2 * Copyright (C) 2007-2008 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/scatterlist.h>
24 #include <linux/iommu-helper.h>
25 #include <asm/proto.h>
26 #include <asm/iommu.h>
28 #include <asm/amd_iommu_types.h>
29 #include <asm/amd_iommu.h>
31 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
33 #define EXIT_LOOP_COUNT 10000000
35 static DEFINE_RWLOCK(amd_iommu_devtable_lock);
37 /* A list of preallocated protection domains */
38 static LIST_HEAD(iommu_pd_list);
39 static DEFINE_SPINLOCK(iommu_pd_list_lock);
42 * general struct to manage commands send to an IOMMU
48 static int dma_ops_unity_map(struct dma_ops_domain *dma_dom,
49 struct unity_map_entry *e);
51 /* returns !0 if the IOMMU is caching non-present entries in its TLB */
52 static int iommu_has_npcache(struct amd_iommu *iommu)
54 return iommu->cap & (1UL << IOMMU_CAP_NPCACHE);
57 /****************************************************************************
59 * Interrupt handling functions
61 ****************************************************************************/
63 static void iommu_print_event(void *__evt)
66 int type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
67 int devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
68 int domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
69 int flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
70 u64 address = (u64)(((u64)event[3]) << 32) | event[2];
72 printk(KERN_ERR "AMD IOMMU: Event logged [");
75 case EVENT_TYPE_ILL_DEV:
76 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
77 "address=0x%016llx flags=0x%04x]\n",
78 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
81 case EVENT_TYPE_IO_FAULT:
82 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
83 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
84 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
85 domid, address, flags);
87 case EVENT_TYPE_DEV_TAB_ERR:
88 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
89 "address=0x%016llx flags=0x%04x]\n",
90 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
93 case EVENT_TYPE_PAGE_TAB_ERR:
94 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
95 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
96 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
97 domid, address, flags);
99 case EVENT_TYPE_ILL_CMD:
100 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
102 case EVENT_TYPE_CMD_HARD_ERR:
103 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
104 "flags=0x%04x]\n", address, flags);
106 case EVENT_TYPE_IOTLB_INV_TO:
107 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
108 "address=0x%016llx]\n",
109 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
112 case EVENT_TYPE_INV_DEV_REQ:
113 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
114 "address=0x%016llx flags=0x%04x]\n",
115 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
119 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
123 static void iommu_poll_events(struct amd_iommu *iommu)
128 spin_lock_irqsave(&iommu->lock, flags);
130 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
131 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
133 while (head != tail) {
134 iommu_print_event(iommu->evt_buf + head);
135 head = (head + EVENT_ENTRY_SIZE) % iommu->evt_buf_size;
138 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
140 spin_unlock_irqrestore(&iommu->lock, flags);
143 irqreturn_t amd_iommu_int_handler(int irq, void *data)
145 struct amd_iommu *iommu;
147 list_for_each_entry(iommu, &amd_iommu_list, list)
148 iommu_poll_events(iommu);
153 /****************************************************************************
155 * IOMMU command queuing functions
157 ****************************************************************************/
160 * Writes the command to the IOMMUs command buffer and informs the
161 * hardware about the new command. Must be called with iommu->lock held.
163 static int __iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
168 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
169 target = iommu->cmd_buf + tail;
170 memcpy_toio(target, cmd, sizeof(*cmd));
171 tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
172 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
175 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
181 * General queuing function for commands. Takes iommu->lock and calls
182 * __iommu_queue_command().
