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);
36 /* A list of preallocated protection domains */
37 static LIST_HEAD(iommu_pd_list);
38 static DEFINE_SPINLOCK(iommu_pd_list_lock);
41 * general struct to manage commands send to an IOMMU
47 static int dma_ops_unity_map(struct dma_ops_domain *dma_dom,
48 struct unity_map_entry *e);
50 /* returns !0 if the IOMMU is caching non-present entries in its TLB */
51 static int iommu_has_npcache(struct amd_iommu *iommu)
53 return iommu->cap & IOMMU_CAP_NPCACHE;
56 /****************************************************************************
58 * Interrupt handling functions
60 ****************************************************************************/
62 static void iommu_print_event(void *__evt)
65 int type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
66 int devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
67 int domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
68 int flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
69 u64 address = (u64)(((u64)event[3]) << 32) | event[2];
71 printk(KERN_ERR "AMD IOMMU: Event logged [");
74 case EVENT_TYPE_ILL_DEV:
75 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
76 "address=0x%016llx flags=0x%04x]\n",
77 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
80 case EVENT_TYPE_IO_FAULT:
81 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
82 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
83 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
84 domid, address, flags);
86 case EVENT_TYPE_DEV_TAB_ERR:
87 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
88 "address=0x%016llx flags=0x%04x]\n",
89 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
92 case EVENT_TYPE_PAGE_TAB_ERR:
93 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
94 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
95 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
96 domid, address, flags);
98 case EVENT_TYPE_ILL_CMD:
99 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
101 case EVENT_TYPE_CMD_HARD_ERR:
102 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
103 "flags=0x%04x]\n", address, flags);
105 case EVENT_TYPE_IOTLB_INV_TO:
106 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
107 "address=0x%016llx]\n",
108 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
111 case EVENT_TYPE_INV_DEV_REQ:
112 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
113 "address=0x%016llx flags=0x%04x]\n",
114 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
118 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
122 static void iommu_poll_events(struct amd_iommu *iommu)
127 spin_lock_irqsave(&iommu->lock, flags);
129 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
130 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
132 while (head != tail) {
133 iommu_print_event(iommu->evt_buf + head);
134 head = (head + EVENT_ENTRY_SIZE) % iommu->evt_buf_size;
137 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
139 spin_unlock_irqrestore(&iommu->lock, flags);
142 irqreturn_t amd_iommu_int_handler(int irq, void *data)
144 struct amd_iommu *iommu;
146 list_for_each_entry(iommu, &amd_iommu_list, list)
147 iommu_poll_events(iommu);
152 /****************************************************************************
154 * IOMMU command queuing functions
156 ****************************************************************************/
159 * Writes the command to the IOMMUs command buffer and informs the
160 * hardware about the new command. Must be called with iommu->lock held.
162 static int __iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
167 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
168 target = iommu->cmd_buf + tail;
169 memcpy_toio(target, cmd, sizeof(*cmd));
170 tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
171 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
174 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
180 * General queuing function for commands. Takes iommu->lock and calls
181 * __iommu_queue_command().
