3261 }
3262
3263 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3264 u64 address)
3265 {
3266 INC_STATS_COUNTER(invalidate_iotlb);
3267
3268 return __flush_pasid(domain, pasid, address, false);
3269 }
3270
3271 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3272 u64 address)
3273 {
3274 struct protection_domain *domain = to_pdomain(dom);
3275 unsigned long flags;
3276 int ret;
3277
3278 spin_lock_irqsave(&domain->lock, flags);
3279 ret = __amd_iommu_flush_page(domain, pasid, address);
3280 spin_unlock_irqrestore(&domain->lock, flags);
3281
3282 return ret;
3283 }
3284 EXPORT_SYMBOL(amd_iommu_flush_page);
3285
3286 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3287 {
3288 INC_STATS_COUNTER(invalidate_iotlb_all);
3289
3290 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3291 true);
3292 }
3293
3294 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3295 {
3296 struct protection_domain *domain = to_pdomain(dom);
3297 unsigned long flags;
3298 int ret;
3299
3300 spin_lock_irqsave(&domain->lock, flags);
3301 ret = __amd_iommu_flush_tlb(domain, pasid);
3302 spin_unlock_irqrestore(&domain->lock, flags);
3303
3304 return ret;
3305 }
3306 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3307
3308 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3309 {
3310 int index;
3311 u64 *pte;
3312
3313 while (true) {
3314
3315 index = (pasid >> (9 * level)) & 0x1ff;
3316 pte = &root[index];
3317
3318 if (level == 0)
3319 break;
3320
3321 if (!(*pte & GCR3_VALID)) {
3322 if (!alloc)
3323 return NULL;
3324
3325 root = (void *)get_zeroed_page(GFP_ATOMIC);
3326 if (root == NULL)
3327 return NULL;
3328
3329 *pte = __pa(root) | GCR3_VALID;
3330 }
3331
3332 root = __va(*pte & PAGE_MASK);
3333
3334 level -= 1;
3335 }
3336
3337 return pte;
3338 }
3339
3340 static int __set_gcr3(struct protection_domain *domain, int pasid,
3341 unsigned long cr3)
3342 {
3343 u64 *pte;
3344
3345 if (domain->mode != PAGE_MODE_NONE)
3346 return -EINVAL;
3347
3348 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3349 if (pte == NULL)
3350 return -ENOMEM;
3351
3352 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3353
3354 return __amd_iommu_flush_tlb(domain, pasid);
3355 }
3356
3357 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3358 {
3359 u64 *pte;
3360
3361 if (domain->mode != PAGE_MODE_NONE)
3362 return -EINVAL;
3363
3364 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3365 if (pte == NULL)
3366 return 0;
3367
3368 *pte = 0;
3369
3370 return __amd_iommu_flush_tlb(domain, pasid);
3371 }
3372
3373 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3374 unsigned long cr3)
3375 {
3376 struct protection_domain *domain = to_pdomain(dom);
3377 unsigned long flags;
3378 int ret;
3379
3380 spin_lock_irqsave(&domain->lock, flags);
3381 ret = __set_gcr3(domain, pasid, cr3);
3382 spin_unlock_irqrestore(&domain->lock, flags);
3383
3384 return ret;
3385 }
3386 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3387
3388 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3389 {
3390 struct protection_domain *domain = to_pdomain(dom);
3391 unsigned long flags;
3392 int ret;
3393
3394 spin_lock_irqsave(&domain->lock, flags);
3395 ret = __clear_gcr3(domain, pasid);
3396 spin_unlock_irqrestore(&domain->lock, flags);
3397
3398 return ret;
