2 * Copyright (C) 2001 Mike Corrigan & Dave Engebretsen, IBM Corporation
4 * Rewrite, cleanup, new allocation schemes, virtual merging:
5 * Copyright (C) 2004 Olof Johansson, IBM Corporation
6 * and Ben. Herrenschmidt, IBM Corporation
8 * Dynamic DMA mapping support, bus-independent parts.
10 * This program is free software; you can redistribute it and/or modify
11 * it under the terms of the GNU General Public License as published by
12 * the Free Software Foundation; either version 2 of the License, or
13 * (at your option) any later version.
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 * GNU General Public License for more details.
20 * You should have received a copy of the GNU General Public License
21 * along with this program; if not, write to the Free Software
22 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
26 #include <linux/init.h>
27 #include <linux/types.h>
28 #include <linux/slab.h>
30 #include <linux/spinlock.h>
31 #include <linux/string.h>
32 #include <linux/dma-mapping.h>
33 #include <linux/bitmap.h>
34 #include <linux/iommu-helper.h>
35 #include <linux/crash_dump.h>
36 #include <linux/hash.h>
37 #include <linux/fault-inject.h>
38 #include <linux/pci.h>
39 #include <linux/iommu.h>
40 #include <linux/sched.h>
43 #include <asm/iommu.h>
44 #include <asm/pci-bridge.h>
45 #include <asm/machdep.h>
46 #include <asm/kdump.h>
47 #include <asm/fadump.h>
55 static void __iommu_free(struct iommu_table *, dma_addr_t, unsigned int);
57 static int __init setup_iommu(char *str)
59 if (!strcmp(str, "novmerge"))
61 else if (!strcmp(str, "vmerge"))
66 __setup("iommu=", setup_iommu);
68 static DEFINE_PER_CPU(unsigned int, iommu_pool_hash);
71 * We precalculate the hash to avoid doing it on every allocation.
73 * The hash is important to spread CPUs across all the pools. For example,
74 * on a POWER7 with 4 way SMT we want interrupts on the primary threads and
75 * with 4 pools all primary threads would map to the same pool.
77 static int __init setup_iommu_pool_hash(void)
81 for_each_possible_cpu(i)
82 per_cpu(iommu_pool_hash, i) = hash_32(i, IOMMU_POOL_HASHBITS);
86 subsys_initcall(setup_iommu_pool_hash);
88 #ifdef CONFIG_FAIL_IOMMU
90 static DECLARE_FAULT_ATTR(fail_iommu);
92 static int __init setup_fail_iommu(char *str)
94 return setup_fault_attr(&fail_iommu, str);
96 __setup("fail_iommu=", setup_fail_iommu);
98 static bool should_fail_iommu(struct device *dev)
100 return dev->archdata.fail_iommu && should_fail(&fail_iommu, 1);
103 static int __init fail_iommu_debugfs(void)
105 struct dentry *dir = fault_create_debugfs_attr("fail_iommu",
108 return PTR_ERR_OR_ZERO(dir);
110 late_initcall(fail_iommu_debugfs);
112 static ssize_t fail_iommu_show(struct device *dev,
113 struct device_attribute *attr, char *buf)
115 return sprintf(buf, "%d\n", dev->archdata.fail_iommu);
118 static ssize_t fail_iommu_store(struct device *dev,
119 struct device_attribute *attr, const char *buf,
124 if (count > 0 && sscanf(buf, "%d", &i) > 0)
125 dev->archdata.fail_iommu = (i == 0) ? 0 : 1;
130 static DEVICE_ATTR(fail_iommu, S_IRUGO|S_IWUSR, fail_iommu_show,
133 static int fail_iommu_bus_notify(struct notifier_block *nb,
134 unsigned long action, void *data)
136 struct device *dev = data;
138 if (action == BUS_NOTIFY_ADD_DEVICE) {
139 if (device_create_file(dev, &dev_attr_fail_iommu))
140 pr_warn("Unable to create IOMMU fault injection sysfs "
142 } else if (action == BUS_NOTIFY_DEL_DEVICE) {
143 device_remove_file(dev, &dev_attr_fail_iommu);
149 static struct notifier_block fail_iommu_bus_notifier = {
150 .notifier_call = fail_iommu_bus_notify
153 static int __init fail_iommu_setup(void)
156 bus_register_notifier(&pci_bus_type, &fail_iommu_bus_notifier);
159 bus_register_notifier(&vio_bus_type, &fail_iommu_bus_notifier);
165 * Must execute after PCI and VIO subsystem have initialised but before
166 * devices are probed.
