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>
37 #include <asm/iommu.h>
38 #include <asm/pci-bridge.h>
39 #include <asm/machdep.h>
40 #include <asm/kdump.h>
44 #ifdef CONFIG_IOMMU_VMERGE
45 static int novmerge = 0;
47 static int novmerge = 1;
50 static int protect4gb = 1;
52 static void __iommu_free(struct iommu_table *, dma_addr_t, unsigned int);
54 static inline unsigned long iommu_num_pages(unsigned long vaddr,
59 npages = IOMMU_PAGE_ALIGN(vaddr + slen) - (vaddr & IOMMU_PAGE_MASK);
60 npages >>= IOMMU_PAGE_SHIFT;
65 static int __init setup_protect4gb(char *str)
67 if (strcmp(str, "on") == 0)
69 else if (strcmp(str, "off") == 0)
75 static int __init setup_iommu(char *str)
77 if (!strcmp(str, "novmerge"))
79 else if (!strcmp(str, "vmerge"))
84 __setup("protect4gb=", setup_protect4gb);
85 __setup("iommu=", setup_iommu);
87 static unsigned long iommu_range_alloc(struct device *dev,
88 struct iommu_table *tbl,
90 unsigned long *handle,
92 unsigned int align_order)
94 unsigned long n, end, start;
96 int largealloc = npages > 15;
98 unsigned long align_mask;
99 unsigned long boundary_size;
101 align_mask = 0xffffffffffffffffl >> (64 - align_order);
103 /* This allocator was derived from x86_64's bit string search */
106 if (unlikely(npages == 0)) {
107 if (printk_ratelimit())
109 return DMA_ERROR_CODE;
112 if (handle && *handle)
115 start = largealloc ? tbl->it_largehint : tbl->it_hint;
117 /* Use only half of the table for small allocs (15 pages or less) */
118 limit = largealloc ? tbl->it_size : tbl->it_halfpoint;
120 if (largealloc && start < tbl->it_halfpoint)
121 start = tbl->it_halfpoint;
123 /* The case below can happen if we have a small segment appended
124 * to a large, or when the previous alloc was at the very end of
125 * the available space. If so, go back to the initial start.
128 start = largealloc ? tbl->it_largehint : tbl->it_hint;
132 if (limit + tbl->it_offset > mask) {
133 limit = mask - tbl->it_offset + 1;
134 /* If we're constrained on address range, first try
135 * at the masked hint to avoid O(n) search complexity,
136 * but on second pass, start at 0.
138 if ((start & mask) >= limit || pass > 0)
145 boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
146 1 << IOMMU_PAGE_SHIFT);
148 boundary_size = ALIGN(1UL << 32, 1 << IOMMU_PAGE_SHIFT);
149 /* 4GB boundary for iseries_hv_alloc and iseries_hv_map */
151 n = iommu_area_alloc(tbl->it_map, limit, start, npages,
152 tbl->it_offset, boundary_size >> IOMMU_PAGE_SHIFT,
155 if (likely(pass < 2)) {
156 /* First failure, just rescan the half of the table.
157 * Second failure, rescan the other half of the table.
