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>
27 #include <asm/amd_iommu_types.h>
28 #include <asm/amd_iommu.h>
30 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
32 #define EXIT_LOOP_COUNT 10000000
34 static DEFINE_RWLOCK(amd_iommu_devtable_lock);
37 * general struct to manage commands send to an IOMMU
43 static int dma_ops_unity_map(struct dma_ops_domain *dma_dom,
44 struct unity_map_entry *e);
46 /* returns !0 if the IOMMU is caching non-present entries in its TLB */
47 static int iommu_has_npcache(struct amd_iommu *iommu)
49 return iommu->cap & IOMMU_CAP_NPCACHE;
52 /****************************************************************************
54 * IOMMU command queuing functions
56 ****************************************************************************/
59 * Writes the command to the IOMMUs command buffer and informs the
60 * hardware about the new command. Must be called with iommu->lock held.
62 static int __iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
67 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
68 target = (iommu->cmd_buf + tail);
69 memcpy_toio(target, cmd, sizeof(*cmd));
70 tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
71 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
74 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
80 * General queuing function for commands. Takes iommu->lock and calls
81 * __iommu_queue_command().
83 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
88 spin_lock_irqsave(&iommu->lock, flags);
89 ret = __iommu_queue_command(iommu, cmd);
90 spin_unlock_irqrestore(&iommu->lock, flags);
96 * This function is called whenever we need to ensure that the IOMMU has
97 * completed execution of all commands we sent. It sends a
98 * COMPLETION_WAIT command and waits for it to finish. The IOMMU informs
99 * us about that by writing a value to a physical address we pass with
102 static int iommu_completion_wait(struct amd_iommu *iommu)
105 struct iommu_cmd cmd;
106 volatile u64 ready = 0;
107 unsigned long ready_phys = virt_to_phys(&ready);
110 memset(&cmd, 0, sizeof(cmd));
111 cmd.data[0] = LOW_U32(ready_phys) | CMD_COMPL_WAIT_STORE_MASK;
112 cmd.data[1] = upper_32_bits(ready_phys);
113 cmd.data[2] = 1; /* value written to 'ready' */
114 CMD_SET_TYPE(&cmd, CMD_COMPL_WAIT);
116 iommu->need_sync = 0;
118 ret = iommu_queue_command(iommu, &cmd);
123 while (!ready && (i < EXIT_LOOP_COUNT)) {
128 if (unlikely((i == EXIT_LOOP_COUNT) && printk_ratelimit()))
129 printk(KERN_WARNING "AMD IOMMU: Completion wait loop failed\n");
135 * Command send function for invalidating a device table entry
137 static int iommu_queue_inv_dev_entry(struct amd_iommu *iommu, u16 devid)
139 struct iommu_cmd cmd;
141 BUG_ON(iommu == NULL);
143 memset(&cmd, 0, sizeof(cmd));
144 CMD_SET_TYPE(&cmd, CMD_INV_DEV_ENTRY);
147 iommu->need_sync = 1;
149 return iommu_queue_command(iommu, &cmd);
153 * Generic command send function for invalidaing TLB entries
155 static int iommu_queue_inv_iommu_pages(struct amd_iommu *iommu,
156 u64 address, u16 domid, int pde, int s)
158 struct iommu_cmd cmd;
160 memset(&cmd, 0, sizeof(cmd));
161 address &= PAGE_MASK;
162 CMD_SET_TYPE(&cmd, CMD_INV_IOMMU_PAGES);
163 cmd.data[1] |= domid;
164 cmd.data[2] = LOW_U32(address);
165 cmd.data[3] = upper_32_bits(address);
166 if (s) /* size bit - we flush more than one 4kb page */
167 cmd.data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
168 if (pde) /* PDE bit - we wan't flush everything not only the PTEs */
169 cmd.data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
171 iommu->need_sync = 1;
173 return iommu_queue_command(iommu, &cmd);
177 * TLB invalidation function which is called from the mapping functions.
178 * It invalidates a single PTE if the range to flush is within a single
179 * page. Otherwise it flushes the whole TLB of the IOMMU.