184 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
189 spin_lock_irqsave(&iommu->lock, flags);
190 ret = __iommu_queue_command(iommu, cmd);
192 iommu->need_sync = 1;
193 spin_unlock_irqrestore(&iommu->lock, flags);
199 * This function is called whenever we need to ensure that the IOMMU has
200 * completed execution of all commands we sent. It sends a
201 * COMPLETION_WAIT command and waits for it to finish. The IOMMU informs
202 * us about that by writing a value to a physical address we pass with
205 static int iommu_completion_wait(struct amd_iommu *iommu)
207 int ret = 0, ready = 0;
209 struct iommu_cmd cmd;
210 unsigned long flags, i = 0;
212 memset(&cmd, 0, sizeof(cmd));
213 cmd.data[0] = CMD_COMPL_WAIT_INT_MASK;
214 CMD_SET_TYPE(&cmd, CMD_COMPL_WAIT);
216 spin_lock_irqsave(&iommu->lock, flags);
218 if (!iommu->need_sync)
221 iommu->need_sync = 0;
223 ret = __iommu_queue_command(iommu, &cmd);
228 while (!ready && (i < EXIT_LOOP_COUNT)) {
230 /* wait for the bit to become one */
231 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
232 ready = status & MMIO_STATUS_COM_WAIT_INT_MASK;
235 /* set bit back to zero */
236 status &= ~MMIO_STATUS_COM_WAIT_INT_MASK;
237 writel(status, iommu->mmio_base + MMIO_STATUS_OFFSET);
239 if (unlikely(i == EXIT_LOOP_COUNT))
240 panic("AMD IOMMU: Completion wait loop failed\n");
243 spin_unlock_irqrestore(&iommu->lock, flags);
249 * Command send function for invalidating a device table entry
251 static int iommu_queue_inv_dev_entry(struct amd_iommu *iommu, u16 devid)
253 struct iommu_cmd cmd;
256 BUG_ON(iommu == NULL);
258 memset(&cmd, 0, sizeof(cmd));
259 CMD_SET_TYPE(&cmd, CMD_INV_DEV_ENTRY);
262 ret = iommu_queue_command(iommu, &cmd);
268 * Generic command send function for invalidaing TLB entries
270 static int iommu_queue_inv_iommu_pages(struct amd_iommu *iommu,
271 u64 address, u16 domid, int pde, int s)
273 struct iommu_cmd cmd;
276 memset(&cmd, 0, sizeof(cmd));
277 address &= PAGE_MASK;
278 CMD_SET_TYPE(&cmd, CMD_INV_IOMMU_PAGES);
279 cmd.data[1] |= domid;
280 cmd.data[2] = lower_32_bits(address);
281 cmd.data[3] = upper_32_bits(address);
282 if (s) /* size bit - we flush more than one 4kb page */
283 cmd.data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
284 if (pde) /* PDE bit - we wan't flush everything not only the PTEs */
285 cmd.data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
287 ret = iommu_queue_command(iommu, &cmd);
293 * TLB invalidation function which is called from the mapping functions.
294 * It invalidates a single PTE if the range to flush is within a single
295 * page. Otherwise it flushes the whole TLB of the IOMMU.
297 static int iommu_flush_pages(struct amd_iommu *iommu, u16 domid,
298 u64 address, size_t size)
301 unsigned pages = iommu_num_pages(address, size, PAGE_SIZE);
303 address &= PAGE_MASK;
307 * If we have to flush more than one page, flush all
308 * TLB entries for this domain
310 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
314 iommu_queue_inv_iommu_pages(iommu, address, domid, 0, s);
319 /* Flush the whole IO/TLB for a given protection domain */
320 static void iommu_flush_tlb(struct amd_iommu *iommu, u16 domid)
322 u64 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
324 iommu_queue_inv_iommu_pages(iommu, address, domid, 0, 1);
327 /****************************************************************************
329 * The functions below are used the create the page table mappings for
330 * unity mapped regions.
332 ****************************************************************************/
335 * Generic mapping functions. It maps a physical address into a DMA
336 * address space. It allocates the page table pages if necessary.
337 * In the future it can be extended to a generic mapping function
338 * supporting all features of AMD IOMMU page tables like level skipping
339 * and full 64 bit address spaces.
