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>
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 to_pages(addr, size) \
33 (round_up(((addr) & ~PAGE_MASK) + (size), PAGE_SIZE) >> PAGE_SHIFT)
35 static DEFINE_RWLOCK(amd_iommu_devtable_lock);
38 * general struct to manage commands send to an IOMMU
44 static int dma_ops_unity_map(struct dma_ops_domain *dma_dom,
45 struct unity_map_entry *e);
47 /* returns !0 if the IOMMU is caching non-present entries in its TLB */
48 static int iommu_has_npcache(struct amd_iommu *iommu)
50 return iommu->cap & IOMMU_CAP_NPCACHE;
53 /****************************************************************************
55 * IOMMU command queuing functions
57 ****************************************************************************/
60 * Writes the command to the IOMMUs command buffer and informs the
61 * hardware about the new command. Must be called with iommu->lock held.
63 static int __iommu_queue_command(struct amd_iommu *iommu, struct command *cmd)
68 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
69 target = (iommu->cmd_buf + tail);
70 memcpy_toio(target, cmd, sizeof(*cmd));
71 tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
72 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
75 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
81 * General queuing function for commands. Takes iommu->lock and calls
82 * __iommu_queue_command().
84 static int iommu_queue_command(struct amd_iommu *iommu, struct command *cmd)
89 spin_lock_irqsave(&iommu->lock, flags);
90 ret = __iommu_queue_command(iommu, cmd);
91 spin_unlock_irqrestore(&iommu->lock, flags);
97 * This function is called whenever we need to ensure that the IOMMU has
98 * completed execution of all commands we sent. It sends a
99 * COMPLETION_WAIT command and waits for it to finish. The IOMMU informs
100 * us about that by writing a value to a physical address we pass with
103 static int iommu_completion_wait(struct amd_iommu *iommu)
107 volatile u64 ready = 0;
108 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);
130 * Command send function for invalidating a device table entry
132 static int iommu_queue_inv_dev_entry(struct amd_iommu *iommu, u16 devid)
136 BUG_ON(iommu == NULL);
138 memset(&cmd, 0, sizeof(cmd));
139 CMD_SET_TYPE(&cmd, CMD_INV_DEV_ENTRY);
142 iommu->need_sync = 1;
144 return iommu_queue_command(iommu, &cmd);
148 * Generic command send function for invalidaing TLB entries
150 static int iommu_queue_inv_iommu_pages(struct amd_iommu *iommu,
151 u64 address, u16 domid, int pde, int s)
155 memset(&cmd, 0, sizeof(cmd));
156 address &= PAGE_MASK;
157 CMD_SET_TYPE(&cmd, CMD_INV_IOMMU_PAGES);
158 cmd.data[1] |= domid;
159 cmd.data[2] = LOW_U32(address);
160 cmd.data[3] = upper_32_bits(address);
161 if (s) /* size bit - we flush more than one 4kb page */
162 cmd.data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
163 if (pde) /* PDE bit - we wan't flush everything not only the PTEs */
164 cmd.data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
166 iommu->need_sync = 1;
168 return iommu_queue_command(iommu, &cmd);
172 * TLB invalidation function which is called from the mapping functions.
173 * It invalidates a single PTE if the range to flush is within a single
174 * page. Otherwise it flushes the whole TLB of the IOMMU.
176 static int iommu_flush_pages(struct amd_iommu *iommu, u16 domid,
177 u64 address, size_t size)
180 unsigned pages = to_pages(address, size);
182 address &= PAGE_MASK;
186 * If we have to flush more than one page, flush all
187 * TLB entries for this domain
189 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
193 iommu_queue_inv_iommu_pages(iommu, address, domid, 0, s);
198 /****************************************************************************
200 * The functions below are used the create the page table mappings for
201 * unity mapped regions.
203 ****************************************************************************/
206 * Generic mapping functions. It maps a physical address into a DMA
207 * address space. It allocates the page table pages if necessary.