183 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
188 spin_lock_irqsave(&iommu->lock, flags);
189 ret = __iommu_queue_command(iommu, cmd);
190 spin_unlock_irqrestore(&iommu->lock, flags);
196 * This function is called whenever we need to ensure that the IOMMU has
197 * completed execution of all commands we sent. It sends a
198 * COMPLETION_WAIT command and waits for it to finish. The IOMMU informs
199 * us about that by writing a value to a physical address we pass with
202 static int iommu_completion_wait(struct amd_iommu *iommu)
204 int ret = 0, ready = 0;
206 struct iommu_cmd cmd;
207 unsigned long flags, i = 0;
209 memset(&cmd, 0, sizeof(cmd));
210 cmd.data[0] = CMD_COMPL_WAIT_INT_MASK;
211 CMD_SET_TYPE(&cmd, CMD_COMPL_WAIT);
213 iommu->need_sync = 0;
215 spin_lock_irqsave(&iommu->lock, flags);
217 ret = __iommu_queue_command(iommu, &cmd);
222 while (!ready && (i < EXIT_LOOP_COUNT)) {
224 /* wait for the bit to become one */
225 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
226 ready = status & MMIO_STATUS_COM_WAIT_INT_MASK;
229 /* set bit back to zero */
230 status &= ~MMIO_STATUS_COM_WAIT_INT_MASK;
231 writel(status, iommu->mmio_base + MMIO_STATUS_OFFSET);
233 if (unlikely((i == EXIT_LOOP_COUNT) && printk_ratelimit()))
234 printk(KERN_WARNING "AMD IOMMU: Completion wait loop failed\n");
236 spin_unlock_irqrestore(&iommu->lock, flags);
242 * Command send function for invalidating a device table entry
244 static int iommu_queue_inv_dev_entry(struct amd_iommu *iommu, u16 devid)
246 struct iommu_cmd cmd;
249 BUG_ON(iommu == NULL);
251 memset(&cmd, 0, sizeof(cmd));
252 CMD_SET_TYPE(&cmd, CMD_INV_DEV_ENTRY);
255 ret = iommu_queue_command(iommu, &cmd);
257 iommu->need_sync = 1;
263 * Generic command send function for invalidaing TLB entries
265 static int iommu_queue_inv_iommu_pages(struct amd_iommu *iommu,
266 u64 address, u16 domid, int pde, int s)
268 struct iommu_cmd cmd;
271 memset(&cmd, 0, sizeof(cmd));
272 address &= PAGE_MASK;
273 CMD_SET_TYPE(&cmd, CMD_INV_IOMMU_PAGES);
274 cmd.data[1] |= domid;
275 cmd.data[2] = lower_32_bits(address);
276 cmd.data[3] = upper_32_bits(address);
277 if (s) /* size bit - we flush more than one 4kb page */
278 cmd.data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
279 if (pde) /* PDE bit - we wan't flush everything not only the PTEs */
280 cmd.data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
282 ret = iommu_queue_command(iommu, &cmd);
284 iommu->need_sync = 1;
290 * TLB invalidation function which is called from the mapping functions.
291 * It invalidates a single PTE if the range to flush is within a single
292 * page. Otherwise it flushes the whole TLB of the IOMMU.
294 static int iommu_flush_pages(struct amd_iommu *iommu, u16 domid,
295 u64 address, size_t size)
298 unsigned pages = iommu_num_pages(address, size, PAGE_SIZE);
300 address &= PAGE_MASK;
304 * If we have to flush more than one page, flush all
305 * TLB entries for this domain
307 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
311 iommu_queue_inv_iommu_pages(iommu, address, domid, 0, s);
316 /* Flush the whole IO/TLB for a given protection domain */
317 static void iommu_flush_tlb(struct amd_iommu *iommu, u16 domid)
319 u64 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
321 iommu_queue_inv_iommu_pages(iommu, address, domid, 0, 1);
324 /****************************************************************************
326 * The functions below are used the create the page table mappings for
327 * unity mapped regions.
329 ****************************************************************************/
332 * Generic mapping functions. It maps a physical address into a DMA
333 * address space. It allocates the page table pages if necessary.
334 * In the future it can be extended to a generic mapping function
335 * supporting all features of AMD IOMMU page tables like level skipping
336 * and full 64 bit address spaces.
338 static int iommu_map(struct protection_domain *dom,
339 unsigned long bus_addr,
340 unsigned long phys_addr,
343 u64 __pte, *pte, *page;
345 bus_addr = PAGE_ALIGN(bus_addr);
346 phys_addr = PAGE_ALIGN(bus_addr);
348 /* only support 512GB address spaces for now */
349 if (bus_addr > IOMMU_MAP_SIZE_L3 || !(prot & IOMMU_PROT_MASK))
352 pte = &dom->pt_root[IOMMU_PTE_L2_INDEX(bus_addr)];
354 if (!IOMMU_PTE_PRESENT(*pte)) {
355 page = (u64 *)get_zeroed_page(GFP_KERNEL);
358 *pte = IOMMU_L2_PDE(virt_to_phys(page));
361 pte = IOMMU_PTE_PAGE(*pte);
362 pte = &pte[IOMMU_PTE_L1_INDEX(bus_addr)];
364 if (!IOMMU_PTE_PRESENT(*pte)) {
365 page = (u64 *)get_zeroed_page(GFP_KERNEL);
368 *pte = IOMMU_L1_PDE(virt_to_phys(page));
371 pte = IOMMU_PTE_PAGE(*pte);
372 pte = &pte[IOMMU_PTE_L0_INDEX(bus_addr)];
374 if (IOMMU_PTE_PRESENT(*pte))
377 __pte = phys_addr | IOMMU_PTE_P;
378 if (prot & IOMMU_PROT_IR)
379 __pte |= IOMMU_PTE_IR;
380 if (prot & IOMMU_PROT_IW)
381 __pte |= IOMMU_PTE_IW;
389 * This function checks if a specific unity mapping entry is needed for
390 * this specific IOMMU.