159 start = (largealloc ^ pass) ? tbl->it_halfpoint : 0;
160 limit = pass ? tbl->it_size : limit;
164 /* Third failure, give up */
165 return DMA_ERROR_CODE;
171 /* Bump the hint to a new block for small allocs. */
173 /* Don't bump to new block to avoid fragmentation */
174 tbl->it_largehint = end;
176 /* Overflow will be taken care of at the next allocation */
177 tbl->it_hint = (end + tbl->it_blocksize - 1) &
178 ~(tbl->it_blocksize - 1);
181 /* Update handle for SG allocations */
188 static dma_addr_t iommu_alloc(struct device *dev, struct iommu_table *tbl,
189 void *page, unsigned int npages,
190 enum dma_data_direction direction,
191 unsigned long mask, unsigned int align_order,
192 struct dma_attrs *attrs)
194 unsigned long entry, flags;
195 dma_addr_t ret = DMA_ERROR_CODE;
198 spin_lock_irqsave(&(tbl->it_lock), flags);
200 entry = iommu_range_alloc(dev, tbl, npages, NULL, mask, align_order);
202 if (unlikely(entry == DMA_ERROR_CODE)) {
203 spin_unlock_irqrestore(&(tbl->it_lock), flags);
204 return DMA_ERROR_CODE;
207 entry += tbl->it_offset; /* Offset into real TCE table */
208 ret = entry << IOMMU_PAGE_SHIFT; /* Set the return dma address */
210 /* Put the TCEs in the HW table */
211 build_fail = ppc_md.tce_build(tbl, entry, npages,
212 (unsigned long)page & IOMMU_PAGE_MASK,
215 /* ppc_md.tce_build() only returns non-zero for transient errors.
216 * Clean up the table bitmap in this case and return
217 * DMA_ERROR_CODE. For all other errors the functionality is
220 if (unlikely(build_fail)) {
221 __iommu_free(tbl, ret, npages);
223 spin_unlock_irqrestore(&(tbl->it_lock), flags);
224 return DMA_ERROR_CODE;
227 /* Flush/invalidate TLB caches if necessary */
228 if (ppc_md.tce_flush)
229 ppc_md.tce_flush(tbl);
231 spin_unlock_irqrestore(&(tbl->it_lock), flags);
233 /* Make sure updates are seen by hardware */
239 static void __iommu_free(struct iommu_table *tbl, dma_addr_t dma_addr,
242 unsigned long entry, free_entry;
244 entry = dma_addr >> IOMMU_PAGE_SHIFT;
245 free_entry = entry - tbl->it_offset;
247 if (((free_entry + npages) > tbl->it_size) ||
248 (entry < tbl->it_offset)) {
249 if (printk_ratelimit()) {
250 printk(KERN_INFO "iommu_free: invalid entry\n");
251 printk(KERN_INFO "\tentry = 0x%lx\n", entry);
252 printk(KERN_INFO "\tdma_addr = 0x%lx\n", (u64)dma_addr);
253 printk(KERN_INFO "\tTable = 0x%lx\n", (u64)tbl);
254 printk(KERN_INFO "\tbus# = 0x%lx\n", (u64)tbl->it_busno);
255 printk(KERN_INFO "\tsize = 0x%lx\n", (u64)tbl->it_size);
256 printk(KERN_INFO "\tstartOff = 0x%lx\n", (u64)tbl->it_offset);
257 printk(KERN_INFO "\tindex = 0x%lx\n", (u64)tbl->it_index);
263 ppc_md.tce_free(tbl, entry, npages);
264 iommu_area_free(tbl->it_map, free_entry, npages);
267 static void iommu_free(struct iommu_table *tbl, dma_addr_t dma_addr,
272 spin_lock_irqsave(&(tbl->it_lock), flags);
274 __iommu_free(tbl, dma_addr, npages);
276 /* Make sure TLB cache is flushed if the HW needs it. We do
277 * not do an mb() here on purpose, it is not needed on any of
278 * the current platforms.
