2 * Copyright (C) 2007-2009 Advanced Micro Devices, Inc.
3 * Author: Joerg Roedel <joerg.roedel@amd.com>
4 * Leo Duran <leo.duran@amd.com>
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.
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.
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
20 #include <linux/pci.h>
21 #include <linux/gfp.h>
22 #include <linux/bitops.h>
23 #include <linux/debugfs.h>
24 #include <linux/scatterlist.h>
25 #include <linux/dma-mapping.h>
26 #include <linux/iommu-helper.h>
27 #include <linux/iommu.h>
28 #include <asm/proto.h>
29 #include <asm/iommu.h>
31 #include <asm/amd_iommu_proto.h>
32 #include <asm/amd_iommu_types.h>
33 #include <asm/amd_iommu.h>
35 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
37 #define EXIT_LOOP_COUNT 10000000
39 static DEFINE_RWLOCK(amd_iommu_devtable_lock);
41 /* A list of preallocated protection domains */
42 static LIST_HEAD(iommu_pd_list);
43 static DEFINE_SPINLOCK(iommu_pd_list_lock);
46 * Domain for untranslated devices - only allocated
47 * if iommu=pt passed on kernel cmd line.
49 static struct protection_domain *pt_domain;
51 static struct iommu_ops amd_iommu_ops;
54 * general struct to manage commands send to an IOMMU
60 static void reset_iommu_command_buffer(struct amd_iommu *iommu);
61 static void update_domain(struct protection_domain *domain);
63 /****************************************************************************
67 ****************************************************************************/
69 static inline u16 get_device_id(struct device *dev)
71 struct pci_dev *pdev = to_pci_dev(dev);
73 return calc_devid(pdev->bus->number, pdev->devfn);
76 static struct iommu_dev_data *get_dev_data(struct device *dev)
78 return dev->archdata.iommu;
82 * In this function the list of preallocated protection domains is traversed to
83 * find the domain for a specific device
85 static struct dma_ops_domain *find_protection_domain(u16 devid)
87 struct dma_ops_domain *entry, *ret = NULL;
89 u16 alias = amd_iommu_alias_table[devid];
91 if (list_empty(&iommu_pd_list))
94 spin_lock_irqsave(&iommu_pd_list_lock, flags);
96 list_for_each_entry(entry, &iommu_pd_list, list) {
97 if (entry->target_dev == devid ||
98 entry->target_dev == alias) {
104 spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
110 * This function checks if the driver got a valid device from the caller to
111 * avoid dereferencing invalid pointers.
113 static bool check_device(struct device *dev)
117 if (!dev || !dev->dma_mask)
120 /* No device or no PCI device */
121 if (!dev || dev->bus != &pci_bus_type)
124 devid = get_device_id(dev);
126 /* Out of our scope? */
127 if (devid > amd_iommu_last_bdf)
130 if (amd_iommu_rlookup_table[devid] == NULL)
136 static int iommu_init_device(struct device *dev)
138 struct iommu_dev_data *dev_data;
139 struct pci_dev *pdev;
142 if (dev->archdata.iommu)
145 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
149 devid = get_device_id(dev);
150 alias = amd_iommu_alias_table[devid];
151 pdev = pci_get_bus_and_slot(PCI_BUS(alias), alias & 0xff);
153 dev_data->alias = &pdev->dev;
155 dev->archdata.iommu = dev_data;
161 static void iommu_uninit_device(struct device *dev)
163 kfree(dev->archdata.iommu);
165 #ifdef CONFIG_AMD_IOMMU_STATS
168 * Initialization code for statistics collection
171 DECLARE_STATS_COUNTER(compl_wait);
172 DECLARE_STATS_COUNTER(cnt_map_single);
173 DECLARE_STATS_COUNTER(cnt_unmap_single);
174 DECLARE_STATS_COUNTER(cnt_map_sg);
175 DECLARE_STATS_COUNTER(cnt_unmap_sg);
176 DECLARE_STATS_COUNTER(cnt_alloc_coherent);
177 DECLARE_STATS_COUNTER(cnt_free_coherent);
178 DECLARE_STATS_COUNTER(cross_page);
179 DECLARE_STATS_COUNTER(domain_flush_single);
180 DECLARE_STATS_COUNTER(domain_flush_all);
181 DECLARE_STATS_COUNTER(alloced_io_mem);
182 DECLARE_STATS_COUNTER(total_map_requests);
184 static struct dentry *stats_dir;
185 static struct dentry *de_fflush;
187 static void amd_iommu_stats_add(struct __iommu_counter *cnt)
189 if (stats_dir == NULL)
192 cnt->dent = debugfs_create_u64(cnt->name, 0444, stats_dir,
196 static void amd_iommu_stats_init(void)
198 stats_dir = debugfs_create_dir("amd-iommu", NULL);
199 if (stats_dir == NULL)
202 de_fflush = debugfs_create_bool("fullflush", 0444, stats_dir,
203 (u32 *)&amd_iommu_unmap_flush);
205 amd_iommu_stats_add(&compl_wait);
206 amd_iommu_stats_add(&cnt_map_single);
207 amd_iommu_stats_add(&cnt_unmap_single);
208 amd_iommu_stats_add(&cnt_map_sg);
209 amd_iommu_stats_add(&cnt_unmap_sg);
210 amd_iommu_stats_add(&cnt_alloc_coherent);
211 amd_iommu_stats_add(&cnt_free_coherent);
212 amd_iommu_stats_add(&cross_page);
213 amd_iommu_stats_add(&domain_flush_single);
214 amd_iommu_stats_add(&domain_flush_all);
215 amd_iommu_stats_add(&alloced_io_mem);
216 amd_iommu_stats_add(&total_map_requests);
221 /****************************************************************************
223 * Interrupt handling functions
225 ****************************************************************************/
227 static void dump_dte_entry(u16 devid)
231 for (i = 0; i < 8; ++i)
232 pr_err("AMD-Vi: DTE[%d]: %08x\n", i,
233 amd_iommu_dev_table[devid].data[i]);
236 static void dump_command(unsigned long phys_addr)
238 struct iommu_cmd *cmd = phys_to_virt(phys_addr);
241 for (i = 0; i < 4; ++i)
242 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
245 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
248 int type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
249 int devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
250 int domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
251 int flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
252 u64 address = (u64)(((u64)event[3]) << 32) | event[2];
254 printk(KERN_ERR "AMD-Vi: Event logged [");
257 case EVENT_TYPE_ILL_DEV:
258 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
259 "address=0x%016llx flags=0x%04x]\n",
260 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
262 dump_dte_entry(devid);
264 case EVENT_TYPE_IO_FAULT:
265 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
266 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
267 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
268 domid, address, flags);
270 case EVENT_TYPE_DEV_TAB_ERR:
271 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
272 "address=0x%016llx flags=0x%04x]\n",
273 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
276 case EVENT_TYPE_PAGE_TAB_ERR:
277 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
278 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
279 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
280 domid, address, flags);
282 case EVENT_TYPE_ILL_CMD:
283 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
284 reset_iommu_command_buffer(iommu);
285 dump_command(address);
287 case EVENT_TYPE_CMD_HARD_ERR:
288 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
289 "flags=0x%04x]\n", address, flags);
291 case EVENT_TYPE_IOTLB_INV_TO:
292 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
293 "address=0x%016llx]\n",
294 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
297 case EVENT_TYPE_INV_DEV_REQ:
298 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
299 "address=0x%016llx flags=0x%04x]\n",
300 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
304 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
308 static void iommu_poll_events(struct amd_iommu *iommu)
313 spin_lock_irqsave(&iommu->lock, flags);
315 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
316 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
318 while (head != tail) {
319 iommu_print_event(iommu, iommu->evt_buf + head);
320 head = (head + EVENT_ENTRY_SIZE) % iommu->evt_buf_size;
323 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
325 spin_unlock_irqrestore(&iommu->lock, flags);
328 irqreturn_t amd_iommu_int_handler(int irq, void *data)
330 struct amd_iommu *iommu;
332 for_each_iommu(iommu)
333 iommu_poll_events(iommu);
338 /****************************************************************************
340 * IOMMU command queuing functions
342 ****************************************************************************/
345 * Writes the command to the IOMMUs command buffer and informs the
346 * hardware about the new command. Must be called with iommu->lock held.
