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)
104 int ret = 0, ready = 0;
106 struct iommu_cmd cmd;
107 unsigned long flags, i = 0;
109 memset(&cmd, 0, sizeof(cmd));
110 cmd.data[0] = CMD_COMPL_WAIT_INT_MASK;
111 CMD_SET_TYPE(&cmd, CMD_COMPL_WAIT);
113 iommu->need_sync = 0;
115 spin_lock_irqsave(&iommu->lock, flags);
117 ret = __iommu_queue_command(iommu, &cmd);
122 while (!ready && (i < EXIT_LOOP_COUNT)) {
124 /* wait for the bit to become one */
125 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
126 ready = status & MMIO_STATUS_COM_WAIT_INT_MASK;
129 /* set bit back to zero */
130 status &= ~MMIO_STATUS_COM_WAIT_INT_MASK;
131 writel(status, iommu->mmio_base + MMIO_STATUS_OFFSET);
133 if (unlikely((i == EXIT_LOOP_COUNT) && printk_ratelimit()))
134 printk(KERN_WARNING "AMD IOMMU: Completion wait loop failed\n");
136 spin_unlock_irqrestore(&iommu->lock, flags);
142 * Command send function for invalidating a device table entry
144 static int iommu_queue_inv_dev_entry(struct amd_iommu *iommu, u16 devid)
146 struct iommu_cmd cmd;
149 BUG_ON(iommu == NULL);
151 memset(&cmd, 0, sizeof(cmd));
152 CMD_SET_TYPE(&cmd, CMD_INV_DEV_ENTRY);
155 ret = iommu_queue_command(iommu, &cmd);
157 iommu->need_sync = 1;
163 * Generic command send function for invalidaing TLB entries
165 static int iommu_queue_inv_iommu_pages(struct amd_iommu *iommu,
166 u64 address, u16 domid, int pde, int s)
168 struct iommu_cmd cmd;
171 memset(&cmd, 0, sizeof(cmd));
172 address &= PAGE_MASK;
173 CMD_SET_TYPE(&cmd, CMD_INV_IOMMU_PAGES);
174 cmd.data[1] |= domid;
175 cmd.data[2] = lower_32_bits(address);
176 cmd.data[3] = upper_32_bits(address);
177 if (s) /* size bit - we flush more than one 4kb page */
178 cmd.data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
179 if (pde) /* PDE bit - we wan't flush everything not only the PTEs */
180 cmd.data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
182 ret = iommu_queue_command(iommu, &cmd);
184 iommu->need_sync = 1;
190 * TLB invalidation function which is called from the mapping functions.
191 * It invalidates a single PTE if the range to flush is within a single
192 * page. Otherwise it flushes the whole TLB of the IOMMU.
194 static int iommu_flush_pages(struct amd_iommu *iommu, u16 domid,
195 u64 address, size_t size)
198 unsigned pages = iommu_num_pages(address, size);
200 address &= PAGE_MASK;
204 * If we have to flush more than one page, flush all
205 * TLB entries for this domain
207 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
211 iommu_queue_inv_iommu_pages(iommu, address, domid, 0, s);
216 /****************************************************************************
218 * The functions below are used the create the page table mappings for
219 * unity mapped regions.
221 ****************************************************************************/
224 * Generic mapping functions. It maps a physical address into a DMA
225 * address space. It allocates the page table pages if necessary.
226 * In the future it can be extended to a generic mapping function
227 * supporting all features of AMD IOMMU page tables like level skipping
228 * and full 64 bit address spaces.