168 arch_initcall(fail_iommu_setup);
170 static inline bool should_fail_iommu(struct device *dev)
176 static unsigned long iommu_range_alloc(struct device *dev,
177 struct iommu_table *tbl,
178 unsigned long npages,
179 unsigned long *handle,
181 unsigned int align_order)
183 unsigned long n, end, start;
185 int largealloc = npages > 15;
187 unsigned long align_mask;
188 unsigned long boundary_size;
190 unsigned int pool_nr;
191 struct iommu_pool *pool;
193 align_mask = 0xffffffffffffffffl >> (64 - align_order);
195 /* This allocator was derived from x86_64's bit string search */
198 if (unlikely(npages == 0)) {
199 if (printk_ratelimit())
201 return DMA_ERROR_CODE;
204 if (should_fail_iommu(dev))
205 return DMA_ERROR_CODE;
208 * We don't need to disable preemption here because any CPU can
209 * safely use any IOMMU pool.
211 pool_nr = __raw_get_cpu_var(iommu_pool_hash) & (tbl->nr_pools - 1);
214 pool = &(tbl->large_pool);
216 pool = &(tbl->pools[pool_nr]);
218 spin_lock_irqsave(&(pool->lock), flags);
221 if ((pass == 0) && handle && *handle &&
222 (*handle >= pool->start) && (*handle < pool->end))
229 /* The case below can happen if we have a small segment appended
230 * to a large, or when the previous alloc was at the very end of
231 * the available space. If so, go back to the initial start.
236 if (limit + tbl->it_offset > mask) {
237 limit = mask - tbl->it_offset + 1;
238 /* If we're constrained on address range, first try
239 * at the masked hint to avoid O(n) search complexity,
240 * but on second pass, start at 0 in pool 0.
242 if ((start & mask) >= limit || pass > 0) {
243 spin_unlock(&(pool->lock));
244 pool = &(tbl->pools[0]);
245 spin_lock(&(pool->lock));
253 boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
254 1 << IOMMU_PAGE_SHIFT);
256 boundary_size = ALIGN(1UL << 32, 1 << IOMMU_PAGE_SHIFT);
257 /* 4GB boundary for iseries_hv_alloc and iseries_hv_map */
259 n = iommu_area_alloc(tbl->it_map, limit, start, npages,
260 tbl->it_offset, boundary_size >> IOMMU_PAGE_SHIFT,
263 if (likely(pass == 0)) {
264 /* First try the pool from the start */
265 pool->hint = pool->start;
269 } else if (pass <= tbl->nr_pools) {
270 /* Now try scanning all the other pools */
271 spin_unlock(&(pool->lock));
272 pool_nr = (pool_nr + 1) & (tbl->nr_pools - 1);
273 pool = &tbl->pools[pool_nr];
274 spin_lock(&(pool->lock));
275 pool->hint = pool->start;
281 spin_unlock_irqrestore(&(pool->lock), flags);
282 return DMA_ERROR_CODE;
288 /* Bump the hint to a new block for small allocs. */
290 /* Don't bump to new block to avoid fragmentation */
293 /* Overflow will be taken care of at the next allocation */
294 pool->hint = (end + tbl->it_blocksize - 1) &
295 ~(tbl->it_blocksize - 1);
298 /* Update handle for SG allocations */
302 spin_unlock_irqrestore(&(pool->lock), flags);
307 static dma_addr_t iommu_alloc(struct device *dev, struct iommu_table *tbl,
308 void *page, unsigned int npages,
309 enum dma_data_direction direction,
310 unsigned long mask, unsigned int align_order,
311 struct dma_attrs *attrs)
314 dma_addr_t ret = DMA_ERROR_CODE;
317 entry = iommu_range_alloc(dev, tbl, npages, NULL, mask, align_order);
319 if (unlikely(entry == DMA_ERROR_CODE))
320 return DMA_ERROR_CODE;
322 entry += tbl->it_offset; /* Offset into real TCE table */
323 ret = entry << IOMMU_PAGE_SHIFT; /* Set the return dma address */
325 /* Put the TCEs in the HW table */
326 build_fail = ppc_md.tce_build(tbl, entry, npages,
327 (unsigned long)page & IOMMU_PAGE_MASK,
330 /* ppc_md.tce_build() only returns non-zero for transient errors.
