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/bitops.h>
34 #include <linux/iommu-helper.h>
35 #include <linux/crash_dump.h>
38 #include <asm/iommu.h>
39 #include <asm/pci-bridge.h>
40 #include <asm/machdep.h>
41 #include <asm/kdump.h>
45 #ifdef CONFIG_IOMMU_VMERGE
46 static int novmerge = 0;
48 static int novmerge = 1;
51 static int protect4gb = 1;
53 static void __iommu_free(struct iommu_table *, dma_addr_t, unsigned int);
55 static int __init setup_protect4gb(char *str)
57 if (strcmp(str, "on") == 0)
59 else if (strcmp(str, "off") == 0)
65 static int __init setup_iommu(char *str)
67 if (!strcmp(str, "novmerge"))
69 else if (!strcmp(str, "vmerge"))
74 __setup("protect4gb=", setup_protect4gb);
75 __setup("iommu=", setup_iommu);
77 static unsigned long iommu_range_alloc(struct device *dev,
78 struct iommu_table *tbl,
80 unsigned long *handle,
82 unsigned int align_order)
84 unsigned long n, end, start;
86 int largealloc = npages > 15;
88 unsigned long align_mask;
89 unsigned long boundary_size;
91 align_mask = 0xffffffffffffffffl >> (64 - align_order);
93 /* This allocator was derived from x86_64's bit string search */
96 if (unlikely(npages == 0)) {
97 if (printk_ratelimit())
99 return DMA_ERROR_CODE;
102 if (handle && *handle)
105 start = largealloc ? tbl->it_largehint : tbl->it_hint;
107 /* Use only half of the table for small allocs (15 pages or less) */
108 limit = largealloc ? tbl->it_size : tbl->it_halfpoint;
110 if (largealloc && start < tbl->it_halfpoint)
111 start = tbl->it_halfpoint;
113 /* The case below can happen if we have a small segment appended
114 * to a large, or when the previous alloc was at the very end of
115 * the available space. If so, go back to the initial start.
118 start = largealloc ? tbl->it_largehint : tbl->it_hint;
122 if (limit + tbl->it_offset > mask) {
123 limit = mask - tbl->it_offset + 1;
124 /* If we're constrained on address range, first try
125 * at the masked hint to avoid O(n) search complexity,
126 * but on second pass, start at 0.
128 if ((start & mask) >= limit || pass > 0)
135 boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
136 1 << IOMMU_PAGE_SHIFT);
138 boundary_size = ALIGN(1UL << 32, 1 << IOMMU_PAGE_SHIFT);
139 /* 4GB boundary for iseries_hv_alloc and iseries_hv_map */
141 n = iommu_area_alloc(tbl->it_map, limit, start, npages,
142 tbl->it_offset, boundary_size >> IOMMU_PAGE_SHIFT,
145 if (likely(pass < 2)) {
146 /* First failure, just rescan the half of the table.
147 * Second failure, rescan the other half of the table.
149 start = (largealloc ^ pass) ? tbl->it_halfpoint : 0;
150 limit = pass ? tbl->it_size : limit;
154 /* Third failure, give up */
155 return DMA_ERROR_CODE;
161 /* Bump the hint to a new block for small allocs. */
163 /* Don't bump to new block to avoid fragmentation */
164 tbl->it_largehint = end;
166 /* Overflow will be taken care of at the next allocation */
167 tbl->it_hint = (end + tbl->it_blocksize - 1) &
168 ~(tbl->it_blocksize - 1);
171 /* Update handle for SG allocations */
178 static dma_addr_t iommu_alloc(struct device *dev, struct iommu_table *tbl,
179 void *page, unsigned int npages,
180 enum dma_data_direction direction,
181 unsigned long mask, unsigned int align_order,
182 struct dma_attrs *attrs)
184 unsigned long entry, flags;
185 dma_addr_t ret = DMA_ERROR_CODE;
188 spin_lock_irqsave(&(tbl->it_lock), flags);
190 entry = iommu_range_alloc(dev, tbl, npages, NULL, mask, align_order);
192 if (unlikely(entry == DMA_ERROR_CODE)) {
193 spin_unlock_irqrestore(&(tbl->it_lock), flags);
194 return DMA_ERROR_CODE;
197 entry += tbl->it_offset; /* Offset into real TCE table */
198 ret = entry << IOMMU_PAGE_SHIFT; /* Set the return dma address */
200 /* Put the TCEs in the HW table */
201 build_fail = ppc_md.tce_build(tbl, entry, npages,
202 (unsigned long)page & IOMMU_PAGE_MASK,
205 /* ppc_md.tce_build() only returns non-zero for transient errors.