280 if (ppc_md.tce_flush)
281 ppc_md.tce_flush(tbl);
283 spin_unlock_irqrestore(&(tbl->it_lock), flags);
286 int iommu_map_sg(struct device *dev, struct iommu_table *tbl,
287 struct scatterlist *sglist, int nelems,
288 unsigned long mask, enum dma_data_direction direction,
289 struct dma_attrs *attrs)
291 dma_addr_t dma_next = 0, dma_addr;
293 struct scatterlist *s, *outs, *segstart;
294 int outcount, incount, i, build_fail = 0;
296 unsigned long handle;
297 unsigned int max_seg_size;
299 BUG_ON(direction == DMA_NONE);
301 if ((nelems == 0) || !tbl)
304 outs = s = segstart = &sglist[0];
309 /* Init first segment length for backout at failure */
310 outs->dma_length = 0;
312 DBG("sg mapping %d elements:\n", nelems);
314 spin_lock_irqsave(&(tbl->it_lock), flags);
316 max_seg_size = dma_get_max_seg_size(dev);
317 for_each_sg(sglist, s, nelems, i) {
318 unsigned long vaddr, npages, entry, slen;
326 /* Allocate iommu entries for that segment */
327 vaddr = (unsigned long) sg_virt(s);
328 npages = iommu_num_pages(vaddr, slen);
330 if (IOMMU_PAGE_SHIFT < PAGE_SHIFT && slen >= PAGE_SIZE &&
331 (vaddr & ~PAGE_MASK) == 0)
332 align = PAGE_SHIFT - IOMMU_PAGE_SHIFT;
333 entry = iommu_range_alloc(dev, tbl, npages, &handle,
334 mask >> IOMMU_PAGE_SHIFT, align);
336 DBG(" - vaddr: %lx, size: %lx\n", vaddr, slen);
339 if (unlikely(entry == DMA_ERROR_CODE)) {
340 if (printk_ratelimit())
341 printk(KERN_INFO "iommu_alloc failed, tbl %p vaddr %lx"
342 " npages %lx\n", tbl, vaddr, npages);
346 /* Convert entry to a dma_addr_t */
347 entry += tbl->it_offset;
348 dma_addr = entry << IOMMU_PAGE_SHIFT;
349 dma_addr |= (s->offset & ~IOMMU_PAGE_MASK);
351 DBG(" - %lu pages, entry: %lx, dma_addr: %lx\n",
352 npages, entry, dma_addr);
354 /* Insert into HW table */
355 build_fail = ppc_md.tce_build(tbl, entry, npages,
356 vaddr & IOMMU_PAGE_MASK,
358 if(unlikely(build_fail))
361 /* If we are in an open segment, try merging */
363 DBG(" - trying merge...\n");
364 /* We cannot merge if:
365 * - allocated dma_addr isn't contiguous to previous allocation
367 if (novmerge || (dma_addr != dma_next) ||
368 (outs->dma_length + s->length > max_seg_size)) {
369 /* Can't merge: create a new segment */
372 outs = sg_next(outs);
373 DBG(" can't merge, new segment.\n");
375 outs->dma_length += s->length;
376 DBG(" merged, new len: %ux\n", outs->dma_length);
381 /* This is a new segment, fill entries */
382 DBG(" - filling new segment.\n");
383 outs->dma_address = dma_addr;
384 outs->dma_length = slen;
387 /* Calculate next page pointer for contiguous check */
388 dma_next = dma_addr + slen;
390 DBG(" - dma next is: %lx\n", dma_next);
393 /* Flush/invalidate TLB caches if necessary */
394 if (ppc_md.tce_flush)
395 ppc_md.tce_flush(tbl);
397 spin_unlock_irqrestore(&(tbl->it_lock), flags);
399 DBG("mapped %d elements:\n", outcount);
401 /* For the sake of iommu_unmap_sg, we clear out the length in the
402 * next entry of the sglist if we didn't fill the list completely
404 if (outcount < incount) {
405 outs = sg_next(outs);
406 outs->dma_address = DMA_ERROR_CODE;
407 outs->dma_length = 0;
410 /* Make sure updates are seen by hardware */
416 for_each_sg(sglist, s, nelems, i) {
417 if (s->dma_length != 0) {
418 unsigned long vaddr, npages;
420 vaddr = s->dma_address & IOMMU_PAGE_MASK;
421 npages = iommu_num_pages(s->dma_address, s->dma_length);
422 __iommu_free(tbl, vaddr, npages);
423 s->dma_address = DMA_ERROR_CODE;
429 spin_unlock_irqrestore(&(tbl->it_lock), flags);
434 void iommu_unmap_sg(struct iommu_table *tbl, struct scatterlist *sglist,
435 int nelems, enum dma_data_direction direction,
436 struct dma_attrs *attrs)
438 struct scatterlist *sg;
441 BUG_ON(direction == DMA_NONE);
446 spin_lock_irqsave(&(tbl->it_lock), flags);
451 dma_addr_t dma_handle = sg->dma_address;
453 if (sg->dma_length == 0)
455 npages = iommu_num_pages(dma_handle, sg->dma_length);
456 __iommu_free(tbl, dma_handle, npages);
460 /* Flush/invalidate TLBs if necessary. As for iommu_free(), we
461 * do not do an mb() here, the affected platforms do not need it
464 if (ppc_md.tce_flush)
465 ppc_md.tce_flush(tbl);
467 spin_unlock_irqrestore(&(tbl->it_lock), flags);
471 * Build a iommu_table structure. This contains a bit map which
472 * is used to manage allocation of the tce space.