208 * In the future it can be extended to a generic mapping function
209 * supporting all features of AMD IOMMU page tables like level skipping
210 * and full 64 bit address spaces.
212 static int iommu_map(struct protection_domain *dom,
213 unsigned long bus_addr,
214 unsigned long phys_addr,
217 u64 __pte, *pte, *page;
219 bus_addr = PAGE_ALIGN(bus_addr);
220 phys_addr = PAGE_ALIGN(bus_addr);
222 /* only support 512GB address spaces for now */
223 if (bus_addr > IOMMU_MAP_SIZE_L3 || !(prot & IOMMU_PROT_MASK))
226 pte = &dom->pt_root[IOMMU_PTE_L2_INDEX(bus_addr)];
228 if (!IOMMU_PTE_PRESENT(*pte)) {
229 page = (u64 *)get_zeroed_page(GFP_KERNEL);
232 *pte = IOMMU_L2_PDE(virt_to_phys(page));
235 pte = IOMMU_PTE_PAGE(*pte);
236 pte = &pte[IOMMU_PTE_L1_INDEX(bus_addr)];
238 if (!IOMMU_PTE_PRESENT(*pte)) {
239 page = (u64 *)get_zeroed_page(GFP_KERNEL);
242 *pte = IOMMU_L1_PDE(virt_to_phys(page));
245 pte = IOMMU_PTE_PAGE(*pte);
246 pte = &pte[IOMMU_PTE_L0_INDEX(bus_addr)];
248 if (IOMMU_PTE_PRESENT(*pte))
251 __pte = phys_addr | IOMMU_PTE_P;
252 if (prot & IOMMU_PROT_IR)
253 __pte |= IOMMU_PTE_IR;
254 if (prot & IOMMU_PROT_IW)
255 __pte |= IOMMU_PTE_IW;
263 * This function checks if a specific unity mapping entry is needed for
264 * this specific IOMMU.
266 static int iommu_for_unity_map(struct amd_iommu *iommu,
267 struct unity_map_entry *entry)
271 for (i = entry->devid_start; i <= entry->devid_end; ++i) {
272 bdf = amd_iommu_alias_table[i];
273 if (amd_iommu_rlookup_table[bdf] == iommu)
281 * Init the unity mappings for a specific IOMMU in the system
283 * Basically iterates over all unity mapping entries and applies them to
284 * the default domain DMA of that IOMMU if necessary.
286 static int iommu_init_unity_mappings(struct amd_iommu *iommu)
288 struct unity_map_entry *entry;
291 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
292 if (!iommu_for_unity_map(iommu, entry))
294 ret = dma_ops_unity_map(iommu->default_dom, entry);
303 * This function actually applies the mapping to the page table of the
306 static int dma_ops_unity_map(struct dma_ops_domain *dma_dom,
307 struct unity_map_entry *e)
312 for (addr = e->address_start; addr < e->address_end;
314 ret = iommu_map(&dma_dom->domain, addr, addr, e->prot);
318 * if unity mapping is in aperture range mark the page
319 * as allocated in the aperture
321 if (addr < dma_dom->aperture_size)
322 __set_bit(addr >> PAGE_SHIFT, dma_dom->bitmap);
329 * Inits the unity mappings required for a specific device
331 static int init_unity_mappings_for_device(struct dma_ops_domain *dma_dom,
334 struct unity_map_entry *e;
337 list_for_each_entry(e, &amd_iommu_unity_map, list) {
338 if (!(devid >= e->devid_start && devid <= e->devid_end))
340 ret = dma_ops_unity_map(dma_dom, e);
348 /****************************************************************************
350 * The next functions belong to the address allocator for the dma_ops
351 * interface functions. They work like the allocators in the other IOMMU
352 * drivers. Its basically a bitmap which marks the allocated pages in
353 * the aperture. Maybe it could be enhanced in the future to a more
354 * efficient allocator.
