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/debugfs.h>
24 #include <linux/scatterlist.h>
25 #include <linux/dma-mapping.h>
26 #include <linux/iommu-helper.h>
27 #include <linux/iommu.h>
28 #include <asm/proto.h>
29 #include <asm/iommu.h>
31 #include <asm/amd_iommu_types.h>
32 #include <asm/amd_iommu.h>
34 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
36 #define EXIT_LOOP_COUNT 10000000
38 static DEFINE_RWLOCK(amd_iommu_devtable_lock);
40 /* A list of preallocated protection domains */
41 static LIST_HEAD(iommu_pd_list);
42 static DEFINE_SPINLOCK(iommu_pd_list_lock);
44 #ifdef CONFIG_IOMMU_API
45 static struct iommu_ops amd_iommu_ops;
49 * general struct to manage commands send to an IOMMU
55 static int dma_ops_unity_map(struct dma_ops_domain *dma_dom,
56 struct unity_map_entry *e);
57 static struct dma_ops_domain *find_protection_domain(u16 devid);
58 static u64* alloc_pte(struct protection_domain *dom,
59 unsigned long address, u64
60 **pte_page, gfp_t gfp);
61 static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
62 unsigned long start_page,
65 #ifndef BUS_NOTIFY_UNBOUND_DRIVER
66 #define BUS_NOTIFY_UNBOUND_DRIVER 0x0005
69 #ifdef CONFIG_AMD_IOMMU_STATS
72 * Initialization code for statistics collection
75 DECLARE_STATS_COUNTER(compl_wait);
76 DECLARE_STATS_COUNTER(cnt_map_single);
77 DECLARE_STATS_COUNTER(cnt_unmap_single);
78 DECLARE_STATS_COUNTER(cnt_map_sg);
79 DECLARE_STATS_COUNTER(cnt_unmap_sg);
80 DECLARE_STATS_COUNTER(cnt_alloc_coherent);
81 DECLARE_STATS_COUNTER(cnt_free_coherent);
82 DECLARE_STATS_COUNTER(cross_page);
83 DECLARE_STATS_COUNTER(domain_flush_single);
84 DECLARE_STATS_COUNTER(domain_flush_all);
85 DECLARE_STATS_COUNTER(alloced_io_mem);
86 DECLARE_STATS_COUNTER(total_map_requests);
88 static struct dentry *stats_dir;
89 static struct dentry *de_isolate;
90 static struct dentry *de_fflush;
92 static void amd_iommu_stats_add(struct __iommu_counter *cnt)
94 if (stats_dir == NULL)
97 cnt->dent = debugfs_create_u64(cnt->name, 0444, stats_dir,
101 static void amd_iommu_stats_init(void)
103 stats_dir = debugfs_create_dir("amd-iommu", NULL);
104 if (stats_dir == NULL)
107 de_isolate = debugfs_create_bool("isolation", 0444, stats_dir,
108 (u32 *)&amd_iommu_isolate);
110 de_fflush = debugfs_create_bool("fullflush", 0444, stats_dir,
111 (u32 *)&amd_iommu_unmap_flush);
113 amd_iommu_stats_add(&compl_wait);
114 amd_iommu_stats_add(&cnt_map_single);
115 amd_iommu_stats_add(&cnt_unmap_single);
116 amd_iommu_stats_add(&cnt_map_sg);
117 amd_iommu_stats_add(&cnt_unmap_sg);
118 amd_iommu_stats_add(&cnt_alloc_coherent);
119 amd_iommu_stats_add(&cnt_free_coherent);
120 amd_iommu_stats_add(&cross_page);
121 amd_iommu_stats_add(&domain_flush_single);
122 amd_iommu_stats_add(&domain_flush_all);
123 amd_iommu_stats_add(&alloced_io_mem);
124 amd_iommu_stats_add(&total_map_requests);
129 /* returns !0 if the IOMMU is caching non-present entries in its TLB */
130 static int iommu_has_npcache(struct amd_iommu *iommu)
132 return iommu->cap & (1UL << IOMMU_CAP_NPCACHE);
135 /****************************************************************************
137 * Interrupt handling functions
139 ****************************************************************************/
141 static void iommu_print_event(void *__evt)
144 int type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
145 int devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
146 int domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
147 int flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
148 u64 address = (u64)(((u64)event[3]) << 32) | event[2];
150 printk(KERN_ERR "AMD IOMMU: Event logged [");
153 case EVENT_TYPE_ILL_DEV:
154 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
155 "address=0x%016llx flags=0x%04x]\n",
156 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
159 case EVENT_TYPE_IO_FAULT:
160 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
161 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
162 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
163 domid, address, flags);
165 case EVENT_TYPE_DEV_TAB_ERR:
166 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
167 "address=0x%016llx flags=0x%04x]\n",
168 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
171 case EVENT_TYPE_PAGE_TAB_ERR:
172 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
173 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
174 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
175 domid, address, flags);
177 case EVENT_TYPE_ILL_CMD:
178 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
180 case EVENT_TYPE_CMD_HARD_ERR:
181 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
182 "flags=0x%04x]\n", address, flags);
184 case EVENT_TYPE_IOTLB_INV_TO:
185 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
186 "address=0x%016llx]\n",
187 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
190 case EVENT_TYPE_INV_DEV_REQ:
191 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
192 "address=0x%016llx flags=0x%04x]\n",
193 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
197 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
201 static void iommu_poll_events(struct amd_iommu *iommu)
206 spin_lock_irqsave(&iommu->lock, flags);
208 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
209 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
211 while (head != tail) {
212 iommu_print_event(iommu->evt_buf + head);
213 head = (head + EVENT_ENTRY_SIZE) % iommu->evt_buf_size;
216 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
218 spin_unlock_irqrestore(&iommu->lock, flags);
221 irqreturn_t amd_iommu_int_handler(int irq, void *data)
223 struct amd_iommu *iommu;
225 for_each_iommu(iommu)
226 iommu_poll_events(iommu);
231 /****************************************************************************
233 * IOMMU command queuing functions
235 ****************************************************************************/
238 * Writes the command to the IOMMUs command buffer and informs the
239 * hardware about the new command. Must be called with iommu->lock held.
241 static int __iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
246 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
247 target = iommu->cmd_buf + tail;
248 memcpy_toio(target, cmd, sizeof(*cmd));
249 tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
250 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
253 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
259 * General queuing function for commands. Takes iommu->lock and calls
260 * __iommu_queue_command().