206 * Clean up the table bitmap in this case and return
207 * DMA_ERROR_CODE. For all other errors the functionality is
210 if (unlikely(build_fail)) {
211 __iommu_free(tbl, ret, npages);
213 spin_unlock_irqrestore(&(tbl->it_lock), flags);
214 return DMA_ERROR_CODE;
217 /* Flush/invalidate TLB caches if necessary */
218 if (ppc_md.tce_flush)
219 ppc_md.tce_flush(tbl);
221 spin_unlock_irqrestore(&(tbl->it_lock), flags);
223 /* Make sure updates are seen by hardware */
229 static void __iommu_free(struct iommu_table *tbl, dma_addr_t dma_addr,
232 unsigned long entry, free_entry;
234 entry = dma_addr >> IOMMU_PAGE_SHIFT;
235 free_entry = entry - tbl->it_offset;
237 if (((free_entry + npages) > tbl->it_size) ||
238 (entry < tbl->it_offset)) {
239 if (printk_ratelimit()) {
240 printk(KERN_INFO "iommu_free: invalid entry\n");
241 printk(KERN_INFO "\tentry = 0x%lx\n", entry);
242 printk(KERN_INFO "\tdma_addr = 0x%llx\n", (u64)dma_addr);
243 printk(KERN_INFO "\tTable = 0x%llx\n", (u64)tbl);
244 printk(KERN_INFO "\tbus# = 0x%llx\n", (u64)tbl->it_busno);
245 printk(KERN_INFO "\tsize = 0x%llx\n", (u64)tbl->it_size);
246 printk(KERN_INFO "\tstartOff = 0x%llx\n", (u64)tbl->it_offset);
247 printk(KERN_INFO "\tindex = 0x%llx\n", (u64)tbl->it_index);
253 ppc_md.tce_free(tbl, entry, npages);
254 iommu_area_free(tbl->it_map, free_entry, npages);
257 static void iommu_free(struct iommu_table *tbl, dma_addr_t dma_addr,
262 spin_lock_irqsave(&(tbl->it_lock), flags);
264 __iommu_free(tbl, dma_addr, npages);
266 /* Make sure TLB cache is flushed if the HW needs it. We do
267 * not do an mb() here on purpose, it is not needed on any of
268 * the current platforms.