230 static int iommu_map(struct protection_domain *dom,
231 unsigned long bus_addr,
232 unsigned long phys_addr,
235 u64 __pte, *pte, *page;
237 bus_addr = PAGE_ALIGN(bus_addr);
238 phys_addr = PAGE_ALIGN(bus_addr);
240 /* only support 512GB address spaces for now */
241 if (bus_addr > IOMMU_MAP_SIZE_L3 || !(prot & IOMMU_PROT_MASK))
244 pte = &dom->pt_root[IOMMU_PTE_L2_INDEX(bus_addr)];
246 if (!IOMMU_PTE_PRESENT(*pte)) {
247 page = (u64 *)get_zeroed_page(GFP_KERNEL);
250 *pte = IOMMU_L2_PDE(virt_to_phys(page));
253 pte = IOMMU_PTE_PAGE(*pte);
254 pte = &pte[IOMMU_PTE_L1_INDEX(bus_addr)];
256 if (!IOMMU_PTE_PRESENT(*pte)) {
257 page = (u64 *)get_zeroed_page(GFP_KERNEL);
260 *pte = IOMMU_L1_PDE(virt_to_phys(page));
263 pte = IOMMU_PTE_PAGE(*pte);
264 pte = &pte[IOMMU_PTE_L0_INDEX(bus_addr)];
266 if (IOMMU_PTE_PRESENT(*pte))
269 __pte = phys_addr | IOMMU_PTE_P;
270 if (prot & IOMMU_PROT_IR)
271 __pte |= IOMMU_PTE_IR;
272 if (prot & IOMMU_PROT_IW)
273 __pte |= IOMMU_PTE_IW;
281 * This function checks if a specific unity mapping entry is needed for
282 * this specific IOMMU.
284 static int iommu_for_unity_map(struct amd_iommu *iommu,
285 struct unity_map_entry *entry)
289 for (i = entry->devid_start; i <= entry->devid_end; ++i) {
290 bdf = amd_iommu_alias_table[i];
291 if (amd_iommu_rlookup_table[bdf] == iommu)
299 * Init the unity mappings for a specific IOMMU in the system
301 * Basically iterates over all unity mapping entries and applies them to
302 * the default domain DMA of that IOMMU if necessary.
304 static int iommu_init_unity_mappings(struct amd_iommu *iommu)
306 struct unity_map_entry *entry;
309 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
310 if (!iommu_for_unity_map(iommu, entry))
312 ret = dma_ops_unity_map(iommu->default_dom, entry);
321 * This function actually applies the mapping to the page table of the
324 static int dma_ops_unity_map(struct dma_ops_domain *dma_dom,
325 struct unity_map_entry *e)
330 for (addr = e->address_start; addr < e->address_end;
332 ret = iommu_map(&dma_dom->domain, addr, addr, e->prot);
336 * if unity mapping is in aperture range mark the page
337 * as allocated in the aperture
339 if (addr < dma_dom->aperture_size)
340 __set_bit(addr >> PAGE_SHIFT, dma_dom->bitmap);
347 * Inits the unity mappings required for a specific device
349 static int init_unity_mappings_for_device(struct dma_ops_domain *dma_dom,
352 struct unity_map_entry *e;
355 list_for_each_entry(e, &amd_iommu_unity_map, list) {
356 if (!(devid >= e->devid_start && devid <= e->devid_end))
358 ret = dma_ops_unity_map(dma_dom, e);
366 /****************************************************************************
368 * The next functions belong to the address allocator for the dma_ops
369 * interface functions. They work like the allocators in the other IOMMU
370 * drivers. Its basically a bitmap which marks the allocated pages in
371 * the aperture. Maybe it could be enhanced in the future to a more
372 * efficient allocator.
374 ****************************************************************************/
375 static unsigned long dma_mask_to_pages(unsigned long mask)
377 return (mask >> PAGE_SHIFT) +
378 (PAGE_ALIGN(mask & ~PAGE_MASK) >> PAGE_SHIFT);
382 * The address allocator core function.