262 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
267 spin_lock_irqsave(&iommu->lock, flags);
268 ret = __iommu_queue_command(iommu, cmd);
270 iommu->need_sync = true;
271 spin_unlock_irqrestore(&iommu->lock, flags);
277 * This function waits until an IOMMU has completed a completion
280 static void __iommu_wait_for_completion(struct amd_iommu *iommu)
286 INC_STATS_COUNTER(compl_wait);
288 while (!ready && (i < EXIT_LOOP_COUNT)) {
290 /* wait for the bit to become one */
291 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
292 ready = status & MMIO_STATUS_COM_WAIT_INT_MASK;
295 /* set bit back to zero */
296 status &= ~MMIO_STATUS_COM_WAIT_INT_MASK;
297 writel(status, iommu->mmio_base + MMIO_STATUS_OFFSET);
299 if (unlikely(i == EXIT_LOOP_COUNT))
300 panic("AMD IOMMU: Completion wait loop failed\n");
304 * This function queues a completion wait command into the command
307 static int __iommu_completion_wait(struct amd_iommu *iommu)
309 struct iommu_cmd cmd;
311 memset(&cmd, 0, sizeof(cmd));
312 cmd.data[0] = CMD_COMPL_WAIT_INT_MASK;
313 CMD_SET_TYPE(&cmd, CMD_COMPL_WAIT);
315 return __iommu_queue_command(iommu, &cmd);
319 * This function is called whenever we need to ensure that the IOMMU has
320 * completed execution of all commands we sent. It sends a
321 * COMPLETION_WAIT command and waits for it to finish. The IOMMU informs
322 * us about that by writing a value to a physical address we pass with
325 static int iommu_completion_wait(struct amd_iommu *iommu)
330 spin_lock_irqsave(&iommu->lock, flags);
332 if (!iommu->need_sync)
335 ret = __iommu_completion_wait(iommu);
337 iommu->need_sync = false;
342 __iommu_wait_for_completion(iommu);
345 spin_unlock_irqrestore(&iommu->lock, flags);
351 * Command send function for invalidating a device table entry
353 static int iommu_queue_inv_dev_entry(struct amd_iommu *iommu, u16 devid)
355 struct iommu_cmd cmd;
358 BUG_ON(iommu == NULL);
360 memset(&cmd, 0, sizeof(cmd));
361 CMD_SET_TYPE(&cmd, CMD_INV_DEV_ENTRY);
364 ret = iommu_queue_command(iommu, &cmd);
369 static void __iommu_build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
370 u16 domid, int pde, int s)
372 memset(cmd, 0, sizeof(*cmd));
373 address &= PAGE_MASK;
374 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
375 cmd->data[1] |= domid;
376 cmd->data[2] = lower_32_bits(address);
377 cmd->data[3] = upper_32_bits(address);
378 if (s) /* size bit - we flush more than one 4kb page */
379 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
380 if (pde) /* PDE bit - we wan't flush everything not only the PTEs */
381 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
385 * Generic command send function for invalidaing TLB entries
387 static int iommu_queue_inv_iommu_pages(struct amd_iommu *iommu,
388 u64 address, u16 domid, int pde, int s)
390 struct iommu_cmd cmd;
393 __iommu_build_inv_iommu_pages(&cmd, address, domid, pde, s);
395 ret = iommu_queue_command(iommu, &cmd);
401 * TLB invalidation function which is called from the mapping functions.
402 * It invalidates a single PTE if the range to flush is within a single
403 * page. Otherwise it flushes the whole TLB of the IOMMU.
405 static int iommu_flush_pages(struct amd_iommu *iommu, u16 domid,
406 u64 address, size_t size)
409 unsigned pages = iommu_num_pages(address, size, PAGE_SIZE);
411 address &= PAGE_MASK;
415 * If we have to flush more than one page, flush all
416 * TLB entries for this domain
418 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
422 iommu_queue_inv_iommu_pages(iommu, address, domid, 0, s);
427 /* Flush the whole IO/TLB for a given protection domain */
428 static void iommu_flush_tlb(struct amd_iommu *iommu, u16 domid)
430 u64 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
432 INC_STATS_COUNTER(domain_flush_single);
434 iommu_queue_inv_iommu_pages(iommu, address, domid, 0, 1);
438 * This function is used to flush the IO/TLB for a given protection domain
439 * on every IOMMU in the system
441 static void iommu_flush_domain(u16 domid)
444 struct amd_iommu *iommu;
445 struct iommu_cmd cmd;
447 INC_STATS_COUNTER(domain_flush_all);
449 __iommu_build_inv_iommu_pages(&cmd, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
452 for_each_iommu(iommu) {
453 spin_lock_irqsave(&iommu->lock, flags);
454 __iommu_queue_command(iommu, &cmd);
455 __iommu_completion_wait(iommu);
456 __iommu_wait_for_completion(iommu);
457 spin_unlock_irqrestore(&iommu->lock, flags);
461 void amd_iommu_flush_all_domains(void)
465 for (i = 1; i < MAX_DOMAIN_ID; ++i) {
466 if (!test_bit(i, amd_iommu_pd_alloc_bitmap))
468 iommu_flush_domain(i);
472 void amd_iommu_flush_all_devices(void)
474 struct amd_iommu *iommu;
477 for (i = 0; i <= amd_iommu_last_bdf; ++i) {
478 if (amd_iommu_pd_table[i] == NULL)
481 iommu = amd_iommu_rlookup_table[i];
485 iommu_queue_inv_dev_entry(iommu, i);
486 iommu_completion_wait(iommu);
490 /****************************************************************************
492 * The functions below are used the create the page table mappings for
493 * unity mapped regions.
495 ****************************************************************************/
498 * Generic mapping functions. It maps a physical address into a DMA
499 * address space. It allocates the page table pages if necessary.
500 * In the future it can be extended to a generic mapping function
501 * supporting all features of AMD IOMMU page tables like level skipping
502 * and full 64 bit address spaces.
504 static int iommu_map_page(struct protection_domain *dom,
505 unsigned long bus_addr,
506 unsigned long phys_addr,
511 bus_addr = PAGE_ALIGN(bus_addr);
512 phys_addr = PAGE_ALIGN(phys_addr);
514 /* only support 512GB address spaces for now */
515 if (bus_addr > IOMMU_MAP_SIZE_L3 || !(prot & IOMMU_PROT_MASK))
518 pte = alloc_pte(dom, bus_addr, NULL, GFP_KERNEL);
520 if (IOMMU_PTE_PRESENT(*pte))
523 __pte = phys_addr | IOMMU_PTE_P;
524 if (prot & IOMMU_PROT_IR)
525 __pte |= IOMMU_PTE_IR;
526 if (prot & IOMMU_PROT_IW)
527 __pte |= IOMMU_PTE_IW;
534 static void iommu_unmap_page(struct protection_domain *dom,
535 unsigned long bus_addr)
539 pte = &dom->pt_root[IOMMU_PTE_L2_INDEX(bus_addr)];
541 if (!IOMMU_PTE_PRESENT(*pte))
544 pte = IOMMU_PTE_PAGE(*pte);
545 pte = &pte[IOMMU_PTE_L1_INDEX(bus_addr)];
547 if (!IOMMU_PTE_PRESENT(*pte))
550 pte = IOMMU_PTE_PAGE(*pte);
551 pte = &pte[IOMMU_PTE_L1_INDEX(bus_addr)];
557 * This function checks if a specific unity mapping entry is needed for
558 * this specific IOMMU.