3399 }
3400 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3401
3402 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3403 int status, int tag)
3404 {
3405 struct iommu_dev_data *dev_data;
3406 struct amd_iommu *iommu;
3407 struct iommu_cmd cmd;
3408
3409 INC_STATS_COUNTER(complete_ppr);
3410
3411 dev_data = get_dev_data(&pdev->dev);
3412 iommu = amd_iommu_rlookup_table[dev_data->devid];
3413
3414 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3415 tag, dev_data->pri_tlp);
3416
3417 return iommu_queue_command(iommu, &cmd);
3418 }
3419 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3420
3421 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3422 {
3423 struct protection_domain *pdomain;
3424
3425 pdomain = get_domain(&pdev->dev);
3426 if (IS_ERR(pdomain))
3427 return NULL;
3428
3429 /* Only return IOMMUv2 domains */
3430 if (!(pdomain->flags & PD_IOMMUV2_MASK))
3431 return NULL;
3432
3433 return &pdomain->domain;
3434 }
3435 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3436
3437 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3438 {
3439 struct iommu_dev_data *dev_data;
3440
3441 if (!amd_iommu_v2_supported())
3442 return;
3443
3444 dev_data = get_dev_data(&pdev->dev);
3445 dev_data->errata |= (1 << erratum);
3446 }
3447 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3448
3449 int amd_iommu_device_info(struct pci_dev *pdev,
3450 struct amd_iommu_device_info *info)
3451 {
3452 int max_pasids;
3453 int pos;
3454
3455 if (pdev == NULL || info == NULL)
3456 return -EINVAL;
3457
3458 if (!amd_iommu_v2_supported())
3459 return -EINVAL;
3460
3461 memset(info, 0, sizeof(*info));
3462
3463 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3464 if (pos)
3465 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3466
3467 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3468 if (pos)
3469 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3470
3471 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3472 if (pos) {
3473 int features;
3474
3475 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3476 max_pasids = min(max_pasids, (1 << 20));
3477
3478 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3479 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3480
3481 features = pci_pasid_features(pdev);
3482 if (features & PCI_PASID_CAP_EXEC)
3483 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3484 if (features & PCI_PASID_CAP_PRIV)
3485 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3486 }
3487
3488 return 0;
3489 }
3490 EXPORT_SYMBOL(amd_iommu_device_info);
3491
3492 #ifdef CONFIG_IRQ_REMAP
3493
3494 /*****************************************************************************
3495 *
3496 * Interrupt Remapping Implementation
3497 *
3498 *****************************************************************************/
3499
3500 union irte {
3501 u32 val;
3502 struct {
3503 u32 valid : 1,
3504 no_fault : 1,
3505 int_type : 3,
3506 rq_eoi : 1,
3507 dm : 1,
3508 rsvd_1 : 1,
3509 destination : 8,
3510 vector : 8,
3511 rsvd_2 : 8;
3512 } fields;
3513 };
3514
3515 struct irq_2_irte {
3516 u16 devid; /* Device ID for IRTE table */
3517 u16 index; /* Index into IRTE table*/
3518 };
3519
3520 struct amd_ir_data {
3521 struct irq_2_irte irq_2_irte;
3522 union irte irte_entry;