341 static int iommu_map_page(struct protection_domain *dom,
342 unsigned long bus_addr,
343 unsigned long phys_addr,
346 u64 __pte, *pte, *page;
348 bus_addr = PAGE_ALIGN(bus_addr);
349 phys_addr = PAGE_ALIGN(phys_addr);
351 /* only support 512GB address spaces for now */
352 if (bus_addr > IOMMU_MAP_SIZE_L3 || !(prot & IOMMU_PROT_MASK))
355 pte = &dom->pt_root[IOMMU_PTE_L2_INDEX(bus_addr)];
357 if (!IOMMU_PTE_PRESENT(*pte)) {
358 page = (u64 *)get_zeroed_page(GFP_KERNEL);
361 *pte = IOMMU_L2_PDE(virt_to_phys(page));
364 pte = IOMMU_PTE_PAGE(*pte);
365 pte = &pte[IOMMU_PTE_L1_INDEX(bus_addr)];
367 if (!IOMMU_PTE_PRESENT(*pte)) {
368 page = (u64 *)get_zeroed_page(GFP_KERNEL);
371 *pte = IOMMU_L1_PDE(virt_to_phys(page));
374 pte = IOMMU_PTE_PAGE(*pte);
375 pte = &pte[IOMMU_PTE_L0_INDEX(bus_addr)];
377 if (IOMMU_PTE_PRESENT(*pte))
380 __pte = phys_addr | IOMMU_PTE_P;
381 if (prot & IOMMU_PROT_IR)
382 __pte |= IOMMU_PTE_IR;
383 if (prot & IOMMU_PROT_IW)
384 __pte |= IOMMU_PTE_IW;
392 * This function checks if a specific unity mapping entry is needed for
393 * this specific IOMMU.
395 static int iommu_for_unity_map(struct amd_iommu *iommu,
396 struct unity_map_entry *entry)
400 for (i = entry->devid_start; i <= entry->devid_end; ++i) {
401 bdf = amd_iommu_alias_table[i];
402 if (amd_iommu_rlookup_table[bdf] == iommu)
410 * Init the unity mappings for a specific IOMMU in the system
412 * Basically iterates over all unity mapping entries and applies them to
413 * the default domain DMA of that IOMMU if necessary.
415 static int iommu_init_unity_mappings(struct amd_iommu *iommu)
417 struct unity_map_entry *entry;
420 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
421 if (!iommu_for_unity_map(iommu, entry))
423 ret = dma_ops_unity_map(iommu->default_dom, entry);
432 * This function actually applies the mapping to the page table of the
435 static int dma_ops_unity_map(struct dma_ops_domain *dma_dom,
436 struct unity_map_entry *e)
441 for (addr = e->address_start; addr < e->address_end;
443 ret = iommu_map_page(&dma_dom->domain, addr, addr, e->prot);
447 * if unity mapping is in aperture range mark the page
448 * as allocated in the aperture
450 if (addr < dma_dom->aperture_size)
451 __set_bit(addr >> PAGE_SHIFT, dma_dom->bitmap);
458 * Inits the unity mappings required for a specific device
460 static int init_unity_mappings_for_device(struct dma_ops_domain *dma_dom,
463 struct unity_map_entry *e;
466 list_for_each_entry(e, &amd_iommu_unity_map, list) {
467 if (!(devid >= e->devid_start && devid <= e->devid_end))
469 ret = dma_ops_unity_map(dma_dom, e);
477 /****************************************************************************
479 * The next functions belong to the address allocator for the dma_ops
480 * interface functions. They work like the allocators in the other IOMMU
481 * drivers. Its basically a bitmap which marks the allocated pages in
482 * the aperture. Maybe it could be enhanced in the future to a more
483 * efficient allocator.
485 ****************************************************************************/
488 * The address allocator core function.
490 * called with domain->lock held
492 static unsigned long dma_ops_alloc_addresses(struct device *dev,
493 struct dma_ops_domain *dom,
495 unsigned long align_mask,
499 unsigned long address;
500 unsigned long boundary_size;
502 boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
503 PAGE_SIZE) >> PAGE_SHIFT;
504 limit = iommu_device_max_index(dom->aperture_size >> PAGE_SHIFT, 0,
505 dma_mask >> PAGE_SHIFT);
507 if (dom->next_bit >= limit) {
509 dom->need_flush = true;
512 address = iommu_area_alloc(dom->bitmap, limit, dom->next_bit, pages,
513 0 , boundary_size, align_mask);
515 address = iommu_area_alloc(dom->bitmap, limit, 0, pages,
516 0, boundary_size, align_mask);
517 dom->need_flush = true;
520 if (likely(address != -1)) {
521 dom->next_bit = address + pages;
522 address <<= PAGE_SHIFT;
524 address = bad_dma_address;
526 WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
532 * The address free function.
