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>
38 #include <asm/iommu.h>
39 #include <asm/pci-bridge.h>
40 #include <asm/machdep.h>
41 #include <asm/kdump.h>
42 #include <asm/fadump.h>
48 static void __iommu_free(struct iommu_table *, dma_addr_t, unsigned int);
50 static int __init setup_iommu(char *str)
52 if (!strcmp(str, "novmerge"))
54 else if (!strcmp(str, "vmerge"))
59 __setup("iommu=", setup_iommu);
61 static unsigned long iommu_range_alloc(struct device *dev,
62 struct iommu_table *tbl,
64 unsigned long *handle,
66 unsigned int align_order)
68 unsigned long n, end, start;
70 int largealloc = npages > 15;
72 unsigned long align_mask;
73 unsigned long boundary_size;
75 align_mask = 0xffffffffffffffffl >> (64 - align_order);
77 /* This allocator was derived from x86_64's bit string search */
80 if (unlikely(npages == 0)) {
81 if (printk_ratelimit())
83 return DMA_ERROR_CODE;
86 if (handle && *handle)
89 start = largealloc ? tbl->it_largehint : tbl->it_hint;
91 /* Use only half of the table for small allocs (15 pages or less) */
92 limit = largealloc ? tbl->it_size : tbl->it_halfpoint;
94 if (largealloc && start < tbl->it_halfpoint)
95 start = tbl->it_halfpoint;
97 /* The case below can happen if we have a small segment appended
98 * to a large, or when the previous alloc was at the very end of
99 * the available space. If so, go back to the initial start.
102 start = largealloc ? tbl->it_largehint : tbl->it_hint;
106 if (limit + tbl->it_offset > mask) {
107 limit = mask - tbl->it_offset + 1;
108 /* If we're constrained on address range, first try
109 * at the masked hint to avoid O(n) search complexity,
110 * but on second pass, start at 0.
112 if ((start & mask) >= limit || pass > 0)
119 boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
120 1 << IOMMU_PAGE_SHIFT);
122 boundary_size = ALIGN(1UL << 32, 1 << IOMMU_PAGE_SHIFT);
123 /* 4GB boundary for iseries_hv_alloc and iseries_hv_map */
125 n = iommu_area_alloc(tbl->it_map, limit, start, npages,
126 tbl->it_offset, boundary_size >> IOMMU_PAGE_SHIFT,
129 if (likely(pass < 2)) {
130 /* First failure, just rescan the half of the table.
131 * Second failure, rescan the other half of the table.
133 start = (largealloc ^ pass) ? tbl->it_halfpoint : 0;
134 limit = pass ? tbl->it_size : limit;
138 /* Third failure, give up */
139 return DMA_ERROR_CODE;
145 /* Bump the hint to a new block for small allocs. */
147 /* Don't bump to new block to avoid fragmentation */
148 tbl->it_largehint = end;
150 /* Overflow will be taken care of at the next allocation */
151 tbl->it_hint = (end + tbl->it_blocksize - 1) &
152 ~(tbl->it_blocksize - 1);
155 /* Update handle for SG allocations */
162 static dma_addr_t iommu_alloc(struct device *dev, struct iommu_table *tbl,
163 void *page, unsigned int npages,
164 enum dma_data_direction direction,
165 unsigned long mask, unsigned int align_order,
166 struct dma_attrs *attrs)
168 unsigned long entry, flags;
169 dma_addr_t ret = DMA_ERROR_CODE;
172 spin_lock_irqsave(&(tbl->it_lock), flags);
174 entry = iommu_range_alloc(dev, tbl, npages, NULL, mask, align_order);
176 if (unlikely(entry == DMA_ERROR_CODE)) {
177 spin_unlock_irqrestore(&(tbl->it_lock), flags);
178 return DMA_ERROR_CODE;
181 entry += tbl->it_offset; /* Offset into real TCE table */
182 ret = entry << IOMMU_PAGE_SHIFT; /* Set the return dma address */
184 /* Put the TCEs in the HW table */
185 build_fail = ppc_md.tce_build(tbl, entry, npages,
186 (unsigned long)page & IOMMU_PAGE_MASK,
189 /* ppc_md.tce_build() only returns non-zero for transient errors.