3523 union {
3524 struct msi_msg msi_entry;
3525 };
3526 };
3527
3528 static struct irq_chip amd_ir_chip;
3529
3530 #define DTE_IRQ_PHYS_ADDR_MASK (((1ULL << 45)-1) << 6)
3531 #define DTE_IRQ_REMAP_INTCTL (2ULL << 60)
3532 #define DTE_IRQ_TABLE_LEN (8ULL << 1)
3533 #define DTE_IRQ_REMAP_ENABLE 1ULL
3534
3535 static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3536 {
3537 u64 dte;
3538
3539 dte = amd_iommu_dev_table[devid].data[2];
3540 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
3541 dte |= virt_to_phys(table->table);
3542 dte |= DTE_IRQ_REMAP_INTCTL;
3543 dte |= DTE_IRQ_TABLE_LEN;
3544 dte |= DTE_IRQ_REMAP_ENABLE;
3545
3546 amd_iommu_dev_table[devid].data[2] = dte;
3547 }
3548
3549 #define IRTE_ALLOCATED (~1U)
3550
3551 static struct irq_remap_table *get_irq_table(u16 devid, bool ioapic)
3552 {
3553 struct irq_remap_table *table = NULL;
3554 struct amd_iommu *iommu;
3555 unsigned long flags;
3556 u16 alias;
3557
3558 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3559
3560 iommu = amd_iommu_rlookup_table[devid];
3561 if (!iommu)
3562 goto out_unlock;
3563
3564 table = irq_lookup_table[devid];
3565 if (table)
3566 goto out;
3567
3568 alias = amd_iommu_alias_table[devid];
3569 table = irq_lookup_table[alias];
3570 if (table) {
3571 irq_lookup_table[devid] = table;
3572 set_dte_irq_entry(devid, table);
3573 iommu_flush_dte(iommu, devid);
3574 goto out;
3575 }
3576
3577 /* Nothing there yet, allocate new irq remapping table */
3578 table = kzalloc(sizeof(*table), GFP_ATOMIC);
3579 if (!table)
3580 goto out;
3581
3582 /* Initialize table spin-lock */
3583 spin_lock_init(&table->lock);
3584
3585 if (ioapic)
3586 /* Keep the first 32 indexes free for IOAPIC interrupts */
3587 table->min_index = 32;
3588
3589 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_ATOMIC);
3590 if (!table->table) {
3591 kfree(table);
3592 table = NULL;
3593 goto out;
3594 }
3595
3596 memset(table->table, 0, MAX_IRQS_PER_TABLE * sizeof(u32));
3597
3598 if (ioapic) {
3599 int i;
3600
3601 for (i = 0; i < 32; ++i)
3602 table->table[i] = IRTE_ALLOCATED;
3603 }
3604
3605 irq_lookup_table[devid] = table;
3606 set_dte_irq_entry(devid, table);
3607 iommu_flush_dte(iommu, devid);
3608 if (devid != alias) {
3609 irq_lookup_table[alias] = table;
3610 set_dte_irq_entry(alias, table);
3611 iommu_flush_dte(iommu, alias);
3612 }
3613
3614 out:
3615 iommu_completion_wait(iommu);
3616
3617 out_unlock:
3618 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3619
3620 return table;
3621 }
3622
3623 static int alloc_irq_index(u16 devid, int count)
3624 {
3625 struct irq_remap_table *table;
3626 unsigned long flags;
3627 int index, c;
3628
3629 table = get_irq_table(devid, false);
3630 if (!table)
3631 return -ENODEV;
3632
3633 spin_lock_irqsave(&table->lock, flags);
3634
3635 /* Scan table for free entries */
3636 for (c = 0, index = table->min_index;
3637 index < MAX_IRQS_PER_TABLE;
3638 ++index) {
3639 if (table->table[index] == 0)
3640 c += 1;
3641 else
3642 c = 0;
3643
3644 if (c == count) {
3645 for (; c != 0; --c)
3646 table->table[index - c + 1] = IRTE_ALLOCATED;
3647
3648 index -= count - 1;
3649 goto out;
3650 }
3651 }
3652
3653 index = -ENOSPC;
3654
3655 out:
3656 spin_unlock_irqrestore(&table->lock, flags);
3657
3658 return index;