331 * Clean up the table bitmap in this case and return
332 * DMA_ERROR_CODE. For all other errors the functionality is
335 if (unlikely(build_fail)) {
336 __iommu_free(tbl, ret, npages);
337 return DMA_ERROR_CODE;
340 /* Flush/invalidate TLB caches if necessary */
341 if (ppc_md.tce_flush)
342 ppc_md.tce_flush(tbl);
344 /* Make sure updates are seen by hardware */
350 static bool iommu_free_check(struct iommu_table *tbl, dma_addr_t dma_addr,
353 unsigned long entry, free_entry;
355 entry = dma_addr >> IOMMU_PAGE_SHIFT;
356 free_entry = entry - tbl->it_offset;
358 if (((free_entry + npages) > tbl->it_size) ||
359 (entry < tbl->it_offset)) {
360 if (printk_ratelimit()) {
361 printk(KERN_INFO "iommu_free: invalid entry\n");
362 printk(KERN_INFO "\tentry = 0x%lx\n", entry);
363 printk(KERN_INFO "\tdma_addr = 0x%llx\n", (u64)dma_addr);
364 printk(KERN_INFO "\tTable = 0x%llx\n", (u64)tbl);
365 printk(KERN_INFO "\tbus# = 0x%llx\n", (u64)tbl->it_busno);
366 printk(KERN_INFO "\tsize = 0x%llx\n", (u64)tbl->it_size);
367 printk(KERN_INFO "\tstartOff = 0x%llx\n", (u64)tbl->it_offset);
368 printk(KERN_INFO "\tindex = 0x%llx\n", (u64)tbl->it_index);
378 static struct iommu_pool *get_pool(struct iommu_table *tbl,
381 struct iommu_pool *p;
382 unsigned long largepool_start = tbl->large_pool.start;
384 /* The large pool is the last pool at the top of the table */
385 if (entry >= largepool_start) {
386 p = &tbl->large_pool;
388 unsigned int pool_nr = entry / tbl->poolsize;
390 BUG_ON(pool_nr > tbl->nr_pools);
391 p = &tbl->pools[pool_nr];
397 static void __iommu_free(struct iommu_table *tbl, dma_addr_t dma_addr,
400 unsigned long entry, free_entry;
402 struct iommu_pool *pool;
404 entry = dma_addr >> IOMMU_PAGE_SHIFT;
405 free_entry = entry - tbl->it_offset;
407 pool = get_pool(tbl, free_entry);
409 if (!iommu_free_check(tbl, dma_addr, npages))
412 ppc_md.tce_free(tbl, entry, npages);
414 spin_lock_irqsave(&(pool->lock), flags);
415 bitmap_clear(tbl->it_map, free_entry, npages);
416 spin_unlock_irqrestore(&(pool->lock), flags);
419 static void iommu_free(struct iommu_table *tbl, dma_addr_t dma_addr,
422 __iommu_free(tbl, dma_addr, npages);
424 /* Make sure TLB cache is flushed if the HW needs it. We do
425 * not do an mb() here on purpose, it is not needed on any of
426 * the current platforms.