348 static int __iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
353 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
354 target = iommu->cmd_buf + tail;
355 memcpy_toio(target, cmd, sizeof(*cmd));
356 tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
357 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
360 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
366 * General queuing function for commands. Takes iommu->lock and calls
367 * __iommu_queue_command().
369 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
374 spin_lock_irqsave(&iommu->lock, flags);
375 ret = __iommu_queue_command(iommu, cmd);
377 iommu->need_sync = true;
378 spin_unlock_irqrestore(&iommu->lock, flags);
384 * This function waits until an IOMMU has completed a completion
387 static void __iommu_wait_for_completion(struct amd_iommu *iommu)
393 INC_STATS_COUNTER(compl_wait);
395 while (!ready && (i < EXIT_LOOP_COUNT)) {
397 /* wait for the bit to become one */
398 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
399 ready = status & MMIO_STATUS_COM_WAIT_INT_MASK;
402 /* set bit back to zero */
403 status &= ~MMIO_STATUS_COM_WAIT_INT_MASK;
404 writel(status, iommu->mmio_base + MMIO_STATUS_OFFSET);
406 if (unlikely(i == EXIT_LOOP_COUNT)) {
407 spin_unlock(&iommu->lock);
408 reset_iommu_command_buffer(iommu);
409 spin_lock(&iommu->lock);
414 * This function queues a completion wait command into the command
417 static int __iommu_completion_wait(struct amd_iommu *iommu)
419 struct iommu_cmd cmd;
421 memset(&cmd, 0, sizeof(cmd));
422 cmd.data[0] = CMD_COMPL_WAIT_INT_MASK;
423 CMD_SET_TYPE(&cmd, CMD_COMPL_WAIT);
425 return __iommu_queue_command(iommu, &cmd);
429 * This function is called whenever we need to ensure that the IOMMU has
430 * completed execution of all commands we sent. It sends a
431 * COMPLETION_WAIT command and waits for it to finish. The IOMMU informs
432 * us about that by writing a value to a physical address we pass with
435 static int iommu_completion_wait(struct amd_iommu *iommu)
440 spin_lock_irqsave(&iommu->lock, flags);
442 if (!iommu->need_sync)
445 ret = __iommu_completion_wait(iommu);
447 iommu->need_sync = false;
452 __iommu_wait_for_completion(iommu);
455 spin_unlock_irqrestore(&iommu->lock, flags);
460 static void iommu_flush_complete(struct protection_domain *domain)
464 for (i = 0; i < amd_iommus_present; ++i) {
465 if (!domain->dev_iommu[i])
469 * Devices of this domain are behind this IOMMU
470 * We need to wait for completion of all commands.
472 iommu_completion_wait(amd_iommus[i]);
477 * Command send function for invalidating a device table entry
479 static int iommu_queue_inv_dev_entry(struct amd_iommu *iommu, u16 devid)
481 struct iommu_cmd cmd;
484 BUG_ON(iommu == NULL);
486 memset(&cmd, 0, sizeof(cmd));
487 CMD_SET_TYPE(&cmd, CMD_INV_DEV_ENTRY);
490 ret = iommu_queue_command(iommu, &cmd);
495 static void __iommu_build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
496 u16 domid, int pde, int s)
498 memset(cmd, 0, sizeof(*cmd));
499 address &= PAGE_MASK;
500 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
501 cmd->data[1] |= domid;
502 cmd->data[2] = lower_32_bits(address);
503 cmd->data[3] = upper_32_bits(address);
504 if (s) /* size bit - we flush more than one 4kb page */
505 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
506 if (pde) /* PDE bit - we wan't flush everything not only the PTEs */
507 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
511 * Generic command send function for invalidaing TLB entries
513 static int iommu_queue_inv_iommu_pages(struct amd_iommu *iommu,
514 u64 address, u16 domid, int pde, int s)
516 struct iommu_cmd cmd;
519 __iommu_build_inv_iommu_pages(&cmd, address, domid, pde, s);
521 ret = iommu_queue_command(iommu, &cmd);
527 * TLB invalidation function which is called from the mapping functions.
528 * It invalidates a single PTE if the range to flush is within a single
529 * page. Otherwise it flushes the whole TLB of the IOMMU.
531 static void __iommu_flush_pages(struct protection_domain *domain,
532 u64 address, size_t size, int pde)
535 unsigned long pages = iommu_num_pages(address, size, PAGE_SIZE);
537 address &= PAGE_MASK;
541 * If we have to flush more than one page, flush all
542 * TLB entries for this domain
544 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
549 for (i = 0; i < amd_iommus_present; ++i) {
550 if (!domain->dev_iommu[i])
554 * Devices of this domain are behind this IOMMU
555 * We need a TLB flush
557 iommu_queue_inv_iommu_pages(amd_iommus[i], address,
564 static void iommu_flush_pages(struct protection_domain *domain,
565 u64 address, size_t size)
567 __iommu_flush_pages(domain, address, size, 0);
570 /* Flush the whole IO/TLB for a given protection domain */
571 static void iommu_flush_tlb(struct protection_domain *domain)
573 __iommu_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
576 /* Flush the whole IO/TLB for a given protection domain - including PDE */
577 static void iommu_flush_tlb_pde(struct protection_domain *domain)
579 __iommu_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
583 * This function flushes all domains that have devices on the given IOMMU
585 static void flush_all_domains_on_iommu(struct amd_iommu *iommu)
587 u64 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
588 struct protection_domain *domain;
591 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
593 list_for_each_entry(domain, &amd_iommu_pd_list, list) {
594 if (domain->dev_iommu[iommu->index] == 0)
597 spin_lock(&domain->lock);
598 iommu_queue_inv_iommu_pages(iommu, address, domain->id, 1, 1);
599 iommu_flush_complete(domain);
600 spin_unlock(&domain->lock);
603 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
607 * This function uses heavy locking and may disable irqs for some time. But
608 * this is no issue because it is only called during resume.