392 static int iommu_for_unity_map(struct amd_iommu *iommu,
393 struct unity_map_entry *entry)
397 for (i = entry->devid_start; i <= entry->devid_end; ++i) {
398 bdf = amd_iommu_alias_table[i];
399 if (amd_iommu_rlookup_table[bdf] == iommu)
407 * Init the unity mappings for a specific IOMMU in the system
409 * Basically iterates over all unity mapping entries and applies them to
410 * the default domain DMA of that IOMMU if necessary.
412 static int iommu_init_unity_mappings(struct amd_iommu *iommu)
414 struct unity_map_entry *entry;
417 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
418 if (!iommu_for_unity_map(iommu, entry))
420 ret = dma_ops_unity_map(iommu->default_dom, entry);
429 * This function actually applies the mapping to the page table of the
432 static int dma_ops_unity_map(struct dma_ops_domain *dma_dom,
433 struct unity_map_entry *e)
438 for (addr = e->address_start; addr < e->address_end;
440 ret = iommu_map(&dma_dom->domain, addr, addr, e->prot);
444 * if unity mapping is in aperture range mark the page
445 * as allocated in the aperture
447 if (addr < dma_dom->aperture_size)
448 __set_bit(addr >> PAGE_SHIFT, dma_dom->bitmap);
455 * Inits the unity mappings required for a specific device
457 static int init_unity_mappings_for_device(struct dma_ops_domain *dma_dom,
460 struct unity_map_entry *e;
463 list_for_each_entry(e, &amd_iommu_unity_map, list) {
464 if (!(devid >= e->devid_start && devid <= e->devid_end))
466 ret = dma_ops_unity_map(dma_dom, e);
474 /****************************************************************************
476 * The next functions belong to the address allocator for the dma_ops
477 * interface functions. They work like the allocators in the other IOMMU
478 * drivers. Its basically a bitmap which marks the allocated pages in
479 * the aperture. Maybe it could be enhanced in the future to a more
480 * efficient allocator.
482 ****************************************************************************/
485 * The address allocator core function.
487 * called with domain->lock held
489 static unsigned long dma_ops_alloc_addresses(struct device *dev,
490 struct dma_ops_domain *dom,
492 unsigned long align_mask,
496 unsigned long address;
497 unsigned long boundary_size;
499 boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
500 PAGE_SIZE) >> PAGE_SHIFT;
501 limit = iommu_device_max_index(dom->aperture_size >> PAGE_SHIFT, 0,
502 dma_mask >> PAGE_SHIFT);
504 if (dom->next_bit >= limit) {
506 dom->need_flush = true;
509 address = iommu_area_alloc(dom->bitmap, limit, dom->next_bit, pages,
510 0 , boundary_size, align_mask);
512 address = iommu_area_alloc(dom->bitmap, limit, 0, pages,
513 0, boundary_size, align_mask);
514 dom->need_flush = true;
517 if (likely(address != -1)) {
518 dom->next_bit = address + pages;
519 address <<= PAGE_SHIFT;
521 address = bad_dma_address;
523 WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
529 * The address free function.