428 if (ppc_md.tce_flush)
429 ppc_md.tce_flush(tbl);
432 int iommu_map_sg(struct device *dev, struct iommu_table *tbl,
433 struct scatterlist *sglist, int nelems,
434 unsigned long mask, enum dma_data_direction direction,
435 struct dma_attrs *attrs)
437 dma_addr_t dma_next = 0, dma_addr;
438 struct scatterlist *s, *outs, *segstart;
439 int outcount, incount, i, build_fail = 0;
441 unsigned long handle;
442 unsigned int max_seg_size;
444 BUG_ON(direction == DMA_NONE);
446 if ((nelems == 0) || !tbl)
449 outs = s = segstart = &sglist[0];
454 /* Init first segment length for backout at failure */
455 outs->dma_length = 0;
457 DBG("sg mapping %d elements:\n", nelems);
459 max_seg_size = dma_get_max_seg_size(dev);
460 for_each_sg(sglist, s, nelems, i) {
461 unsigned long vaddr, npages, entry, slen;
469 /* Allocate iommu entries for that segment */
470 vaddr = (unsigned long) sg_virt(s);
471 npages = iommu_num_pages(vaddr, slen, IOMMU_PAGE_SIZE);
473 if (IOMMU_PAGE_SHIFT < PAGE_SHIFT && slen >= PAGE_SIZE &&
474 (vaddr & ~PAGE_MASK) == 0)
475 align = PAGE_SHIFT - IOMMU_PAGE_SHIFT;
476 entry = iommu_range_alloc(dev, tbl, npages, &handle,
477 mask >> IOMMU_PAGE_SHIFT, align);
479 DBG(" - vaddr: %lx, size: %lx\n", vaddr, slen);
482 if (unlikely(entry == DMA_ERROR_CODE)) {
483 if (printk_ratelimit())
484 dev_info(dev, "iommu_alloc failed, tbl %p "
485 "vaddr %lx npages %lu\n", tbl, vaddr,
490 /* Convert entry to a dma_addr_t */
491 entry += tbl->it_offset;
492 dma_addr = entry << IOMMU_PAGE_SHIFT;
493 dma_addr |= (s->offset & ~IOMMU_PAGE_MASK);
495 DBG(" - %lu pages, entry: %lx, dma_addr: %lx\n",
496 npages, entry, dma_addr);
498 /* Insert into HW table */
499 build_fail = ppc_md.tce_build(tbl, entry, npages,
500 vaddr & IOMMU_PAGE_MASK,
502 if(unlikely(build_fail))
505 /* If we are in an open segment, try merging */
507 DBG(" - trying merge...\n");
508 /* We cannot merge if:
509 * - allocated dma_addr isn't contiguous to previous allocation
511 if (novmerge || (dma_addr != dma_next) ||
512 (outs->dma_length + s->length > max_seg_size)) {
513 /* Can't merge: create a new segment */
516 outs = sg_next(outs);
517 DBG(" can't merge, new segment.\n");
519 outs->dma_length += s->length;
520 DBG(" merged, new len: %ux\n", outs->dma_length);
525 /* This is a new segment, fill entries */
526 DBG(" - filling new segment.\n");
527 outs->dma_address = dma_addr;
528 outs->dma_length = slen;
531 /* Calculate next page pointer for contiguous check */
532 dma_next = dma_addr + slen;
534 DBG(" - dma next is: %lx\n", dma_next);
537 /* Flush/invalidate TLB caches if necessary */
538 if (ppc_md.tce_flush)
539 ppc_md.tce_flush(tbl);
541 DBG("mapped %d elements:\n", outcount);
543 /* For the sake of iommu_unmap_sg, we clear out the length in the
544 * next entry of the sglist if we didn't fill the list completely
546 if (outcount < incount) {
547 outs = sg_next(outs);
548 outs->dma_address = DMA_ERROR_CODE;
549 outs->dma_length = 0;
552 /* Make sure updates are seen by hardware */
558 for_each_sg(sglist, s, nelems, i) {
559 if (s->dma_length != 0) {
560 unsigned long vaddr, npages;
562 vaddr = s->dma_address & IOMMU_PAGE_MASK;
563 npages = iommu_num_pages(s->dma_address, s->dma_length,
565 __iommu_free(tbl, vaddr, npages);
566 s->dma_address = DMA_ERROR_CODE;
576 void iommu_unmap_sg(struct iommu_table *tbl, struct scatterlist *sglist,
577 int nelems, enum dma_data_direction direction,
578 struct dma_attrs *attrs)
580 struct scatterlist *sg;
582 BUG_ON(direction == DMA_NONE);
590 dma_addr_t dma_handle = sg->dma_address;
592 if (sg->dma_length == 0)
594 npages = iommu_num_pages(dma_handle, sg->dma_length,
596 __iommu_free(tbl, dma_handle, npages);
600 /* Flush/invalidate TLBs if necessary. As for iommu_free(), we
601 * do not do an mb() here, the affected platforms do not need it
604 if (ppc_md.tce_flush)
605 ppc_md.tce_flush(tbl);
608 static void iommu_table_clear(struct iommu_table *tbl)
611 * In case of firmware assisted dump system goes through clean
612 * reboot process at the time of system crash. Hence it's safe to
613 * clear the TCE entries if firmware assisted dump is active.