3659 }
3660
3661 static int modify_irte(u16 devid, int index, union irte irte)
3662 {
3663 struct irq_remap_table *table;
3664 struct amd_iommu *iommu;
3665 unsigned long flags;
3666
3667 iommu = amd_iommu_rlookup_table[devid];
3668 if (iommu == NULL)
3669 return -EINVAL;
3670
3671 table = get_irq_table(devid, false);
3672 if (!table)
3673 return -ENOMEM;
3674
3675 spin_lock_irqsave(&table->lock, flags);
3676 table->table[index] = irte.val;
3677 spin_unlock_irqrestore(&table->lock, flags);
3678
3679 iommu_flush_irt(iommu, devid);
3680 iommu_completion_wait(iommu);
3681
3682 return 0;
3683 }
3684
3685 static void free_irte(u16 devid, int index)
3686 {
3687 struct irq_remap_table *table;
3688 struct amd_iommu *iommu;
3689 unsigned long flags;
3690
3691 iommu = amd_iommu_rlookup_table[devid];
3692 if (iommu == NULL)
3693 return;
3694
3695 table = get_irq_table(devid, false);
3696 if (!table)
3697 return;
3698
3699 spin_lock_irqsave(&table->lock, flags);
3700 table->table[index] = 0;
3701 spin_unlock_irqrestore(&table->lock, flags);
3702
3703 iommu_flush_irt(iommu, devid);
3704 iommu_completion_wait(iommu);
3705 }
3706
3707 static int get_devid(struct irq_alloc_info *info)
3708 {
3709 int devid = -1;
3710
3711 switch (info->type) {
3712 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3713 devid = get_ioapic_devid(info->ioapic_id);
3714 break;
3715 case X86_IRQ_ALLOC_TYPE_HPET:
3716 devid = get_hpet_devid(info->hpet_id);
3717 break;
3718 case X86_IRQ_ALLOC_TYPE_MSI:
3719 case X86_IRQ_ALLOC_TYPE_MSIX:
3720 devid = get_device_id(&info->msi_dev->dev);
3721 break;
3722 default:
3723 BUG_ON(1);
3724 break;
3725 }
3726
3727 return devid;
3728 }
3729
3730 static struct irq_domain *get_ir_irq_domain(struct irq_alloc_info *info)
3731 {
3732 struct amd_iommu *iommu;
3733 int devid;
3734
3735 if (!info)
3736 return NULL;
3737
3738 devid = get_devid(info);
3739 if (devid >= 0) {
3740 iommu = amd_iommu_rlookup_table[devid];
3741 if (iommu)
3742 return iommu->ir_domain;
3743 }
3744
3745 return NULL;
3746 }
3747
3748 static struct irq_domain *get_irq_domain(struct irq_alloc_info *info)
3749 {
3750 struct amd_iommu *iommu;
3751 int devid;
3752
3753 if (!info)
3754 return NULL;
3755
3756 switch (info->type) {
3757 case X86_IRQ_ALLOC_TYPE_MSI:
3758 case X86_IRQ_ALLOC_TYPE_MSIX:
3759 devid = get_device_id(&info->msi_dev->dev);
3760 if (devid >= 0) {
3761 iommu = amd_iommu_rlookup_table[devid];
3762 if (iommu)
3763 return iommu->msi_domain;
3764 }
3765 break;
3766 default:
3767 break;
3768 }
3769
3770 return NULL;
3771 }
3772
3773 struct irq_remap_ops amd_iommu_irq_ops = {
3774 .prepare = amd_iommu_prepare,
3775 .enable = amd_iommu_enable,
3776 .disable = amd_iommu_disable,
3777 .reenable = amd_iommu_reenable,
3778 .enable_faulting = amd_iommu_enable_faulting,
3779 .get_ir_irq_domain = get_ir_irq_domain,
3780 .get_irq_domain = get_irq_domain,
3781 };
3782
3783 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
3784 struct irq_cfg *irq_cfg,
3785 struct irq_alloc_info *info,
3786 int devid, int index, int sub_handle)
3787 {
3788 struct irq_2_irte *irte_info = &data->irq_2_irte;
3789 struct msi_msg *msg = &data->msi_entry;
3790 union irte *irte = &data->irte_entry;
3791 struct IO_APIC_route_entry *entry;
3792
3793 data->irq_2_irte.devid = devid;