270 if (ppc_md.tce_flush)
271 ppc_md.tce_flush(tbl);
273 spin_unlock_irqrestore(&(tbl->it_lock), flags);
276 int iommu_map_sg(struct device *dev, struct iommu_table *tbl,
277 struct scatterlist *sglist, int nelems,
278 unsigned long mask, enum dma_data_direction direction,
279 struct dma_attrs *attrs)
281 dma_addr_t dma_next = 0, dma_addr;
283 struct scatterlist *s, *outs, *segstart;
284 int outcount, incount, i, build_fail = 0;
286 unsigned long handle;
287 unsigned int max_seg_size;
289 BUG_ON(direction == DMA_NONE);
291 if ((nelems == 0) || !tbl)
294 outs = s = segstart = &sglist[0];
299 /* Init first segment length for backout at failure */
300 outs->dma_length = 0;
302 DBG("sg mapping %d elements:\n", nelems);
304 spin_lock_irqsave(&(tbl->it_lock), flags);
306 max_seg_size = dma_get_max_seg_size(dev);
307 for_each_sg(sglist, s, nelems, i) {
308 unsigned long vaddr, npages, entry, slen;
316 /* Allocate iommu entries for that segment */
317 vaddr = (unsigned long) sg_virt(s);
318 npages = iommu_num_pages(vaddr, slen, IOMMU_PAGE_SIZE);
320 if (IOMMU_PAGE_SHIFT < PAGE_SHIFT && slen >= PAGE_SIZE &&
321 (vaddr & ~PAGE_MASK) == 0)
322 align = PAGE_SHIFT - IOMMU_PAGE_SHIFT;
323 entry = iommu_range_alloc(dev, tbl, npages, &handle,
324 mask >> IOMMU_PAGE_SHIFT, align);
326 DBG(" - vaddr: %lx, size: %lx\n", vaddr, slen);
329 if (unlikely(entry == DMA_ERROR_CODE)) {
330 if (printk_ratelimit())
331 printk(KERN_INFO "iommu_alloc failed, tbl %p vaddr %lx"
332 " npages %lx\n", tbl, vaddr, npages);
336 /* Convert entry to a dma_addr_t */
337 entry += tbl->it_offset;
338 dma_addr = entry << IOMMU_PAGE_SHIFT;
339 dma_addr |= (s->offset & ~IOMMU_PAGE_MASK);
341 DBG(" - %lu pages, entry: %lx, dma_addr: %lx\n",
342 npages, entry, dma_addr);
344 /* Insert into HW table */
345 build_fail = ppc_md.tce_build(tbl, entry, npages,
346 vaddr & IOMMU_PAGE_MASK,
348 if(unlikely(build_fail))
351 /* If we are in an open segment, try merging */
353 DBG(" - trying merge...\n");
354 /* We cannot merge if:
355 * - allocated dma_addr isn't contiguous to previous allocation
357 if (novmerge || (dma_addr != dma_next) ||
358 (outs->dma_length + s->length > max_seg_size)) {
359 /* Can't merge: create a new segment */
362 outs = sg_next(outs);
363 DBG(" can't merge, new segment.\n");
365 outs->dma_length += s->length;
366 DBG(" merged, new len: %ux\n", outs->dma_length);
371 /* This is a new segment, fill entries */
372 DBG(" - filling new segment.\n");
373 outs->dma_address = dma_addr;
374 outs->dma_length = slen;
377 /* Calculate next page pointer for contiguous check */
378 dma_next = dma_addr + slen;
380 DBG(" - dma next is: %lx\n", dma_next);
383 /* Flush/invalidate TLB caches if necessary */
384 if (ppc_md.tce_flush)
385 ppc_md.tce_flush(tbl);
387 spin_unlock_irqrestore(&(tbl->it_lock), flags);
389 DBG("mapped %d elements:\n", outcount);
391 /* For the sake of iommu_unmap_sg, we clear out the length in the
392 * next entry of the sglist if we didn't fill the list completely
394 if (outcount < incount) {
395 outs = sg_next(outs);
396 outs->dma_address = DMA_ERROR_CODE;
397 outs->dma_length = 0;
400 /* Make sure updates are seen by hardware */
406 for_each_sg(sglist, s, nelems, i) {
407 if (s->dma_length != 0) {
408 unsigned long vaddr, npages;
410 vaddr = s->dma_address & IOMMU_PAGE_MASK;
411 npages = iommu_num_pages(s->dma_address, s->dma_length,
413 __iommu_free(tbl, vaddr, npages);
414 s->dma_address = DMA_ERROR_CODE;