610 void amd_iommu_flush_all_domains(void)
612 struct protection_domain *domain;
615 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
617 list_for_each_entry(domain, &amd_iommu_pd_list, list) {
618 spin_lock(&domain->lock);
619 iommu_flush_tlb_pde(domain);
620 iommu_flush_complete(domain);
621 spin_unlock(&domain->lock);
624 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
627 static void flush_all_devices_for_iommu(struct amd_iommu *iommu)
631 for (i = 0; i <= amd_iommu_last_bdf; ++i) {
632 if (iommu != amd_iommu_rlookup_table[i])
635 iommu_queue_inv_dev_entry(iommu, i);
636 iommu_completion_wait(iommu);
640 static void flush_devices_by_domain(struct protection_domain *domain)
642 struct amd_iommu *iommu;
645 for (i = 0; i <= amd_iommu_last_bdf; ++i) {
646 if ((domain == NULL && amd_iommu_pd_table[i] == NULL) ||
647 (amd_iommu_pd_table[i] != domain))
650 iommu = amd_iommu_rlookup_table[i];
654 iommu_queue_inv_dev_entry(iommu, i);
655 iommu_completion_wait(iommu);
659 static void reset_iommu_command_buffer(struct amd_iommu *iommu)
661 pr_err("AMD-Vi: Resetting IOMMU command buffer\n");
663 if (iommu->reset_in_progress)
664 panic("AMD-Vi: ILLEGAL_COMMAND_ERROR while resetting command buffer\n");
666 iommu->reset_in_progress = true;
668 amd_iommu_reset_cmd_buffer(iommu);
669 flush_all_devices_for_iommu(iommu);
670 flush_all_domains_on_iommu(iommu);
672 iommu->reset_in_progress = false;
675 void amd_iommu_flush_all_devices(void)
677 flush_devices_by_domain(NULL);
680 /****************************************************************************
682 * The functions below are used the create the page table mappings for
683 * unity mapped regions.
685 ****************************************************************************/
688 * This function is used to add another level to an IO page table. Adding
689 * another level increases the size of the address space by 9 bits to a size up
692 static bool increase_address_space(struct protection_domain *domain,
697 if (domain->mode == PAGE_MODE_6_LEVEL)
698 /* address space already 64 bit large */
701 pte = (void *)get_zeroed_page(gfp);
705 *pte = PM_LEVEL_PDE(domain->mode,
706 virt_to_phys(domain->pt_root));
707 domain->pt_root = pte;
709 domain->updated = true;
714 static u64 *alloc_pte(struct protection_domain *domain,
715 unsigned long address,
723 while (address > PM_LEVEL_SIZE(domain->mode))
724 increase_address_space(domain, gfp);
726 level = domain->mode - 1;
727 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
729 while (level > end_lvl) {
730 if (!IOMMU_PTE_PRESENT(*pte)) {
731 page = (u64 *)get_zeroed_page(gfp);
734 *pte = PM_LEVEL_PDE(level, virt_to_phys(page));
739 pte = IOMMU_PTE_PAGE(*pte);
741 if (pte_page && level == end_lvl)
744 pte = &pte[PM_LEVEL_INDEX(level, address)];
751 * This function checks if there is a PTE for a given dma address. If
752 * there is one, it returns the pointer to it.
754 static u64 *fetch_pte(struct protection_domain *domain,
755 unsigned long address, int map_size)
760 level = domain->mode - 1;
761 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
763 while (level > map_size) {
764 if (!IOMMU_PTE_PRESENT(*pte))
769 pte = IOMMU_PTE_PAGE(*pte);
770 pte = &pte[PM_LEVEL_INDEX(level, address)];
772 if ((PM_PTE_LEVEL(*pte) == 0) && level != map_size) {
782 * Generic mapping functions. It maps a physical address into a DMA
783 * address space. It allocates the page table pages if necessary.
784 * In the future it can be extended to a generic mapping function
785 * supporting all features of AMD IOMMU page tables like level skipping
786 * and full 64 bit address spaces.
788 static int iommu_map_page(struct protection_domain *dom,
789 unsigned long bus_addr,
790 unsigned long phys_addr,
796 bus_addr = PAGE_ALIGN(bus_addr);
797 phys_addr = PAGE_ALIGN(phys_addr);
799 BUG_ON(!PM_ALIGNED(map_size, bus_addr));
800 BUG_ON(!PM_ALIGNED(map_size, phys_addr));
802 if (!(prot & IOMMU_PROT_MASK))
805 pte = alloc_pte(dom, bus_addr, map_size, NULL, GFP_KERNEL);
807 if (IOMMU_PTE_PRESENT(*pte))
810 __pte = phys_addr | IOMMU_PTE_P;
811 if (prot & IOMMU_PROT_IR)
812 __pte |= IOMMU_PTE_IR;
813 if (prot & IOMMU_PROT_IW)
814 __pte |= IOMMU_PTE_IW;
823 static void iommu_unmap_page(struct protection_domain *dom,
824 unsigned long bus_addr, int map_size)
826 u64 *pte = fetch_pte(dom, bus_addr, map_size);
833 * This function checks if a specific unity mapping entry is needed for
834 * this specific IOMMU.
836 static int iommu_for_unity_map(struct amd_iommu *iommu,
837 struct unity_map_entry *entry)
841 for (i = entry->devid_start; i <= entry->devid_end; ++i) {
842 bdf = amd_iommu_alias_table[i];
843 if (amd_iommu_rlookup_table[bdf] == iommu)
851 * This function actually applies the mapping to the page table of the
854 static int dma_ops_unity_map(struct dma_ops_domain *dma_dom,
855 struct unity_map_entry *e)
860 for (addr = e->address_start; addr < e->address_end;
862 ret = iommu_map_page(&dma_dom->domain, addr, addr, e->prot,
867 * if unity mapping is in aperture range mark the page
868 * as allocated in the aperture
870 if (addr < dma_dom->aperture_size)
871 __set_bit(addr >> PAGE_SHIFT,
872 dma_dom->aperture[0]->bitmap);
879 * Init the unity mappings for a specific IOMMU in the system
881 * Basically iterates over all unity mapping entries and applies them to
882 * the default domain DMA of that IOMMU if necessary.
884 static int iommu_init_unity_mappings(struct amd_iommu *iommu)
886 struct unity_map_entry *entry;
889 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
890 if (!iommu_for_unity_map(iommu, entry))
892 ret = dma_ops_unity_map(iommu->default_dom, entry);
901 * Inits the unity mappings required for a specific device
903 static int init_unity_mappings_for_device(struct dma_ops_domain *dma_dom,
906 struct unity_map_entry *e;
909 list_for_each_entry(e, &amd_iommu_unity_map, list) {
910 if (!(devid >= e->devid_start && devid <= e->devid_end))
912 ret = dma_ops_unity_map(dma_dom, e);
920 /****************************************************************************
922 * The next functions belong to the address allocator for the dma_ops
923 * interface functions. They work like the allocators in the other IOMMU
924 * drivers. Its basically a bitmap which marks the allocated pages in
925 * the aperture. Maybe it could be enhanced in the future to a more
926 * efficient allocator.
928 ****************************************************************************/
931 * The address allocator core functions.