534 * called with domain->lock held
536 static void dma_ops_free_addresses(struct dma_ops_domain *dom,
537 unsigned long address,
540 address >>= PAGE_SHIFT;
541 iommu_area_free(dom->bitmap, address, pages);
543 if (address >= dom->next_bit)
544 dom->need_flush = true;
547 /****************************************************************************
549 * The next functions belong to the domain allocation. A domain is
550 * allocated for every IOMMU as the default domain. If device isolation
551 * is enabled, every device get its own domain. The most important thing
552 * about domains is the page table mapping the DMA address space they
555 ****************************************************************************/
557 static u16 domain_id_alloc(void)
562 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
563 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
565 if (id > 0 && id < MAX_DOMAIN_ID)
566 __set_bit(id, amd_iommu_pd_alloc_bitmap);
569 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
575 * Used to reserve address ranges in the aperture (e.g. for exclusion
578 static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
579 unsigned long start_page,
582 unsigned int last_page = dom->aperture_size >> PAGE_SHIFT;
584 if (start_page + pages > last_page)
585 pages = last_page - start_page;
587 iommu_area_reserve(dom->bitmap, start_page, pages);
590 static void dma_ops_free_pagetable(struct dma_ops_domain *dma_dom)
595 p1 = dma_dom->domain.pt_root;
600 for (i = 0; i < 512; ++i) {
601 if (!IOMMU_PTE_PRESENT(p1[i]))
604 p2 = IOMMU_PTE_PAGE(p1[i]);
605 for (j = 0; j < 512; ++j) {
606 if (!IOMMU_PTE_PRESENT(p2[j]))
608 p3 = IOMMU_PTE_PAGE(p2[j]);
609 free_page((unsigned long)p3);
612 free_page((unsigned long)p2);
615 free_page((unsigned long)p1);
619 * Free a domain, only used if something went wrong in the
620 * allocation path and we need to free an already allocated page table
622 static void dma_ops_domain_free(struct dma_ops_domain *dom)
627 dma_ops_free_pagetable(dom);
629 kfree(dom->pte_pages);
637 * Allocates a new protection domain usable for the dma_ops functions.
638 * It also intializes the page table and the address allocator data
639 * structures required for the dma_ops interface
641 static struct dma_ops_domain *dma_ops_domain_alloc(struct amd_iommu *iommu,
644 struct dma_ops_domain *dma_dom;
645 unsigned i, num_pte_pages;
650 * Currently the DMA aperture must be between 32 MB and 1GB in size
652 if ((order < 25) || (order > 30))
655 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
659 spin_lock_init(&dma_dom->domain.lock);
661 dma_dom->domain.id = domain_id_alloc();
662 if (dma_dom->domain.id == 0)
664 dma_dom->domain.mode = PAGE_MODE_3_LEVEL;
665 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
666 dma_dom->domain.priv = dma_dom;
667 if (!dma_dom->domain.pt_root)
669 dma_dom->aperture_size = (1ULL << order);
670 dma_dom->bitmap = kzalloc(dma_dom->aperture_size / (PAGE_SIZE * 8),
672 if (!dma_dom->bitmap)
675 * mark the first page as allocated so we never return 0 as
676 * a valid dma-address. So we can use 0 as error value
678 dma_dom->bitmap[0] = 1;
679 dma_dom->next_bit = 0;
681 dma_dom->need_flush = false;
682 dma_dom->target_dev = 0xffff;
684 /* Intialize the exclusion range if necessary */
685 if (iommu->exclusion_start &&
686 iommu->exclusion_start < dma_dom->aperture_size) {
687 unsigned long startpage = iommu->exclusion_start >> PAGE_SHIFT;
688 int pages = iommu_num_pages(iommu->exclusion_start,
689 iommu->exclusion_length,
691 dma_ops_reserve_addresses(dma_dom, startpage, pages);
695 * At the last step, build the page tables so we don't need to
696 * allocate page table pages in the dma_ops mapping/unmapping
699 num_pte_pages = dma_dom->aperture_size / (PAGE_SIZE * 512);
700 dma_dom->pte_pages = kzalloc(num_pte_pages * sizeof(void *),
702 if (!dma_dom->pte_pages)
705 l2_pde = (u64 *)get_zeroed_page(GFP_KERNEL);
709 dma_dom->domain.pt_root[0] = IOMMU_L2_PDE(virt_to_phys(l2_pde));
711 for (i = 0; i < num_pte_pages; ++i) {
712 dma_dom->pte_pages[i] = (u64 *)get_zeroed_page(GFP_KERNEL);
713 if (!dma_dom->pte_pages[i])
715 address = virt_to_phys(dma_dom->pte_pages[i]);
716 l2_pde[i] = IOMMU_L1_PDE(address);
722 dma_ops_domain_free(dma_dom);
728 * Find out the protection domain structure for a given PCI device. This
729 * will give us the pointer to the page table root for example.
