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)
106 struct iommu_cmd cmd;
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 ret = iommu_queue_command(iommu, &cmd);
120 while (!ready && (i < EXIT_LOOP_COUNT)) {
122 /* wait for the bit to become one */
123 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
124 ready = status & MMIO_STATUS_COM_WAIT_INT_MASK;
127 /* set bit back to zero */
128 status &= ~MMIO_STATUS_COM_WAIT_INT_MASK;
129 writel(status, iommu->mmio_base + MMIO_STATUS_OFFSET);
131 if (unlikely((i == EXIT_LOOP_COUNT) && printk_ratelimit()))
132 printk(KERN_WARNING "AMD IOMMU: Completion wait loop failed\n");
138 * Command send function for invalidating a device table entry
140 static int iommu_queue_inv_dev_entry(struct amd_iommu *iommu, u16 devid)
142 struct iommu_cmd cmd;
144 BUG_ON(iommu == NULL);
146 memset(&cmd, 0, sizeof(cmd));
147 CMD_SET_TYPE(&cmd, CMD_INV_DEV_ENTRY);
150 iommu->need_sync = 1;
152 return iommu_queue_command(iommu, &cmd);
156 * Generic command send function for invalidaing TLB entries
158 static int iommu_queue_inv_iommu_pages(struct amd_iommu *iommu,
159 u64 address, u16 domid, int pde, int s)
161 struct iommu_cmd cmd;
163 memset(&cmd, 0, sizeof(cmd));
164 address &= PAGE_MASK;
165 CMD_SET_TYPE(&cmd, CMD_INV_IOMMU_PAGES);
166 cmd.data[1] |= domid;
167 cmd.data[2] = lower_32_bits(address);
168 cmd.data[3] = upper_32_bits(address);
169 if (s) /* size bit - we flush more than one 4kb page */
170 cmd.data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
171 if (pde) /* PDE bit - we wan't flush everything not only the PTEs */
172 cmd.data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
174 iommu->need_sync = 1;
176 return iommu_queue_command(iommu, &cmd);
180 * TLB invalidation function which is called from the mapping functions.
181 * It invalidates a single PTE if the range to flush is within a single
182 * page. Otherwise it flushes the whole TLB of the IOMMU.
184 static int iommu_flush_pages(struct amd_iommu *iommu, u16 domid,
185 u64 address, size_t size)
188 unsigned pages = iommu_num_pages(address, size);
190 address &= PAGE_MASK;
194 * If we have to flush more than one page, flush all
195 * TLB entries for this domain
197 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
201 iommu_queue_inv_iommu_pages(iommu, address, domid, 0, s);
206 /****************************************************************************
208 * The functions below are used the create the page table mappings for
209 * unity mapped regions.
211 ****************************************************************************/
214 * Generic mapping functions. It maps a physical address into a DMA
215 * address space. It allocates the page table pages if necessary.
216 * In the future it can be extended to a generic mapping function
217 * supporting all features of AMD IOMMU page tables like level skipping
218 * and full 64 bit address spaces.
220 static int iommu_map(struct protection_domain *dom,
221 unsigned long bus_addr,
222 unsigned long phys_addr,
225 u64 __pte, *pte, *page;
227 bus_addr = PAGE_ALIGN(bus_addr);
228 phys_addr = PAGE_ALIGN(bus_addr);
230 /* only support 512GB address spaces for now */
231 if (bus_addr > IOMMU_MAP_SIZE_L3 || !(prot & IOMMU_PROT_MASK))
234 pte = &dom->pt_root[IOMMU_PTE_L2_INDEX(bus_addr)];
236 if (!IOMMU_PTE_PRESENT(*pte)) {
237 page = (u64 *)get_zeroed_page(GFP_KERNEL);
240 *pte = IOMMU_L2_PDE(virt_to_phys(page));
243 pte = IOMMU_PTE_PAGE(*pte);
244 pte = &pte[IOMMU_PTE_L1_INDEX(bus_addr)];
246 if (!IOMMU_PTE_PRESENT(*pte)) {
247 page = (u64 *)get_zeroed_page(GFP_KERNEL);
250 *pte = IOMMU_L1_PDE(virt_to_phys(page));
253 pte = IOMMU_PTE_PAGE(*pte);
254 pte = &pte[IOMMU_PTE_L0_INDEX(bus_addr)];
256 if (IOMMU_PTE_PRESENT(*pte))
259 __pte = phys_addr | IOMMU_PTE_P;
260 if (prot & IOMMU_PROT_IR)
261 __pte |= IOMMU_PTE_IR;
262 if (prot & IOMMU_PROT_IW)
263 __pte |= IOMMU_PTE_IW;
271 * This function checks if a specific unity mapping entry is needed for
272 * this specific IOMMU.