933 * called with domain->lock held
937 * Used to reserve address ranges in the aperture (e.g. for exclusion
940 static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
941 unsigned long start_page,
944 unsigned int i, last_page = dom->aperture_size >> PAGE_SHIFT;
946 if (start_page + pages > last_page)
947 pages = last_page - start_page;
949 for (i = start_page; i < start_page + pages; ++i) {
950 int index = i / APERTURE_RANGE_PAGES;
951 int page = i % APERTURE_RANGE_PAGES;
952 __set_bit(page, dom->aperture[index]->bitmap);
957 * This function is used to add a new aperture range to an existing
958 * aperture in case of dma_ops domain allocation or address allocation
961 static int alloc_new_range(struct dma_ops_domain *dma_dom,
962 bool populate, gfp_t gfp)
964 int index = dma_dom->aperture_size >> APERTURE_RANGE_SHIFT;
965 struct amd_iommu *iommu;
968 #ifdef CONFIG_IOMMU_STRESS
972 if (index >= APERTURE_MAX_RANGES)
975 dma_dom->aperture[index] = kzalloc(sizeof(struct aperture_range), gfp);
976 if (!dma_dom->aperture[index])
979 dma_dom->aperture[index]->bitmap = (void *)get_zeroed_page(gfp);
980 if (!dma_dom->aperture[index]->bitmap)
983 dma_dom->aperture[index]->offset = dma_dom->aperture_size;
986 unsigned long address = dma_dom->aperture_size;
987 int i, num_ptes = APERTURE_RANGE_PAGES / 512;
990 for (i = 0; i < num_ptes; ++i) {
991 pte = alloc_pte(&dma_dom->domain, address, PM_MAP_4k,
996 dma_dom->aperture[index]->pte_pages[i] = pte_page;
998 address += APERTURE_RANGE_SIZE / 64;
1002 dma_dom->aperture_size += APERTURE_RANGE_SIZE;
1004 /* Intialize the exclusion range if necessary */
1005 for_each_iommu(iommu) {
1006 if (iommu->exclusion_start &&
1007 iommu->exclusion_start >= dma_dom->aperture[index]->offset
1008 && iommu->exclusion_start < dma_dom->aperture_size) {
1009 unsigned long startpage;
1010 int pages = iommu_num_pages(iommu->exclusion_start,
1011 iommu->exclusion_length,
1013 startpage = iommu->exclusion_start >> PAGE_SHIFT;
1014 dma_ops_reserve_addresses(dma_dom, startpage, pages);
1019 * Check for areas already mapped as present in the new aperture
1020 * range and mark those pages as reserved in the allocator. Such
1021 * mappings may already exist as a result of requested unity
1022 * mappings for devices.
1024 for (i = dma_dom->aperture[index]->offset;
1025 i < dma_dom->aperture_size;
1027 u64 *pte = fetch_pte(&dma_dom->domain, i, PM_MAP_4k);
1028 if (!pte || !IOMMU_PTE_PRESENT(*pte))
1031 dma_ops_reserve_addresses(dma_dom, i << PAGE_SHIFT, 1);
1034 update_domain(&dma_dom->domain);
1039 update_domain(&dma_dom->domain);
1041 free_page((unsigned long)dma_dom->aperture[index]->bitmap);
1043 kfree(dma_dom->aperture[index]);
1044 dma_dom->aperture[index] = NULL;
1049 static unsigned long dma_ops_area_alloc(struct device *dev,
1050 struct dma_ops_domain *dom,
1052 unsigned long align_mask,
1054 unsigned long start)
1056 unsigned long next_bit = dom->next_address % APERTURE_RANGE_SIZE;
1057 int max_index = dom->aperture_size >> APERTURE_RANGE_SHIFT;
1058 int i = start >> APERTURE_RANGE_SHIFT;
1059 unsigned long boundary_size;
1060 unsigned long address = -1;
1061 unsigned long limit;
1063 next_bit >>= PAGE_SHIFT;
1065 boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
1066 PAGE_SIZE) >> PAGE_SHIFT;
1068 for (;i < max_index; ++i) {
1069 unsigned long offset = dom->aperture[i]->offset >> PAGE_SHIFT;
1071 if (dom->aperture[i]->offset >= dma_mask)
1074 limit = iommu_device_max_index(APERTURE_RANGE_PAGES, offset,
1075 dma_mask >> PAGE_SHIFT);
1077 address = iommu_area_alloc(dom->aperture[i]->bitmap,
1078 limit, next_bit, pages, 0,
1079 boundary_size, align_mask);
1080 if (address != -1) {
1081 address = dom->aperture[i]->offset +
1082 (address << PAGE_SHIFT);
1083 dom->next_address = address + (pages << PAGE_SHIFT);
1093 static unsigned long dma_ops_alloc_addresses(struct device *dev,
1094 struct dma_ops_domain *dom,
1096 unsigned long align_mask,
1099 unsigned long address;
1101 #ifdef CONFIG_IOMMU_STRESS
1102 dom->next_address = 0;
1103 dom->need_flush = true;
1106 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1107 dma_mask, dom->next_address);
1109 if (address == -1) {
1110 dom->next_address = 0;
1111 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1113 dom->need_flush = true;
1116 if (unlikely(address == -1))
1117 address = DMA_ERROR_CODE;
1119 WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
1125 * The address free function.
1127 * called with domain->lock held
1129 static void dma_ops_free_addresses(struct dma_ops_domain *dom,
1130 unsigned long address,
1133 unsigned i = address >> APERTURE_RANGE_SHIFT;
1134 struct aperture_range *range = dom->aperture[i];
1136 BUG_ON(i >= APERTURE_MAX_RANGES || range == NULL);
1138 #ifdef CONFIG_IOMMU_STRESS
1143 if (address >= dom->next_address)
1144 dom->need_flush = true;
1146 address = (address % APERTURE_RANGE_SIZE) >> PAGE_SHIFT;
1148 iommu_area_free(range->bitmap, address, pages);
1152 /****************************************************************************
1154 * The next functions belong to the domain allocation. A domain is
1155 * allocated for every IOMMU as the default domain. If device isolation
1156 * is enabled, every device get its own domain. The most important thing
1157 * about domains is the page table mapping the DMA address space they
1160 ****************************************************************************/
1163 * This function adds a protection domain to the global protection domain list
1165 static void add_domain_to_list(struct protection_domain *domain)
1167 unsigned long flags;
1169 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1170 list_add(&domain->list, &amd_iommu_pd_list);
1171 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1175 * This function removes a protection domain to the global
1176 * protection domain list
1178 static void del_domain_from_list(struct protection_domain *domain)
1180 unsigned long flags;
1182 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1183 list_del(&domain->list);
1184 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1187 static u16 domain_id_alloc(void)
1189 unsigned long flags;
1192 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1193 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1195 if (id > 0 && id < MAX_DOMAIN_ID)
1196 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1199 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1204 static void domain_id_free(int id)
1206 unsigned long flags;
1208 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1209 if (id > 0 && id < MAX_DOMAIN_ID)
1210 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1211 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1214 static void free_pagetable(struct protection_domain *domain)
1219 p1 = domain->pt_root;
1224 for (i = 0; i < 512; ++i) {
1225 if (!IOMMU_PTE_PRESENT(p1[i]))
1228 p2 = IOMMU_PTE_PAGE(p1[i]);
1229 for (j = 0; j < 512; ++j) {
1230 if (!IOMMU_PTE_PRESENT(p2[j]))
1232 p3 = IOMMU_PTE_PAGE(p2[j]);
1233 free_page((unsigned long)p3);
1236 free_page((unsigned long)p2);
1239 free_page((unsigned long)p1);
1241 domain->pt_root = NULL;
1245 * Free a domain, only used if something went wrong in the
1246 * allocation path and we need to free an already allocated page table
1248 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1255 del_domain_from_list(&dom->domain);
1257 free_pagetable(&dom->domain);
1259 for (i = 0; i < APERTURE_MAX_RANGES; ++i) {
1260 if (!dom->aperture[i])
1262 free_page((unsigned long)dom->aperture[i]->bitmap);
1263 kfree(dom->aperture[i]);
1270 * Allocates a new protection domain usable for the dma_ops functions.