181 static int iommu_flush_pages(struct amd_iommu *iommu, u16 domid,
182 u64 address, size_t size)
185 unsigned pages = iommu_num_pages(address, size);
187 address &= PAGE_MASK;
191 * If we have to flush more than one page, flush all
192 * TLB entries for this domain
194 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
198 iommu_queue_inv_iommu_pages(iommu, address, domid, 0, s);
203 /****************************************************************************
205 * The functions below are used the create the page table mappings for
206 * unity mapped regions.
208 ****************************************************************************/
211 * Generic mapping functions. It maps a physical address into a DMA
212 * address space. It allocates the page table pages if necessary.
213 * In the future it can be extended to a generic mapping function
214 * supporting all features of AMD IOMMU page tables like level skipping
215 * and full 64 bit address spaces.
217 static int iommu_map(struct protection_domain *dom,
218 unsigned long bus_addr,
219 unsigned long phys_addr,
222 u64 __pte, *pte, *page;
224 bus_addr = PAGE_ALIGN(bus_addr);
225 phys_addr = PAGE_ALIGN(bus_addr);
227 /* only support 512GB address spaces for now */
228 if (bus_addr > IOMMU_MAP_SIZE_L3 || !(prot & IOMMU_PROT_MASK))
231 pte = &dom->pt_root[IOMMU_PTE_L2_INDEX(bus_addr)];
233 if (!IOMMU_PTE_PRESENT(*pte)) {
234 page = (u64 *)get_zeroed_page(GFP_KERNEL);
237 *pte = IOMMU_L2_PDE(virt_to_phys(page));
240 pte = IOMMU_PTE_PAGE(*pte);
241 pte = &pte[IOMMU_PTE_L1_INDEX(bus_addr)];
243 if (!IOMMU_PTE_PRESENT(*pte)) {
244 page = (u64 *)get_zeroed_page(GFP_KERNEL);
247 *pte = IOMMU_L1_PDE(virt_to_phys(page));
250 pte = IOMMU_PTE_PAGE(*pte);
251 pte = &pte[IOMMU_PTE_L0_INDEX(bus_addr)];
253 if (IOMMU_PTE_PRESENT(*pte))
256 __pte = phys_addr | IOMMU_PTE_P;
257 if (prot & IOMMU_PROT_IR)
258 __pte |= IOMMU_PTE_IR;
259 if (prot & IOMMU_PROT_IW)
260 __pte |= IOMMU_PTE_IW;
268 * This function checks if a specific unity mapping entry is needed for
269 * this specific IOMMU.
271 static int iommu_for_unity_map(struct amd_iommu *iommu,
272 struct unity_map_entry *entry)
276 for (i = entry->devid_start; i <= entry->devid_end; ++i) {
277 bdf = amd_iommu_alias_table[i];
278 if (amd_iommu_rlookup_table[bdf] == iommu)
286 * Init the unity mappings for a specific IOMMU in the system
288 * Basically iterates over all unity mapping entries and applies them to
289 * the default domain DMA of that IOMMU if necessary.
291 static int iommu_init_unity_mappings(struct amd_iommu *iommu)
293 struct unity_map_entry *entry;
296 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
297 if (!iommu_for_unity_map(iommu, entry))
299 ret = dma_ops_unity_map(iommu->default_dom, entry);
308 * This function actually applies the mapping to the page table of the
311 static int dma_ops_unity_map(struct dma_ops_domain *dma_dom,
312 struct unity_map_entry *e)
317 for (addr = e->address_start; addr < e->address_end;
319 ret = iommu_map(&dma_dom->domain, addr, addr, e->prot);
323 * if unity mapping is in aperture range mark the page
324 * as allocated in the aperture
326 if (addr < dma_dom->aperture_size)
327 __set_bit(addr >> PAGE_SHIFT, dma_dom->bitmap);
334 * Inits the unity mappings required for a specific device
336 static int init_unity_mappings_for_device(struct dma_ops_domain *dma_dom,
339 struct unity_map_entry *e;
342 list_for_each_entry(e, &amd_iommu_unity_map, list) {
343 if (!(devid >= e->devid_start && devid <= e->devid_end))
345 ret = dma_ops_unity_map(dma_dom, e);
353 /****************************************************************************
355 * The next functions belong to the address allocator for the dma_ops
356 * interface functions. They work like the allocators in the other IOMMU
357 * drivers. Its basically a bitmap which marks the allocated pages in
358 * the aperture. Maybe it could be enhanced in the future to a more
359 * efficient allocator.