560 static int iommu_for_unity_map(struct amd_iommu *iommu,
561 struct unity_map_entry *entry)
565 for (i = entry->devid_start; i <= entry->devid_end; ++i) {
566 bdf = amd_iommu_alias_table[i];
567 if (amd_iommu_rlookup_table[bdf] == iommu)
575 * Init the unity mappings for a specific IOMMU in the system
577 * Basically iterates over all unity mapping entries and applies them to
578 * the default domain DMA of that IOMMU if necessary.
580 static int iommu_init_unity_mappings(struct amd_iommu *iommu)
582 struct unity_map_entry *entry;
585 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
586 if (!iommu_for_unity_map(iommu, entry))
588 ret = dma_ops_unity_map(iommu->default_dom, entry);
597 * This function actually applies the mapping to the page table of the
600 static int dma_ops_unity_map(struct dma_ops_domain *dma_dom,
601 struct unity_map_entry *e)
606 for (addr = e->address_start; addr < e->address_end;
608 ret = iommu_map_page(&dma_dom->domain, addr, addr, e->prot);
612 * if unity mapping is in aperture range mark the page
613 * as allocated in the aperture
615 if (addr < dma_dom->aperture_size)
616 __set_bit(addr >> PAGE_SHIFT,
617 dma_dom->aperture[0]->bitmap);
624 * Inits the unity mappings required for a specific device
626 static int init_unity_mappings_for_device(struct dma_ops_domain *dma_dom,
629 struct unity_map_entry *e;
632 list_for_each_entry(e, &amd_iommu_unity_map, list) {
633 if (!(devid >= e->devid_start && devid <= e->devid_end))
635 ret = dma_ops_unity_map(dma_dom, e);
643 /****************************************************************************
645 * The next functions belong to the address allocator for the dma_ops
646 * interface functions. They work like the allocators in the other IOMMU
647 * drivers. Its basically a bitmap which marks the allocated pages in
648 * the aperture. Maybe it could be enhanced in the future to a more
649 * efficient allocator.
651 ****************************************************************************/
654 * The address allocator core functions.
656 * called with domain->lock held
660 * This function checks if there is a PTE for a given dma address. If
661 * there is one, it returns the pointer to it.
663 static u64* fetch_pte(struct protection_domain *domain,
664 unsigned long address)
668 pte = &domain->pt_root[IOMMU_PTE_L2_INDEX(address)];
670 if (!IOMMU_PTE_PRESENT(*pte))
673 pte = IOMMU_PTE_PAGE(*pte);
674 pte = &pte[IOMMU_PTE_L1_INDEX(address)];
676 if (!IOMMU_PTE_PRESENT(*pte))
679 pte = IOMMU_PTE_PAGE(*pte);
680 pte = &pte[IOMMU_PTE_L0_INDEX(address)];
686 * This function is used to add a new aperture range to an existing
687 * aperture in case of dma_ops domain allocation or address allocation
690 static int alloc_new_range(struct amd_iommu *iommu,
691 struct dma_ops_domain *dma_dom,
692 bool populate, gfp_t gfp)
694 int index = dma_dom->aperture_size >> APERTURE_RANGE_SHIFT;
697 #ifdef CONFIG_IOMMU_STRESS
701 if (index >= APERTURE_MAX_RANGES)
704 dma_dom->aperture[index] = kzalloc(sizeof(struct aperture_range), gfp);
705 if (!dma_dom->aperture[index])
708 dma_dom->aperture[index]->bitmap = (void *)get_zeroed_page(gfp);
709 if (!dma_dom->aperture[index]->bitmap)
712 dma_dom->aperture[index]->offset = dma_dom->aperture_size;
715 unsigned long address = dma_dom->aperture_size;
716 int i, num_ptes = APERTURE_RANGE_PAGES / 512;
719 for (i = 0; i < num_ptes; ++i) {
720 pte = alloc_pte(&dma_dom->domain, address,
725 dma_dom->aperture[index]->pte_pages[i] = pte_page;
727 address += APERTURE_RANGE_SIZE / 64;
731 dma_dom->aperture_size += APERTURE_RANGE_SIZE;
733 /* Intialize the exclusion range if necessary */
734 if (iommu->exclusion_start &&
735 iommu->exclusion_start >= dma_dom->aperture[index]->offset &&
736 iommu->exclusion_start < dma_dom->aperture_size) {
737 unsigned long startpage = iommu->exclusion_start >> PAGE_SHIFT;
738 int pages = iommu_num_pages(iommu->exclusion_start,
739 iommu->exclusion_length,
741 dma_ops_reserve_addresses(dma_dom, startpage, pages);
745 * Check for areas already mapped as present in the new aperture
746 * range and mark those pages as reserved in the allocator. Such
747 * mappings may already exist as a result of requested unity
748 * mappings for devices.