474 struct iommu_table *iommu_init_table(struct iommu_table *tbl, int nid)
477 static int welcomed = 0;
480 /* Set aside 1/4 of the table for large allocations. */
481 tbl->it_halfpoint = tbl->it_size * 3 / 4;
483 /* number of bytes needed for the bitmap */
484 sz = (tbl->it_size + 7) >> 3;
486 page = alloc_pages_node(nid, GFP_ATOMIC, get_order(sz));
488 panic("iommu_init_table: Can't allocate %ld bytes\n", sz);
489 tbl->it_map = page_address(page);
490 memset(tbl->it_map, 0, sz);
493 tbl->it_largehint = tbl->it_halfpoint;
494 spin_lock_init(&tbl->it_lock);
496 #ifdef CONFIG_CRASH_DUMP
497 if (ppc_md.tce_get) {
499 unsigned long tceval;
500 unsigned long tcecount = 0;
503 * Reserve the existing mappings left by the first kernel.
505 for (index = 0; index < tbl->it_size; index++) {
506 tceval = ppc_md.tce_get(tbl, index + tbl->it_offset);
508 * Freed TCE entry contains 0x7fffffffffffffff on JS20
510 if (tceval && (tceval != 0x7fffffffffffffffUL)) {
511 __set_bit(index, tbl->it_map);
515 if ((tbl->it_size - tcecount) < KDUMP_MIN_TCE_ENTRIES) {
516 printk(KERN_WARNING "TCE table is full; ");
517 printk(KERN_WARNING "freeing %d entries for the kdump boot\n",
518 KDUMP_MIN_TCE_ENTRIES);
519 for (index = tbl->it_size - KDUMP_MIN_TCE_ENTRIES;
520 index < tbl->it_size; index++)
521 __clear_bit(index, tbl->it_map);
525 /* Clear the hardware table in case firmware left allocations in it */
526 ppc_md.tce_free(tbl, tbl->it_offset, tbl->it_size);
530 printk(KERN_INFO "IOMMU table initialized, virtual merging %s\n",
531 novmerge ? "disabled" : "enabled");
538 void iommu_free_table(struct iommu_table *tbl, const char *node_name)
540 unsigned long bitmap_sz, i;
543 if (!tbl || !tbl->it_map) {
544 printk(KERN_ERR "%s: expected TCE map for %s\n", __func__,
549 /* verify that table contains no entries */
550 /* it_size is in entries, and we're examining 64 at a time */
551 for (i = 0; i < (tbl->it_size/64); i++) {
552 if (tbl->it_map[i] != 0) {
553 printk(KERN_WARNING "%s: Unexpected TCEs for %s\n",
554 __func__, node_name);
559 /* calculate bitmap size in bytes */
560 bitmap_sz = (tbl->it_size + 7) / 8;
563 order = get_order(bitmap_sz);
564 free_pages((unsigned long) tbl->it_map, order);
570 /* Creates TCEs for a user provided buffer. The user buffer must be
571 * contiguous real kernel storage (not vmalloc). The address of the buffer
572 * passed here is the kernel (virtual) address of the buffer. The buffer
573 * need not be page aligned, the dma_addr_t returned will point to the same
574 * byte within the page as vaddr.