531 * called with domain->lock held
533 static void dma_ops_free_addresses(struct dma_ops_domain *dom,
534 unsigned long address,
537 address >>= PAGE_SHIFT;
538 iommu_area_free(dom->bitmap, address, pages);
541 /****************************************************************************
543 * The next functions belong to the domain allocation. A domain is
544 * allocated for every IOMMU as the default domain. If device isolation
545 * is enabled, every device get its own domain. The most important thing
546 * about domains is the page table mapping the DMA address space they
549 ****************************************************************************/
551 static u16 domain_id_alloc(void)
556 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
557 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
559 if (id > 0 && id < MAX_DOMAIN_ID)
560 __set_bit(id, amd_iommu_pd_alloc_bitmap);
563 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
569 * Used to reserve address ranges in the aperture (e.g. for exclusion
572 static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
573 unsigned long start_page,
576 unsigned int last_page = dom->aperture_size >> PAGE_SHIFT;
578 if (start_page + pages > last_page)
579 pages = last_page - start_page;
581 iommu_area_reserve(dom->bitmap, start_page, pages);
584 static void dma_ops_free_pagetable(struct dma_ops_domain *dma_dom)
589 p1 = dma_dom->domain.pt_root;
594 for (i = 0; i < 512; ++i) {
595 if (!IOMMU_PTE_PRESENT(p1[i]))
598 p2 = IOMMU_PTE_PAGE(p1[i]);
599 for (j = 0; j < 512; ++i) {
600 if (!IOMMU_PTE_PRESENT(p2[j]))
602 p3 = IOMMU_PTE_PAGE(p2[j]);
603 free_page((unsigned long)p3);
606 free_page((unsigned long)p2);
609 free_page((unsigned long)p1);
613 * Free a domain, only used if something went wrong in the
614 * allocation path and we need to free an already allocated page table
616 static void dma_ops_domain_free(struct dma_ops_domain *dom)
621 dma_ops_free_pagetable(dom);
623 kfree(dom->pte_pages);
631 * Allocates a new protection domain usable for the dma_ops functions.
632 * It also intializes the page table and the address allocator data
633 * structures required for the dma_ops interface
635 static struct dma_ops_domain *dma_ops_domain_alloc(struct amd_iommu *iommu,
638 struct dma_ops_domain *dma_dom;
639 unsigned i, num_pte_pages;
644 * Currently the DMA aperture must be between 32 MB and 1GB in size
646 if ((order < 25) || (order > 30))
649 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
653 spin_lock_init(&dma_dom->domain.lock);
655 dma_dom->domain.id = domain_id_alloc();
656 if (dma_dom->domain.id == 0)
658 dma_dom->domain.mode = PAGE_MODE_3_LEVEL;
659 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
660 dma_dom->domain.priv = dma_dom;
661 if (!dma_dom->domain.pt_root)
663 dma_dom->aperture_size = (1ULL << order);
664 dma_dom->bitmap = kzalloc(dma_dom->aperture_size / (PAGE_SIZE * 8),
666 if (!dma_dom->bitmap)
669 * mark the first page as allocated so we never return 0 as
670 * a valid dma-address. So we can use 0 as error value
672 dma_dom->bitmap[0] = 1;
673 dma_dom->next_bit = 0;
675 dma_dom->need_flush = false;
676 dma_dom->target_dev = 0xffff;
678 /* Intialize the exclusion range if necessary */
679 if (iommu->exclusion_start &&
680 iommu->exclusion_start < dma_dom->aperture_size) {
681 unsigned long startpage = iommu->exclusion_start >> PAGE_SHIFT;
682 int pages = iommu_num_pages(iommu->exclusion_start,
683 iommu->exclusion_length,
685 dma_ops_reserve_addresses(dma_dom, startpage, pages);
689 * At the last step, build the page tables so we don't need to
690 * allocate page table pages in the dma_ops mapping/unmapping
693 num_pte_pages = dma_dom->aperture_size / (PAGE_SIZE * 512);
694 dma_dom->pte_pages = kzalloc(num_pte_pages * sizeof(void *),
696 if (!dma_dom->pte_pages)
699 l2_pde = (u64 *)get_zeroed_page(GFP_KERNEL);
703 dma_dom->domain.pt_root[0] = IOMMU_L2_PDE(virt_to_phys(l2_pde));
705 for (i = 0; i < num_pte_pages; ++i) {
706 dma_dom->pte_pages[i] = (u64 *)get_zeroed_page(GFP_KERNEL);
707 if (!dma_dom->pte_pages[i])
709 address = virt_to_phys(dma_dom->pte_pages[i]);
710 l2_pde[i] = IOMMU_L1_PDE(address);
716 dma_ops_domain_free(dma_dom);
722 * Find out the protection domain structure for a given PCI device. This
723 * will give us the pointer to the page table root for example.