384 * called with domain->lock held
386 static unsigned long dma_ops_alloc_addresses(struct device *dev,
387 struct dma_ops_domain *dom,
390 unsigned long limit = dma_mask_to_pages(*dev->dma_mask);
391 unsigned long address;
392 unsigned long size = dom->aperture_size >> PAGE_SHIFT;
393 unsigned long boundary_size;
395 boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
396 PAGE_SIZE) >> PAGE_SHIFT;
397 limit = limit < size ? limit : size;
399 if (dom->next_bit >= limit)
402 address = iommu_area_alloc(dom->bitmap, limit, dom->next_bit, pages,
403 0 , boundary_size, 0);
405 address = iommu_area_alloc(dom->bitmap, limit, 0, pages,
406 0, boundary_size, 0);
408 if (likely(address != -1)) {
409 dom->next_bit = address + pages;
410 address <<= PAGE_SHIFT;
412 address = bad_dma_address;
414 WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
420 * The address free function.
422 * called with domain->lock held
424 static void dma_ops_free_addresses(struct dma_ops_domain *dom,
425 unsigned long address,
428 address >>= PAGE_SHIFT;
429 iommu_area_free(dom->bitmap, address, pages);
432 /****************************************************************************
434 * The next functions belong to the domain allocation. A domain is
435 * allocated for every IOMMU as the default domain. If device isolation
436 * is enabled, every device get its own domain. The most important thing
437 * about domains is the page table mapping the DMA address space they
440 ****************************************************************************/
442 static u16 domain_id_alloc(void)
447 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
448 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
450 if (id > 0 && id < MAX_DOMAIN_ID)
451 __set_bit(id, amd_iommu_pd_alloc_bitmap);
454 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
460 * Used to reserve address ranges in the aperture (e.g. for exclusion
463 static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
464 unsigned long start_page,
467 unsigned int last_page = dom->aperture_size >> PAGE_SHIFT;
469 if (start_page + pages > last_page)
470 pages = last_page - start_page;
472 set_bit_string(dom->bitmap, start_page, pages);
475 static void dma_ops_free_pagetable(struct dma_ops_domain *dma_dom)
480 p1 = dma_dom->domain.pt_root;
485 for (i = 0; i < 512; ++i) {
486 if (!IOMMU_PTE_PRESENT(p1[i]))
489 p2 = IOMMU_PTE_PAGE(p1[i]);
490 for (j = 0; j < 512; ++i) {
491 if (!IOMMU_PTE_PRESENT(p2[j]))
493 p3 = IOMMU_PTE_PAGE(p2[j]);
494 free_page((unsigned long)p3);
497 free_page((unsigned long)p2);
500 free_page((unsigned long)p1);
504 * Free a domain, only used if something went wrong in the
505 * allocation path and we need to free an already allocated page table
507 static void dma_ops_domain_free(struct dma_ops_domain *dom)
512 dma_ops_free_pagetable(dom);
514 kfree(dom->pte_pages);
522 * Allocates a new protection domain usable for the dma_ops functions.