190 * Clean up the table bitmap in this case and return
191 * DMA_ERROR_CODE. For all other errors the functionality is
194 if (unlikely(build_fail)) {
195 __iommu_free(tbl, ret, npages);
197 spin_unlock_irqrestore(&(tbl->it_lock), flags);
198 return DMA_ERROR_CODE;
201 /* Flush/invalidate TLB caches if necessary */
202 if (ppc_md.tce_flush)
203 ppc_md.tce_flush(tbl);
205 spin_unlock_irqrestore(&(tbl->it_lock), flags);
207 /* Make sure updates are seen by hardware */
213 static void __iommu_free(struct iommu_table *tbl, dma_addr_t dma_addr,
216 unsigned long entry, free_entry;
218 entry = dma_addr >> IOMMU_PAGE_SHIFT;
219 free_entry = entry - tbl->it_offset;
221 if (((free_entry + npages) > tbl->it_size) ||
222 (entry < tbl->it_offset)) {
223 if (printk_ratelimit()) {
224 printk(KERN_INFO "iommu_free: invalid entry\n");
225 printk(KERN_INFO "\tentry = 0x%lx\n", entry);
226 printk(KERN_INFO "\tdma_addr = 0x%llx\n", (u64)dma_addr);
227 printk(KERN_INFO "\tTable = 0x%llx\n", (u64)tbl);
228 printk(KERN_INFO "\tbus# = 0x%llx\n", (u64)tbl->it_busno);
229 printk(KERN_INFO "\tsize = 0x%llx\n", (u64)tbl->it_size);
230 printk(KERN_INFO "\tstartOff = 0x%llx\n", (u64)tbl->it_offset);
231 printk(KERN_INFO "\tindex = 0x%llx\n", (u64)tbl->it_index);
237 ppc_md.tce_free(tbl, entry, npages);
238 bitmap_clear(tbl->it_map, free_entry, npages);
241 static void iommu_free(struct iommu_table *tbl, dma_addr_t dma_addr,
246 spin_lock_irqsave(&(tbl->it_lock), flags);
248 __iommu_free(tbl, dma_addr, npages);
250 /* Make sure TLB cache is flushed if the HW needs it. We do
251 * not do an mb() here on purpose, it is not needed on any of
252 * the current platforms.
254 if (ppc_md.tce_flush)
255 ppc_md.tce_flush(tbl);
257 spin_unlock_irqrestore(&(tbl->it_lock), flags);
260 int iommu_map_sg(struct device *dev, struct iommu_table *tbl,
261 struct scatterlist *sglist, int nelems,
262 unsigned long mask, enum dma_data_direction direction,
263 struct dma_attrs *attrs)
265 dma_addr_t dma_next = 0, dma_addr;
267 struct scatterlist *s, *outs, *segstart;
268 int outcount, incount, i, build_fail = 0;
270 unsigned long handle;
271 unsigned int max_seg_size;
273 BUG_ON(direction == DMA_NONE);
275 if ((nelems == 0) || !tbl)
278 outs = s = segstart = &sglist[0];
283 /* Init first segment length for backout at failure */
284 outs->dma_length = 0;
286 DBG("sg mapping %d elements:\n", nelems);
288 spin_lock_irqsave(&(tbl->it_lock), flags);
290 max_seg_size = dma_get_max_seg_size(dev);
291 for_each_sg(sglist, s, nelems, i) {
292 unsigned long vaddr, npages, entry, slen;
300 /* Allocate iommu entries for that segment */
301 vaddr = (unsigned long) sg_virt(s);
302 npages = iommu_num_pages(vaddr, slen, IOMMU_PAGE_SIZE);
304 if (IOMMU_PAGE_SHIFT < PAGE_SHIFT && slen >= PAGE_SIZE &&
305 (vaddr & ~PAGE_MASK) == 0)
306 align = PAGE_SHIFT - IOMMU_PAGE_SHIFT;
307 entry = iommu_range_alloc(dev, tbl, npages, &handle,
308 mask >> IOMMU_PAGE_SHIFT, align);
310 DBG(" - vaddr: %lx, size: %lx\n", vaddr, slen);
313 if (unlikely(entry == DMA_ERROR_CODE)) {