725 static struct protection_domain *domain_for_device(u16 devid)
727 struct protection_domain *dom;
730 read_lock_irqsave(&amd_iommu_devtable_lock, flags);
731 dom = amd_iommu_pd_table[devid];
732 read_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
738 * If a device is not yet associated with a domain, this function does
739 * assigns it visible for the hardware
741 static void set_device_domain(struct amd_iommu *iommu,
742 struct protection_domain *domain,
747 u64 pte_root = virt_to_phys(domain->pt_root);
749 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
750 << DEV_ENTRY_MODE_SHIFT;
751 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
753 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
754 amd_iommu_dev_table[devid].data[0] = lower_32_bits(pte_root);
755 amd_iommu_dev_table[devid].data[1] = upper_32_bits(pte_root);
756 amd_iommu_dev_table[devid].data[2] = domain->id;
758 amd_iommu_pd_table[devid] = domain;
759 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
761 iommu_queue_inv_dev_entry(iommu, devid);
763 iommu->need_sync = 1;
766 /*****************************************************************************
768 * The next functions belong to the dma_ops mapping/unmapping code.
770 *****************************************************************************/
773 * This function checks if the driver got a valid device from the caller to
774 * avoid dereferencing invalid pointers.
776 static bool check_device(struct device *dev)
778 if (!dev || !dev->dma_mask)
785 * In this function the list of preallocated protection domains is traversed to
786 * find the domain for a specific device
788 static struct dma_ops_domain *find_protection_domain(u16 devid)
790 struct dma_ops_domain *entry, *ret = NULL;
793 if (list_empty(&iommu_pd_list))
796 spin_lock_irqsave(&iommu_pd_list_lock, flags);
798 list_for_each_entry(entry, &iommu_pd_list, list) {
799 if (entry->target_dev == devid) {
801 list_del(&ret->list);
806 spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
812 * In the dma_ops path we only have the struct device. This function
813 * finds the corresponding IOMMU, the protection domain and the
814 * requestor id for a given device.
815 * If the device is not yet associated with a domain this is also done
818 static int get_device_resources(struct device *dev,
819 struct amd_iommu **iommu,
820 struct protection_domain **domain,
823 struct dma_ops_domain *dma_dom;
824 struct pci_dev *pcidev;
831 if (dev->bus != &pci_bus_type)
834 pcidev = to_pci_dev(dev);
835 _bdf = calc_devid(pcidev->bus->number, pcidev->devfn);
837 /* device not translated by any IOMMU in the system? */
838 if (_bdf > amd_iommu_last_bdf)
841 *bdf = amd_iommu_alias_table[_bdf];
843 *iommu = amd_iommu_rlookup_table[*bdf];
846 *domain = domain_for_device(*bdf);
847 if (*domain == NULL) {
848 dma_dom = find_protection_domain(*bdf);
850 dma_dom = (*iommu)->default_dom;
851 *domain = &dma_dom->domain;
852 set_device_domain(*iommu, *domain, *bdf);
853 printk(KERN_INFO "AMD IOMMU: Using protection domain %d for "
854 "device ", (*domain)->id);
855 print_devid(_bdf, 1);
862 * This is the generic map function. It maps one 4kb page at paddr to
863 * the given address in the DMA address space for the domain.