731 static struct protection_domain *domain_for_device(u16 devid)
733 struct protection_domain *dom;
736 read_lock_irqsave(&amd_iommu_devtable_lock, flags);
737 dom = amd_iommu_pd_table[devid];
738 read_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
744 * If a device is not yet associated with a domain, this function does
745 * assigns it visible for the hardware
747 static void set_device_domain(struct amd_iommu *iommu,
748 struct protection_domain *domain,
753 u64 pte_root = virt_to_phys(domain->pt_root);
755 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
756 << DEV_ENTRY_MODE_SHIFT;
757 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
759 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
760 amd_iommu_dev_table[devid].data[0] = lower_32_bits(pte_root);
761 amd_iommu_dev_table[devid].data[1] = upper_32_bits(pte_root);
762 amd_iommu_dev_table[devid].data[2] = domain->id;
764 amd_iommu_pd_table[devid] = domain;
765 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
767 iommu_queue_inv_dev_entry(iommu, devid);
770 /*****************************************************************************
772 * The next functions belong to the dma_ops mapping/unmapping code.
774 *****************************************************************************/
777 * This function checks if the driver got a valid device from the caller to
778 * avoid dereferencing invalid pointers.
780 static bool check_device(struct device *dev)
782 if (!dev || !dev->dma_mask)
789 * In this function the list of preallocated protection domains is traversed to
790 * find the domain for a specific device
792 static struct dma_ops_domain *find_protection_domain(u16 devid)
794 struct dma_ops_domain *entry, *ret = NULL;
797 if (list_empty(&iommu_pd_list))
800 spin_lock_irqsave(&iommu_pd_list_lock, flags);
802 list_for_each_entry(entry, &iommu_pd_list, list) {
803 if (entry->target_dev == devid) {
805 list_del(&ret->list);
810 spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
816 * In the dma_ops path we only have the struct device. This function
817 * finds the corresponding IOMMU, the protection domain and the
818 * requestor id for a given device.
819 * If the device is not yet associated with a domain this is also done
822 static int get_device_resources(struct device *dev,
823 struct amd_iommu **iommu,
824 struct protection_domain **domain,
827 struct dma_ops_domain *dma_dom;
828 struct pci_dev *pcidev;
835 if (dev->bus != &pci_bus_type)
838 pcidev = to_pci_dev(dev);
839 _bdf = calc_devid(pcidev->bus->number, pcidev->devfn);
841 /* device not translated by any IOMMU in the system? */
842 if (_bdf > amd_iommu_last_bdf)
845 *bdf = amd_iommu_alias_table[_bdf];
847 *iommu = amd_iommu_rlookup_table[*bdf];
850 *domain = domain_for_device(*bdf);
851 if (*domain == NULL) {
852 dma_dom = find_protection_domain(*bdf);
854 dma_dom = (*iommu)->default_dom;
855 *domain = &dma_dom->domain;
856 set_device_domain(*iommu, *domain, *bdf);
857 printk(KERN_INFO "AMD IOMMU: Using protection domain %d for "
858 "device ", (*domain)->id);
859 print_devid(_bdf, 1);
862 if (domain_for_device(_bdf) == NULL)
863 set_device_domain(*iommu, *domain, _bdf);
869 * This is the generic map function. It maps one 4kb page at paddr to
870 * the given address in the DMA address space for the domain.