274 static int iommu_for_unity_map(struct amd_iommu *iommu,
275 struct unity_map_entry *entry)
279 for (i = entry->devid_start; i <= entry->devid_end; ++i) {
280 bdf = amd_iommu_alias_table[i];
281 if (amd_iommu_rlookup_table[bdf] == iommu)
289 * Init the unity mappings for a specific IOMMU in the system
291 * Basically iterates over all unity mapping entries and applies them to
292 * the default domain DMA of that IOMMU if necessary.
294 static int iommu_init_unity_mappings(struct amd_iommu *iommu)
296 struct unity_map_entry *entry;
299 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
300 if (!iommu_for_unity_map(iommu, entry))
302 ret = dma_ops_unity_map(iommu->default_dom, entry);
311 * This function actually applies the mapping to the page table of the
314 static int dma_ops_unity_map(struct dma_ops_domain *dma_dom,
315 struct unity_map_entry *e)
320 for (addr = e->address_start; addr < e->address_end;
322 ret = iommu_map(&dma_dom->domain, addr, addr, e->prot);
326 * if unity mapping is in aperture range mark the page
327 * as allocated in the aperture
329 if (addr < dma_dom->aperture_size)
330 __set_bit(addr >> PAGE_SHIFT, dma_dom->bitmap);
337 * Inits the unity mappings required for a specific device
339 static int init_unity_mappings_for_device(struct dma_ops_domain *dma_dom,
342 struct unity_map_entry *e;
345 list_for_each_entry(e, &amd_iommu_unity_map, list) {
346 if (!(devid >= e->devid_start && devid <= e->devid_end))
348 ret = dma_ops_unity_map(dma_dom, e);
356 /****************************************************************************
358 * The next functions belong to the address allocator for the dma_ops
359 * interface functions. They work like the allocators in the other IOMMU
360 * drivers. Its basically a bitmap which marks the allocated pages in
361 * the aperture. Maybe it could be enhanced in the future to a more
362 * efficient allocator.
364 ****************************************************************************/
365 static unsigned long dma_mask_to_pages(unsigned long mask)
367 return (mask >> PAGE_SHIFT) +
368 (PAGE_ALIGN(mask & ~PAGE_MASK) >> PAGE_SHIFT);
372 * The address allocator core function.
374 * called with domain->lock held
376 static unsigned long dma_ops_alloc_addresses(struct device *dev,
377 struct dma_ops_domain *dom,
380 unsigned long limit = dma_mask_to_pages(*dev->dma_mask);
381 unsigned long address;
382 unsigned long size = dom->aperture_size >> PAGE_SHIFT;
383 unsigned long boundary_size;
385 boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
386 PAGE_SIZE) >> PAGE_SHIFT;
387 limit = limit < size ? limit : size;
389 if (dom->next_bit >= limit)
392 address = iommu_area_alloc(dom->bitmap, limit, dom->next_bit, pages,
393 0 , boundary_size, 0);
395 address = iommu_area_alloc(dom->bitmap, limit, 0, pages,
396 0, boundary_size, 0);
398 if (likely(address != -1)) {
399 dom->next_bit = address + pages;
400 address <<= PAGE_SHIFT;
402 address = bad_dma_address;
404 WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
410 * The address free function.