1271 * It also intializes the page table and the address allocator data
1272 * structures required for the dma_ops interface
1274 static struct dma_ops_domain *dma_ops_domain_alloc(void)
1276 struct dma_ops_domain *dma_dom;
1278 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1282 spin_lock_init(&dma_dom->domain.lock);
1284 dma_dom->domain.id = domain_id_alloc();
1285 if (dma_dom->domain.id == 0)
1287 dma_dom->domain.mode = PAGE_MODE_2_LEVEL;
1288 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1289 dma_dom->domain.flags = PD_DMA_OPS_MASK;
1290 dma_dom->domain.priv = dma_dom;
1291 if (!dma_dom->domain.pt_root)
1294 dma_dom->need_flush = false;
1295 dma_dom->target_dev = 0xffff;
1297 add_domain_to_list(&dma_dom->domain);
1299 if (alloc_new_range(dma_dom, true, GFP_KERNEL))
1303 * mark the first page as allocated so we never return 0 as
1304 * a valid dma-address. So we can use 0 as error value
1306 dma_dom->aperture[0]->bitmap[0] = 1;
1307 dma_dom->next_address = 0;
1313 dma_ops_domain_free(dma_dom);
1319 * little helper function to check whether a given protection domain is a
1322 static bool dma_ops_domain(struct protection_domain *domain)
1324 return domain->flags & PD_DMA_OPS_MASK;
1327 static void set_dte_entry(u16 devid, struct protection_domain *domain)
1329 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
1330 u64 pte_root = virt_to_phys(domain->pt_root);
1332 BUG_ON(amd_iommu_pd_table[devid] != NULL);
1334 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
1335 << DEV_ENTRY_MODE_SHIFT;
1336 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
1338 amd_iommu_dev_table[devid].data[2] = domain->id;
1339 amd_iommu_dev_table[devid].data[1] = upper_32_bits(pte_root);
1340 amd_iommu_dev_table[devid].data[0] = lower_32_bits(pte_root);
1342 amd_iommu_pd_table[devid] = domain;
1344 /* Do reference counting */
1345 domain->dev_iommu[iommu->index] += 1;
1346 domain->dev_cnt += 1;
1348 /* Flush the changes DTE entry */
1349 iommu_queue_inv_dev_entry(iommu, devid);
1352 static void clear_dte_entry(u16 devid)
1354 struct protection_domain *domain = amd_iommu_pd_table[devid];
1355 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
1357 BUG_ON(domain == NULL);
1359 /* remove domain from the lookup table */
1360 amd_iommu_pd_table[devid] = NULL;
1362 /* remove entry from the device table seen by the hardware */
1363 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
1364 amd_iommu_dev_table[devid].data[1] = 0;
1365 amd_iommu_dev_table[devid].data[2] = 0;
1367 amd_iommu_apply_erratum_63(devid);
1369 /* decrease reference counters */
1370 domain->dev_iommu[iommu->index] -= 1;
1371 domain->dev_cnt -= 1;
1373 iommu_queue_inv_dev_entry(iommu, devid);
1377 * If a device is not yet associated with a domain, this function does
1378 * assigns it visible for the hardware
1380 static int __attach_device(struct device *dev,
1381 struct protection_domain *domain)
1383 struct iommu_dev_data *dev_data, *alias_data;
1386 devid = get_device_id(dev);
1387 alias = amd_iommu_alias_table[devid];
1388 dev_data = get_dev_data(dev);
1389 alias_data = get_dev_data(dev_data->alias);
1394 spin_lock(&domain->lock);
1396 /* Some sanity checks */
1397 if (alias_data->domain != NULL &&
1398 alias_data->domain != domain)
1401 if (dev_data->domain != NULL &&
1402 dev_data->domain != domain)
1405 /* Do real assignment */
1406 if (alias != devid &&
1407 alias_data->domain == NULL) {
1408 alias_data->domain = domain;
1409 set_dte_entry(alias, domain);
1412 if (dev_data->domain == NULL) {
1413 dev_data->domain = domain;
1414 set_dte_entry(devid, domain);
1418 spin_unlock(&domain->lock);
1424 * If a device is not yet associated with a domain, this function does
1425 * assigns it visible for the hardware
1427 static int attach_device(struct device *dev,
1428 struct protection_domain *domain)
1430 unsigned long flags;
1433 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1434 ret = __attach_device(dev, domain);
1435 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1438 * We might boot into a crash-kernel here. The crashed kernel
1439 * left the caches in the IOMMU dirty. So we have to flush
1440 * here to evict all dirty stuff.
1442 iommu_flush_tlb_pde(domain);
1448 * Removes a device from a protection domain (unlocked)
1450 static void __detach_device(struct device *dev)
1452 u16 devid = get_device_id(dev);
1453 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
1454 struct iommu_dev_data *dev_data = get_dev_data(dev);
1458 clear_dte_entry(devid);
1459 dev_data->domain = NULL;
1462 * If we run in passthrough mode the device must be assigned to the
1463 * passthrough domain if it is detached from any other domain
1465 if (iommu_pass_through)
1466 __attach_device(dev, pt_domain);
1470 * Removes a device from a protection domain (with devtable_lock held)
1472 static void detach_device(struct device *dev)
1474 unsigned long flags;
1476 /* lock device table */
1477 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1478 __detach_device(dev);
1479 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1483 * Find out the protection domain structure for a given PCI device. This
1484 * will give us the pointer to the page table root for example.
1486 static struct protection_domain *domain_for_device(struct device *dev)
1488 struct protection_domain *dom;
1489 struct iommu_dev_data *dev_data, *alias_data;
1490 unsigned long flags;
1493 devid = get_device_id(dev);
1494 alias = amd_iommu_alias_table[devid];
1495 dev_data = get_dev_data(dev);
1496 alias_data = get_dev_data(dev_data->alias);
1500 read_lock_irqsave(&amd_iommu_devtable_lock, flags);
1501 dom = dev_data->domain;
1503 alias_data->domain != NULL) {
1504 __attach_device(dev, alias_data->domain);
1505 dom = alias_data->domain;
1508 read_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1513 static int device_change_notifier(struct notifier_block *nb,
1514 unsigned long action, void *data)
1516 struct device *dev = data;
1518 struct protection_domain *domain;
1519 struct dma_ops_domain *dma_domain;
1520 struct amd_iommu *iommu;
1521 unsigned long flags;
1523 if (!check_device(dev))
1526 devid = get_device_id(dev);
1527 iommu = amd_iommu_rlookup_table[devid];
1530 case BUS_NOTIFY_UNBOUND_DRIVER:
1532 domain = domain_for_device(dev);
1536 if (iommu_pass_through)
1540 case BUS_NOTIFY_ADD_DEVICE:
1542 iommu_init_device(dev);
1544 domain = domain_for_device(dev);
1546 /* allocate a protection domain if a device is added */
1547 dma_domain = find_protection_domain(devid);
1550 dma_domain = dma_ops_domain_alloc();
1553 dma_domain->target_dev = devid;
1555 spin_lock_irqsave(&iommu_pd_list_lock, flags);
1556 list_add_tail(&dma_domain->list, &iommu_pd_list);
1557 spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
1560 case BUS_NOTIFY_DEL_DEVICE:
1562 iommu_uninit_device(dev);
1568 iommu_queue_inv_dev_entry(iommu, devid);
1569 iommu_completion_wait(iommu);
1575 static struct notifier_block device_nb = {
1576 .notifier_call = device_change_notifier,
1579 /*****************************************************************************
1581 * The next functions belong to the dma_ops mapping/unmapping code.