523 * It also intializes the page table and the address allocator data
524 * structures required for the dma_ops interface
526 static struct dma_ops_domain *dma_ops_domain_alloc(struct amd_iommu *iommu,
529 struct dma_ops_domain *dma_dom;
530 unsigned i, num_pte_pages;
535 * Currently the DMA aperture must be between 32 MB and 1GB in size
537 if ((order < 25) || (order > 30))
540 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
544 spin_lock_init(&dma_dom->domain.lock);
546 dma_dom->domain.id = domain_id_alloc();
547 if (dma_dom->domain.id == 0)
549 dma_dom->domain.mode = PAGE_MODE_3_LEVEL;
550 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
551 dma_dom->domain.priv = dma_dom;
552 if (!dma_dom->domain.pt_root)
554 dma_dom->aperture_size = (1ULL << order);
555 dma_dom->bitmap = kzalloc(dma_dom->aperture_size / (PAGE_SIZE * 8),
557 if (!dma_dom->bitmap)
560 * mark the first page as allocated so we never return 0 as
561 * a valid dma-address. So we can use 0 as error value
563 dma_dom->bitmap[0] = 1;
564 dma_dom->next_bit = 0;
566 /* Intialize the exclusion range if necessary */
567 if (iommu->exclusion_start &&
568 iommu->exclusion_start < dma_dom->aperture_size) {
569 unsigned long startpage = iommu->exclusion_start >> PAGE_SHIFT;
570 int pages = iommu_num_pages(iommu->exclusion_start,
571 iommu->exclusion_length);
572 dma_ops_reserve_addresses(dma_dom, startpage, pages);
576 * At the last step, build the page tables so we don't need to
577 * allocate page table pages in the dma_ops mapping/unmapping
580 num_pte_pages = dma_dom->aperture_size / (PAGE_SIZE * 512);
581 dma_dom->pte_pages = kzalloc(num_pte_pages * sizeof(void *),
583 if (!dma_dom->pte_pages)
586 l2_pde = (u64 *)get_zeroed_page(GFP_KERNEL);
590 dma_dom->domain.pt_root[0] = IOMMU_L2_PDE(virt_to_phys(l2_pde));
592 for (i = 0; i < num_pte_pages; ++i) {
593 dma_dom->pte_pages[i] = (u64 *)get_zeroed_page(GFP_KERNEL);
594 if (!dma_dom->pte_pages[i])
596 address = virt_to_phys(dma_dom->pte_pages[i]);
597 l2_pde[i] = IOMMU_L1_PDE(address);
603 dma_ops_domain_free(dma_dom);
609 * Find out the protection domain structure for a given PCI device. This
610 * will give us the pointer to the page table root for example.
612 static struct protection_domain *domain_for_device(u16 devid)
614 struct protection_domain *dom;
617 read_lock_irqsave(&amd_iommu_devtable_lock, flags);
618 dom = amd_iommu_pd_table[devid];
619 read_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
625 * If a device is not yet associated with a domain, this function does
626 * assigns it visible for the hardware
628 static void set_device_domain(struct amd_iommu *iommu,
629 struct protection_domain *domain,
634 u64 pte_root = virt_to_phys(domain->pt_root);
636 pte_root |= (domain->mode & 0x07) << 9;
637 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | 2;
639 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
640 amd_iommu_dev_table[devid].data[0] = pte_root;
641 amd_iommu_dev_table[devid].data[1] = pte_root >> 32;
642 amd_iommu_dev_table[devid].data[2] = domain->id;
644 amd_iommu_pd_table[devid] = domain;
645 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
647 iommu_queue_inv_dev_entry(iommu, devid);
649 iommu->need_sync = 1;
652 /*****************************************************************************
654 * The next functions belong to the dma_ops mapping/unmapping code.
656 *****************************************************************************/
659 * In the dma_ops path we only have the struct device. This function
660 * finds the corresponding IOMMU, the protection domain and the
661 * requestor id for a given device.
662 * If the device is not yet associated with a domain this is also done
665 static int get_device_resources(struct device *dev,
666 struct amd_iommu **iommu,
667 struct protection_domain **domain,
670 struct dma_ops_domain *dma_dom;
671 struct pci_dev *pcidev;
674 BUG_ON(!dev || dev->bus != &pci_bus_type || !dev->dma_mask);
676 pcidev = to_pci_dev(dev);
677 _bdf = calc_devid(pcidev->bus->number, pcidev->devfn);
679 /* device not translated by any IOMMU in the system? */
680 if (_bdf > amd_iommu_last_bdf) {
687 *bdf = amd_iommu_alias_table[_bdf];
689 *iommu = amd_iommu_rlookup_table[*bdf];
692 dma_dom = (*iommu)->default_dom;
693 *domain = domain_for_device(*bdf);
694 if (*domain == NULL) {
695 *domain = &dma_dom->domain;
696 set_device_domain(*iommu, *domain, *bdf);
697 printk(KERN_INFO "AMD IOMMU: Using protection domain %d for "
698 "device ", (*domain)->id);
699 print_devid(_bdf, 1);
706 * This is the generic map function. It maps one 4kb page at paddr to
707 * the given address in the DMA address space for the domain.