865 static dma_addr_t dma_ops_domain_map(struct amd_iommu *iommu,
866 struct dma_ops_domain *dom,
867 unsigned long address,
873 WARN_ON(address > dom->aperture_size);
877 pte = dom->pte_pages[IOMMU_PTE_L1_INDEX(address)];
878 pte += IOMMU_PTE_L0_INDEX(address);
880 __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
882 if (direction == DMA_TO_DEVICE)
883 __pte |= IOMMU_PTE_IR;
884 else if (direction == DMA_FROM_DEVICE)
885 __pte |= IOMMU_PTE_IW;
886 else if (direction == DMA_BIDIRECTIONAL)
887 __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
893 return (dma_addr_t)address;
897 * The generic unmapping function for on page in the DMA address space.
899 static void dma_ops_domain_unmap(struct amd_iommu *iommu,
900 struct dma_ops_domain *dom,
901 unsigned long address)
905 if (address >= dom->aperture_size)
908 WARN_ON(address & 0xfffULL || address > dom->aperture_size);
910 pte = dom->pte_pages[IOMMU_PTE_L1_INDEX(address)];
911 pte += IOMMU_PTE_L0_INDEX(address);
919 * This function contains common code for mapping of a physically
920 * contiguous memory region into DMA address space. It is uses by all
921 * mapping functions provided by this IOMMU driver.
922 * Must be called with the domain lock held.
924 static dma_addr_t __map_single(struct device *dev,
925 struct amd_iommu *iommu,
926 struct dma_ops_domain *dma_dom,
933 dma_addr_t offset = paddr & ~PAGE_MASK;
934 dma_addr_t address, start;
936 unsigned long align_mask = 0;
939 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
943 align_mask = (1UL << get_order(size)) - 1;
945 address = dma_ops_alloc_addresses(dev, dma_dom, pages, align_mask,
947 if (unlikely(address == bad_dma_address))
951 for (i = 0; i < pages; ++i) {
952 dma_ops_domain_map(iommu, dma_dom, start, paddr, dir);
958 if (unlikely(dma_dom->need_flush && !amd_iommu_unmap_flush)) {
959 iommu_flush_tlb(iommu, dma_dom->domain.id);
960 dma_dom->need_flush = false;
961 } else if (unlikely(iommu_has_npcache(iommu)))
962 iommu_flush_pages(iommu, dma_dom->domain.id, address, size);
969 * Does the reverse of the __map_single function. Must be called with
970 * the domain lock held too
972 static void __unmap_single(struct amd_iommu *iommu,
973 struct dma_ops_domain *dma_dom,
981 if ((dma_addr == 0) || (dma_addr + size > dma_dom->aperture_size))
984 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
985 dma_addr &= PAGE_MASK;
988 for (i = 0; i < pages; ++i) {
989 dma_ops_domain_unmap(iommu, dma_dom, start);
993 dma_ops_free_addresses(dma_dom, dma_addr, pages);
995 if (amd_iommu_unmap_flush)
996 iommu_flush_pages(iommu, dma_dom->domain.id, dma_addr, size);
1000 * The exported map_single function for dma_ops.
1002 static dma_addr_t map_single(struct device *dev, phys_addr_t paddr,
1003 size_t size, int dir)
1005 unsigned long flags;
1006 struct amd_iommu *iommu;
1007 struct protection_domain *domain;
1012 if (!check_device(dev))
1013 return bad_dma_address;
1015 dma_mask = *dev->dma_mask;
1017 get_device_resources(dev, &iommu, &domain, &devid);
1019 if (iommu == NULL || domain == NULL)
1020 /* device not handled by any AMD IOMMU */
1021 return (dma_addr_t)paddr;
1023 spin_lock_irqsave(&domain->lock, flags);
1024 addr = __map_single(dev, iommu, domain->priv, paddr, size, dir, false,
1026 if (addr == bad_dma_address)
1029 if (unlikely(iommu->need_sync))
1030 iommu_completion_wait(iommu);
1033 spin_unlock_irqrestore(&domain->lock, flags);
1039 * The exported unmap_single function for dma_ops.