412 * called with domain->lock held
414 static void dma_ops_free_addresses(struct dma_ops_domain *dom,
415 unsigned long address,
418 address >>= PAGE_SHIFT;
419 iommu_area_free(dom->bitmap, address, pages);
422 /****************************************************************************
424 * The next functions belong to the domain allocation. A domain is
425 * allocated for every IOMMU as the default domain. If device isolation
426 * is enabled, every device get its own domain. The most important thing
427 * about domains is the page table mapping the DMA address space they
430 ****************************************************************************/
432 static u16 domain_id_alloc(void)
437 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
438 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
440 if (id > 0 && id < MAX_DOMAIN_ID)
441 __set_bit(id, amd_iommu_pd_alloc_bitmap);
444 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
450 * Used to reserve address ranges in the aperture (e.g. for exclusion
453 static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
454 unsigned long start_page,
457 unsigned int last_page = dom->aperture_size >> PAGE_SHIFT;
459 if (start_page + pages > last_page)
460 pages = last_page - start_page;
462 set_bit_string(dom->bitmap, start_page, pages);
465 static void dma_ops_free_pagetable(struct dma_ops_domain *dma_dom)
470 p1 = dma_dom->domain.pt_root;
475 for (i = 0; i < 512; ++i) {
476 if (!IOMMU_PTE_PRESENT(p1[i]))
479 p2 = IOMMU_PTE_PAGE(p1[i]);
480 for (j = 0; j < 512; ++i) {
481 if (!IOMMU_PTE_PRESENT(p2[j]))
483 p3 = IOMMU_PTE_PAGE(p2[j]);
484 free_page((unsigned long)p3);
487 free_page((unsigned long)p2);
490 free_page((unsigned long)p1);
494 * Free a domain, only used if something went wrong in the
495 * allocation path and we need to free an already allocated page table
497 static void dma_ops_domain_free(struct dma_ops_domain *dom)
502 dma_ops_free_pagetable(dom);
504 kfree(dom->pte_pages);
512 * Allocates a new protection domain usable for the dma_ops functions.
513 * It also intializes the page table and the address allocator data
514 * structures required for the dma_ops interface
516 static struct dma_ops_domain *dma_ops_domain_alloc(struct amd_iommu *iommu,
519 struct dma_ops_domain *dma_dom;
520 unsigned i, num_pte_pages;
525 * Currently the DMA aperture must be between 32 MB and 1GB in size
527 if ((order < 25) || (order > 30))
530 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
534 spin_lock_init(&dma_dom->domain.lock);
536 dma_dom->domain.id = domain_id_alloc();
537 if (dma_dom->domain.id == 0)
539 dma_dom->domain.mode = PAGE_MODE_3_LEVEL;
540 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
541 dma_dom->domain.priv = dma_dom;
542 if (!dma_dom->domain.pt_root)
544 dma_dom->aperture_size = (1ULL << order);
545 dma_dom->bitmap = kzalloc(dma_dom->aperture_size / (PAGE_SIZE * 8),
547 if (!dma_dom->bitmap)
550 * mark the first page as allocated so we never return 0 as
551 * a valid dma-address. So we can use 0 as error value
553 dma_dom->bitmap[0] = 1;
554 dma_dom->next_bit = 0;
556 /* Intialize the exclusion range if necessary */
557 if (iommu->exclusion_start &&
558 iommu->exclusion_start < dma_dom->aperture_size) {
559 unsigned long startpage = iommu->exclusion_start >> PAGE_SHIFT;
560 int pages = iommu_num_pages(iommu->exclusion_start,
561 iommu->exclusion_length);
562 dma_ops_reserve_addresses(dma_dom, startpage, pages);
566 * At the last step, build the page tables so we don't need to
567 * allocate page table pages in the dma_ops mapping/unmapping
570 num_pte_pages = dma_dom->aperture_size / (PAGE_SIZE * 512);
571 dma_dom->pte_pages = kzalloc(num_pte_pages * sizeof(void *),
573 if (!dma_dom->pte_pages)
576 l2_pde = (u64 *)get_zeroed_page(GFP_KERNEL);
580 dma_dom->domain.pt_root[0] = IOMMU_L2_PDE(virt_to_phys(l2_pde));
582 for (i = 0; i < num_pte_pages; ++i) {
583 dma_dom->pte_pages[i] = (u64 *)get_zeroed_page(GFP_KERNEL);
584 if (!dma_dom->pte_pages[i])
586 address = virt_to_phys(dma_dom->pte_pages[i]);
587 l2_pde[i] = IOMMU_L1_PDE(address);
593 dma_ops_domain_free(dma_dom);
599 * Find out the protection domain structure for a given PCI device. This
600 * will give us the pointer to the page table root for example.
602 static struct protection_domain *domain_for_device(u16 devid)
604 struct protection_domain *dom;
607 read_lock_irqsave(&amd_iommu_devtable_lock, flags);
608 dom = amd_iommu_pd_table[devid];
609 read_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
615 * If a device is not yet associated with a domain, this function does
616 * assigns it visible for the hardware
618 static void set_device_domain(struct amd_iommu *iommu,
619 struct protection_domain *domain,
624 u64 pte_root = virt_to_phys(domain->pt_root);
626 pte_root |= (domain->mode & 0x07) << 9;
627 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | 2;
629 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
630 amd_iommu_dev_table[devid].data[0] = pte_root;
631 amd_iommu_dev_table[devid].data[1] = pte_root >> 32;
632 amd_iommu_dev_table[devid].data[2] = domain->id;
634 amd_iommu_pd_table[devid] = domain;
635 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
637 iommu_queue_inv_dev_entry(iommu, devid);
639 iommu->need_sync = 1;
642 /*****************************************************************************
644 * The next functions belong to the dma_ops mapping/unmapping code.