356 ****************************************************************************/
357 static unsigned long dma_mask_to_pages(unsigned long mask)
359 return (mask >> PAGE_SHIFT) +
360 (PAGE_ALIGN(mask & ~PAGE_MASK) >> PAGE_SHIFT);
364 * The address allocator core function.
366 * called with domain->lock held
368 static unsigned long dma_ops_alloc_addresses(struct device *dev,
369 struct dma_ops_domain *dom,
372 unsigned long limit = dma_mask_to_pages(*dev->dma_mask);
373 unsigned long address;
374 unsigned long size = dom->aperture_size >> PAGE_SHIFT;
375 unsigned long boundary_size;
377 boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
378 PAGE_SIZE) >> PAGE_SHIFT;
379 limit = limit < size ? limit : size;
381 if (dom->next_bit >= limit)
384 address = iommu_area_alloc(dom->bitmap, limit, dom->next_bit, pages,
385 0 , boundary_size, 0);
387 address = iommu_area_alloc(dom->bitmap, limit, 0, pages,
388 0, boundary_size, 0);
390 if (likely(address != -1)) {
391 dom->next_bit = address + pages;
392 address <<= PAGE_SHIFT;
394 address = bad_dma_address;
396 WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
402 * The address free function.
404 * called with domain->lock held
406 static void dma_ops_free_addresses(struct dma_ops_domain *dom,
407 unsigned long address,
410 address >>= PAGE_SHIFT;
411 iommu_area_free(dom->bitmap, address, pages);
414 /****************************************************************************
416 * The next functions belong to the domain allocation. A domain is
417 * allocated for every IOMMU as the default domain. If device isolation
418 * is enabled, every device get its own domain. The most important thing
419 * about domains is the page table mapping the DMA address space they
422 ****************************************************************************/
424 static u16 domain_id_alloc(void)
429 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
430 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
432 if (id > 0 && id < MAX_DOMAIN_ID)
433 __set_bit(id, amd_iommu_pd_alloc_bitmap);
436 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
442 * Used to reserve address ranges in the aperture (e.g. for exclusion
445 static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
446 unsigned long start_page,
449 unsigned int last_page = dom->aperture_size >> PAGE_SHIFT;
451 if (start_page + pages > last_page)
452 pages = last_page - start_page;
454 set_bit_string(dom->bitmap, start_page, pages);
457 static void dma_ops_free_pagetable(struct dma_ops_domain *dma_dom)
462 p1 = dma_dom->domain.pt_root;
467 for (i = 0; i < 512; ++i) {
468 if (!IOMMU_PTE_PRESENT(p1[i]))
471 p2 = IOMMU_PTE_PAGE(p1[i]);
472 for (j = 0; j < 512; ++i) {
473 if (!IOMMU_PTE_PRESENT(p2[j]))
475 p3 = IOMMU_PTE_PAGE(p2[j]);
476 free_page((unsigned long)p3);
479 free_page((unsigned long)p2);
482 free_page((unsigned long)p1);
486 * Free a domain, only used if something went wrong in the
487 * allocation path and we need to free an already allocated page table
489 static void dma_ops_domain_free(struct dma_ops_domain *dom)
494 dma_ops_free_pagetable(dom);
496 kfree(dom->pte_pages);
504 * Allocates a new protection domain usable for the dma_ops functions.