709 static dma_addr_t dma_ops_domain_map(struct amd_iommu *iommu,
710 struct dma_ops_domain *dom,
711 unsigned long address,
717 WARN_ON(address > dom->aperture_size);
721 pte = dom->pte_pages[IOMMU_PTE_L1_INDEX(address)];
722 pte += IOMMU_PTE_L0_INDEX(address);
724 __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
726 if (direction == DMA_TO_DEVICE)
727 __pte |= IOMMU_PTE_IR;
728 else if (direction == DMA_FROM_DEVICE)
729 __pte |= IOMMU_PTE_IW;
730 else if (direction == DMA_BIDIRECTIONAL)
731 __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
737 return (dma_addr_t)address;
741 * The generic unmapping function for on page in the DMA address space.
743 static void dma_ops_domain_unmap(struct amd_iommu *iommu,
744 struct dma_ops_domain *dom,
745 unsigned long address)
749 if (address >= dom->aperture_size)
752 WARN_ON(address & 0xfffULL || address > dom->aperture_size);
754 pte = dom->pte_pages[IOMMU_PTE_L1_INDEX(address)];
755 pte += IOMMU_PTE_L0_INDEX(address);
763 * This function contains common code for mapping of a physically
764 * contiguous memory region into DMA address space. It is uses by all
765 * mapping functions provided by this IOMMU driver.
766 * Must be called with the domain lock held.
768 static dma_addr_t __map_single(struct device *dev,
769 struct amd_iommu *iommu,
770 struct dma_ops_domain *dma_dom,
775 dma_addr_t offset = paddr & ~PAGE_MASK;
776 dma_addr_t address, start;
780 pages = iommu_num_pages(paddr, size);
783 address = dma_ops_alloc_addresses(dev, dma_dom, pages);
784 if (unlikely(address == bad_dma_address))
788 for (i = 0; i < pages; ++i) {
789 dma_ops_domain_map(iommu, dma_dom, start, paddr, dir);
800 * Does the reverse of the __map_single function. Must be called with
801 * the domain lock held too
803 static void __unmap_single(struct amd_iommu *iommu,
804 struct dma_ops_domain *dma_dom,
812 if ((dma_addr == 0) || (dma_addr + size > dma_dom->aperture_size))
815 pages = iommu_num_pages(dma_addr, size);
816 dma_addr &= PAGE_MASK;
819 for (i = 0; i < pages; ++i) {
820 dma_ops_domain_unmap(iommu, dma_dom, start);
824 dma_ops_free_addresses(dma_dom, dma_addr, pages);
828 * The exported map_single function for dma_ops.
830 static dma_addr_t map_single(struct device *dev, phys_addr_t paddr,
831 size_t size, int dir)
834 struct amd_iommu *iommu;
835 struct protection_domain *domain;
839 get_device_resources(dev, &iommu, &domain, &devid);
841 if (iommu == NULL || domain == NULL)
842 /* device not handled by any AMD IOMMU */
843 return (dma_addr_t)paddr;
845 spin_lock_irqsave(&domain->lock, flags);
846 addr = __map_single(dev, iommu, domain->priv, paddr, size, dir);
847 if (addr == bad_dma_address)
850 if (iommu_has_npcache(iommu))
851 iommu_flush_pages(iommu, domain->id, addr, size);
853 if (iommu->need_sync)
854 iommu_completion_wait(iommu);
857 spin_unlock_irqrestore(&domain->lock, flags);
863 * The exported unmap_single function for dma_ops.