646 *****************************************************************************/
649 * In the dma_ops path we only have the struct device. This function
650 * finds the corresponding IOMMU, the protection domain and the
651 * requestor id for a given device.
652 * If the device is not yet associated with a domain this is also done
655 static int get_device_resources(struct device *dev,
656 struct amd_iommu **iommu,
657 struct protection_domain **domain,
660 struct dma_ops_domain *dma_dom;
661 struct pci_dev *pcidev;
664 BUG_ON(!dev || dev->bus != &pci_bus_type || !dev->dma_mask);
666 pcidev = to_pci_dev(dev);
667 _bdf = calc_devid(pcidev->bus->number, pcidev->devfn);
669 /* device not translated by any IOMMU in the system? */
670 if (_bdf > amd_iommu_last_bdf) {
677 *bdf = amd_iommu_alias_table[_bdf];
679 *iommu = amd_iommu_rlookup_table[*bdf];
682 dma_dom = (*iommu)->default_dom;
683 *domain = domain_for_device(*bdf);
684 if (*domain == NULL) {
685 *domain = &dma_dom->domain;
686 set_device_domain(*iommu, *domain, *bdf);
687 printk(KERN_INFO "AMD IOMMU: Using protection domain %d for "
688 "device ", (*domain)->id);
689 print_devid(_bdf, 1);
696 * This is the generic map function. It maps one 4kb page at paddr to
697 * the given address in the DMA address space for the domain.
699 static dma_addr_t dma_ops_domain_map(struct amd_iommu *iommu,
700 struct dma_ops_domain *dom,
701 unsigned long address,
707 WARN_ON(address > dom->aperture_size);
711 pte = dom->pte_pages[IOMMU_PTE_L1_INDEX(address)];
712 pte += IOMMU_PTE_L0_INDEX(address);
714 __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
716 if (direction == DMA_TO_DEVICE)
717 __pte |= IOMMU_PTE_IR;
718 else if (direction == DMA_FROM_DEVICE)
719 __pte |= IOMMU_PTE_IW;
720 else if (direction == DMA_BIDIRECTIONAL)
721 __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
727 return (dma_addr_t)address;
731 * The generic unmapping function for on page in the DMA address space.
733 static void dma_ops_domain_unmap(struct amd_iommu *iommu,
734 struct dma_ops_domain *dom,
735 unsigned long address)
739 if (address >= dom->aperture_size)
742 WARN_ON(address & 0xfffULL || address > dom->aperture_size);
744 pte = dom->pte_pages[IOMMU_PTE_L1_INDEX(address)];
745 pte += IOMMU_PTE_L0_INDEX(address);
753 * This function contains common code for mapping of a physically
754 * contiguous memory region into DMA address space. It is uses by all
755 * mapping functions provided by this IOMMU driver.
756 * Must be called with the domain lock held.
758 static dma_addr_t __map_single(struct device *dev,
759 struct amd_iommu *iommu,
760 struct dma_ops_domain *dma_dom,
765 dma_addr_t offset = paddr & ~PAGE_MASK;
766 dma_addr_t address, start;
770 pages = iommu_num_pages(paddr, size);
773 address = dma_ops_alloc_addresses(dev, dma_dom, pages);
774 if (unlikely(address == bad_dma_address))
778 for (i = 0; i < pages; ++i) {
779 dma_ops_domain_map(iommu, dma_dom, start, paddr, dir);
790 * Does the reverse of the __map_single function. Must be called with
791 * the domain lock held too
793 static void __unmap_single(struct amd_iommu *iommu,
794 struct dma_ops_domain *dma_dom,
802 if ((dma_addr == 0) || (dma_addr + size > dma_dom->aperture_size))
805 pages = iommu_num_pages(dma_addr, size);
806 dma_addr &= PAGE_MASK;
809 for (i = 0; i < pages; ++i) {
810 dma_ops_domain_unmap(iommu, dma_dom, start);
814 dma_ops_free_addresses(dma_dom, dma_addr, pages);
818 * The exported map_single function for dma_ops.