505 * It also intializes the page table and the address allocator data
506 * structures required for the dma_ops interface
508 static struct dma_ops_domain *dma_ops_domain_alloc(struct amd_iommu *iommu,
511 struct dma_ops_domain *dma_dom;
512 unsigned i, num_pte_pages;
517 * Currently the DMA aperture must be between 32 MB and 1GB in size
519 if ((order < 25) || (order > 30))
522 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
526 spin_lock_init(&dma_dom->domain.lock);
528 dma_dom->domain.id = domain_id_alloc();
529 if (dma_dom->domain.id == 0)
531 dma_dom->domain.mode = PAGE_MODE_3_LEVEL;
532 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
533 dma_dom->domain.priv = dma_dom;
534 if (!dma_dom->domain.pt_root)
536 dma_dom->aperture_size = (1ULL << order);
537 dma_dom->bitmap = kzalloc(dma_dom->aperture_size / (PAGE_SIZE * 8),
539 if (!dma_dom->bitmap)
542 * mark the first page as allocated so we never return 0 as
543 * a valid dma-address. So we can use 0 as error value
545 dma_dom->bitmap[0] = 1;
546 dma_dom->next_bit = 0;
548 /* Intialize the exclusion range if necessary */
549 if (iommu->exclusion_start &&
550 iommu->exclusion_start < dma_dom->aperture_size) {
551 unsigned long startpage = iommu->exclusion_start >> PAGE_SHIFT;
552 int pages = to_pages(iommu->exclusion_start,
553 iommu->exclusion_length);
554 dma_ops_reserve_addresses(dma_dom, startpage, pages);
558 * At the last step, build the page tables so we don't need to
559 * allocate page table pages in the dma_ops mapping/unmapping
562 num_pte_pages = dma_dom->aperture_size / (PAGE_SIZE * 512);
563 dma_dom->pte_pages = kzalloc(num_pte_pages * sizeof(void *),
565 if (!dma_dom->pte_pages)
568 l2_pde = (u64 *)get_zeroed_page(GFP_KERNEL);
572 dma_dom->domain.pt_root[0] = IOMMU_L2_PDE(virt_to_phys(l2_pde));
574 for (i = 0; i < num_pte_pages; ++i) {
575 dma_dom->pte_pages[i] = (u64 *)get_zeroed_page(GFP_KERNEL);
576 if (!dma_dom->pte_pages[i])
578 address = virt_to_phys(dma_dom->pte_pages[i]);
579 l2_pde[i] = IOMMU_L1_PDE(address);
585 dma_ops_domain_free(dma_dom);
591 * Find out the protection domain structure for a given PCI device. This
592 * will give us the pointer to the page table root for example.
594 static struct protection_domain *domain_for_device(u16 devid)
596 struct protection_domain *dom;
599 read_lock_irqsave(&amd_iommu_devtable_lock, flags);
600 dom = amd_iommu_pd_table[devid];
601 read_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
607 * If a device is not yet associated with a domain, this function does
608 * assigns it visible for the hardware
610 static void set_device_domain(struct amd_iommu *iommu,
611 struct protection_domain *domain,
616 u64 pte_root = virt_to_phys(domain->pt_root);
618 pte_root |= (domain->mode & 0x07) << 9;
619 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | 2;
621 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
622 amd_iommu_dev_table[devid].data[0] = pte_root;
623 amd_iommu_dev_table[devid].data[1] = pte_root >> 32;
624 amd_iommu_dev_table[devid].data[2] = domain->id;
626 amd_iommu_pd_table[devid] = domain;
627 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
629 iommu_queue_inv_dev_entry(iommu, devid);
631 iommu->need_sync = 1;
634 /*****************************************************************************
636 * The next functions belong to the dma_ops mapping/unmapping code.
638 *****************************************************************************/
641 * In the dma_ops path we only have the struct device. This function
642 * finds the corresponding IOMMU, the protection domain and the
643 * requestor id for a given device.
644 * If the device is not yet associated with a domain this is also done
647 static int get_device_resources(struct device *dev,
648 struct amd_iommu **iommu,
649 struct protection_domain **domain,
652 struct dma_ops_domain *dma_dom;
653 struct pci_dev *pcidev;
656 BUG_ON(!dev || dev->bus != &pci_bus_type || !dev->dma_mask);
658 pcidev = to_pci_dev(dev);
659 _bdf = (pcidev->bus->number << 8) | pcidev->devfn;
661 /* device not translated by any IOMMU in the system? */
662 if (_bdf >= amd_iommu_last_bdf) {
669 *bdf = amd_iommu_alias_table[_bdf];
671 *iommu = amd_iommu_rlookup_table[*bdf];
674 dma_dom = (*iommu)->default_dom;
675 *domain = domain_for_device(*bdf);
676 if (*domain == NULL) {
677 *domain = &dma_dom->domain;
678 set_device_domain(*iommu, *domain, *bdf);
679 printk(KERN_INFO "AMD IOMMU: Using protection domain %d for "
680 "device ", (*domain)->id);
681 print_devid(_bdf, 1);
688 * This is the generic map function. It maps one 4kb page at paddr to
689 * the given address in the DMA address space for the domain.