615 if (!is_kdump_kernel() || is_fadump_active()) {
616 /* Clear the table in case firmware left allocations in it */
617 ppc_md.tce_free(tbl, tbl->it_offset, tbl->it_size);
621 #ifdef CONFIG_CRASH_DUMP
622 if (ppc_md.tce_get) {
623 unsigned long index, tceval, tcecount = 0;
625 /* Reserve the existing mappings left by the first kernel. */
626 for (index = 0; index < tbl->it_size; index++) {
627 tceval = ppc_md.tce_get(tbl, index + tbl->it_offset);
629 * Freed TCE entry contains 0x7fffffffffffffff on JS20
631 if (tceval && (tceval != 0x7fffffffffffffffUL)) {
632 __set_bit(index, tbl->it_map);
637 if ((tbl->it_size - tcecount) < KDUMP_MIN_TCE_ENTRIES) {
638 printk(KERN_WARNING "TCE table is full; freeing ");
639 printk(KERN_WARNING "%d entries for the kdump boot\n",
640 KDUMP_MIN_TCE_ENTRIES);
641 for (index = tbl->it_size - KDUMP_MIN_TCE_ENTRIES;
642 index < tbl->it_size; index++)
643 __clear_bit(index, tbl->it_map);
650 * Build a iommu_table structure. This contains a bit map which
651 * is used to manage allocation of the tce space.
653 struct iommu_table *iommu_init_table(struct iommu_table *tbl, int nid)
656 static int welcomed = 0;
659 struct iommu_pool *p;
661 /* number of bytes needed for the bitmap */
662 sz = BITS_TO_LONGS(tbl->it_size) * sizeof(unsigned long);
664 page = alloc_pages_node(nid, GFP_KERNEL, get_order(sz));
666 panic("iommu_init_table: Can't allocate %ld bytes\n", sz);
667 tbl->it_map = page_address(page);
668 memset(tbl->it_map, 0, sz);
671 * Reserve page 0 so it will not be used for any mappings.
672 * This avoids buggy drivers that consider page 0 to be invalid
673 * to crash the machine or even lose data.
675 if (tbl->it_offset == 0)
676 set_bit(0, tbl->it_map);
678 /* We only split the IOMMU table if we have 1GB or more of space */
679 if ((tbl->it_size << IOMMU_PAGE_SHIFT) >= (1UL * 1024 * 1024 * 1024))
680 tbl->nr_pools = IOMMU_NR_POOLS;
684 /* We reserve the top 1/4 of the table for large allocations */
685 tbl->poolsize = (tbl->it_size * 3 / 4) / tbl->nr_pools;
687 for (i = 0; i < tbl->nr_pools; i++) {
689 spin_lock_init(&(p->lock));
690 p->start = tbl->poolsize * i;
692 p->end = p->start + tbl->poolsize;
695 p = &tbl->large_pool;
696 spin_lock_init(&(p->lock));
697 p->start = tbl->poolsize * i;
699 p->end = tbl->it_size;
701 iommu_table_clear(tbl);
704 printk(KERN_INFO "IOMMU table initialized, virtual merging %s\n",
705 novmerge ? "disabled" : "enabled");
712 void iommu_free_table(struct iommu_table *tbl, const char *node_name)
714 unsigned long bitmap_sz;
717 if (!tbl || !tbl->it_map) {
718 printk(KERN_ERR "%s: expected TCE map for %s\n", __func__,
724 * In case we have reserved the first bit, we should not emit
727 if (tbl->it_offset == 0)
728 clear_bit(0, tbl->it_map);
730 #ifdef CONFIG_IOMMU_API
732 iommu_group_put(tbl->it_group);
733 BUG_ON(tbl->it_group);
737 /* verify that table contains no entries */
738 if (!bitmap_empty(tbl->it_map, tbl->it_size))
739 pr_warn("%s: Unexpected TCEs for %s\n", __func__, node_name);
741 /* calculate bitmap size in bytes */
742 bitmap_sz = BITS_TO_LONGS(tbl->it_size) * sizeof(unsigned long);
745 order = get_order(bitmap_sz);
746 free_pages((unsigned long) tbl->it_map, order);
752 /* Creates TCEs for a user provided buffer. The user buffer must be
753 * contiguous real kernel storage (not vmalloc). The address passed here
754 * comprises a page address and offset into that page. The dma_addr_t
755 * returned will point to the same byte within the page as was passed in.