750 for (i = dma_dom->aperture[index]->offset;
751 i < dma_dom->aperture_size;
753 u64 *pte = fetch_pte(&dma_dom->domain, i);
754 if (!pte || !IOMMU_PTE_PRESENT(*pte))
757 dma_ops_reserve_addresses(dma_dom, i << PAGE_SHIFT, 1);
763 free_page((unsigned long)dma_dom->aperture[index]->bitmap);
765 kfree(dma_dom->aperture[index]);
766 dma_dom->aperture[index] = NULL;
771 static unsigned long dma_ops_area_alloc(struct device *dev,
772 struct dma_ops_domain *dom,
774 unsigned long align_mask,
778 unsigned long next_bit = dom->next_address % APERTURE_RANGE_SIZE;
779 int max_index = dom->aperture_size >> APERTURE_RANGE_SHIFT;
780 int i = start >> APERTURE_RANGE_SHIFT;
781 unsigned long boundary_size;
782 unsigned long address = -1;
785 next_bit >>= PAGE_SHIFT;
787 boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
788 PAGE_SIZE) >> PAGE_SHIFT;
790 for (;i < max_index; ++i) {
791 unsigned long offset = dom->aperture[i]->offset >> PAGE_SHIFT;
793 if (dom->aperture[i]->offset >= dma_mask)
796 limit = iommu_device_max_index(APERTURE_RANGE_PAGES, offset,
797 dma_mask >> PAGE_SHIFT);
799 address = iommu_area_alloc(dom->aperture[i]->bitmap,
800 limit, next_bit, pages, 0,
801 boundary_size, align_mask);
803 address = dom->aperture[i]->offset +
804 (address << PAGE_SHIFT);
805 dom->next_address = address + (pages << PAGE_SHIFT);
815 static unsigned long dma_ops_alloc_addresses(struct device *dev,
816 struct dma_ops_domain *dom,
818 unsigned long align_mask,
821 unsigned long address;
823 #ifdef CONFIG_IOMMU_STRESS
824 dom->next_address = 0;
825 dom->need_flush = true;
828 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
829 dma_mask, dom->next_address);
832 dom->next_address = 0;
833 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
835 dom->need_flush = true;
838 if (unlikely(address == -1))
839 address = bad_dma_address;
841 WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
847 * The address free function.
849 * called with domain->lock held
851 static void dma_ops_free_addresses(struct dma_ops_domain *dom,
852 unsigned long address,
855 unsigned i = address >> APERTURE_RANGE_SHIFT;
856 struct aperture_range *range = dom->aperture[i];
858 BUG_ON(i >= APERTURE_MAX_RANGES || range == NULL);
860 #ifdef CONFIG_IOMMU_STRESS
865 if (address >= dom->next_address)
866 dom->need_flush = true;
868 address = (address % APERTURE_RANGE_SIZE) >> PAGE_SHIFT;
870 iommu_area_free(range->bitmap, address, pages);
874 /****************************************************************************
876 * The next functions belong to the domain allocation. A domain is
877 * allocated for every IOMMU as the default domain. If device isolation
878 * is enabled, every device get its own domain. The most important thing
879 * about domains is the page table mapping the DMA address space they
882 ****************************************************************************/
884 static u16 domain_id_alloc(void)
889 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
890 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
892 if (id > 0 && id < MAX_DOMAIN_ID)
893 __set_bit(id, amd_iommu_pd_alloc_bitmap);
896 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
901 static void domain_id_free(int id)
905 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
906 if (id > 0 && id < MAX_DOMAIN_ID)
907 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
908 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
912 * Used to reserve address ranges in the aperture (e.g. for exclusion
915 static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
916 unsigned long start_page,
919 unsigned int i, last_page = dom->aperture_size >> PAGE_SHIFT;
921 if (start_page + pages > last_page)
922 pages = last_page - start_page;
924 for (i = start_page; i < start_page + pages; ++i) {
925 int index = i / APERTURE_RANGE_PAGES;
926 int page = i % APERTURE_RANGE_PAGES;
927 __set_bit(page, dom->aperture[index]->bitmap);
931 static void free_pagetable(struct protection_domain *domain)
936 p1 = domain->pt_root;
941 for (i = 0; i < 512; ++i) {
942 if (!IOMMU_PTE_PRESENT(p1[i]))
945 p2 = IOMMU_PTE_PAGE(p1[i]);
946 for (j = 0; j < 512; ++j) {
947 if (!IOMMU_PTE_PRESENT(p2[j]))
949 p3 = IOMMU_PTE_PAGE(p2[j]);
950 free_page((unsigned long)p3);
953 free_page((unsigned long)p2);
956 free_page((unsigned long)p1);
958 domain->pt_root = NULL;
962 * Free a domain, only used if something went wrong in the
963 * allocation path and we need to free an already allocated page table
965 static void dma_ops_domain_free(struct dma_ops_domain *dom)
972 free_pagetable(&dom->domain);
974 for (i = 0; i < APERTURE_MAX_RANGES; ++i) {
975 if (!dom->aperture[i])
977 free_page((unsigned long)dom->aperture[i]->bitmap);
978 kfree(dom->aperture[i]);
985 * Allocates a new protection domain usable for the dma_ops functions.
986 * It also intializes the page table and the address allocator data
987 * structures required for the dma_ops interface
989 static struct dma_ops_domain *dma_ops_domain_alloc(struct amd_iommu *iommu)
991 struct dma_ops_domain *dma_dom;
993 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
997 spin_lock_init(&dma_dom->domain.lock);
999 dma_dom->domain.id = domain_id_alloc();
1000 if (dma_dom->domain.id == 0)
1002 dma_dom->domain.mode = PAGE_MODE_3_LEVEL;
1003 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1004 dma_dom->domain.flags = PD_DMA_OPS_MASK;
1005 dma_dom->domain.priv = dma_dom;
1006 if (!dma_dom->domain.pt_root)
1009 dma_dom->need_flush = false;
1010 dma_dom->target_dev = 0xffff;
1012 if (alloc_new_range(iommu, dma_dom, true, GFP_KERNEL))
1016 * mark the first page as allocated so we never return 0 as
1017 * a valid dma-address. So we can use 0 as error value
1019 dma_dom->aperture[0]->bitmap[0] = 1;
1020 dma_dom->next_address = 0;
1026 dma_ops_domain_free(dma_dom);
1032 * little helper function to check whether a given protection domain is a
1035 static bool dma_ops_domain(struct protection_domain *domain)
1037 return domain->flags & PD_DMA_OPS_MASK;
1041 * Find out the protection domain structure for a given PCI device. This
1042 * will give us the pointer to the page table root for example.