691 static dma_addr_t dma_ops_domain_map(struct amd_iommu *iommu,
692 struct dma_ops_domain *dom,
693 unsigned long address,
699 WARN_ON(address > dom->aperture_size);
703 pte = dom->pte_pages[IOMMU_PTE_L1_INDEX(address)];
704 pte += IOMMU_PTE_L0_INDEX(address);
706 __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
708 if (direction == DMA_TO_DEVICE)
709 __pte |= IOMMU_PTE_IR;
710 else if (direction == DMA_FROM_DEVICE)
711 __pte |= IOMMU_PTE_IW;
712 else if (direction == DMA_BIDIRECTIONAL)
713 __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
719 return (dma_addr_t)address;
723 * The generic unmapping function for on page in the DMA address space.
725 static void dma_ops_domain_unmap(struct amd_iommu *iommu,
726 struct dma_ops_domain *dom,
727 unsigned long address)
731 if (address >= dom->aperture_size)
734 WARN_ON(address & 0xfffULL || address > dom->aperture_size);
736 pte = dom->pte_pages[IOMMU_PTE_L1_INDEX(address)];
737 pte += IOMMU_PTE_L0_INDEX(address);
745 * This function contains common code for mapping of a physically
746 * contiguous memory region into DMA address space. It is uses by all
747 * mapping functions provided by this IOMMU driver.
748 * Must be called with the domain lock held.
750 static dma_addr_t __map_single(struct device *dev,
751 struct amd_iommu *iommu,
752 struct dma_ops_domain *dma_dom,
757 dma_addr_t offset = paddr & ~PAGE_MASK;
758 dma_addr_t address, start;
762 pages = to_pages(paddr, size);
765 address = dma_ops_alloc_addresses(dev, dma_dom, pages);
766 if (unlikely(address == bad_dma_address))
770 for (i = 0; i < pages; ++i) {
771 dma_ops_domain_map(iommu, dma_dom, start, paddr, dir);
782 * Does the reverse of the __map_single function. Must be called with
783 * the domain lock held too
785 static void __unmap_single(struct amd_iommu *iommu,
786 struct dma_ops_domain *dma_dom,
794 if ((dma_addr == 0) || (dma_addr + size > dma_dom->aperture_size))
797 pages = to_pages(dma_addr, size);
798 dma_addr &= PAGE_MASK;
801 for (i = 0; i < pages; ++i) {
802 dma_ops_domain_unmap(iommu, dma_dom, start);
806 dma_ops_free_addresses(dma_dom, dma_addr, pages);
810 * The exported map_single function for dma_ops.
812 static dma_addr_t map_single(struct device *dev, phys_addr_t paddr,
813 size_t size, int dir)
816 struct amd_iommu *iommu;
817 struct protection_domain *domain;
821 get_device_resources(dev, &iommu, &domain, &devid);
823 if (iommu == NULL || domain == NULL)
824 /* device not handled by any AMD IOMMU */
825 return (dma_addr_t)paddr;
827 spin_lock_irqsave(&domain->lock, flags);
828 addr = __map_single(dev, iommu, domain->priv, paddr, size, dir);
829 if (addr == bad_dma_address)
832 if (iommu_has_npcache(iommu))
833 iommu_flush_pages(iommu, domain->id, addr, size);
835 if (iommu->need_sync)
836 iommu_completion_wait(iommu);
839 spin_unlock_irqrestore(&domain->lock, flags);
845 * The exported unmap_single function for dma_ops.