361 ****************************************************************************/
362 static unsigned long dma_mask_to_pages(unsigned long mask)
364 return (mask >> PAGE_SHIFT) +
365 (PAGE_ALIGN(mask & ~PAGE_MASK) >> PAGE_SHIFT);
369 * The address allocator core function.
371 * called with domain->lock held
373 static unsigned long dma_ops_alloc_addresses(struct device *dev,
374 struct dma_ops_domain *dom,
377 unsigned long limit = dma_mask_to_pages(*dev->dma_mask);
378 unsigned long address;
379 unsigned long size = dom->aperture_size >> PAGE_SHIFT;
380 unsigned long boundary_size;
382 boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
383 PAGE_SIZE) >> PAGE_SHIFT;
384 limit = limit < size ? limit : size;
386 if (dom->next_bit >= limit)
389 address = iommu_area_alloc(dom->bitmap, limit, dom->next_bit, pages,
390 0 , boundary_size, 0);
392 address = iommu_area_alloc(dom->bitmap, limit, 0, pages,
393 0, boundary_size, 0);
395 if (likely(address != -1)) {
396 dom->next_bit = address + pages;
397 address <<= PAGE_SHIFT;
399 address = bad_dma_address;
401 WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
407 * The address free function.
409 * called with domain->lock held
411 static void dma_ops_free_addresses(struct dma_ops_domain *dom,
412 unsigned long address,
415 address >>= PAGE_SHIFT;
416 iommu_area_free(dom->bitmap, address, pages);
419 /****************************************************************************
421 * The next functions belong to the domain allocation. A domain is
422 * allocated for every IOMMU as the default domain. If device isolation
423 * is enabled, every device get its own domain. The most important thing
424 * about domains is the page table mapping the DMA address space they
427 ****************************************************************************/
429 static u16 domain_id_alloc(void)
434 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
435 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
437 if (id > 0 && id < MAX_DOMAIN_ID)
438 __set_bit(id, amd_iommu_pd_alloc_bitmap);
441 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
447 * Used to reserve address ranges in the aperture (e.g. for exclusion
450 static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
451 unsigned long start_page,
454 unsigned int last_page = dom->aperture_size >> PAGE_SHIFT;
456 if (start_page + pages > last_page)
457 pages = last_page - start_page;
459 set_bit_string(dom->bitmap, start_page, pages);
462 static void dma_ops_free_pagetable(struct dma_ops_domain *dma_dom)
467 p1 = dma_dom->domain.pt_root;
472 for (i = 0; i < 512; ++i) {
473 if (!IOMMU_PTE_PRESENT(p1[i]))
476 p2 = IOMMU_PTE_PAGE(p1[i]);
477 for (j = 0; j < 512; ++i) {
478 if (!IOMMU_PTE_PRESENT(p2[j]))
480 p3 = IOMMU_PTE_PAGE(p2[j]);
481 free_page((unsigned long)p3);
484 free_page((unsigned long)p2);
487 free_page((unsigned long)p1);
491 * Free a domain, only used if something went wrong in the
492 * allocation path and we need to free an already allocated page table
494 static void dma_ops_domain_free(struct dma_ops_domain *dom)
499 dma_ops_free_pagetable(dom);
501 kfree(dom->pte_pages);
509 * Allocates a new protection domain usable for the dma_ops functions.