420 spin_unlock_irqrestore(&(tbl->it_lock), flags);
425 void iommu_unmap_sg(struct iommu_table *tbl, struct scatterlist *sglist,
426 int nelems, enum dma_data_direction direction,
427 struct dma_attrs *attrs)
429 struct scatterlist *sg;
432 BUG_ON(direction == DMA_NONE);
437 spin_lock_irqsave(&(tbl->it_lock), flags);
442 dma_addr_t dma_handle = sg->dma_address;
444 if (sg->dma_length == 0)
446 npages = iommu_num_pages(dma_handle, sg->dma_length,
448 __iommu_free(tbl, dma_handle, npages);
452 /* Flush/invalidate TLBs if necessary. As for iommu_free(), we
453 * do not do an mb() here, the affected platforms do not need it
456 if (ppc_md.tce_flush)
457 ppc_md.tce_flush(tbl);
459 spin_unlock_irqrestore(&(tbl->it_lock), flags);
462 static void iommu_table_clear(struct iommu_table *tbl)
464 if (!is_kdump_kernel()) {
465 /* Clear the table in case firmware left allocations in it */
466 ppc_md.tce_free(tbl, tbl->it_offset, tbl->it_size);
470 #ifdef CONFIG_CRASH_DUMP
471 if (ppc_md.tce_get) {
472 unsigned long index, tceval, tcecount = 0;
474 /* Reserve the existing mappings left by the first kernel. */
475 for (index = 0; index < tbl->it_size; index++) {
476 tceval = ppc_md.tce_get(tbl, index + tbl->it_offset);
478 * Freed TCE entry contains 0x7fffffffffffffff on JS20
480 if (tceval && (tceval != 0x7fffffffffffffffUL)) {
481 __set_bit(index, tbl->it_map);
486 if ((tbl->it_size - tcecount) < KDUMP_MIN_TCE_ENTRIES) {
487 printk(KERN_WARNING "TCE table is full; freeing ");
488 printk(KERN_WARNING "%d entries for the kdump boot\n",
489 KDUMP_MIN_TCE_ENTRIES);
490 for (index = tbl->it_size - KDUMP_MIN_TCE_ENTRIES;
491 index < tbl->it_size; index++)
492 __clear_bit(index, tbl->it_map);
499 * Build a iommu_table structure. This contains a bit map which
500 * is used to manage allocation of the tce space.
502 struct iommu_table *iommu_init_table(struct iommu_table *tbl, int nid)
505 static int welcomed = 0;
508 /* Set aside 1/4 of the table for large allocations. */
509 tbl->it_halfpoint = tbl->it_size * 3 / 4;
511 /* number of bytes needed for the bitmap */
512 sz = (tbl->it_size + 7) >> 3;
514 page = alloc_pages_node(nid, GFP_ATOMIC, get_order(sz));
516 panic("iommu_init_table: Can't allocate %ld bytes\n", sz);
517 tbl->it_map = page_address(page);
518 memset(tbl->it_map, 0, sz);
521 tbl->it_largehint = tbl->it_halfpoint;
522 spin_lock_init(&tbl->it_lock);
524 iommu_table_clear(tbl);
527 printk(KERN_INFO "IOMMU table initialized, virtual merging %s\n",
528 novmerge ? "disabled" : "enabled");
535 void iommu_free_table(struct iommu_table *tbl, const char *node_name)
537 unsigned long bitmap_sz, i;
540 if (!tbl || !tbl->it_map) {
541 printk(KERN_ERR "%s: expected TCE map for %s\n", __func__,
546 /* verify that table contains no entries */
547 /* it_size is in entries, and we're examining 64 at a time */
548 for (i = 0; i < (tbl->it_size/64); i++) {
549 if (tbl->it_map[i] != 0) {
550 printk(KERN_WARNING "%s: Unexpected TCEs for %s\n",
551 __func__, node_name);
556 /* calculate bitmap size in bytes */
557 bitmap_sz = (tbl->it_size + 7) / 8;
560 order = get_order(bitmap_sz);
561 free_pages((unsigned long) tbl->it_map, order);
567 /* Creates TCEs for a user provided buffer. The user buffer must be
568 * contiguous real kernel storage (not vmalloc). The address passed here
569 * comprises a page address and offset into that page. The dma_addr_t
570 * returned will point to the same byte within the page as was passed in.