847 static void unmap_single(struct device *dev, dma_addr_t dma_addr,
848 size_t size, int dir)
851 struct amd_iommu *iommu;
852 struct protection_domain *domain;
855 if (!get_device_resources(dev, &iommu, &domain, &devid))
856 /* device not handled by any AMD IOMMU */
859 spin_lock_irqsave(&domain->lock, flags);
861 __unmap_single(iommu, domain->priv, dma_addr, size, dir);
863 iommu_flush_pages(iommu, domain->id, dma_addr, size);
865 if (iommu->need_sync)
866 iommu_completion_wait(iommu);
868 spin_unlock_irqrestore(&domain->lock, flags);
872 * This is a special map_sg function which is used if we should map a
873 * device which is not handled by an AMD IOMMU in the system.
875 static int map_sg_no_iommu(struct device *dev, struct scatterlist *sglist,
878 struct scatterlist *s;
881 for_each_sg(sglist, s, nelems, i) {
882 s->dma_address = (dma_addr_t)sg_phys(s);
883 s->dma_length = s->length;
890 * The exported map_sg function for dma_ops (handles scatter-gather
893 static int map_sg(struct device *dev, struct scatterlist *sglist,
897 struct amd_iommu *iommu;
898 struct protection_domain *domain;
901 struct scatterlist *s;
903 int mapped_elems = 0;
905 get_device_resources(dev, &iommu, &domain, &devid);
907 if (!iommu || !domain)
908 return map_sg_no_iommu(dev, sglist, nelems, dir);
910 spin_lock_irqsave(&domain->lock, flags);
912 for_each_sg(sglist, s, nelems, i) {
915 s->dma_address = __map_single(dev, iommu, domain->priv,
916 paddr, s->length, dir);
918 if (s->dma_address) {
919 s->dma_length = s->length;
923 if (iommu_has_npcache(iommu))
924 iommu_flush_pages(iommu, domain->id, s->dma_address,
928 if (iommu->need_sync)
929 iommu_completion_wait(iommu);
932 spin_unlock_irqrestore(&domain->lock, flags);
936 for_each_sg(sglist, s, mapped_elems, i) {
938 __unmap_single(iommu, domain->priv, s->dma_address,
940 s->dma_address = s->dma_length = 0;
949 * The exported map_sg function for dma_ops (handles scatter-gather
952 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
956 struct amd_iommu *iommu;
957 struct protection_domain *domain;
958 struct scatterlist *s;
962 if (!get_device_resources(dev, &iommu, &domain, &devid))
965 spin_lock_irqsave(&domain->lock, flags);
967 for_each_sg(sglist, s, nelems, i) {
968 __unmap_single(iommu, domain->priv, s->dma_address,
970 iommu_flush_pages(iommu, domain->id, s->dma_address,
972 s->dma_address = s->dma_length = 0;
975 if (iommu->need_sync)
976 iommu_completion_wait(iommu);
978 spin_unlock_irqrestore(&domain->lock, flags);
982 * The exported alloc_coherent function for dma_ops.
984 static void *alloc_coherent(struct device *dev, size_t size,
985 dma_addr_t *dma_addr, gfp_t flag)
989 struct amd_iommu *iommu;
990 struct protection_domain *domain;
994 virt_addr = (void *)__get_free_pages(flag, get_order(size));
998 memset(virt_addr, 0, size);
999 paddr = virt_to_phys(virt_addr);
1001 get_device_resources(dev, &iommu, &domain, &devid);
1003 if (!iommu || !domain) {
1004 *dma_addr = (dma_addr_t)paddr;
1008 spin_lock_irqsave(&domain->lock, flags);
1010 *dma_addr = __map_single(dev, iommu, domain->priv, paddr,
1011 size, DMA_BIDIRECTIONAL);
1013 if (*dma_addr == bad_dma_address) {
1014 free_pages((unsigned long)virt_addr, get_order(size));
1019 if (iommu_has_npcache(iommu))
1020 iommu_flush_pages(iommu, domain->id, *dma_addr, size);
1022 if (iommu->need_sync)
1023 iommu_completion_wait(iommu);
1026 spin_unlock_irqrestore(&domain->lock, flags);
1032 * The exported free_coherent function for dma_ops.