510 * It also intializes the page table and the address allocator data
511 * structures required for the dma_ops interface
513 static struct dma_ops_domain *dma_ops_domain_alloc(struct amd_iommu *iommu,
516 struct dma_ops_domain *dma_dom;
517 unsigned i, num_pte_pages;
522 * Currently the DMA aperture must be between 32 MB and 1GB in size
524 if ((order < 25) || (order > 30))
527 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
531 spin_lock_init(&dma_dom->domain.lock);
533 dma_dom->domain.id = domain_id_alloc();
534 if (dma_dom->domain.id == 0)
536 dma_dom->domain.mode = PAGE_MODE_3_LEVEL;
537 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
538 dma_dom->domain.priv = dma_dom;
539 if (!dma_dom->domain.pt_root)
541 dma_dom->aperture_size = (1ULL << order);
542 dma_dom->bitmap = kzalloc(dma_dom->aperture_size / (PAGE_SIZE * 8),
544 if (!dma_dom->bitmap)
547 * mark the first page as allocated so we never return 0 as
548 * a valid dma-address. So we can use 0 as error value
550 dma_dom->bitmap[0] = 1;
551 dma_dom->next_bit = 0;
553 /* Intialize the exclusion range if necessary */
554 if (iommu->exclusion_start &&
555 iommu->exclusion_start < dma_dom->aperture_size) {
556 unsigned long startpage = iommu->exclusion_start >> PAGE_SHIFT;
557 int pages = iommu_num_pages(iommu->exclusion_start,
558 iommu->exclusion_length);
559 dma_ops_reserve_addresses(dma_dom, startpage, pages);
563 * At the last step, build the page tables so we don't need to
564 * allocate page table pages in the dma_ops mapping/unmapping
567 num_pte_pages = dma_dom->aperture_size / (PAGE_SIZE * 512);
568 dma_dom->pte_pages = kzalloc(num_pte_pages * sizeof(void *),
570 if (!dma_dom->pte_pages)
573 l2_pde = (u64 *)get_zeroed_page(GFP_KERNEL);
577 dma_dom->domain.pt_root[0] = IOMMU_L2_PDE(virt_to_phys(l2_pde));
579 for (i = 0; i < num_pte_pages; ++i) {
580 dma_dom->pte_pages[i] = (u64 *)get_zeroed_page(GFP_KERNEL);
581 if (!dma_dom->pte_pages[i])
583 address = virt_to_phys(dma_dom->pte_pages[i]);
584 l2_pde[i] = IOMMU_L1_PDE(address);
590 dma_ops_domain_free(dma_dom);
596 * Find out the protection domain structure for a given PCI device. This
597 * will give us the pointer to the page table root for example.
599 static struct protection_domain *domain_for_device(u16 devid)
601 struct protection_domain *dom;
604 read_lock_irqsave(&amd_iommu_devtable_lock, flags);
605 dom = amd_iommu_pd_table[devid];
606 read_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
612 * If a device is not yet associated with a domain, this function does
613 * assigns it visible for the hardware
615 static void set_device_domain(struct amd_iommu *iommu,
616 struct protection_domain *domain,
621 u64 pte_root = virt_to_phys(domain->pt_root);
623 pte_root |= (domain->mode & 0x07) << 9;
624 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | 2;
626 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
627 amd_iommu_dev_table[devid].data[0] = pte_root;
628 amd_iommu_dev_table[devid].data[1] = pte_root >> 32;
629 amd_iommu_dev_table[devid].data[2] = domain->id;
631 amd_iommu_pd_table[devid] = domain;
632 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
634 iommu_queue_inv_dev_entry(iommu, devid);
636 iommu->need_sync = 1;
639 /*****************************************************************************
641 * The next functions belong to the dma_ops mapping/unmapping code.
643 *****************************************************************************/
646 * In the dma_ops path we only have the struct device. This function
647 * finds the corresponding IOMMU, the protection domain and the
648 * requestor id for a given device.