314 if (printk_ratelimit())
315 dev_info(dev, "iommu_alloc failed, tbl %p "
316 "vaddr %lx npages %lu\n", tbl, vaddr,
321 /* Convert entry to a dma_addr_t */
322 entry += tbl->it_offset;
323 dma_addr = entry << IOMMU_PAGE_SHIFT;
324 dma_addr |= (s->offset & ~IOMMU_PAGE_MASK);
326 DBG(" - %lu pages, entry: %lx, dma_addr: %lx\n",
327 npages, entry, dma_addr);
329 /* Insert into HW table */
330 build_fail = ppc_md.tce_build(tbl, entry, npages,
331 vaddr & IOMMU_PAGE_MASK,
333 if(unlikely(build_fail))
336 /* If we are in an open segment, try merging */
338 DBG(" - trying merge...\n");
339 /* We cannot merge if:
340 * - allocated dma_addr isn't contiguous to previous allocation
342 if (novmerge || (dma_addr != dma_next) ||
343 (outs->dma_length + s->length > max_seg_size)) {
344 /* Can't merge: create a new segment */
347 outs = sg_next(outs);
348 DBG(" can't merge, new segment.\n");
350 outs->dma_length += s->length;
351 DBG(" merged, new len: %ux\n", outs->dma_length);
356 /* This is a new segment, fill entries */
357 DBG(" - filling new segment.\n");
358 outs->dma_address = dma_addr;
359 outs->dma_length = slen;
362 /* Calculate next page pointer for contiguous check */
363 dma_next = dma_addr + slen;
365 DBG(" - dma next is: %lx\n", dma_next);
368 /* Flush/invalidate TLB caches if necessary */
369 if (ppc_md.tce_flush)
370 ppc_md.tce_flush(tbl);
372 spin_unlock_irqrestore(&(tbl->it_lock), flags);
374 DBG("mapped %d elements:\n", outcount);
376 /* For the sake of iommu_unmap_sg, we clear out the length in the
377 * next entry of the sglist if we didn't fill the list completely
379 if (outcount < incount) {
380 outs = sg_next(outs);
381 outs->dma_address = DMA_ERROR_CODE;
382 outs->dma_length = 0;
385 /* Make sure updates are seen by hardware */
391 for_each_sg(sglist, s, nelems, i) {
392 if (s->dma_length != 0) {
393 unsigned long vaddr, npages;
395 vaddr = s->dma_address & IOMMU_PAGE_MASK;
396 npages = iommu_num_pages(s->dma_address, s->dma_length,
398 __iommu_free(tbl, vaddr, npages);
399 s->dma_address = DMA_ERROR_CODE;
405 spin_unlock_irqrestore(&(tbl->it_lock), flags);
410 void iommu_unmap_sg(struct iommu_table *tbl, struct scatterlist *sglist,
411 int nelems, enum dma_data_direction direction,
412 struct dma_attrs *attrs)
414 struct scatterlist *sg;
417 BUG_ON(direction == DMA_NONE);
422 spin_lock_irqsave(&(tbl->it_lock), flags);
427 dma_addr_t dma_handle = sg->dma_address;
429 if (sg->dma_length == 0)
431 npages = iommu_num_pages(dma_handle, sg->dma_length,
433 __iommu_free(tbl, dma_handle, npages);
437 /* Flush/invalidate TLBs if necessary. As for iommu_free(), we
438 * do not do an mb() here, the affected platforms do not need it
441 if (ppc_md.tce_flush)
442 ppc_md.tce_flush(tbl);
444 spin_unlock_irqrestore(&(tbl->it_lock), flags);
447 static void iommu_table_clear(struct iommu_table *tbl)
450 * In case of firmware assisted dump system goes through clean
451 * reboot process at the time of system crash. Hence it's safe to
452 * clear the TCE entries if firmware assisted dump is active.