820 static dma_addr_t map_single(struct device *dev, phys_addr_t paddr,
821 size_t size, int dir)
824 struct amd_iommu *iommu;
825 struct protection_domain *domain;
829 get_device_resources(dev, &iommu, &domain, &devid);
831 if (iommu == NULL || domain == NULL)
832 /* device not handled by any AMD IOMMU */
833 return (dma_addr_t)paddr;
835 spin_lock_irqsave(&domain->lock, flags);
836 addr = __map_single(dev, iommu, domain->priv, paddr, size, dir);
837 if (addr == bad_dma_address)
840 if (iommu_has_npcache(iommu))
841 iommu_flush_pages(iommu, domain->id, addr, size);
843 if (iommu->need_sync)
844 iommu_completion_wait(iommu);
847 spin_unlock_irqrestore(&domain->lock, flags);
853 * The exported unmap_single function for dma_ops.
855 static void unmap_single(struct device *dev, dma_addr_t dma_addr,
856 size_t size, int dir)
859 struct amd_iommu *iommu;
860 struct protection_domain *domain;
863 if (!get_device_resources(dev, &iommu, &domain, &devid))
864 /* device not handled by any AMD IOMMU */
867 spin_lock_irqsave(&domain->lock, flags);
869 __unmap_single(iommu, domain->priv, dma_addr, size, dir);
871 iommu_flush_pages(iommu, domain->id, dma_addr, size);
873 if (iommu->need_sync)
874 iommu_completion_wait(iommu);
876 spin_unlock_irqrestore(&domain->lock, flags);
880 * This is a special map_sg function which is used if we should map a
881 * device which is not handled by an AMD IOMMU in the system.
883 static int map_sg_no_iommu(struct device *dev, struct scatterlist *sglist,
886 struct scatterlist *s;
889 for_each_sg(sglist, s, nelems, i) {
890 s->dma_address = (dma_addr_t)sg_phys(s);
891 s->dma_length = s->length;
898 * The exported map_sg function for dma_ops (handles scatter-gather
901 static int map_sg(struct device *dev, struct scatterlist *sglist,
905 struct amd_iommu *iommu;
906 struct protection_domain *domain;
909 struct scatterlist *s;
911 int mapped_elems = 0;
913 get_device_resources(dev, &iommu, &domain, &devid);
915 if (!iommu || !domain)
916 return map_sg_no_iommu(dev, sglist, nelems, dir);
918 spin_lock_irqsave(&domain->lock, flags);
920 for_each_sg(sglist, s, nelems, i) {
923 s->dma_address = __map_single(dev, iommu, domain->priv,
924 paddr, s->length, dir);
926 if (s->dma_address) {
927 s->dma_length = s->length;
931 if (iommu_has_npcache(iommu))
932 iommu_flush_pages(iommu, domain->id, s->dma_address,
936 if (iommu->need_sync)
937 iommu_completion_wait(iommu);
940 spin_unlock_irqrestore(&domain->lock, flags);
944 for_each_sg(sglist, s, mapped_elems, i) {
946 __unmap_single(iommu, domain->priv, s->dma_address,
948 s->dma_address = s->dma_length = 0;
957 * The exported map_sg function for dma_ops (handles scatter-gather
960 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
964 struct amd_iommu *iommu;
965 struct protection_domain *domain;
966 struct scatterlist *s;
970 if (!get_device_resources(dev, &iommu, &domain, &devid))
973 spin_lock_irqsave(&domain->lock, flags);
975 for_each_sg(sglist, s, nelems, i) {
976 __unmap_single(iommu, domain->priv, s->dma_address,
978 iommu_flush_pages(iommu, domain->id, s->dma_address,
980 s->dma_address = s->dma_length = 0;
983 if (iommu->need_sync)
984 iommu_completion_wait(iommu);
986 spin_unlock_irqrestore(&domain->lock, flags);
990 * The exported alloc_coherent function for dma_ops.