757 dma_addr_t iommu_map_page(struct device *dev, struct iommu_table *tbl,
758 struct page *page, unsigned long offset, size_t size,
759 unsigned long mask, enum dma_data_direction direction,
760 struct dma_attrs *attrs)
762 dma_addr_t dma_handle = DMA_ERROR_CODE;
765 unsigned int npages, align;
767 BUG_ON(direction == DMA_NONE);
769 vaddr = page_address(page) + offset;
770 uaddr = (unsigned long)vaddr;
771 npages = iommu_num_pages(uaddr, size, IOMMU_PAGE_SIZE);
775 if (IOMMU_PAGE_SHIFT < PAGE_SHIFT && size >= PAGE_SIZE &&
776 ((unsigned long)vaddr & ~PAGE_MASK) == 0)
777 align = PAGE_SHIFT - IOMMU_PAGE_SHIFT;
779 dma_handle = iommu_alloc(dev, tbl, vaddr, npages, direction,
780 mask >> IOMMU_PAGE_SHIFT, align,
782 if (dma_handle == DMA_ERROR_CODE) {
783 if (printk_ratelimit()) {
784 dev_info(dev, "iommu_alloc failed, tbl %p "
785 "vaddr %p npages %d\n", tbl, vaddr,
789 dma_handle |= (uaddr & ~IOMMU_PAGE_MASK);
795 void iommu_unmap_page(struct iommu_table *tbl, dma_addr_t dma_handle,
796 size_t size, enum dma_data_direction direction,
797 struct dma_attrs *attrs)
801 BUG_ON(direction == DMA_NONE);
804 npages = iommu_num_pages(dma_handle, size, IOMMU_PAGE_SIZE);
805 iommu_free(tbl, dma_handle, npages);
809 /* Allocates a contiguous real buffer and creates mappings over it.
810 * Returns the virtual address of the buffer and sets dma_handle
811 * to the dma address (mapping) of the first page.
813 void *iommu_alloc_coherent(struct device *dev, struct iommu_table *tbl,
814 size_t size, dma_addr_t *dma_handle,
815 unsigned long mask, gfp_t flag, int node)
820 unsigned int nio_pages, io_order;
823 size = PAGE_ALIGN(size);
824 order = get_order(size);
827 * Client asked for way too much space. This is checked later
828 * anyway. It is easier to debug here for the drivers than in
831 if (order >= IOMAP_MAX_ORDER) {
832 dev_info(dev, "iommu_alloc_consistent size too large: 0x%lx\n",
840 /* Alloc enough pages (and possibly more) */
841 page = alloc_pages_node(node, flag, order);
844 ret = page_address(page);
845 memset(ret, 0, size);
847 /* Set up tces to cover the allocated range */
848 nio_pages = size >> IOMMU_PAGE_SHIFT;
849 io_order = get_iommu_order(size);
850 mapping = iommu_alloc(dev, tbl, ret, nio_pages, DMA_BIDIRECTIONAL,
851 mask >> IOMMU_PAGE_SHIFT, io_order, NULL);
852 if (mapping == DMA_ERROR_CODE) {
853 free_pages((unsigned long)ret, order);
856 *dma_handle = mapping;
860 void iommu_free_coherent(struct iommu_table *tbl, size_t size,
861 void *vaddr, dma_addr_t dma_handle)
864 unsigned int nio_pages;
866 size = PAGE_ALIGN(size);
867 nio_pages = size >> IOMMU_PAGE_SHIFT;
868 iommu_free(tbl, dma_handle, nio_pages);
869 size = PAGE_ALIGN(size);
870 free_pages((unsigned long)vaddr, get_order(size));
874 #ifdef CONFIG_IOMMU_API