1044 static struct protection_domain *domain_for_device(u16 devid)
1046 struct protection_domain *dom;
1047 unsigned long flags;
1049 read_lock_irqsave(&amd_iommu_devtable_lock, flags);
1050 dom = amd_iommu_pd_table[devid];
1051 read_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1057 * If a device is not yet associated with a domain, this function does
1058 * assigns it visible for the hardware
1060 static void attach_device(struct amd_iommu *iommu,
1061 struct protection_domain *domain,
1064 unsigned long flags;
1065 u64 pte_root = virt_to_phys(domain->pt_root);
1067 domain->dev_cnt += 1;
1069 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
1070 << DEV_ENTRY_MODE_SHIFT;
1071 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
1073 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1074 amd_iommu_dev_table[devid].data[0] = lower_32_bits(pte_root);
1075 amd_iommu_dev_table[devid].data[1] = upper_32_bits(pte_root);
1076 amd_iommu_dev_table[devid].data[2] = domain->id;
1078 amd_iommu_pd_table[devid] = domain;
1079 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1081 iommu_queue_inv_dev_entry(iommu, devid);
1085 * Removes a device from a protection domain (unlocked)
1087 static void __detach_device(struct protection_domain *domain, u16 devid)
1091 spin_lock(&domain->lock);
1093 /* remove domain from the lookup table */
1094 amd_iommu_pd_table[devid] = NULL;
1096 /* remove entry from the device table seen by the hardware */
1097 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
1098 amd_iommu_dev_table[devid].data[1] = 0;
1099 amd_iommu_dev_table[devid].data[2] = 0;
1101 /* decrease reference counter */
1102 domain->dev_cnt -= 1;
1105 spin_unlock(&domain->lock);
1109 * Removes a device from a protection domain (with devtable_lock held)
1111 static void detach_device(struct protection_domain *domain, u16 devid)
1113 unsigned long flags;
1115 /* lock device table */
1116 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1117 __detach_device(domain, devid);
1118 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1121 static int device_change_notifier(struct notifier_block *nb,
1122 unsigned long action, void *data)
1124 struct device *dev = data;
1125 struct pci_dev *pdev = to_pci_dev(dev);
1126 u16 devid = calc_devid(pdev->bus->number, pdev->devfn);
1127 struct protection_domain *domain;
1128 struct dma_ops_domain *dma_domain;
1129 struct amd_iommu *iommu;
1130 unsigned long flags;
1132 if (devid > amd_iommu_last_bdf)
1135 devid = amd_iommu_alias_table[devid];
1137 iommu = amd_iommu_rlookup_table[devid];
1141 domain = domain_for_device(devid);
1143 if (domain && !dma_ops_domain(domain))
1144 WARN_ONCE(1, "AMD IOMMU WARNING: device %s already bound "
1145 "to a non-dma-ops domain\n", dev_name(dev));
1148 case BUS_NOTIFY_BOUND_DRIVER:
1151 dma_domain = find_protection_domain(devid);
1153 dma_domain = iommu->default_dom;
1154 attach_device(iommu, &dma_domain->domain, devid);
1155 DUMP_printk(KERN_INFO "AMD IOMMU: Using protection domain "
1156 "%d for device %s\n",
1157 dma_domain->domain.id, dev_name(dev));
1159 case BUS_NOTIFY_UNBOUND_DRIVER:
1162 detach_device(domain, devid);
1164 case BUS_NOTIFY_ADD_DEVICE:
1165 /* allocate a protection domain if a device is added */
1166 dma_domain = find_protection_domain(devid);
1169 dma_domain = dma_ops_domain_alloc(iommu);
1172 dma_domain->target_dev = devid;
1174 spin_lock_irqsave(&iommu_pd_list_lock, flags);
1175 list_add_tail(&dma_domain->list, &iommu_pd_list);
1176 spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
1183 iommu_queue_inv_dev_entry(iommu, devid);
1184 iommu_completion_wait(iommu);
1190 struct notifier_block device_nb = {
1191 .notifier_call = device_change_notifier,
1194 /*****************************************************************************
1196 * The next functions belong to the dma_ops mapping/unmapping code.
1198 *****************************************************************************/
1201 * This function checks if the driver got a valid device from the caller to
1202 * avoid dereferencing invalid pointers.
1204 static bool check_device(struct device *dev)
1206 if (!dev || !dev->dma_mask)
1213 * In this function the list of preallocated protection domains is traversed to
1214 * find the domain for a specific device
1216 static struct dma_ops_domain *find_protection_domain(u16 devid)
1218 struct dma_ops_domain *entry, *ret = NULL;
1219 unsigned long flags;
1221 if (list_empty(&iommu_pd_list))
1224 spin_lock_irqsave(&iommu_pd_list_lock, flags);
1226 list_for_each_entry(entry, &iommu_pd_list, list) {
1227 if (entry->target_dev == devid) {
1233 spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
1239 * In the dma_ops path we only have the struct device. This function
1240 * finds the corresponding IOMMU, the protection domain and the
1241 * requestor id for a given device.
1242 * If the device is not yet associated with a domain this is also done
1245 static int get_device_resources(struct device *dev,
1246 struct amd_iommu **iommu,
1247 struct protection_domain **domain,
1250 struct dma_ops_domain *dma_dom;
1251 struct pci_dev *pcidev;
1258 if (dev->bus != &pci_bus_type)
1261 pcidev = to_pci_dev(dev);
1262 _bdf = calc_devid(pcidev->bus->number, pcidev->devfn);
1264 /* device not translated by any IOMMU in the system? */
1265 if (_bdf > amd_iommu_last_bdf)
1268 *bdf = amd_iommu_alias_table[_bdf];
1270 *iommu = amd_iommu_rlookup_table[*bdf];
1273 *domain = domain_for_device(*bdf);
1274 if (*domain == NULL) {
1275 dma_dom = find_protection_domain(*bdf);
1277 dma_dom = (*iommu)->default_dom;
1278 *domain = &dma_dom->domain;
1279 attach_device(*iommu, *domain, *bdf);
1280 DUMP_printk(KERN_INFO "AMD IOMMU: Using protection domain "
1281 "%d for device %s\n",
1282 (*domain)->id, dev_name(dev));
1285 if (domain_for_device(_bdf) == NULL)
1286 attach_device(*iommu, *domain, _bdf);
1292 * If the pte_page is not yet allocated this function is called
1294 static u64* alloc_pte(struct protection_domain *dom,
1295 unsigned long address, u64 **pte_page, gfp_t gfp)
1299 pte = &dom->pt_root[IOMMU_PTE_L2_INDEX(address)];
1301 if (!IOMMU_PTE_PRESENT(*pte)) {
1302 page = (u64 *)get_zeroed_page(gfp);
1305 *pte = IOMMU_L2_PDE(virt_to_phys(page));
1308 pte = IOMMU_PTE_PAGE(*pte);
1309 pte = &pte[IOMMU_PTE_L1_INDEX(address)];
1311 if (!IOMMU_PTE_PRESENT(*pte)) {
1312 page = (u64 *)get_zeroed_page(gfp);
1315 *pte = IOMMU_L1_PDE(virt_to_phys(page));
1318 pte = IOMMU_PTE_PAGE(*pte);
1323 pte = &pte[IOMMU_PTE_L0_INDEX(address)];
1329 * This function fetches the PTE for a given address in the aperture
1331 static u64* dma_ops_get_pte(struct dma_ops_domain *dom,
1332 unsigned long address)
1334 struct aperture_range *aperture;
1335 u64 *pte, *pte_page;
1337 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
1341 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
1343 pte = alloc_pte(&dom->domain, address, &pte_page, GFP_ATOMIC);
1344 aperture->pte_pages[APERTURE_PAGE_INDEX(address)] = pte_page;
1346 pte += IOMMU_PTE_L0_INDEX(address);
1352 * This is the generic map function. It maps one 4kb page at paddr to
1353 * the given address in the DMA address space for the domain.