872 static dma_addr_t dma_ops_domain_map(struct amd_iommu *iommu,
873 struct dma_ops_domain *dom,
874 unsigned long address,
880 WARN_ON(address > dom->aperture_size);
884 pte = dom->pte_pages[IOMMU_PTE_L1_INDEX(address)];
885 pte += IOMMU_PTE_L0_INDEX(address);
887 __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
889 if (direction == DMA_TO_DEVICE)
890 __pte |= IOMMU_PTE_IR;
891 else if (direction == DMA_FROM_DEVICE)
892 __pte |= IOMMU_PTE_IW;
893 else if (direction == DMA_BIDIRECTIONAL)
894 __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
900 return (dma_addr_t)address;
904 * The generic unmapping function for on page in the DMA address space.
906 static void dma_ops_domain_unmap(struct amd_iommu *iommu,
907 struct dma_ops_domain *dom,
908 unsigned long address)
912 if (address >= dom->aperture_size)
915 WARN_ON(address & ~PAGE_MASK || address >= dom->aperture_size);
917 pte = dom->pte_pages[IOMMU_PTE_L1_INDEX(address)];
918 pte += IOMMU_PTE_L0_INDEX(address);
926 * This function contains common code for mapping of a physically
927 * contiguous memory region into DMA address space. It is used by all
928 * mapping functions provided with this IOMMU driver.
929 * Must be called with the domain lock held.
931 static dma_addr_t __map_single(struct device *dev,
932 struct amd_iommu *iommu,
933 struct dma_ops_domain *dma_dom,
940 dma_addr_t offset = paddr & ~PAGE_MASK;
941 dma_addr_t address, start;
943 unsigned long align_mask = 0;
946 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
950 align_mask = (1UL << get_order(size)) - 1;
952 address = dma_ops_alloc_addresses(dev, dma_dom, pages, align_mask,
954 if (unlikely(address == bad_dma_address))
958 for (i = 0; i < pages; ++i) {
959 dma_ops_domain_map(iommu, dma_dom, start, paddr, dir);
965 if (unlikely(dma_dom->need_flush && !amd_iommu_unmap_flush)) {
966 iommu_flush_tlb(iommu, dma_dom->domain.id);
967 dma_dom->need_flush = false;
968 } else if (unlikely(iommu_has_npcache(iommu)))
969 iommu_flush_pages(iommu, dma_dom->domain.id, address, size);
976 * Does the reverse of the __map_single function. Must be called with
977 * the domain lock held too
979 static void __unmap_single(struct amd_iommu *iommu,
980 struct dma_ops_domain *dma_dom,
988 if ((dma_addr == bad_dma_address) ||
989 (dma_addr + size > dma_dom->aperture_size))
992 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
993 dma_addr &= PAGE_MASK;
996 for (i = 0; i < pages; ++i) {
997 dma_ops_domain_unmap(iommu, dma_dom, start);
1001 dma_ops_free_addresses(dma_dom, dma_addr, pages);
1003 if (amd_iommu_unmap_flush || dma_dom->need_flush) {
1004 iommu_flush_pages(iommu, dma_dom->domain.id, dma_addr, size);
1005 dma_dom->need_flush = false;
1010 * The exported map_single function for dma_ops.
1012 static dma_addr_t map_single(struct device *dev, phys_addr_t paddr,
1013 size_t size, int dir)
1015 unsigned long flags;
1016 struct amd_iommu *iommu;
1017 struct protection_domain *domain;
1022 if (!check_device(dev))
1023 return bad_dma_address;
1025 dma_mask = *dev->dma_mask;
1027 get_device_resources(dev, &iommu, &domain, &devid);
1029 if (iommu == NULL || domain == NULL)
1030 /* device not handled by any AMD IOMMU */
1031 return (dma_addr_t)paddr;
1033 spin_lock_irqsave(&domain->lock, flags);
1034 addr = __map_single(dev, iommu, domain->priv, paddr, size, dir, false,
1036 if (addr == bad_dma_address)
1039 iommu_completion_wait(iommu);
1042 spin_unlock_irqrestore(&domain->lock, flags);
1048 * The exported unmap_single function for dma_ops.