1033 * FIXME: fix the generic x86 DMA layer so that it actually calls that
1036 static void free_coherent(struct device *dev, size_t size,
1037 void *virt_addr, dma_addr_t dma_addr)
1039 unsigned long flags;
1040 struct amd_iommu *iommu;
1041 struct protection_domain *domain;
1044 get_device_resources(dev, &iommu, &domain, &devid);
1046 if (!iommu || !domain)
1049 spin_lock_irqsave(&domain->lock, flags);
1051 __unmap_single(iommu, domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
1052 iommu_flush_pages(iommu, domain->id, dma_addr, size);
1054 if (iommu->need_sync)
1055 iommu_completion_wait(iommu);
1057 spin_unlock_irqrestore(&domain->lock, flags);
1060 free_pages((unsigned long)virt_addr, get_order(size));
1064 * The function for pre-allocating protection domains.
1066 * If the driver core informs the DMA layer if a driver grabs a device
1067 * we don't need to preallocate the protection domains anymore.
1068 * For now we have to.
1070 void prealloc_protection_domains(void)
1072 struct pci_dev *dev = NULL;
1073 struct dma_ops_domain *dma_dom;
1074 struct amd_iommu *iommu;
1075 int order = amd_iommu_aperture_order;
1078 while ((dev = pci_get_device(PCI_ANY_ID, PCI_ANY_ID, dev)) != NULL) {
1079 devid = (dev->bus->number << 8) | dev->devfn;
1080 if (devid >= amd_iommu_last_bdf)
1082 devid = amd_iommu_alias_table[devid];
1083 if (domain_for_device(devid))
1085 iommu = amd_iommu_rlookup_table[devid];
1088 dma_dom = dma_ops_domain_alloc(iommu, order);
1091 init_unity_mappings_for_device(dma_dom, devid);
1092 set_device_domain(iommu, &dma_dom->domain, devid);
1093 printk(KERN_INFO "AMD IOMMU: Allocated domain %d for device ",
1094 dma_dom->domain.id);
1095 print_devid(devid, 1);
1099 static struct dma_mapping_ops amd_iommu_dma_ops = {
1100 .alloc_coherent = alloc_coherent,
1101 .free_coherent = free_coherent,
1102 .map_single = map_single,
1103 .unmap_single = unmap_single,
1105 .unmap_sg = unmap_sg,
1109 * The function which clues the AMD IOMMU driver into dma_ops.
1111 int __init amd_iommu_init_dma_ops(void)
1113 struct amd_iommu *iommu;
1114 int order = amd_iommu_aperture_order;
1118 * first allocate a default protection domain for every IOMMU we
1119 * found in the system. Devices not assigned to any other
1120 * protection domain will be assigned to the default one.
1122 list_for_each_entry(iommu, &amd_iommu_list, list) {
1123 iommu->default_dom = dma_ops_domain_alloc(iommu, order);
1124 if (iommu->default_dom == NULL)
1126 ret = iommu_init_unity_mappings(iommu);
1132 * If device isolation is enabled, pre-allocate the protection
1133 * domains for each device.
1135 if (amd_iommu_isolate)
1136 prealloc_protection_domains();
1140 bad_dma_address = 0;
1141 #ifdef CONFIG_GART_IOMMU
1142 gart_iommu_aperture_disabled = 1;
1143 gart_iommu_aperture = 0;
1146 /* Make the driver finally visible to the drivers */
1147 dma_ops = &amd_iommu_dma_ops;
1153 list_for_each_entry(iommu, &amd_iommu_list, list) {
1154 if (iommu->default_dom)
1155 dma_ops_domain_free(iommu->default_dom);