649 * If the device is not yet associated with a domain this is also done
652 static int get_device_resources(struct device *dev,
653 struct amd_iommu **iommu,
654 struct protection_domain **domain,
657 struct dma_ops_domain *dma_dom;
658 struct pci_dev *pcidev;
661 BUG_ON(!dev || dev->bus != &pci_bus_type || !dev->dma_mask);
663 pcidev = to_pci_dev(dev);
664 _bdf = calc_devid(pcidev->bus->number, pcidev->devfn);
666 /* device not translated by any IOMMU in the system? */
667 if (_bdf > amd_iommu_last_bdf) {
674 *bdf = amd_iommu_alias_table[_bdf];
676 *iommu = amd_iommu_rlookup_table[*bdf];
679 dma_dom = (*iommu)->default_dom;
680 *domain = domain_for_device(*bdf);
681 if (*domain == NULL) {
682 *domain = &dma_dom->domain;
683 set_device_domain(*iommu, *domain, *bdf);
684 printk(KERN_INFO "AMD IOMMU: Using protection domain %d for "
685 "device ", (*domain)->id);
686 print_devid(_bdf, 1);
693 * This is the generic map function. It maps one 4kb page at paddr to
694 * the given address in the DMA address space for the domain.
696 static dma_addr_t dma_ops_domain_map(struct amd_iommu *iommu,
697 struct dma_ops_domain *dom,
698 unsigned long address,
704 WARN_ON(address > dom->aperture_size);
708 pte = dom->pte_pages[IOMMU_PTE_L1_INDEX(address)];
709 pte += IOMMU_PTE_L0_INDEX(address);
711 __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
713 if (direction == DMA_TO_DEVICE)
714 __pte |= IOMMU_PTE_IR;
715 else if (direction == DMA_FROM_DEVICE)
716 __pte |= IOMMU_PTE_IW;
717 else if (direction == DMA_BIDIRECTIONAL)
718 __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
724 return (dma_addr_t)address;
728 * The generic unmapping function for on page in the DMA address space.
730 static void dma_ops_domain_unmap(struct amd_iommu *iommu,
731 struct dma_ops_domain *dom,
732 unsigned long address)
736 if (address >= dom->aperture_size)
739 WARN_ON(address & 0xfffULL || address > dom->aperture_size);
741 pte = dom->pte_pages[IOMMU_PTE_L1_INDEX(address)];
742 pte += IOMMU_PTE_L0_INDEX(address);
750 * This function contains common code for mapping of a physically
751 * contiguous memory region into DMA address space. It is uses by all
752 * mapping functions provided by this IOMMU driver.
753 * Must be called with the domain lock held.
755 static dma_addr_t __map_single(struct device *dev,
756 struct amd_iommu *iommu,
757 struct dma_ops_domain *dma_dom,
762 dma_addr_t offset = paddr & ~PAGE_MASK;
763 dma_addr_t address, start;
767 pages = iommu_num_pages(paddr, size);
770 address = dma_ops_alloc_addresses(dev, dma_dom, pages);
771 if (unlikely(address == bad_dma_address))
775 for (i = 0; i < pages; ++i) {
776 dma_ops_domain_map(iommu, dma_dom, start, paddr, dir);
787 * Does the reverse of the __map_single function. Must be called with
788 * the domain lock held too
790 static void __unmap_single(struct amd_iommu *iommu,
791 struct dma_ops_domain *dma_dom,
799 if ((dma_addr == 0) || (dma_addr + size > dma_dom->aperture_size))
802 pages = iommu_num_pages(dma_addr, size);
803 dma_addr &= PAGE_MASK;
806 for (i = 0; i < pages; ++i) {
807 dma_ops_domain_unmap(iommu, dma_dom, start);
811 dma_ops_free_addresses(dma_dom, dma_addr, pages);
815 * The exported map_single function for dma_ops.
817 static dma_addr_t map_single(struct device *dev, phys_addr_t paddr,
818 size_t size, int dir)
821 struct amd_iommu *iommu;
822 struct protection_domain *domain;
826 get_device_resources(dev, &iommu, &domain, &devid);
828 if (iommu == NULL || domain == NULL)
829 /* device not handled by any AMD IOMMU */
830 return (dma_addr_t)paddr;
832 spin_lock_irqsave(&domain->lock, flags);
833 addr = __map_single(dev, iommu, domain->priv, paddr, size, dir);
834 if (addr == bad_dma_address)
837 if (iommu_has_npcache(iommu))
838 iommu_flush_pages(iommu, domain->id, addr, size);
840 if (iommu->need_sync)
841 iommu_completion_wait(iommu);
844 spin_unlock_irqrestore(&domain->lock, flags);
850 * The exported unmap_single function for dma_ops.