1041 static void unmap_single(struct device *dev, dma_addr_t dma_addr,
1042 size_t size, int dir)
1044 unsigned long flags;
1045 struct amd_iommu *iommu;
1046 struct protection_domain *domain;
1049 if (!check_device(dev) ||
1050 !get_device_resources(dev, &iommu, &domain, &devid))
1051 /* device not handled by any AMD IOMMU */
1054 spin_lock_irqsave(&domain->lock, flags);
1056 __unmap_single(iommu, domain->priv, dma_addr, size, dir);
1058 if (unlikely(iommu->need_sync))
1059 iommu_completion_wait(iommu);
1061 spin_unlock_irqrestore(&domain->lock, flags);
1065 * This is a special map_sg function which is used if we should map a
1066 * device which is not handled by an AMD IOMMU in the system.
1068 static int map_sg_no_iommu(struct device *dev, struct scatterlist *sglist,
1069 int nelems, int dir)
1071 struct scatterlist *s;
1074 for_each_sg(sglist, s, nelems, i) {
1075 s->dma_address = (dma_addr_t)sg_phys(s);
1076 s->dma_length = s->length;
1083 * The exported map_sg function for dma_ops (handles scatter-gather
1086 static int map_sg(struct device *dev, struct scatterlist *sglist,
1087 int nelems, int dir)
1089 unsigned long flags;
1090 struct amd_iommu *iommu;
1091 struct protection_domain *domain;
1094 struct scatterlist *s;
1096 int mapped_elems = 0;
1099 if (!check_device(dev))
1102 dma_mask = *dev->dma_mask;
1104 get_device_resources(dev, &iommu, &domain, &devid);
1106 if (!iommu || !domain)
1107 return map_sg_no_iommu(dev, sglist, nelems, dir);
1109 spin_lock_irqsave(&domain->lock, flags);
1111 for_each_sg(sglist, s, nelems, i) {
1114 s->dma_address = __map_single(dev, iommu, domain->priv,
1115 paddr, s->length, dir, false,
1118 if (s->dma_address) {
1119 s->dma_length = s->length;
1125 if (unlikely(iommu->need_sync))
1126 iommu_completion_wait(iommu);
1129 spin_unlock_irqrestore(&domain->lock, flags);
1131 return mapped_elems;
1133 for_each_sg(sglist, s, mapped_elems, i) {
1135 __unmap_single(iommu, domain->priv, s->dma_address,
1136 s->dma_length, dir);
1137 s->dma_address = s->dma_length = 0;
1146 * The exported map_sg function for dma_ops (handles scatter-gather
1149 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
1150 int nelems, int dir)
1152 unsigned long flags;
1153 struct amd_iommu *iommu;
1154 struct protection_domain *domain;
1155 struct scatterlist *s;
1159 if (!check_device(dev) ||
1160 !get_device_resources(dev, &iommu, &domain, &devid))
1163 spin_lock_irqsave(&domain->lock, flags);
1165 for_each_sg(sglist, s, nelems, i) {
1166 __unmap_single(iommu, domain->priv, s->dma_address,
1167 s->dma_length, dir);
1168 s->dma_address = s->dma_length = 0;
1171 if (unlikely(iommu->need_sync))
1172 iommu_completion_wait(iommu);
1174 spin_unlock_irqrestore(&domain->lock, flags);
1178 * The exported alloc_coherent function for dma_ops.
1180 static void *alloc_coherent(struct device *dev, size_t size,
1181 dma_addr_t *dma_addr, gfp_t flag)
1183 unsigned long flags;
1185 struct amd_iommu *iommu;
1186 struct protection_domain *domain;
1189 u64 dma_mask = dev->coherent_dma_mask;
1191 if (!check_device(dev))
1194 if (!get_device_resources(dev, &iommu, &domain, &devid))
1195 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
1198 virt_addr = (void *)__get_free_pages(flag, get_order(size));
1202 paddr = virt_to_phys(virt_addr);
1204 if (!iommu || !domain) {
1205 *dma_addr = (dma_addr_t)paddr;
1210 dma_mask = *dev->dma_mask;
1212 spin_lock_irqsave(&domain->lock, flags);
1214 *dma_addr = __map_single(dev, iommu, domain->priv, paddr,
1215 size, DMA_BIDIRECTIONAL, true, dma_mask);
1217 if (*dma_addr == bad_dma_address) {
1218 free_pages((unsigned long)virt_addr, get_order(size));
1223 if (unlikely(iommu->need_sync))
1224 iommu_completion_wait(iommu);
1227 spin_unlock_irqrestore(&domain->lock, flags);
1233 * The exported free_coherent function for dma_ops.