865 static void unmap_single(struct device *dev, dma_addr_t dma_addr,
866 size_t size, int dir)
869 struct amd_iommu *iommu;
870 struct protection_domain *domain;
873 if (!get_device_resources(dev, &iommu, &domain, &devid))
874 /* device not handled by any AMD IOMMU */
877 spin_lock_irqsave(&domain->lock, flags);
879 __unmap_single(iommu, domain->priv, dma_addr, size, dir);
881 iommu_flush_pages(iommu, domain->id, dma_addr, size);
883 if (iommu->need_sync)
884 iommu_completion_wait(iommu);
886 spin_unlock_irqrestore(&domain->lock, flags);
890 * This is a special map_sg function which is used if we should map a
891 * device which is not handled by an AMD IOMMU in the system.
893 static int map_sg_no_iommu(struct device *dev, struct scatterlist *sglist,
896 struct scatterlist *s;
899 for_each_sg(sglist, s, nelems, i) {
900 s->dma_address = (dma_addr_t)sg_phys(s);
901 s->dma_length = s->length;
908 * The exported map_sg function for dma_ops (handles scatter-gather
911 static int map_sg(struct device *dev, struct scatterlist *sglist,
915 struct amd_iommu *iommu;
916 struct protection_domain *domain;
919 struct scatterlist *s;
921 int mapped_elems = 0;
923 get_device_resources(dev, &iommu, &domain, &devid);
925 if (!iommu || !domain)
926 return map_sg_no_iommu(dev, sglist, nelems, dir);
928 spin_lock_irqsave(&domain->lock, flags);
930 for_each_sg(sglist, s, nelems, i) {
933 s->dma_address = __map_single(dev, iommu, domain->priv,
934 paddr, s->length, dir);
936 if (s->dma_address) {
937 s->dma_length = s->length;
941 if (iommu_has_npcache(iommu))
942 iommu_flush_pages(iommu, domain->id, s->dma_address,
946 if (iommu->need_sync)
947 iommu_completion_wait(iommu);
950 spin_unlock_irqrestore(&domain->lock, flags);
954 for_each_sg(sglist, s, mapped_elems, i) {
956 __unmap_single(iommu, domain->priv, s->dma_address,
958 s->dma_address = s->dma_length = 0;
967 * The exported map_sg function for dma_ops (handles scatter-gather
970 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
974 struct amd_iommu *iommu;
975 struct protection_domain *domain;
976 struct scatterlist *s;
980 if (!get_device_resources(dev, &iommu, &domain, &devid))
983 spin_lock_irqsave(&domain->lock, flags);
985 for_each_sg(sglist, s, nelems, i) {
986 __unmap_single(iommu, domain->priv, s->dma_address,
988 iommu_flush_pages(iommu, domain->id, s->dma_address,
990 s->dma_address = s->dma_length = 0;
993 if (iommu->need_sync)
994 iommu_completion_wait(iommu);
996 spin_unlock_irqrestore(&domain->lock, flags);
1000 * The exported alloc_coherent function for dma_ops.
1002 static void *alloc_coherent(struct device *dev, size_t size,
1003 dma_addr_t *dma_addr, gfp_t flag)
1005 unsigned long flags;
1007 struct amd_iommu *iommu;
1008 struct protection_domain *domain;
1012 virt_addr = (void *)__get_free_pages(flag, get_order(size));
1016 memset(virt_addr, 0, size);
1017 paddr = virt_to_phys(virt_addr);
1019 get_device_resources(dev, &iommu, &domain, &devid);
1021 if (!iommu || !domain) {
1022 *dma_addr = (dma_addr_t)paddr;
1026 spin_lock_irqsave(&domain->lock, flags);
1028 *dma_addr = __map_single(dev, iommu, domain->priv, paddr,
1029 size, DMA_BIDIRECTIONAL);
1031 if (*dma_addr == bad_dma_address) {
1032 free_pages((unsigned long)virt_addr, get_order(size));
1037 if (iommu_has_npcache(iommu))
1038 iommu_flush_pages(iommu, domain->id, *dma_addr, size);
1040 if (iommu->need_sync)
1041 iommu_completion_wait(iommu);
1044 spin_unlock_irqrestore(&domain->lock, flags);
1050 * The exported free_coherent function for dma_ops.