572 dma_addr_t iommu_map_page(struct device *dev, struct iommu_table *tbl,
573 struct page *page, unsigned long offset, size_t size,
574 unsigned long mask, enum dma_data_direction direction,
575 struct dma_attrs *attrs)
577 dma_addr_t dma_handle = DMA_ERROR_CODE;
580 unsigned int npages, align;
582 BUG_ON(direction == DMA_NONE);
584 vaddr = page_address(page) + offset;
585 uaddr = (unsigned long)vaddr;
586 npages = iommu_num_pages(uaddr, size, IOMMU_PAGE_SIZE);
590 if (IOMMU_PAGE_SHIFT < PAGE_SHIFT && size >= PAGE_SIZE &&
591 ((unsigned long)vaddr & ~PAGE_MASK) == 0)
592 align = PAGE_SHIFT - IOMMU_PAGE_SHIFT;
594 dma_handle = iommu_alloc(dev, tbl, vaddr, npages, direction,
595 mask >> IOMMU_PAGE_SHIFT, align,
597 if (dma_handle == DMA_ERROR_CODE) {
598 if (printk_ratelimit()) {
599 printk(KERN_INFO "iommu_alloc failed, "
600 "tbl %p vaddr %p npages %d\n",
604 dma_handle |= (uaddr & ~IOMMU_PAGE_MASK);
610 void iommu_unmap_page(struct iommu_table *tbl, dma_addr_t dma_handle,
611 size_t size, enum dma_data_direction direction,
612 struct dma_attrs *attrs)
616 BUG_ON(direction == DMA_NONE);
619 npages = iommu_num_pages(dma_handle, size, IOMMU_PAGE_SIZE);
620 iommu_free(tbl, dma_handle, npages);
624 /* Allocates a contiguous real buffer and creates mappings over it.
625 * Returns the virtual address of the buffer and sets dma_handle
626 * to the dma address (mapping) of the first page.
628 void *iommu_alloc_coherent(struct device *dev, struct iommu_table *tbl,
629 size_t size, dma_addr_t *dma_handle,
630 unsigned long mask, gfp_t flag, int node)
635 unsigned int nio_pages, io_order;
638 size = PAGE_ALIGN(size);
639 order = get_order(size);
642 * Client asked for way too much space. This is checked later
643 * anyway. It is easier to debug here for the drivers than in
646 if (order >= IOMAP_MAX_ORDER) {
647 printk("iommu_alloc_consistent size too large: 0x%lx\n", size);
654 /* Alloc enough pages (and possibly more) */
655 page = alloc_pages_node(node, flag, order);
658 ret = page_address(page);
659 memset(ret, 0, size);
661 /* Set up tces to cover the allocated range */
662 nio_pages = size >> IOMMU_PAGE_SHIFT;
663 io_order = get_iommu_order(size);
664 mapping = iommu_alloc(dev, tbl, ret, nio_pages, DMA_BIDIRECTIONAL,
665 mask >> IOMMU_PAGE_SHIFT, io_order, NULL);
666 if (mapping == DMA_ERROR_CODE) {
667 free_pages((unsigned long)ret, order);
670 *dma_handle = mapping;
674 void iommu_free_coherent(struct iommu_table *tbl, size_t size,
675 void *vaddr, dma_addr_t dma_handle)
678 unsigned int nio_pages;
680 size = PAGE_ALIGN(size);
681 nio_pages = size >> IOMMU_PAGE_SHIFT;
682 iommu_free(tbl, dma_handle, nio_pages);
683 size = PAGE_ALIGN(size);
684 free_pages((unsigned long)vaddr, get_order(size));