992 static void *alloc_coherent(struct device *dev, size_t size,
993 dma_addr_t *dma_addr, gfp_t flag)
997 struct amd_iommu *iommu;
998 struct protection_domain *domain;
1002 virt_addr = (void *)__get_free_pages(flag, get_order(size));
1006 memset(virt_addr, 0, size);
1007 paddr = virt_to_phys(virt_addr);
1009 get_device_resources(dev, &iommu, &domain, &devid);
1011 if (!iommu || !domain) {
1012 *dma_addr = (dma_addr_t)paddr;
1016 spin_lock_irqsave(&domain->lock, flags);
1018 *dma_addr = __map_single(dev, iommu, domain->priv, paddr,
1019 size, DMA_BIDIRECTIONAL);
1021 if (*dma_addr == bad_dma_address) {
1022 free_pages((unsigned long)virt_addr, get_order(size));
1027 if (iommu_has_npcache(iommu))
1028 iommu_flush_pages(iommu, domain->id, *dma_addr, size);
1030 if (iommu->need_sync)
1031 iommu_completion_wait(iommu);
1034 spin_unlock_irqrestore(&domain->lock, flags);
1040 * The exported free_coherent function for dma_ops.
1041 * FIXME: fix the generic x86 DMA layer so that it actually calls that
1044 static void free_coherent(struct device *dev, size_t size,
1045 void *virt_addr, dma_addr_t dma_addr)
1047 unsigned long flags;
1048 struct amd_iommu *iommu;
1049 struct protection_domain *domain;
1052 get_device_resources(dev, &iommu, &domain, &devid);
1054 if (!iommu || !domain)
1057 spin_lock_irqsave(&domain->lock, flags);
1059 __unmap_single(iommu, domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
1060 iommu_flush_pages(iommu, domain->id, dma_addr, size);
1062 if (iommu->need_sync)
1063 iommu_completion_wait(iommu);
1065 spin_unlock_irqrestore(&domain->lock, flags);
1068 free_pages((unsigned long)virt_addr, get_order(size));
1072 * The function for pre-allocating protection domains.
1074 * If the driver core informs the DMA layer if a driver grabs a device
1075 * we don't need to preallocate the protection domains anymore.
1076 * For now we have to.
1078 void prealloc_protection_domains(void)
1080 struct pci_dev *dev = NULL;
1081 struct dma_ops_domain *dma_dom;
1082 struct amd_iommu *iommu;
1083 int order = amd_iommu_aperture_order;
1086 while ((dev = pci_get_device(PCI_ANY_ID, PCI_ANY_ID, dev)) != NULL) {
1087 devid = (dev->bus->number << 8) | dev->devfn;
1088 if (devid > amd_iommu_last_bdf)
1090 devid = amd_iommu_alias_table[devid];
1091 if (domain_for_device(devid))
1093 iommu = amd_iommu_rlookup_table[devid];
1096 dma_dom = dma_ops_domain_alloc(iommu, order);
1099 init_unity_mappings_for_device(dma_dom, devid);
1100 set_device_domain(iommu, &dma_dom->domain, devid);
1101 printk(KERN_INFO "AMD IOMMU: Allocated domain %d for device ",
1102 dma_dom->domain.id);
1103 print_devid(devid, 1);
1107 static struct dma_mapping_ops amd_iommu_dma_ops = {
1108 .alloc_coherent = alloc_coherent,
1109 .free_coherent = free_coherent,
1110 .map_single = map_single,
1111 .unmap_single = unmap_single,
1113 .unmap_sg = unmap_sg,
1117 * The function which clues the AMD IOMMU driver into dma_ops.
1119 int __init amd_iommu_init_dma_ops(void)
1121 struct amd_iommu *iommu;
1122 int order = amd_iommu_aperture_order;
1126 * first allocate a default protection domain for every IOMMU we
1127 * found in the system. Devices not assigned to any other
1128 * protection domain will be assigned to the default one.
1130 list_for_each_entry(iommu, &amd_iommu_list, list) {
1131 iommu->default_dom = dma_ops_domain_alloc(iommu, order);
1132 if (iommu->default_dom == NULL)
1134 ret = iommu_init_unity_mappings(iommu);
1140 * If device isolation is enabled, pre-allocate the protection
1141 * domains for each device.
1143 if (amd_iommu_isolate)
1144 prealloc_protection_domains();
1148 bad_dma_address = 0;
1149 #ifdef CONFIG_GART_IOMMU
1150 gart_iommu_aperture_disabled = 1;
1151 gart_iommu_aperture = 0;
1154 /* Make the driver finally visible to the drivers */
1155 dma_ops = &amd_iommu_dma_ops;
1161 list_for_each_entry(iommu, &amd_iommu_list, list) {
1162 if (iommu->default_dom)
1163 dma_ops_domain_free(iommu->default_dom);