1051 * FIXME: fix the generic x86 DMA layer so that it actually calls that
1054 static void free_coherent(struct device *dev, size_t size,
1055 void *virt_addr, dma_addr_t dma_addr)
1057 unsigned long flags;
1058 struct amd_iommu *iommu;
1059 struct protection_domain *domain;
1062 get_device_resources(dev, &iommu, &domain, &devid);
1064 if (!iommu || !domain)
1067 spin_lock_irqsave(&domain->lock, flags);
1069 __unmap_single(iommu, domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
1070 iommu_flush_pages(iommu, domain->id, dma_addr, size);
1072 if (iommu->need_sync)
1073 iommu_completion_wait(iommu);
1075 spin_unlock_irqrestore(&domain->lock, flags);
1078 free_pages((unsigned long)virt_addr, get_order(size));
1082 * The function for pre-allocating protection domains.
1084 * If the driver core informs the DMA layer if a driver grabs a device
1085 * we don't need to preallocate the protection domains anymore.
1086 * For now we have to.
1088 void prealloc_protection_domains(void)
1090 struct pci_dev *dev = NULL;
1091 struct dma_ops_domain *dma_dom;
1092 struct amd_iommu *iommu;
1093 int order = amd_iommu_aperture_order;
1096 while ((dev = pci_get_device(PCI_ANY_ID, PCI_ANY_ID, dev)) != NULL) {
1097 devid = (dev->bus->number << 8) | dev->devfn;
1098 if (devid > amd_iommu_last_bdf)
1100 devid = amd_iommu_alias_table[devid];
1101 if (domain_for_device(devid))
1103 iommu = amd_iommu_rlookup_table[devid];
1106 dma_dom = dma_ops_domain_alloc(iommu, order);
1109 init_unity_mappings_for_device(dma_dom, devid);
1110 set_device_domain(iommu, &dma_dom->domain, devid);
1111 printk(KERN_INFO "AMD IOMMU: Allocated domain %d for device ",
1112 dma_dom->domain.id);
1113 print_devid(devid, 1);
1117 static struct dma_mapping_ops amd_iommu_dma_ops = {
1118 .alloc_coherent = alloc_coherent,
1119 .free_coherent = free_coherent,
1120 .map_single = map_single,
1121 .unmap_single = unmap_single,
1123 .unmap_sg = unmap_sg,
1127 * The function which clues the AMD IOMMU driver into dma_ops.
1129 int __init amd_iommu_init_dma_ops(void)
1131 struct amd_iommu *iommu;
1132 int order = amd_iommu_aperture_order;
1136 * first allocate a default protection domain for every IOMMU we
1137 * found in the system. Devices not assigned to any other
1138 * protection domain will be assigned to the default one.
1140 list_for_each_entry(iommu, &amd_iommu_list, list) {
1141 iommu->default_dom = dma_ops_domain_alloc(iommu, order);
1142 if (iommu->default_dom == NULL)
1144 ret = iommu_init_unity_mappings(iommu);
1150 * If device isolation is enabled, pre-allocate the protection
1151 * domains for each device.
1153 if (amd_iommu_isolate)
1154 prealloc_protection_domains();
1158 bad_dma_address = 0;
1159 #ifdef CONFIG_GART_IOMMU
1160 gart_iommu_aperture_disabled = 1;
1161 gart_iommu_aperture = 0;
1164 /* Make the driver finally visible to the drivers */
1165 dma_ops = &amd_iommu_dma_ops;
1171 list_for_each_entry(iommu, &amd_iommu_list, list) {
1172 if (iommu->default_dom)
1173 dma_ops_domain_free(iommu->default_dom);