1583 *****************************************************************************/
1586 * In the dma_ops path we only have the struct device. This function
1587 * finds the corresponding IOMMU, the protection domain and the
1588 * requestor id for a given device.
1589 * If the device is not yet associated with a domain this is also done
1592 static struct protection_domain *get_domain(struct device *dev)
1594 struct protection_domain *domain;
1595 struct dma_ops_domain *dma_dom;
1596 u16 devid = get_device_id(dev);
1598 if (!check_device(dev))
1599 return ERR_PTR(-EINVAL);
1601 domain = domain_for_device(dev);
1602 if (domain != NULL && !dma_ops_domain(domain))
1603 return ERR_PTR(-EBUSY);
1608 /* Device not bount yet - bind it */
1609 dma_dom = find_protection_domain(devid);
1611 dma_dom = amd_iommu_rlookup_table[devid]->default_dom;
1612 attach_device(dev, &dma_dom->domain);
1613 DUMP_printk("Using protection domain %d for device %s\n",
1614 dma_dom->domain.id, dev_name(dev));
1616 return &dma_dom->domain;
1619 static void update_device_table(struct protection_domain *domain)
1621 unsigned long flags;
1624 for (i = 0; i <= amd_iommu_last_bdf; ++i) {
1625 if (amd_iommu_pd_table[i] != domain)
1627 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1628 set_dte_entry(i, domain);
1629 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1633 static void update_domain(struct protection_domain *domain)
1635 if (!domain->updated)
1638 update_device_table(domain);
1639 flush_devices_by_domain(domain);
1640 iommu_flush_tlb_pde(domain);
1642 domain->updated = false;
1646 * This function fetches the PTE for a given address in the aperture
1648 static u64* dma_ops_get_pte(struct dma_ops_domain *dom,
1649 unsigned long address)
1651 struct aperture_range *aperture;
1652 u64 *pte, *pte_page;
1654 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
1658 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
1660 pte = alloc_pte(&dom->domain, address, PM_MAP_4k, &pte_page,
1662 aperture->pte_pages[APERTURE_PAGE_INDEX(address)] = pte_page;
1664 pte += PM_LEVEL_INDEX(0, address);
1666 update_domain(&dom->domain);
1672 * This is the generic map function. It maps one 4kb page at paddr to
1673 * the given address in the DMA address space for the domain.
1675 static dma_addr_t dma_ops_domain_map(struct dma_ops_domain *dom,
1676 unsigned long address,
1682 WARN_ON(address > dom->aperture_size);
1686 pte = dma_ops_get_pte(dom, address);
1688 return DMA_ERROR_CODE;
1690 __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
1692 if (direction == DMA_TO_DEVICE)
1693 __pte |= IOMMU_PTE_IR;
1694 else if (direction == DMA_FROM_DEVICE)
1695 __pte |= IOMMU_PTE_IW;
1696 else if (direction == DMA_BIDIRECTIONAL)
1697 __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
1703 return (dma_addr_t)address;
1707 * The generic unmapping function for on page in the DMA address space.
1709 static void dma_ops_domain_unmap(struct dma_ops_domain *dom,
1710 unsigned long address)
1712 struct aperture_range *aperture;
1715 if (address >= dom->aperture_size)
1718 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
1722 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
1726 pte += PM_LEVEL_INDEX(0, address);
1734 * This function contains common code for mapping of a physically
1735 * contiguous memory region into DMA address space. It is used by all
1736 * mapping functions provided with this IOMMU driver.
1737 * Must be called with the domain lock held.
1739 static dma_addr_t __map_single(struct device *dev,
1740 struct dma_ops_domain *dma_dom,
1747 dma_addr_t offset = paddr & ~PAGE_MASK;
1748 dma_addr_t address, start, ret;
1750 unsigned long align_mask = 0;
1753 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
1756 INC_STATS_COUNTER(total_map_requests);
1759 INC_STATS_COUNTER(cross_page);
1762 align_mask = (1UL << get_order(size)) - 1;
1765 address = dma_ops_alloc_addresses(dev, dma_dom, pages, align_mask,
1767 if (unlikely(address == DMA_ERROR_CODE)) {
1769 * setting next_address here will let the address
1770 * allocator only scan the new allocated range in the
1771 * first run. This is a small optimization.
1773 dma_dom->next_address = dma_dom->aperture_size;
1775 if (alloc_new_range(dma_dom, false, GFP_ATOMIC))
1779 * aperture was sucessfully enlarged by 128 MB, try
1786 for (i = 0; i < pages; ++i) {
1787 ret = dma_ops_domain_map(dma_dom, start, paddr, dir);
1788 if (ret == DMA_ERROR_CODE)
1796 ADD_STATS_COUNTER(alloced_io_mem, size);
1798 if (unlikely(dma_dom->need_flush && !amd_iommu_unmap_flush)) {
1799 iommu_flush_tlb(&dma_dom->domain);
1800 dma_dom->need_flush = false;
1801 } else if (unlikely(amd_iommu_np_cache))
1802 iommu_flush_pages(&dma_dom->domain, address, size);
1809 for (--i; i >= 0; --i) {
1811 dma_ops_domain_unmap(dma_dom, start);
1814 dma_ops_free_addresses(dma_dom, address, pages);
1816 return DMA_ERROR_CODE;
1820 * Does the reverse of the __map_single function. Must be called with
1821 * the domain lock held too
1823 static void __unmap_single(struct dma_ops_domain *dma_dom,
1824 dma_addr_t dma_addr,
1828 dma_addr_t i, start;
1831 if ((dma_addr == DMA_ERROR_CODE) ||
1832 (dma_addr + size > dma_dom->aperture_size))
1835 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
1836 dma_addr &= PAGE_MASK;
1839 for (i = 0; i < pages; ++i) {
1840 dma_ops_domain_unmap(dma_dom, start);
1844 SUB_STATS_COUNTER(alloced_io_mem, size);
1846 dma_ops_free_addresses(dma_dom, dma_addr, pages);
1848 if (amd_iommu_unmap_flush || dma_dom->need_flush) {
1849 iommu_flush_pages(&dma_dom->domain, dma_addr, size);
1850 dma_dom->need_flush = false;
1855 * The exported map_single function for dma_ops.