1235 static void free_coherent(struct device *dev, size_t size,
1236 void *virt_addr, dma_addr_t dma_addr)
1238 unsigned long flags;
1239 struct amd_iommu *iommu;
1240 struct protection_domain *domain;
1243 if (!check_device(dev))
1246 get_device_resources(dev, &iommu, &domain, &devid);
1248 if (!iommu || !domain)
1251 spin_lock_irqsave(&domain->lock, flags);
1253 __unmap_single(iommu, domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
1255 if (unlikely(iommu->need_sync))
1256 iommu_completion_wait(iommu);
1258 spin_unlock_irqrestore(&domain->lock, flags);
1261 free_pages((unsigned long)virt_addr, get_order(size));
1265 * This function is called by the DMA layer to find out if we can handle a
1266 * particular device. It is part of the dma_ops.
1268 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
1271 struct pci_dev *pcidev;
1273 /* No device or no PCI device */
1274 if (!dev || dev->bus != &pci_bus_type)
1277 pcidev = to_pci_dev(dev);
1279 bdf = calc_devid(pcidev->bus->number, pcidev->devfn);
1281 /* Out of our scope? */
1282 if (bdf > amd_iommu_last_bdf)
1289 * The function for pre-allocating protection domains.
1291 * If the driver core informs the DMA layer if a driver grabs a device
1292 * we don't need to preallocate the protection domains anymore.
1293 * For now we have to.
1295 void prealloc_protection_domains(void)
1297 struct pci_dev *dev = NULL;
1298 struct dma_ops_domain *dma_dom;
1299 struct amd_iommu *iommu;
1300 int order = amd_iommu_aperture_order;
1303 while ((dev = pci_get_device(PCI_ANY_ID, PCI_ANY_ID, dev)) != NULL) {
1304 devid = (dev->bus->number << 8) | dev->devfn;
1305 if (devid > amd_iommu_last_bdf)
1307 devid = amd_iommu_alias_table[devid];
1308 if (domain_for_device(devid))
1310 iommu = amd_iommu_rlookup_table[devid];
1313 dma_dom = dma_ops_domain_alloc(iommu, order);
1316 init_unity_mappings_for_device(dma_dom, devid);
1317 dma_dom->target_dev = devid;
1319 list_add_tail(&dma_dom->list, &iommu_pd_list);
1323 static struct dma_mapping_ops amd_iommu_dma_ops = {
1324 .alloc_coherent = alloc_coherent,
1325 .free_coherent = free_coherent,
1326 .map_single = map_single,
1327 .unmap_single = unmap_single,
1329 .unmap_sg = unmap_sg,
1330 .dma_supported = amd_iommu_dma_supported,
1334 * The function which clues the AMD IOMMU driver into dma_ops.
1336 int __init amd_iommu_init_dma_ops(void)
1338 struct amd_iommu *iommu;
1339 int order = amd_iommu_aperture_order;
1343 * first allocate a default protection domain for every IOMMU we
1344 * found in the system. Devices not assigned to any other
1345 * protection domain will be assigned to the default one.
1347 list_for_each_entry(iommu, &amd_iommu_list, list) {
1348 iommu->default_dom = dma_ops_domain_alloc(iommu, order);
1349 if (iommu->default_dom == NULL)
1351 ret = iommu_init_unity_mappings(iommu);
1357 * If device isolation is enabled, pre-allocate the protection
1358 * domains for each device.
1360 if (amd_iommu_isolate)
1361 prealloc_protection_domains();
1365 bad_dma_address = 0;
1366 #ifdef CONFIG_GART_IOMMU
1367 gart_iommu_aperture_disabled = 1;
1368 gart_iommu_aperture = 0;
1371 /* Make the driver finally visible to the drivers */
1372 dma_ops = &amd_iommu_dma_ops;
1378 list_for_each_entry(iommu, &amd_iommu_list, list) {
1379 if (iommu->default_dom)
1380 dma_ops_domain_free(iommu->default_dom);