454 if (!is_kdump_kernel() || is_fadump_active()) {
455 /* Clear the table in case firmware left allocations in it */
456 ppc_md.tce_free(tbl, tbl->it_offset, tbl->it_size);
460 #ifdef CONFIG_CRASH_DUMP
461 if (ppc_md.tce_get) {
462 unsigned long index, tceval, tcecount = 0;
464 /* Reserve the existing mappings left by the first kernel. */
465 for (index = 0; index < tbl->it_size; index++) {
466 tceval = ppc_md.tce_get(tbl, index + tbl->it_offset);
468 * Freed TCE entry contains 0x7fffffffffffffff on JS20
470 if (tceval && (tceval != 0x7fffffffffffffffUL)) {
471 __set_bit(index, tbl->it_map);
476 if ((tbl->it_size - tcecount) < KDUMP_MIN_TCE_ENTRIES) {
477 printk(KERN_WARNING "TCE table is full; freeing ");
478 printk(KERN_WARNING "%d entries for the kdump boot\n",
479 KDUMP_MIN_TCE_ENTRIES);
480 for (index = tbl->it_size - KDUMP_MIN_TCE_ENTRIES;
481 index < tbl->it_size; index++)
482 __clear_bit(index, tbl->it_map);
489 * Build a iommu_table structure. This contains a bit map which
490 * is used to manage allocation of the tce space.
492 struct iommu_table *iommu_init_table(struct iommu_table *tbl, int nid)
495 static int welcomed = 0;
498 /* Set aside 1/4 of the table for large allocations. */
499 tbl->it_halfpoint = tbl->it_size * 3 / 4;
501 /* number of bytes needed for the bitmap */
502 sz = (tbl->it_size + 7) >> 3;
504 page = alloc_pages_node(nid, GFP_ATOMIC, get_order(sz));
506 panic("iommu_init_table: Can't allocate %ld bytes\n", sz);
507 tbl->it_map = page_address(page);
508 memset(tbl->it_map, 0, sz);
511 * Reserve page 0 so it will not be used for any mappings.
512 * This avoids buggy drivers that consider page 0 to be invalid
513 * to crash the machine or even lose data.
515 if (tbl->it_offset == 0)
516 set_bit(0, tbl->it_map);
519 tbl->it_largehint = tbl->it_halfpoint;
520 spin_lock_init(&tbl->it_lock);
522 iommu_table_clear(tbl);
525 printk(KERN_INFO "IOMMU table initialized, virtual merging %s\n",
526 novmerge ? "disabled" : "enabled");
533 void iommu_free_table(struct iommu_table *tbl, const char *node_name)
535 unsigned long bitmap_sz, i;
538 if (!tbl || !tbl->it_map) {
539 printk(KERN_ERR "%s: expected TCE map for %s\n", __func__,
544 /* verify that table contains no entries */
545 /* it_size is in entries, and we're examining 64 at a time */
546 for (i = 0; i < (tbl->it_size/64); i++) {
547 if (tbl->it_map[i] != 0) {
548 printk(KERN_WARNING "%s: Unexpected TCEs for %s\n",
549 __func__, node_name);
554 /* calculate bitmap size in bytes */
555 bitmap_sz = (tbl->it_size + 7) / 8;
558 order = get_order(bitmap_sz);
559 free_pages((unsigned long) tbl->it_map, order);
565 /* Creates TCEs for a user provided buffer. The user buffer must be
566 * contiguous real kernel storage (not vmalloc). The address passed here
567 * comprises a page address and offset into that page. The dma_addr_t
568 * returned will point to the same byte within the page as was passed in.