1050 static void unmap_single(struct device *dev, dma_addr_t dma_addr,
1051 size_t size, int dir)
1053 unsigned long flags;
1054 struct amd_iommu *iommu;
1055 struct protection_domain *domain;
1058 if (!check_device(dev) ||
1059 !get_device_resources(dev, &iommu, &domain, &devid))
1060 /* device not handled by any AMD IOMMU */
1063 spin_lock_irqsave(&domain->lock, flags);
1065 __unmap_single(iommu, domain->priv, dma_addr, size, dir);
1067 iommu_completion_wait(iommu);
1069 spin_unlock_irqrestore(&domain->lock, flags);
1073 * This is a special map_sg function which is used if we should map a
1074 * device which is not handled by an AMD IOMMU in the system.
1076 static int map_sg_no_iommu(struct device *dev, struct scatterlist *sglist,
1077 int nelems, int dir)
1079 struct scatterlist *s;
1082 for_each_sg(sglist, s, nelems, i) {
1083 s->dma_address = (dma_addr_t)sg_phys(s);
1084 s->dma_length = s->length;
1091 * The exported map_sg function for dma_ops (handles scatter-gather
1094 static int map_sg(struct device *dev, struct scatterlist *sglist,
1095 int nelems, int dir)
1097 unsigned long flags;
1098 struct amd_iommu *iommu;
1099 struct protection_domain *domain;
1102 struct scatterlist *s;
1104 int mapped_elems = 0;
1107 if (!check_device(dev))
1110 dma_mask = *dev->dma_mask;
1112 get_device_resources(dev, &iommu, &domain, &devid);
1114 if (!iommu || !domain)
1115 return map_sg_no_iommu(dev, sglist, nelems, dir);
1117 spin_lock_irqsave(&domain->lock, flags);
1119 for_each_sg(sglist, s, nelems, i) {
1122 s->dma_address = __map_single(dev, iommu, domain->priv,
1123 paddr, s->length, dir, false,
1126 if (s->dma_address) {
1127 s->dma_length = s->length;
1133 iommu_completion_wait(iommu);
1136 spin_unlock_irqrestore(&domain->lock, flags);
1138 return mapped_elems;
1140 for_each_sg(sglist, s, mapped_elems, i) {
1142 __unmap_single(iommu, domain->priv, s->dma_address,
1143 s->dma_length, dir);
1144 s->dma_address = s->dma_length = 0;
1153 * The exported map_sg function for dma_ops (handles scatter-gather
1156 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
1157 int nelems, int dir)
1159 unsigned long flags;
1160 struct amd_iommu *iommu;
1161 struct protection_domain *domain;
1162 struct scatterlist *s;
1166 if (!check_device(dev) ||
1167 !get_device_resources(dev, &iommu, &domain, &devid))
1170 spin_lock_irqsave(&domain->lock, flags);
1172 for_each_sg(sglist, s, nelems, i) {
1173 __unmap_single(iommu, domain->priv, s->dma_address,
1174 s->dma_length, dir);
1175 s->dma_address = s->dma_length = 0;
1178 iommu_completion_wait(iommu);
1180 spin_unlock_irqrestore(&domain->lock, flags);
1184 * The exported alloc_coherent function for dma_ops.
1186 static void *alloc_coherent(struct device *dev, size_t size,
1187 dma_addr_t *dma_addr, gfp_t flag)
1189 unsigned long flags;
1191 struct amd_iommu *iommu;
1192 struct protection_domain *domain;
1195 u64 dma_mask = dev->coherent_dma_mask;
1197 if (!check_device(dev))
1200 if (!get_device_resources(dev, &iommu, &domain, &devid))
1201 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
1204 virt_addr = (void *)__get_free_pages(flag, get_order(size));
1208 paddr = virt_to_phys(virt_addr);
1210 if (!iommu || !domain) {
1211 *dma_addr = (dma_addr_t)paddr;
1216 dma_mask = *dev->dma_mask;
1218 spin_lock_irqsave(&domain->lock, flags);
1220 *dma_addr = __map_single(dev, iommu, domain->priv, paddr,
1221 size, DMA_BIDIRECTIONAL, true, dma_mask);
1223 if (*dma_addr == bad_dma_address) {
1224 free_pages((unsigned long)virt_addr, get_order(size));
1229 iommu_completion_wait(iommu);
1232 spin_unlock_irqrestore(&domain->lock, flags);
1238 * The exported free_coherent function for dma_ops.