1857 static dma_addr_t map_page(struct device *dev, struct page *page,
1858 unsigned long offset, size_t size,
1859 enum dma_data_direction dir,
1860 struct dma_attrs *attrs)
1862 unsigned long flags;
1863 struct protection_domain *domain;
1866 phys_addr_t paddr = page_to_phys(page) + offset;
1868 INC_STATS_COUNTER(cnt_map_single);
1870 domain = get_domain(dev);
1871 if (PTR_ERR(domain) == -EINVAL)
1872 return (dma_addr_t)paddr;
1873 else if (IS_ERR(domain))
1874 return DMA_ERROR_CODE;
1876 dma_mask = *dev->dma_mask;
1878 spin_lock_irqsave(&domain->lock, flags);
1880 addr = __map_single(dev, domain->priv, paddr, size, dir, false,
1882 if (addr == DMA_ERROR_CODE)
1885 iommu_flush_complete(domain);
1888 spin_unlock_irqrestore(&domain->lock, flags);
1894 * The exported unmap_single function for dma_ops.
1896 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
1897 enum dma_data_direction dir, struct dma_attrs *attrs)
1899 unsigned long flags;
1900 struct protection_domain *domain;
1902 INC_STATS_COUNTER(cnt_unmap_single);
1904 domain = get_domain(dev);
1908 spin_lock_irqsave(&domain->lock, flags);
1910 __unmap_single(domain->priv, dma_addr, size, dir);
1912 iommu_flush_complete(domain);
1914 spin_unlock_irqrestore(&domain->lock, flags);
1918 * This is a special map_sg function which is used if we should map a
1919 * device which is not handled by an AMD IOMMU in the system.
1921 static int map_sg_no_iommu(struct device *dev, struct scatterlist *sglist,
1922 int nelems, int dir)
1924 struct scatterlist *s;
1927 for_each_sg(sglist, s, nelems, i) {
1928 s->dma_address = (dma_addr_t)sg_phys(s);
1929 s->dma_length = s->length;
1936 * The exported map_sg function for dma_ops (handles scatter-gather
1939 static int map_sg(struct device *dev, struct scatterlist *sglist,
1940 int nelems, enum dma_data_direction dir,
1941 struct dma_attrs *attrs)
1943 unsigned long flags;
1944 struct protection_domain *domain;
1946 struct scatterlist *s;
1948 int mapped_elems = 0;
1951 INC_STATS_COUNTER(cnt_map_sg);
1953 domain = get_domain(dev);
1954 if (PTR_ERR(domain) == -EINVAL)
1955 return map_sg_no_iommu(dev, sglist, nelems, dir);
1956 else if (IS_ERR(domain))
1959 dma_mask = *dev->dma_mask;
1961 spin_lock_irqsave(&domain->lock, flags);
1963 for_each_sg(sglist, s, nelems, i) {
1966 s->dma_address = __map_single(dev, domain->priv,
1967 paddr, s->length, dir, false,
1970 if (s->dma_address) {
1971 s->dma_length = s->length;
1977 iommu_flush_complete(domain);
1980 spin_unlock_irqrestore(&domain->lock, flags);
1982 return mapped_elems;
1984 for_each_sg(sglist, s, mapped_elems, i) {
1986 __unmap_single(domain->priv, s->dma_address,
1987 s->dma_length, dir);
1988 s->dma_address = s->dma_length = 0;
1997 * The exported map_sg function for dma_ops (handles scatter-gather
2000 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2001 int nelems, enum dma_data_direction dir,
2002 struct dma_attrs *attrs)
2004 unsigned long flags;
2005 struct protection_domain *domain;
2006 struct scatterlist *s;
2009 INC_STATS_COUNTER(cnt_unmap_sg);
2011 domain = get_domain(dev);
2015 spin_lock_irqsave(&domain->lock, flags);
2017 for_each_sg(sglist, s, nelems, i) {
2018 __unmap_single(domain->priv, s->dma_address,
2019 s->dma_length, dir);
2020 s->dma_address = s->dma_length = 0;
2023 iommu_flush_complete(domain);
2025 spin_unlock_irqrestore(&domain->lock, flags);
2029 * The exported alloc_coherent function for dma_ops.
2031 static void *alloc_coherent(struct device *dev, size_t size,
2032 dma_addr_t *dma_addr, gfp_t flag)
2034 unsigned long flags;
2036 struct protection_domain *domain;
2038 u64 dma_mask = dev->coherent_dma_mask;
2040 INC_STATS_COUNTER(cnt_alloc_coherent);
2042 domain = get_domain(dev);
2043 if (PTR_ERR(domain) == -EINVAL) {
2044 virt_addr = (void *)__get_free_pages(flag, get_order(size));
2045 *dma_addr = __pa(virt_addr);
2047 } else if (IS_ERR(domain))
2050 dma_mask = dev->coherent_dma_mask;
2051 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2054 virt_addr = (void *)__get_free_pages(flag, get_order(size));
2058 paddr = virt_to_phys(virt_addr);
2061 dma_mask = *dev->dma_mask;
2063 spin_lock_irqsave(&domain->lock, flags);
2065 *dma_addr = __map_single(dev, domain->priv, paddr,
2066 size, DMA_BIDIRECTIONAL, true, dma_mask);
2068 if (*dma_addr == DMA_ERROR_CODE) {
2069 spin_unlock_irqrestore(&domain->lock, flags);
2073 iommu_flush_complete(domain);
2075 spin_unlock_irqrestore(&domain->lock, flags);
2081 free_pages((unsigned long)virt_addr, get_order(size));
2087 * The exported free_coherent function for dma_ops.
2089 static void free_coherent(struct device *dev, size_t size,
2090 void *virt_addr, dma_addr_t dma_addr)
2092 unsigned long flags;
2093 struct protection_domain *domain;
2095 INC_STATS_COUNTER(cnt_free_coherent);
2097 domain = get_domain(dev);
2101 spin_lock_irqsave(&domain->lock, flags);
2103 __unmap_single(domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
2105 iommu_flush_complete(domain);
2107 spin_unlock_irqrestore(&domain->lock, flags);
2110 free_pages((unsigned long)virt_addr, get_order(size));
2114 * This function is called by the DMA layer to find out if we can handle a
2115 * particular device. It is part of the dma_ops.
2117 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2119 return check_device(dev);
2123 * The function for pre-allocating protection domains.
2125 * If the driver core informs the DMA layer if a driver grabs a device
2126 * we don't need to preallocate the protection domains anymore.
2127 * For now we have to.
2129 static void prealloc_protection_domains(void)
2131 struct pci_dev *dev = NULL;
2132 struct dma_ops_domain *dma_dom;
2135 while ((dev = pci_get_device(PCI_ANY_ID, PCI_ANY_ID, dev)) != NULL) {
2137 /* Do we handle this device? */
2138 if (!check_device(&dev->dev))
2141 iommu_init_device(&dev->dev);
2143 /* Is there already any domain for it? */
2144 if (domain_for_device(&dev->dev))
2147 devid = get_device_id(&dev->dev);
2149 dma_dom = dma_ops_domain_alloc();
2152 init_unity_mappings_for_device(dma_dom, devid);
2153 dma_dom->target_dev = devid;
2155 attach_device(&dev->dev, &dma_dom->domain);
2157 list_add_tail(&dma_dom->list, &iommu_pd_list);
2161 static struct dma_map_ops amd_iommu_dma_ops = {
2162 .alloc_coherent = alloc_coherent,
2163 .free_coherent = free_coherent,
2164 .map_page = map_page,
2165 .unmap_page = unmap_page,
2167 .unmap_sg = unmap_sg,
2168 .dma_supported = amd_iommu_dma_supported,
2172 * The function which clues the AMD IOMMU driver into dma_ops.