878 static void group_release(void *iommu_data)
880 struct iommu_table *tbl = iommu_data;
881 tbl->it_group = NULL;
884 void iommu_register_group(struct iommu_table *tbl,
885 int pci_domain_number, unsigned long pe_num)
887 struct iommu_group *grp;
890 grp = iommu_group_alloc();
892 pr_warn("powerpc iommu api: cannot create new group, err=%ld\n",
897 iommu_group_set_iommudata(grp, tbl, group_release);
898 name = kasprintf(GFP_KERNEL, "domain%d-pe%lx",
899 pci_domain_number, pe_num);
902 iommu_group_set_name(grp, name);
906 enum dma_data_direction iommu_tce_direction(unsigned long tce)
908 if ((tce & TCE_PCI_READ) && (tce & TCE_PCI_WRITE))
909 return DMA_BIDIRECTIONAL;
910 else if (tce & TCE_PCI_READ)
911 return DMA_TO_DEVICE;
912 else if (tce & TCE_PCI_WRITE)
913 return DMA_FROM_DEVICE;
917 EXPORT_SYMBOL_GPL(iommu_tce_direction);
919 void iommu_flush_tce(struct iommu_table *tbl)
921 /* Flush/invalidate TLB caches if necessary */
922 if (ppc_md.tce_flush)
923 ppc_md.tce_flush(tbl);
925 /* Make sure updates are seen by hardware */
928 EXPORT_SYMBOL_GPL(iommu_flush_tce);
930 int iommu_tce_clear_param_check(struct iommu_table *tbl,
931 unsigned long ioba, unsigned long tce_value,
932 unsigned long npages)
934 /* ppc_md.tce_free() does not support any value but 0 */
938 if (ioba & ~IOMMU_PAGE_MASK)
941 ioba >>= IOMMU_PAGE_SHIFT;
942 if (ioba < tbl->it_offset)
945 if ((ioba + npages) > (tbl->it_offset + tbl->it_size))
950 EXPORT_SYMBOL_GPL(iommu_tce_clear_param_check);
952 int iommu_tce_put_param_check(struct iommu_table *tbl,
953 unsigned long ioba, unsigned long tce)
955 if (!(tce & (TCE_PCI_WRITE | TCE_PCI_READ)))
958 if (tce & ~(IOMMU_PAGE_MASK | TCE_PCI_WRITE | TCE_PCI_READ))
961 if (ioba & ~IOMMU_PAGE_MASK)
964 ioba >>= IOMMU_PAGE_SHIFT;
965 if (ioba < tbl->it_offset)
968 if ((ioba + 1) > (tbl->it_offset + tbl->it_size))
973 EXPORT_SYMBOL_GPL(iommu_tce_put_param_check);
975 unsigned long iommu_clear_tce(struct iommu_table *tbl, unsigned long entry)
977 unsigned long oldtce;
978 struct iommu_pool *pool = get_pool(tbl, entry);
980 spin_lock(&(pool->lock));
982 oldtce = ppc_md.tce_get(tbl, entry);
983 if (oldtce & (TCE_PCI_WRITE | TCE_PCI_READ))
984 ppc_md.tce_free(tbl, entry, 1);
988 spin_unlock(&(pool->lock));
992 EXPORT_SYMBOL_GPL(iommu_clear_tce);
994 int iommu_clear_tces_and_put_pages(struct iommu_table *tbl,
995 unsigned long entry, unsigned long pages)
997 unsigned long oldtce;
1000 for ( ; pages; --pages, ++entry) {
1001 oldtce = iommu_clear_tce(tbl, entry);
1005 page = pfn_to_page(oldtce >> PAGE_SHIFT);
1008 if (oldtce & TCE_PCI_WRITE)
1016 EXPORT_SYMBOL_GPL(iommu_clear_tces_and_put_pages);
1019 * hwaddr is a kernel virtual address here (0xc... bazillion),
1020 * tce_build converts it to a physical address.