3794 data->irq_2_irte.index = index + sub_handle;
3795
3796 /* Setup IRTE for IOMMU */
3797 irte->val = 0;
3798 irte->fields.vector = irq_cfg->vector;
3799 irte->fields.int_type = apic->irq_delivery_mode;
3800 irte->fields.destination = irq_cfg->dest_apicid;
3801 irte->fields.dm = apic->irq_dest_mode;
3802 irte->fields.valid = 1;
3803
3804 switch (info->type) {
3805 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3806 /* Setup IOAPIC entry */
3807 entry = info->ioapic_entry;
3808 info->ioapic_entry = NULL;
3809 memset(entry, 0, sizeof(*entry));
3810 entry->vector = index;
3811 entry->mask = 0;
3812 entry->trigger = info->ioapic_trigger;
3813 entry->polarity = info->ioapic_polarity;
3814 /* Mask level triggered irqs. */
3815 if (info->ioapic_trigger)
3816 entry->mask = 1;
3817 break;
3818
3819 case X86_IRQ_ALLOC_TYPE_HPET:
3820 case X86_IRQ_ALLOC_TYPE_MSI:
3821 case X86_IRQ_ALLOC_TYPE_MSIX:
3822 msg->address_hi = MSI_ADDR_BASE_HI;
3823 msg->address_lo = MSI_ADDR_BASE_LO;
3824 msg->data = irte_info->index;
3825 break;
3826
3827 default:
3828 BUG_ON(1);
3829 break;
3830 }
3831 }
3832
3833 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
3834 unsigned int nr_irqs, void *arg)
3835 {
3836 struct irq_alloc_info *info = arg;
3837 struct irq_data *irq_data;
3838 struct amd_ir_data *data;
3839 struct irq_cfg *cfg;
3840 int i, ret, devid;
3841 int index = -1;
3842
3843 if (!info)
3844 return -EINVAL;
3845 if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_MSI &&
3846 info->type != X86_IRQ_ALLOC_TYPE_MSIX)
3847 return -EINVAL;
3848
3849 /*
3850 * With IRQ remapping enabled, don't need contiguous CPU vectors
3851 * to support multiple MSI interrupts.
3852 */
3853 if (info->type == X86_IRQ_ALLOC_TYPE_MSI)
3854 info->flags &= ~X86_IRQ_ALLOC_CONTIGUOUS_VECTORS;
3855
3856 devid = get_devid(info);
3857 if (devid < 0)
3858 return -EINVAL;
3859
3860 ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
3861 if (ret < 0)
3862 return ret;
3863
3864 if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
3865 if (get_irq_table(devid, true))
3866 index = info->ioapic_pin;
3867 else
3868 ret = -ENOMEM;
3869 } else {
3870 index = alloc_irq_index(devid, nr_irqs);
3871 }
3872 if (index < 0) {
3873 pr_warn("Failed to allocate IRTE\n");
3874 goto out_free_parent;
3875 }
3876
3877 for (i = 0; i < nr_irqs; i++) {
3878 irq_data = irq_domain_get_irq_data(domain, virq + i);
3879 cfg = irqd_cfg(irq_data);
3880 if (!irq_data || !cfg) {
3881 ret = -EINVAL;
3882 goto out_free_data;
3883 }
3884
3885 ret = -ENOMEM;
3886 data = kzalloc(sizeof(*data), GFP_KERNEL);
3887 if (!data)
3888 goto out_free_data;
3889
3890 irq_data->hwirq = (devid << 16) + i;
3891 irq_data->chip_data = data;
3892 irq_data->chip = &amd_ir_chip;
3893 irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
3894 irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
3895 }
3896
3897 return 0;
3898
3899 out_free_data:
3900 for (i--; i >= 0; i--) {
3901 irq_data = irq_domain_get_irq_data(domain, virq + i);
3902 if (irq_data)
3903 kfree(irq_data->chip_data);
3904 }
3905 for (i = 0; i < nr_irqs; i++)
3906 free_irte(devid, index + i);
3907 out_free_parent:
3908 irq_domain_free_irqs_common(domain, virq, nr_irqs);
3909 return ret;
3910 }
3911
3912 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