2174 int __init amd_iommu_init_dma_ops(void)
2176 struct amd_iommu *iommu;
2180 * first allocate a default protection domain for every IOMMU we
2181 * found in the system. Devices not assigned to any other
2182 * protection domain will be assigned to the default one.
2184 for_each_iommu(iommu) {
2185 iommu->default_dom = dma_ops_domain_alloc();
2186 if (iommu->default_dom == NULL)
2188 iommu->default_dom->domain.flags |= PD_DEFAULT_MASK;
2189 ret = iommu_init_unity_mappings(iommu);
2195 * Pre-allocate the protection domains for each device.
2197 prealloc_protection_domains();
2201 #ifdef CONFIG_GART_IOMMU
2202 gart_iommu_aperture_disabled = 1;
2203 gart_iommu_aperture = 0;
2206 /* Make the driver finally visible to the drivers */
2207 dma_ops = &amd_iommu_dma_ops;
2209 register_iommu(&amd_iommu_ops);
2211 bus_register_notifier(&pci_bus_type, &device_nb);
2213 amd_iommu_stats_init();
2219 for_each_iommu(iommu) {
2220 if (iommu->default_dom)
2221 dma_ops_domain_free(iommu->default_dom);
2227 /*****************************************************************************
2229 * The following functions belong to the exported interface of AMD IOMMU
2231 * This interface allows access to lower level functions of the IOMMU
2232 * like protection domain handling and assignement of devices to domains
2233 * which is not possible with the dma_ops interface.
2235 *****************************************************************************/
2237 static void cleanup_domain(struct protection_domain *domain)
2239 unsigned long flags;
2242 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2244 for (devid = 0; devid <= amd_iommu_last_bdf; ++devid)
2245 if (amd_iommu_pd_table[devid] == domain)
2246 clear_dte_entry(devid);
2248 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2251 static void protection_domain_free(struct protection_domain *domain)
2256 del_domain_from_list(domain);
2259 domain_id_free(domain->id);
2264 static struct protection_domain *protection_domain_alloc(void)
2266 struct protection_domain *domain;
2268 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2272 spin_lock_init(&domain->lock);
2273 domain->id = domain_id_alloc();
2277 add_domain_to_list(domain);
2287 static int amd_iommu_domain_init(struct iommu_domain *dom)
2289 struct protection_domain *domain;
2291 domain = protection_domain_alloc();
2295 domain->mode = PAGE_MODE_3_LEVEL;
2296 domain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2297 if (!domain->pt_root)
2305 protection_domain_free(domain);
2310 static void amd_iommu_domain_destroy(struct iommu_domain *dom)
2312 struct protection_domain *domain = dom->priv;
2317 if (domain->dev_cnt > 0)
2318 cleanup_domain(domain);
2320 BUG_ON(domain->dev_cnt != 0);
2322 free_pagetable(domain);
2324 domain_id_free(domain->id);
2331 static void amd_iommu_detach_device(struct iommu_domain *dom,
2334 struct iommu_dev_data *dev_data = dev->archdata.iommu;
2335 struct amd_iommu *iommu;
2338 if (!check_device(dev))
2341 devid = get_device_id(dev);
2343 if (dev_data->domain != NULL)
2346 iommu = amd_iommu_rlookup_table[devid];
2350 iommu_queue_inv_dev_entry(iommu, devid);
2351 iommu_completion_wait(iommu);
2354 static int amd_iommu_attach_device(struct iommu_domain *dom,
2357 struct protection_domain *domain = dom->priv;
2358 struct iommu_dev_data *dev_data;
2359 struct amd_iommu *iommu;
2363 if (!check_device(dev))
2366 dev_data = dev->archdata.iommu;
2368 devid = get_device_id(dev);
2370 iommu = amd_iommu_rlookup_table[devid];
2374 if (dev_data->domain)
2377 ret = attach_device(dev, domain);
2379 iommu_completion_wait(iommu);
2384 static int amd_iommu_map_range(struct iommu_domain *dom,
2385 unsigned long iova, phys_addr_t paddr,
2386 size_t size, int iommu_prot)
2388 struct protection_domain *domain = dom->priv;
2389 unsigned long i, npages = iommu_num_pages(paddr, size, PAGE_SIZE);
2393 if (iommu_prot & IOMMU_READ)
2394 prot |= IOMMU_PROT_IR;
2395 if (iommu_prot & IOMMU_WRITE)
2396 prot |= IOMMU_PROT_IW;
2401 for (i = 0; i < npages; ++i) {
2402 ret = iommu_map_page(domain, iova, paddr, prot, PM_MAP_4k);
2413 static void amd_iommu_unmap_range(struct iommu_domain *dom,
2414 unsigned long iova, size_t size)
2417 struct protection_domain *domain = dom->priv;
2418 unsigned long i, npages = iommu_num_pages(iova, size, PAGE_SIZE);
2422 for (i = 0; i < npages; ++i) {
2423 iommu_unmap_page(domain, iova, PM_MAP_4k);
2427 iommu_flush_tlb_pde(domain);
2430 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
2433 struct protection_domain *domain = dom->priv;
2434 unsigned long offset = iova & ~PAGE_MASK;
2438 pte = fetch_pte(domain, iova, PM_MAP_4k);
2440 if (!pte || !IOMMU_PTE_PRESENT(*pte))
2443 paddr = *pte & IOMMU_PAGE_MASK;
2449 static int amd_iommu_domain_has_cap(struct iommu_domain *domain,
2455 static struct iommu_ops amd_iommu_ops = {
2456 .domain_init = amd_iommu_domain_init,
2457 .domain_destroy = amd_iommu_domain_destroy,
2458 .attach_dev = amd_iommu_attach_device,
2459 .detach_dev = amd_iommu_detach_device,
2460 .map = amd_iommu_map_range,
2461 .unmap = amd_iommu_unmap_range,
2462 .iova_to_phys = amd_iommu_iova_to_phys,
2463 .domain_has_cap = amd_iommu_domain_has_cap,
2466 /*****************************************************************************
2468 * The next functions do a basic initialization of IOMMU for pass through
2471 * In passthrough mode the IOMMU is initialized and enabled but not used for
2472 * DMA-API translation.
2474 *****************************************************************************/
2476 int __init amd_iommu_init_passthrough(void)
2478 struct amd_iommu *iommu;
2479 struct pci_dev *dev = NULL;
2482 /* allocate passthroug domain */
2483 pt_domain = protection_domain_alloc();
2487 pt_domain->mode |= PAGE_MODE_NONE;
2489 while ((dev = pci_get_device(PCI_ANY_ID, PCI_ANY_ID, dev)) != NULL) {
2491 if (!check_device(&dev->dev))
2494 devid = get_device_id(&dev->dev);
2496 iommu = amd_iommu_rlookup_table[devid];
2500 attach_device(&dev->dev, pt_domain);
2503 pr_info("AMD-Vi: Initialized for Passthrough Mode\n");