1355 static dma_addr_t dma_ops_domain_map(struct amd_iommu *iommu,
1356 struct dma_ops_domain *dom,
1357 unsigned long address,
1363 WARN_ON(address > dom->aperture_size);
1367 pte = dma_ops_get_pte(dom, address);
1369 return bad_dma_address;
1371 __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
1373 if (direction == DMA_TO_DEVICE)
1374 __pte |= IOMMU_PTE_IR;
1375 else if (direction == DMA_FROM_DEVICE)
1376 __pte |= IOMMU_PTE_IW;
1377 else if (direction == DMA_BIDIRECTIONAL)
1378 __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
1384 return (dma_addr_t)address;
1388 * The generic unmapping function for on page in the DMA address space.
1390 static void dma_ops_domain_unmap(struct amd_iommu *iommu,
1391 struct dma_ops_domain *dom,
1392 unsigned long address)
1394 struct aperture_range *aperture;
1397 if (address >= dom->aperture_size)
1400 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
1404 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
1408 pte += IOMMU_PTE_L0_INDEX(address);
1416 * This function contains common code for mapping of a physically
1417 * contiguous memory region into DMA address space. It is used by all
1418 * mapping functions provided with this IOMMU driver.
1419 * Must be called with the domain lock held.
1421 static dma_addr_t __map_single(struct device *dev,
1422 struct amd_iommu *iommu,
1423 struct dma_ops_domain *dma_dom,
1430 dma_addr_t offset = paddr & ~PAGE_MASK;
1431 dma_addr_t address, start, ret;
1433 unsigned long align_mask = 0;
1436 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
1439 INC_STATS_COUNTER(total_map_requests);
1442 INC_STATS_COUNTER(cross_page);
1445 align_mask = (1UL << get_order(size)) - 1;
1448 address = dma_ops_alloc_addresses(dev, dma_dom, pages, align_mask,
1450 if (unlikely(address == bad_dma_address)) {
1452 * setting next_address here will let the address
1453 * allocator only scan the new allocated range in the
1454 * first run. This is a small optimization.
1456 dma_dom->next_address = dma_dom->aperture_size;
1458 if (alloc_new_range(iommu, dma_dom, false, GFP_ATOMIC))
1462 * aperture was sucessfully enlarged by 128 MB, try
1469 for (i = 0; i < pages; ++i) {
1470 ret = dma_ops_domain_map(iommu, dma_dom, start, paddr, dir);
1471 if (ret == bad_dma_address)
1479 ADD_STATS_COUNTER(alloced_io_mem, size);
1481 if (unlikely(dma_dom->need_flush && !amd_iommu_unmap_flush)) {
1482 iommu_flush_tlb(iommu, dma_dom->domain.id);
1483 dma_dom->need_flush = false;
1484 } else if (unlikely(iommu_has_npcache(iommu)))
1485 iommu_flush_pages(iommu, dma_dom->domain.id, address, size);
1492 for (--i; i >= 0; --i) {
1494 dma_ops_domain_unmap(iommu, dma_dom, start);
1497 dma_ops_free_addresses(dma_dom, address, pages);
1499 return bad_dma_address;
1503 * Does the reverse of the __map_single function. Must be called with
1504 * the domain lock held too
1506 static void __unmap_single(struct amd_iommu *iommu,
1507 struct dma_ops_domain *dma_dom,
1508 dma_addr_t dma_addr,
1512 dma_addr_t i, start;
1515 if ((dma_addr == bad_dma_address) ||
1516 (dma_addr + size > dma_dom->aperture_size))
1519 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
1520 dma_addr &= PAGE_MASK;
1523 for (i = 0; i < pages; ++i) {
1524 dma_ops_domain_unmap(iommu, dma_dom, start);
1528 SUB_STATS_COUNTER(alloced_io_mem, size);
1530 dma_ops_free_addresses(dma_dom, dma_addr, pages);
1532 if (amd_iommu_unmap_flush || dma_dom->need_flush) {
1533 iommu_flush_pages(iommu, dma_dom->domain.id, dma_addr, size);
1534 dma_dom->need_flush = false;
1539 * The exported map_single function for dma_ops.
1541 static dma_addr_t map_page(struct device *dev, struct page *page,
1542 unsigned long offset, size_t size,
1543 enum dma_data_direction dir,
1544 struct dma_attrs *attrs)
1546 unsigned long flags;
1547 struct amd_iommu *iommu;
1548 struct protection_domain *domain;
1552 phys_addr_t paddr = page_to_phys(page) + offset;
1554 INC_STATS_COUNTER(cnt_map_single);
1556 if (!check_device(dev))
1557 return bad_dma_address;
1559 dma_mask = *dev->dma_mask;
1561 get_device_resources(dev, &iommu, &domain, &devid);
1563 if (iommu == NULL || domain == NULL)
1564 /* device not handled by any AMD IOMMU */
1565 return (dma_addr_t)paddr;
1567 if (!dma_ops_domain(domain))
1568 return bad_dma_address;
1570 spin_lock_irqsave(&domain->lock, flags);
1571 addr = __map_single(dev, iommu, domain->priv, paddr, size, dir, false,
1573 if (addr == bad_dma_address)
1576 iommu_completion_wait(iommu);
1579 spin_unlock_irqrestore(&domain->lock, flags);
1585 * The exported unmap_single function for dma_ops.
1587 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
1588 enum dma_data_direction dir, struct dma_attrs *attrs)
1590 unsigned long flags;
1591 struct amd_iommu *iommu;
1592 struct protection_domain *domain;
1595 INC_STATS_COUNTER(cnt_unmap_single);
1597 if (!check_device(dev) ||
1598 !get_device_resources(dev, &iommu, &domain, &devid))
1599 /* device not handled by any AMD IOMMU */
1602 if (!dma_ops_domain(domain))
1605 spin_lock_irqsave(&domain->lock, flags);
1607 __unmap_single(iommu, domain->priv, dma_addr, size, dir);
1609 iommu_completion_wait(iommu);
1611 spin_unlock_irqrestore(&domain->lock, flags);
1615 * This is a special map_sg function which is used if we should map a
1616 * device which is not handled by an AMD IOMMU in the system.