3913 unsigned int nr_irqs)
3914 {
3915 struct irq_2_irte *irte_info;
3916 struct irq_data *irq_data;
3917 struct amd_ir_data *data;
3918 int i;
3919
3920 for (i = 0; i < nr_irqs; i++) {
3921 irq_data = irq_domain_get_irq_data(domain, virq + i);
3922 if (irq_data && irq_data->chip_data) {
3923 data = irq_data->chip_data;
3924 irte_info = &data->irq_2_irte;
3925 free_irte(irte_info->devid, irte_info->index);
3926 kfree(data);
3927 }
3928 }
3929 irq_domain_free_irqs_common(domain, virq, nr_irqs);
3930 }
3931
3932 static void irq_remapping_activate(struct irq_domain *domain,
3933 struct irq_data *irq_data)
3934 {
3935 struct amd_ir_data *data = irq_data->chip_data;
3936 struct irq_2_irte *irte_info = &data->irq_2_irte;
3937
3938 modify_irte(irte_info->devid, irte_info->index, data->irte_entry);
3939 }
3940
3941 static void irq_remapping_deactivate(struct irq_domain *domain,
3942 struct irq_data *irq_data)
3943 {
3944 struct amd_ir_data *data = irq_data->chip_data;
3945 struct irq_2_irte *irte_info = &data->irq_2_irte;
3946 union irte entry;
3947
3948 entry.val = 0;
3949 modify_irte(irte_info->devid, irte_info->index, data->irte_entry);
3950 }
3951
3952 static struct irq_domain_ops amd_ir_domain_ops = {
3953 .alloc = irq_remapping_alloc,
3954 .free = irq_remapping_free,
3955 .activate = irq_remapping_activate,
3956 .deactivate = irq_remapping_deactivate,
3957 };
3958
3959 static int amd_ir_set_affinity(struct irq_data *data,
3960 const struct cpumask *mask, bool force)
3961 {
3962 struct amd_ir_data *ir_data = data->chip_data;
3963 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
3964 struct irq_cfg *cfg = irqd_cfg(data);
3965 struct irq_data *parent = data->parent_data;
3966 int ret;
3967
3968 ret = parent->chip->irq_set_affinity(parent, mask, force);
3969 if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
3970 return ret;
3971
3972 /*
3973 * Atomically updates the IRTE with the new destination, vector
3974 * and flushes the interrupt entry cache.
3975 */
3976 ir_data->irte_entry.fields.vector = cfg->vector;
3977 ir_data->irte_entry.fields.destination = cfg->dest_apicid;
3978 modify_irte(irte_info->devid, irte_info->index, ir_data->irte_entry);
3979
3980 /*
3981 * After this point, all the interrupts will start arriving
3982 * at the new destination. So, time to cleanup the previous
3983 * vector allocation.
3984 */
3985 send_cleanup_vector(cfg);
3986
3987 return IRQ_SET_MASK_OK_DONE;
3988 }
3989
3990 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
3991 {
3992 struct amd_ir_data *ir_data = irq_data->chip_data;
3993
3994 *msg = ir_data->msi_entry;
3995 }
3996
3997 static struct irq_chip amd_ir_chip = {
3998 .irq_ack = ir_ack_apic_edge,
3999 .irq_set_affinity = amd_ir_set_affinity,
4000 .irq_compose_msi_msg = ir_compose_msi_msg,
4001 };
4002
4003 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
4004 {
4005 iommu->ir_domain = irq_domain_add_tree(NULL, &amd_ir_domain_ops, iommu);
4006 if (!iommu->ir_domain)
4007 return -ENOMEM;
4008
4009 iommu->ir_domain->parent = arch_get_ir_parent_domain();
4010 iommu->msi_domain = arch_create_msi_irq_domain(iommu->ir_domain);
4011
4012 return 0;
4013 }
4014 #endif
Linux kernel - Deliverables
This page took 0.174668 seconds and 5 git commands to generate.