852 static void unmap_single(struct device *dev, dma_addr_t dma_addr,
853 size_t size, int dir)
856 struct amd_iommu *iommu;
857 struct protection_domain *domain;
860 if (!get_device_resources(dev, &iommu, &domain, &devid))
861 /* device not handled by any AMD IOMMU */
864 spin_lock_irqsave(&domain->lock, flags);
866 __unmap_single(iommu, domain->priv, dma_addr, size, dir);
868 iommu_flush_pages(iommu, domain->id, dma_addr, size);
870 if (iommu->need_sync)
871 iommu_completion_wait(iommu);
873 spin_unlock_irqrestore(&domain->lock, flags);
877 * This is a special map_sg function which is used if we should map a
878 * device which is not handled by an AMD IOMMU in the system.
880 static int map_sg_no_iommu(struct device *dev, struct scatterlist *sglist,
883 struct scatterlist *s;
886 for_each_sg(sglist, s, nelems, i) {
887 s->dma_address = (dma_addr_t)sg_phys(s);
888 s->dma_length = s->length;
895 * The exported map_sg function for dma_ops (handles scatter-gather
898 static int map_sg(struct device *dev, struct scatterlist *sglist,
902 struct amd_iommu *iommu;
903 struct protection_domain *domain;
906 struct scatterlist *s;
908 int mapped_elems = 0;
910 get_device_resources(dev, &iommu, &domain, &devid);
912 if (!iommu || !domain)
913 return map_sg_no_iommu(dev, sglist, nelems, dir);
915 spin_lock_irqsave(&domain->lock, flags);
917 for_each_sg(sglist, s, nelems, i) {
920 s->dma_address = __map_single(dev, iommu, domain->priv,
921 paddr, s->length, dir);
923 if (s->dma_address) {
924 s->dma_length = s->length;
928 if (iommu_has_npcache(iommu))
929 iommu_flush_pages(iommu, domain->id, s->dma_address,
933 if (iommu->need_sync)
934 iommu_completion_wait(iommu);
937 spin_unlock_irqrestore(&domain->lock, flags);
941 for_each_sg(sglist, s, mapped_elems, i) {
943 __unmap_single(iommu, domain->priv, s->dma_address,
945 s->dma_address = s->dma_length = 0;
954 * The exported map_sg function for dma_ops (handles scatter-gather
957 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
961 struct amd_iommu *iommu;
962 struct protection_domain *domain;
963 struct scatterlist *s;
967 if (!get_device_resources(dev, &iommu, &domain, &devid))
970 spin_lock_irqsave(&domain->lock, flags);
972 for_each_sg(sglist, s, nelems, i) {
973 __unmap_single(iommu, domain->priv, s->dma_address,
975 iommu_flush_pages(iommu, domain->id, s->dma_address,
977 s->dma_address = s->dma_length = 0;
980 if (iommu->need_sync)
981 iommu_completion_wait(iommu);
983 spin_unlock_irqrestore(&domain->lock, flags);
987 * The exported alloc_coherent function for dma_ops.
989 static void *alloc_coherent(struct device *dev, size_t size,
990 dma_addr_t *dma_addr, gfp_t flag)
994 struct amd_iommu *iommu;
995 struct protection_domain *domain;
999 virt_addr = (void *)__get_free_pages(flag, get_order(size));
1003 memset(virt_addr, 0, size);
1004 paddr = virt_to_phys(virt_addr);
1006 get_device_resources(dev, &iommu, &domain, &devid);
1008 if (!iommu || !domain) {
1009 *dma_addr = (dma_addr_t)paddr;
1013 spin_lock_irqsave(&domain->lock, flags);
1015 *dma_addr = __map_single(dev, iommu, domain->priv, paddr,
1016 size, DMA_BIDIRECTIONAL);
1018 if (*dma_addr == bad_dma_address) {
1019 free_pages((unsigned long)virt_addr, get_order(size));
1024 if (iommu_has_npcache(iommu))
1025 iommu_flush_pages(iommu, domain->id, *dma_addr, size);
1027 if (iommu->need_sync)
1028 iommu_completion_wait(iommu);
1031 spin_unlock_irqrestore(&domain->lock, flags);
1037 * The exported free_coherent function for dma_ops.