1618 static int map_sg_no_iommu(struct device *dev, struct scatterlist *sglist,
1619 int nelems, int dir)
1621 struct scatterlist *s;
1624 for_each_sg(sglist, s, nelems, i) {
1625 s->dma_address = (dma_addr_t)sg_phys(s);
1626 s->dma_length = s->length;
1633 * The exported map_sg function for dma_ops (handles scatter-gather
1636 static int map_sg(struct device *dev, struct scatterlist *sglist,
1637 int nelems, enum dma_data_direction dir,
1638 struct dma_attrs *attrs)
1640 unsigned long flags;
1641 struct amd_iommu *iommu;
1642 struct protection_domain *domain;
1645 struct scatterlist *s;
1647 int mapped_elems = 0;
1650 INC_STATS_COUNTER(cnt_map_sg);
1652 if (!check_device(dev))
1655 dma_mask = *dev->dma_mask;
1657 get_device_resources(dev, &iommu, &domain, &devid);
1659 if (!iommu || !domain)
1660 return map_sg_no_iommu(dev, sglist, nelems, dir);
1662 if (!dma_ops_domain(domain))
1665 spin_lock_irqsave(&domain->lock, flags);
1667 for_each_sg(sglist, s, nelems, i) {
1670 s->dma_address = __map_single(dev, iommu, domain->priv,
1671 paddr, s->length, dir, false,
1674 if (s->dma_address) {
1675 s->dma_length = s->length;
1681 iommu_completion_wait(iommu);
1684 spin_unlock_irqrestore(&domain->lock, flags);
1686 return mapped_elems;
1688 for_each_sg(sglist, s, mapped_elems, i) {
1690 __unmap_single(iommu, domain->priv, s->dma_address,
1691 s->dma_length, dir);
1692 s->dma_address = s->dma_length = 0;
1701 * The exported map_sg function for dma_ops (handles scatter-gather
1704 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
1705 int nelems, enum dma_data_direction dir,
1706 struct dma_attrs *attrs)
1708 unsigned long flags;
1709 struct amd_iommu *iommu;
1710 struct protection_domain *domain;
1711 struct scatterlist *s;
1715 INC_STATS_COUNTER(cnt_unmap_sg);
1717 if (!check_device(dev) ||
1718 !get_device_resources(dev, &iommu, &domain, &devid))
1721 if (!dma_ops_domain(domain))
1724 spin_lock_irqsave(&domain->lock, flags);
1726 for_each_sg(sglist, s, nelems, i) {
1727 __unmap_single(iommu, domain->priv, s->dma_address,
1728 s->dma_length, dir);
1729 s->dma_address = s->dma_length = 0;
1732 iommu_completion_wait(iommu);
1734 spin_unlock_irqrestore(&domain->lock, flags);
1738 * The exported alloc_coherent function for dma_ops.
1740 static void *alloc_coherent(struct device *dev, size_t size,
1741 dma_addr_t *dma_addr, gfp_t flag)
1743 unsigned long flags;
1745 struct amd_iommu *iommu;
1746 struct protection_domain *domain;
1749 u64 dma_mask = dev->coherent_dma_mask;
1751 INC_STATS_COUNTER(cnt_alloc_coherent);
1753 if (!check_device(dev))
1756 if (!get_device_resources(dev, &iommu, &domain, &devid))
1757 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
1760 virt_addr = (void *)__get_free_pages(flag, get_order(size));
1764 paddr = virt_to_phys(virt_addr);
1766 if (!iommu || !domain) {
1767 *dma_addr = (dma_addr_t)paddr;
1771 if (!dma_ops_domain(domain))
1775 dma_mask = *dev->dma_mask;
1777 spin_lock_irqsave(&domain->lock, flags);
1779 *dma_addr = __map_single(dev, iommu, domain->priv, paddr,
1780 size, DMA_BIDIRECTIONAL, true, dma_mask);
1782 if (*dma_addr == bad_dma_address)
1785 iommu_completion_wait(iommu);
1787 spin_unlock_irqrestore(&domain->lock, flags);
1793 free_pages((unsigned long)virt_addr, get_order(size));
1799 * The exported free_coherent function for dma_ops.
1801 static void free_coherent(struct device *dev, size_t size,
1802 void *virt_addr, dma_addr_t dma_addr)
1804 unsigned long flags;
1805 struct amd_iommu *iommu;
1806 struct protection_domain *domain;
1809 INC_STATS_COUNTER(cnt_free_coherent);
1811 if (!check_device(dev))
1814 get_device_resources(dev, &iommu, &domain, &devid);
1816 if (!iommu || !domain)
1819 if (!dma_ops_domain(domain))
1822 spin_lock_irqsave(&domain->lock, flags);
1824 __unmap_single(iommu, domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
1826 iommu_completion_wait(iommu);
1828 spin_unlock_irqrestore(&domain->lock, flags);
1831 free_pages((unsigned long)virt_addr, get_order(size));
1835 * This function is called by the DMA layer to find out if we can handle a
1836 * particular device. It is part of the dma_ops.
1838 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
1841 struct pci_dev *pcidev;
1843 /* No device or no PCI device */
1844 if (!dev || dev->bus != &pci_bus_type)
1847 pcidev = to_pci_dev(dev);
1849 bdf = calc_devid(pcidev->bus->number, pcidev->devfn);
1851 /* Out of our scope? */
1852 if (bdf > amd_iommu_last_bdf)
1859 * The function for pre-allocating protection domains.
1861 * If the driver core informs the DMA layer if a driver grabs a device
1862 * we don't need to preallocate the protection domains anymore.
1863 * For now we have to.
1865 static void prealloc_protection_domains(void)
1867 struct pci_dev *dev = NULL;
1868 struct dma_ops_domain *dma_dom;
1869 struct amd_iommu *iommu;
1872 while ((dev = pci_get_device(PCI_ANY_ID, PCI_ANY_ID, dev)) != NULL) {
1873 devid = calc_devid(dev->bus->number, dev->devfn);
1874 if (devid > amd_iommu_last_bdf)
1876 devid = amd_iommu_alias_table[devid];
1877 if (domain_for_device(devid))
1879 iommu = amd_iommu_rlookup_table[devid];
1882 dma_dom = dma_ops_domain_alloc(iommu);
1885 init_unity_mappings_for_device(dma_dom, devid);
1886 dma_dom->target_dev = devid;
1888 list_add_tail(&dma_dom->list, &iommu_pd_list);
1892 static struct dma_map_ops amd_iommu_dma_ops = {
1893 .alloc_coherent = alloc_coherent,
1894 .free_coherent = free_coherent,
1895 .map_page = map_page,
1896 .unmap_page = unmap_page,
1898 .unmap_sg = unmap_sg,
1899 .dma_supported = amd_iommu_dma_supported,
1903 * The function which clues the AMD IOMMU driver into dma_ops.