576 dma_addr_t iommu_map_single(struct device *dev, struct iommu_table *tbl,
577 void *vaddr, size_t size, unsigned long mask,
578 enum dma_data_direction direction, struct dma_attrs *attrs)
580 dma_addr_t dma_handle = DMA_ERROR_CODE;
582 unsigned int npages, align;
584 BUG_ON(direction == DMA_NONE);
586 uaddr = (unsigned long)vaddr;
587 npages = iommu_num_pages(uaddr, size);
591 if (IOMMU_PAGE_SHIFT < PAGE_SHIFT && size >= PAGE_SIZE &&
592 ((unsigned long)vaddr & ~PAGE_MASK) == 0)
593 align = PAGE_SHIFT - IOMMU_PAGE_SHIFT;
595 dma_handle = iommu_alloc(dev, tbl, vaddr, npages, direction,
596 mask >> IOMMU_PAGE_SHIFT, align,
598 if (dma_handle == DMA_ERROR_CODE) {
599 if (printk_ratelimit()) {
600 printk(KERN_INFO "iommu_alloc failed, "
601 "tbl %p vaddr %p npages %d\n",
605 dma_handle |= (uaddr & ~IOMMU_PAGE_MASK);
611 void iommu_unmap_single(struct iommu_table *tbl, dma_addr_t dma_handle,
612 size_t size, enum dma_data_direction direction,
613 struct dma_attrs *attrs)
617 BUG_ON(direction == DMA_NONE);
620 npages = iommu_num_pages(dma_handle, size);
621 iommu_free(tbl, dma_handle, npages);
625 /* Allocates a contiguous real buffer and creates mappings over it.
626 * Returns the virtual address of the buffer and sets dma_handle
627 * to the dma address (mapping) of the first page.
629 void *iommu_alloc_coherent(struct device *dev, struct iommu_table *tbl,
630 size_t size, dma_addr_t *dma_handle,
631 unsigned long mask, gfp_t flag, int node)
636 unsigned int nio_pages, io_order;
639 size = PAGE_ALIGN(size);
640 order = get_order(size);
643 * Client asked for way too much space. This is checked later
644 * anyway. It is easier to debug here for the drivers than in
647 if (order >= IOMAP_MAX_ORDER) {
648 printk("iommu_alloc_consistent size too large: 0x%lx\n", size);
655 /* Alloc enough pages (and possibly more) */
656 page = alloc_pages_node(node, flag, order);
659 ret = page_address(page);
660 memset(ret, 0, size);
662 /* Set up tces to cover the allocated range */
663 nio_pages = size >> IOMMU_PAGE_SHIFT;
664 io_order = get_iommu_order(size);
665 mapping = iommu_alloc(dev, tbl, ret, nio_pages, DMA_BIDIRECTIONAL,
666 mask >> IOMMU_PAGE_SHIFT, io_order, NULL);
667 if (mapping == DMA_ERROR_CODE) {
668 free_pages((unsigned long)ret, order);
671 *dma_handle = mapping;
675 void iommu_free_coherent(struct iommu_table *tbl, size_t size,
676 void *vaddr, dma_addr_t dma_handle)
679 unsigned int nio_pages;
681 size = PAGE_ALIGN(size);
682 nio_pages = size >> IOMMU_PAGE_SHIFT;
683 iommu_free(tbl, dma_handle, nio_pages);
684 size = PAGE_ALIGN(size);
685 free_pages((unsigned long)vaddr, get_order(size));