1038 * FIXME: fix the generic x86 DMA layer so that it actually calls that
1041 static void free_coherent(struct device *dev, size_t size,
1042 void *virt_addr, dma_addr_t dma_addr)
1044 unsigned long flags;
1045 struct amd_iommu *iommu;
1046 struct protection_domain *domain;
1049 get_device_resources(dev, &iommu, &domain, &devid);
1051 if (!iommu || !domain)
1054 spin_lock_irqsave(&domain->lock, flags);
1056 __unmap_single(iommu, domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
1057 iommu_flush_pages(iommu, domain->id, dma_addr, size);
1059 if (iommu->need_sync)
1060 iommu_completion_wait(iommu);
1062 spin_unlock_irqrestore(&domain->lock, flags);
1065 free_pages((unsigned long)virt_addr, get_order(size));
1069 * The function for pre-allocating protection domains.
1071 * If the driver core informs the DMA layer if a driver grabs a device
1072 * we don't need to preallocate the protection domains anymore.
1073 * For now we have to.
1075 void prealloc_protection_domains(void)
1077 struct pci_dev *dev = NULL;
1078 struct dma_ops_domain *dma_dom;
1079 struct amd_iommu *iommu;
1080 int order = amd_iommu_aperture_order;
1083 while ((dev = pci_get_device(PCI_ANY_ID, PCI_ANY_ID, dev)) != NULL) {
1084 devid = (dev->bus->number << 8) | dev->devfn;
1085 if (devid > amd_iommu_last_bdf)
1087 devid = amd_iommu_alias_table[devid];
1088 if (domain_for_device(devid))
1090 iommu = amd_iommu_rlookup_table[devid];
1093 dma_dom = dma_ops_domain_alloc(iommu, order);
1096 init_unity_mappings_for_device(dma_dom, devid);
1097 set_device_domain(iommu, &dma_dom->domain, devid);
1098 printk(KERN_INFO "AMD IOMMU: Allocated domain %d for device ",
1099 dma_dom->domain.id);
1100 print_devid(devid, 1);
1104 static struct dma_mapping_ops amd_iommu_dma_ops = {
1105 .alloc_coherent = alloc_coherent,
1106 .free_coherent = free_coherent,
1107 .map_single = map_single,
1108 .unmap_single = unmap_single,
1110 .unmap_sg = unmap_sg,
1114 * The function which clues the AMD IOMMU driver into dma_ops.
1116 int __init amd_iommu_init_dma_ops(void)
1118 struct amd_iommu *iommu;
1119 int order = amd_iommu_aperture_order;
1123 * first allocate a default protection domain for every IOMMU we
1124 * found in the system. Devices not assigned to any other
1125 * protection domain will be assigned to the default one.
1127 list_for_each_entry(iommu, &amd_iommu_list, list) {
1128 iommu->default_dom = dma_ops_domain_alloc(iommu, order);
1129 if (iommu->default_dom == NULL)
1131 ret = iommu_init_unity_mappings(iommu);
1137 * If device isolation is enabled, pre-allocate the protection
1138 * domains for each device.
1140 if (amd_iommu_isolate)
1141 prealloc_protection_domains();
1145 bad_dma_address = 0;
1146 #ifdef CONFIG_GART_IOMMU
1147 gart_iommu_aperture_disabled = 1;
1148 gart_iommu_aperture = 0;
1151 /* Make the driver finally visible to the drivers */
1152 dma_ops = &amd_iommu_dma_ops;
1158 list_for_each_entry(iommu, &amd_iommu_list, list) {
1159 if (iommu->default_dom)
1160 dma_ops_domain_free(iommu->default_dom);