1240 static void free_coherent(struct device *dev, size_t size,
1241 void *virt_addr, dma_addr_t dma_addr)
1243 unsigned long flags;
1244 struct amd_iommu *iommu;
1245 struct protection_domain *domain;
1248 if (!check_device(dev))
1251 get_device_resources(dev, &iommu, &domain, &devid);
1253 if (!iommu || !domain)
1256 spin_lock_irqsave(&domain->lock, flags);
1258 __unmap_single(iommu, domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
1260 iommu_completion_wait(iommu);
1262 spin_unlock_irqrestore(&domain->lock, flags);
1265 free_pages((unsigned long)virt_addr, get_order(size));
1269 * This function is called by the DMA layer to find out if we can handle a
1270 * particular device. It is part of the dma_ops.
1272 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
1275 struct pci_dev *pcidev;
1277 /* No device or no PCI device */
1278 if (!dev || dev->bus != &pci_bus_type)
1281 pcidev = to_pci_dev(dev);
1283 bdf = calc_devid(pcidev->bus->number, pcidev->devfn);
1285 /* Out of our scope? */
1286 if (bdf > amd_iommu_last_bdf)
1293 * The function for pre-allocating protection domains.
1295 * If the driver core informs the DMA layer if a driver grabs a device
1296 * we don't need to preallocate the protection domains anymore.
1297 * For now we have to.
1299 void prealloc_protection_domains(void)
1301 struct pci_dev *dev = NULL;
1302 struct dma_ops_domain *dma_dom;
1303 struct amd_iommu *iommu;
1304 int order = amd_iommu_aperture_order;
1307 while ((dev = pci_get_device(PCI_ANY_ID, PCI_ANY_ID, dev)) != NULL) {
1308 devid = (dev->bus->number << 8) | dev->devfn;
1309 if (devid > amd_iommu_last_bdf)
1311 devid = amd_iommu_alias_table[devid];
1312 if (domain_for_device(devid))
1314 iommu = amd_iommu_rlookup_table[devid];
1317 dma_dom = dma_ops_domain_alloc(iommu, order);
1320 init_unity_mappings_for_device(dma_dom, devid);
1321 dma_dom->target_dev = devid;
1323 list_add_tail(&dma_dom->list, &iommu_pd_list);
1327 static struct dma_mapping_ops amd_iommu_dma_ops = {
1328 .alloc_coherent = alloc_coherent,
1329 .free_coherent = free_coherent,
1330 .map_single = map_single,
1331 .unmap_single = unmap_single,
1333 .unmap_sg = unmap_sg,
1334 .dma_supported = amd_iommu_dma_supported,
1338 * The function which clues the AMD IOMMU driver into dma_ops.
1340 int __init amd_iommu_init_dma_ops(void)
1342 struct amd_iommu *iommu;
1343 int order = amd_iommu_aperture_order;
1347 * first allocate a default protection domain for every IOMMU we
1348 * found in the system. Devices not assigned to any other
1349 * protection domain will be assigned to the default one.
1351 list_for_each_entry(iommu, &amd_iommu_list, list) {
1352 iommu->default_dom = dma_ops_domain_alloc(iommu, order);
1353 if (iommu->default_dom == NULL)
1355 ret = iommu_init_unity_mappings(iommu);
1361 * If device isolation is enabled, pre-allocate the protection
1362 * domains for each device.
1364 if (amd_iommu_isolate)
1365 prealloc_protection_domains();
1369 bad_dma_address = 0;
1370 #ifdef CONFIG_GART_IOMMU
1371 gart_iommu_aperture_disabled = 1;
1372 gart_iommu_aperture = 0;
1375 /* Make the driver finally visible to the drivers */
1376 dma_ops = &amd_iommu_dma_ops;
1382 list_for_each_entry(iommu, &amd_iommu_list, list) {
1383 if (iommu->default_dom)
1384 dma_ops_domain_free(iommu->default_dom);