1905 int __init amd_iommu_init_dma_ops(void)
1907 struct amd_iommu *iommu;
1911 * first allocate a default protection domain for every IOMMU we
1912 * found in the system. Devices not assigned to any other
1913 * protection domain will be assigned to the default one.
1915 for_each_iommu(iommu) {
1916 iommu->default_dom = dma_ops_domain_alloc(iommu);
1917 if (iommu->default_dom == NULL)
1919 iommu->default_dom->domain.flags |= PD_DEFAULT_MASK;
1920 ret = iommu_init_unity_mappings(iommu);
1926 * If device isolation is enabled, pre-allocate the protection
1927 * domains for each device.
1929 if (amd_iommu_isolate)
1930 prealloc_protection_domains();
1934 bad_dma_address = 0;
1935 #ifdef CONFIG_GART_IOMMU
1936 gart_iommu_aperture_disabled = 1;
1937 gart_iommu_aperture = 0;
1940 /* Make the driver finally visible to the drivers */
1941 dma_ops = &amd_iommu_dma_ops;
1943 register_iommu(&amd_iommu_ops);
1945 bus_register_notifier(&pci_bus_type, &device_nb);
1947 amd_iommu_stats_init();
1953 for_each_iommu(iommu) {
1954 if (iommu->default_dom)
1955 dma_ops_domain_free(iommu->default_dom);
1961 /*****************************************************************************
1963 * The following functions belong to the exported interface of AMD IOMMU
1965 * This interface allows access to lower level functions of the IOMMU
1966 * like protection domain handling and assignement of devices to domains
1967 * which is not possible with the dma_ops interface.
1969 *****************************************************************************/
1971 static void cleanup_domain(struct protection_domain *domain)
1973 unsigned long flags;
1976 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1978 for (devid = 0; devid <= amd_iommu_last_bdf; ++devid)
1979 if (amd_iommu_pd_table[devid] == domain)
1980 __detach_device(domain, devid);
1982 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1985 static int amd_iommu_domain_init(struct iommu_domain *dom)
1987 struct protection_domain *domain;
1989 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
1993 spin_lock_init(&domain->lock);
1994 domain->mode = PAGE_MODE_3_LEVEL;
1995 domain->id = domain_id_alloc();
1998 domain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1999 if (!domain->pt_root)
2012 static void amd_iommu_domain_destroy(struct iommu_domain *dom)
2014 struct protection_domain *domain = dom->priv;
2019 if (domain->dev_cnt > 0)
2020 cleanup_domain(domain);
2022 BUG_ON(domain->dev_cnt != 0);
2024 free_pagetable(domain);
2026 domain_id_free(domain->id);
2033 static void amd_iommu_detach_device(struct iommu_domain *dom,
2036 struct protection_domain *domain = dom->priv;
2037 struct amd_iommu *iommu;
2038 struct pci_dev *pdev;
2041 if (dev->bus != &pci_bus_type)
2044 pdev = to_pci_dev(dev);
2046 devid = calc_devid(pdev->bus->number, pdev->devfn);
2049 detach_device(domain, devid);
2051 iommu = amd_iommu_rlookup_table[devid];
2055 iommu_queue_inv_dev_entry(iommu, devid);
2056 iommu_completion_wait(iommu);
2059 static int amd_iommu_attach_device(struct iommu_domain *dom,
2062 struct protection_domain *domain = dom->priv;
2063 struct protection_domain *old_domain;
2064 struct amd_iommu *iommu;
2065 struct pci_dev *pdev;
2068 if (dev->bus != &pci_bus_type)
2071 pdev = to_pci_dev(dev);
2073 devid = calc_devid(pdev->bus->number, pdev->devfn);
2075 if (devid >= amd_iommu_last_bdf ||
2076 devid != amd_iommu_alias_table[devid])
2079 iommu = amd_iommu_rlookup_table[devid];
2083 old_domain = domain_for_device(devid);
2087 attach_device(iommu, domain, devid);
2089 iommu_completion_wait(iommu);
2094 static int amd_iommu_map_range(struct iommu_domain *dom,
2095 unsigned long iova, phys_addr_t paddr,
2096 size_t size, int iommu_prot)
2098 struct protection_domain *domain = dom->priv;
2099 unsigned long i, npages = iommu_num_pages(paddr, size, PAGE_SIZE);
2103 if (iommu_prot & IOMMU_READ)
2104 prot |= IOMMU_PROT_IR;
2105 if (iommu_prot & IOMMU_WRITE)
2106 prot |= IOMMU_PROT_IW;
2111 for (i = 0; i < npages; ++i) {
2112 ret = iommu_map_page(domain, iova, paddr, prot);
2123 static void amd_iommu_unmap_range(struct iommu_domain *dom,
2124 unsigned long iova, size_t size)
2127 struct protection_domain *domain = dom->priv;
2128 unsigned long i, npages = iommu_num_pages(iova, size, PAGE_SIZE);
2132 for (i = 0; i < npages; ++i) {
2133 iommu_unmap_page(domain, iova);
2137 iommu_flush_domain(domain->id);
2140 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
2143 struct protection_domain *domain = dom->priv;
2144 unsigned long offset = iova & ~PAGE_MASK;
2148 pte = &domain->pt_root[IOMMU_PTE_L2_INDEX(iova)];
2150 if (!IOMMU_PTE_PRESENT(*pte))
2153 pte = IOMMU_PTE_PAGE(*pte);
2154 pte = &pte[IOMMU_PTE_L1_INDEX(iova)];
2156 if (!IOMMU_PTE_PRESENT(*pte))
2159 pte = IOMMU_PTE_PAGE(*pte);
2160 pte = &pte[IOMMU_PTE_L0_INDEX(iova)];
2162 if (!IOMMU_PTE_PRESENT(*pte))
2165 paddr = *pte & IOMMU_PAGE_MASK;
2171 static int amd_iommu_domain_has_cap(struct iommu_domain *domain,
2177 static struct iommu_ops amd_iommu_ops = {
2178 .domain_init = amd_iommu_domain_init,
2179 .domain_destroy = amd_iommu_domain_destroy,
2180 .attach_dev = amd_iommu_attach_device,
2181 .detach_dev = amd_iommu_detach_device,
2182 .map = amd_iommu_map_range,
2183 .unmap = amd_iommu_unmap_range,
2184 .iova_to_phys = amd_iommu_iova_to_phys,
2185 .domain_has_cap = amd_iommu_domain_has_cap,