1022 int iommu_tce_build(struct iommu_table *tbl, unsigned long entry,
1023 unsigned long hwaddr, enum dma_data_direction direction)
1026 unsigned long oldtce;
1027 struct iommu_pool *pool = get_pool(tbl, entry);
1029 spin_lock(&(pool->lock));
1031 oldtce = ppc_md.tce_get(tbl, entry);
1032 /* Add new entry if it is not busy */
1033 if (!(oldtce & (TCE_PCI_WRITE | TCE_PCI_READ)))
1034 ret = ppc_md.tce_build(tbl, entry, 1, hwaddr, direction, NULL);
1036 spin_unlock(&(pool->lock));
1038 /* if (unlikely(ret))
1039 pr_err("iommu_tce: %s failed on hwaddr=%lx ioba=%lx kva=%lx ret=%d\n",
1040 __func__, hwaddr, entry << IOMMU_PAGE_SHIFT,
1045 EXPORT_SYMBOL_GPL(iommu_tce_build);
1047 int iommu_put_tce_user_mode(struct iommu_table *tbl, unsigned long entry,
1051 struct page *page = NULL;
1052 unsigned long hwaddr, offset = tce & IOMMU_PAGE_MASK & ~PAGE_MASK;
1053 enum dma_data_direction direction = iommu_tce_direction(tce);
1055 ret = get_user_pages_fast(tce & PAGE_MASK, 1,
1056 direction != DMA_TO_DEVICE, &page);
1057 if (unlikely(ret != 1)) {
1058 /* pr_err("iommu_tce: get_user_pages_fast failed tce=%lx ioba=%lx ret=%d\n",
1059 tce, entry << IOMMU_PAGE_SHIFT, ret); */
1062 hwaddr = (unsigned long) page_address(page) + offset;
1064 ret = iommu_tce_build(tbl, entry, hwaddr, direction);
1069 pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%d\n",
1070 __func__, entry << IOMMU_PAGE_SHIFT, tce, ret);
1074 EXPORT_SYMBOL_GPL(iommu_put_tce_user_mode);
1076 int iommu_take_ownership(struct iommu_table *tbl)
1078 unsigned long sz = (tbl->it_size + 7) >> 3;
1080 if (tbl->it_offset == 0)
1081 clear_bit(0, tbl->it_map);
1083 if (!bitmap_empty(tbl->it_map, tbl->it_size)) {
1084 pr_err("iommu_tce: it_map is not empty");
1088 memset(tbl->it_map, 0xff, sz);
1089 iommu_clear_tces_and_put_pages(tbl, tbl->it_offset, tbl->it_size);
1093 EXPORT_SYMBOL_GPL(iommu_take_ownership);
1095 void iommu_release_ownership(struct iommu_table *tbl)
1097 unsigned long sz = (tbl->it_size + 7) >> 3;
1099 iommu_clear_tces_and_put_pages(tbl, tbl->it_offset, tbl->it_size);
1100 memset(tbl->it_map, 0, sz);
1102 /* Restore bit#0 set by iommu_init_table() */
1103 if (tbl->it_offset == 0)
1104 set_bit(0, tbl->it_map);
1106 EXPORT_SYMBOL_GPL(iommu_release_ownership);
1108 static int iommu_add_device(struct device *dev)
1110 struct iommu_table *tbl;
1113 if (WARN_ON(dev->iommu_group)) {
1114 pr_warn("iommu_tce: device %s is already in iommu group %d, skipping\n",
1116 iommu_group_id(dev->iommu_group));
1120 tbl = get_iommu_table_base(dev);
1121 if (!tbl || !tbl->it_group) {
1122 pr_debug("iommu_tce: skipping device %s with no tbl\n",
1127 pr_debug("iommu_tce: adding %s to iommu group %d\n",
1128 dev_name(dev), iommu_group_id(tbl->it_group));
1130 ret = iommu_group_add_device(tbl->it_group, dev);
1132 pr_err("iommu_tce: %s has not been added, ret=%d\n",
1133 dev_name(dev), ret);
1138 static void iommu_del_device(struct device *dev)
1140 iommu_group_remove_device(dev);
1143 static int iommu_bus_notifier(struct notifier_block *nb,
1144 unsigned long action, void *data)
1146 struct device *dev = data;
1149 case BUS_NOTIFY_ADD_DEVICE:
1150 return iommu_add_device(dev);
1151 case BUS_NOTIFY_DEL_DEVICE:
1152 iommu_del_device(dev);
1159 static struct notifier_block tce_iommu_bus_nb = {
1160 .notifier_call = iommu_bus_notifier,
1163 static int __init tce_iommu_init(void)
1165 struct pci_dev *pdev = NULL;
1167 BUILD_BUG_ON(PAGE_SIZE < IOMMU_PAGE_SIZE);
1169 for_each_pci_dev(pdev)
1170 iommu_add_device(&pdev->dev);
1172 bus_register_notifier(&pci_bus_type, &tce_iommu_bus_nb);
1176 subsys_initcall_sync(tce_iommu_init);
1180 void iommu_register_group(struct iommu_table *tbl,
1181 int pci_domain_number, unsigned long pe_num)
1185 #endif /* CONFIG_IOMMU_API */