570 dma_addr_t iommu_map_page(struct device *dev, struct iommu_table *tbl,
571 struct page *page, unsigned long offset, size_t size,
572 unsigned long mask, enum dma_data_direction direction,
573 struct dma_attrs *attrs)
575 dma_addr_t dma_handle = DMA_ERROR_CODE;
578 unsigned int npages, align;
580 BUG_ON(direction == DMA_NONE);
582 vaddr = page_address(page) + offset;
583 uaddr = (unsigned long)vaddr;
584 npages = iommu_num_pages(uaddr, size, IOMMU_PAGE_SIZE);
588 if (IOMMU_PAGE_SHIFT < PAGE_SHIFT && size >= PAGE_SIZE &&
589 ((unsigned long)vaddr & ~PAGE_MASK) == 0)
590 align = PAGE_SHIFT - IOMMU_PAGE_SHIFT;
592 dma_handle = iommu_alloc(dev, tbl, vaddr, npages, direction,
593 mask >> IOMMU_PAGE_SHIFT, align,
595 if (dma_handle == DMA_ERROR_CODE) {
596 if (printk_ratelimit()) {
597 dev_info(dev, "iommu_alloc failed, tbl %p "
598 "vaddr %p npages %d\n", tbl, vaddr,
602 dma_handle |= (uaddr & ~IOMMU_PAGE_MASK);
608 void iommu_unmap_page(struct iommu_table *tbl, dma_addr_t dma_handle,
609 size_t size, enum dma_data_direction direction,
610 struct dma_attrs *attrs)
614 BUG_ON(direction == DMA_NONE);
617 npages = iommu_num_pages(dma_handle, size, IOMMU_PAGE_SIZE);
618 iommu_free(tbl, dma_handle, npages);
622 /* Allocates a contiguous real buffer and creates mappings over it.
623 * Returns the virtual address of the buffer and sets dma_handle
624 * to the dma address (mapping) of the first page.
626 void *iommu_alloc_coherent(struct device *dev, struct iommu_table *tbl,
627 size_t size, dma_addr_t *dma_handle,
628 unsigned long mask, gfp_t flag, int node)
633 unsigned int nio_pages, io_order;
636 size = PAGE_ALIGN(size);
637 order = get_order(size);
640 * Client asked for way too much space. This is checked later
641 * anyway. It is easier to debug here for the drivers than in
644 if (order >= IOMAP_MAX_ORDER) {
645 dev_info(dev, "iommu_alloc_consistent size too large: 0x%lx\n",
653 /* Alloc enough pages (and possibly more) */
654 page = alloc_pages_node(node, flag, order);
657 ret = page_address(page);
658 memset(ret, 0, size);
660 /* Set up tces to cover the allocated range */
661 nio_pages = size >> IOMMU_PAGE_SHIFT;
662 io_order = get_iommu_order(size);
663 mapping = iommu_alloc(dev, tbl, ret, nio_pages, DMA_BIDIRECTIONAL,
664 mask >> IOMMU_PAGE_SHIFT, io_order, NULL);
665 if (mapping == DMA_ERROR_CODE) {
666 free_pages((unsigned long)ret, order);
669 *dma_handle = mapping;
673 void iommu_free_coherent(struct iommu_table *tbl, size_t size,
674 void *vaddr, dma_addr_t dma_handle)
677 unsigned int nio_pages;
679 size = PAGE_ALIGN(size);
680 nio_pages = size >> IOMMU_PAGE_SHIFT;
681 iommu_free(tbl, dma_handle, nio_pages);
682 size = PAGE_ALIGN(size);
683 free_pages((unsigned long)vaddr, get_order(size));