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/init.h>
34 #include <linux/bitops.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 inline unsigned long iommu_num_pages(unsigned long vaddr,
57 npages = IOMMU_PAGE_ALIGN(vaddr + slen) - (vaddr & IOMMU_PAGE_MASK);
58 npages >>= IOMMU_PAGE_SHIFT;
63 static int __init setup_protect4gb(char *str)
65 if (strcmp(str, "on") == 0)
67 else if (strcmp(str, "off") == 0)
73 static int __init setup_iommu(char *str)
75 if (!strcmp(str, "novmerge"))
77 else if (!strcmp(str, "vmerge"))
82 __setup("protect4gb=", setup_protect4gb);
83 __setup("iommu=", setup_iommu);
85 static unsigned long iommu_range_alloc(struct iommu_table *tbl,
87 unsigned long *handle,
89 unsigned int align_order)
91 unsigned long n, end, i, start;
93 int largealloc = npages > 15;
95 unsigned long align_mask;
97 align_mask = 0xffffffffffffffffl >> (64 - align_order);
99 /* This allocator was derived from x86_64's bit string search */
102 if (unlikely(npages == 0)) {
103 if (printk_ratelimit())
105 return DMA_ERROR_CODE;
108 if (handle && *handle)
111 start = largealloc ? tbl->it_largehint : tbl->it_hint;
113 /* Use only half of the table for small allocs (15 pages or less) */
114 limit = largealloc ? tbl->it_size : tbl->it_halfpoint;
116 if (largealloc && start < tbl->it_halfpoint)
117 start = tbl->it_halfpoint;
119 /* The case below can happen if we have a small segment appended
120 * to a large, or when the previous alloc was at the very end of
121 * the available space. If so, go back to the initial start.
124 start = largealloc ? tbl->it_largehint : tbl->it_hint;
128 if (limit + tbl->it_offset > mask) {
129 limit = mask - tbl->it_offset + 1;
130 /* If we're constrained on address range, first try
131 * at the masked hint to avoid O(n) search complexity,
132 * but on second pass, start at 0.
134 if ((start & mask) >= limit || pass > 0)
140 n = find_next_zero_bit(tbl->it_map, limit, start);
142 /* Align allocation */
143 n = (n + align_mask) & ~align_mask;
147 if (unlikely(end >= limit)) {
148 if (likely(pass < 2)) {
149 /* First failure, just rescan the half of the table.
150 * Second failure, rescan the other half of the table.
152 start = (largealloc ^ pass) ? tbl->it_halfpoint : 0;
153 limit = pass ? tbl->it_size : limit;
157 /* Third failure, give up */
158 return DMA_ERROR_CODE;
162 for (i = n; i < end; i++)
163 if (test_bit(i, tbl->it_map)) {
168 for (i = n; i < end; i++)
169 __set_bit(i, tbl->it_map);
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 iommu_table *tbl, void *page,
189 unsigned int npages, enum dma_data_direction direction,
190 unsigned long mask, unsigned int align_order)
192 unsigned long entry, flags;
193 dma_addr_t ret = DMA_ERROR_CODE;
195 spin_lock_irqsave(&(tbl->it_lock), flags);
197 entry = iommu_range_alloc(tbl, npages, NULL, mask, align_order);
199 if (unlikely(entry == DMA_ERROR_CODE)) {
200 spin_unlock_irqrestore(&(tbl->it_lock), flags);
201 return DMA_ERROR_CODE;
204 entry += tbl->it_offset; /* Offset into real TCE table */
205 ret = entry << IOMMU_PAGE_SHIFT; /* Set the return dma address */
207 /* Put the TCEs in the HW table */
208 ppc_md.tce_build(tbl, entry, npages, (unsigned long)page & IOMMU_PAGE_MASK,
212 /* Flush/invalidate TLB caches if necessary */
213 if (ppc_md.tce_flush)
214 ppc_md.tce_flush(tbl);
216 spin_unlock_irqrestore(&(tbl->it_lock), flags);
218 /* Make sure updates are seen by hardware */
224 static void __iommu_free(struct iommu_table *tbl, dma_addr_t dma_addr,
227 unsigned long entry, free_entry;
230 entry = dma_addr >> IOMMU_PAGE_SHIFT;
231 free_entry = entry - tbl->it_offset;
233 if (((free_entry + npages) > tbl->it_size) ||
234 (entry < tbl->it_offset)) {
235 if (printk_ratelimit()) {
236 printk(KERN_INFO "iommu_free: invalid entry\n");
237 printk(KERN_INFO "\tentry = 0x%lx\n", entry);
238 printk(KERN_INFO "\tdma_addr = 0x%lx\n", (u64)dma_addr);
239 printk(KERN_INFO "\tTable = 0x%lx\n", (u64)tbl);
240 printk(KERN_INFO "\tbus# = 0x%lx\n", (u64)tbl->it_busno);
241 printk(KERN_INFO "\tsize = 0x%lx\n", (u64)tbl->it_size);
242 printk(KERN_INFO "\tstartOff = 0x%lx\n", (u64)tbl->it_offset);
243 printk(KERN_INFO "\tindex = 0x%lx\n", (u64)tbl->it_index);
249 ppc_md.tce_free(tbl, entry, npages);
251 for (i = 0; i < npages; i++)
252 __clear_bit(free_entry+i, tbl->it_map);
255 static void iommu_free(struct iommu_table *tbl, dma_addr_t dma_addr,
260 spin_lock_irqsave(&(tbl->it_lock), flags);
262 __iommu_free(tbl, dma_addr, npages);
264 /* Make sure TLB cache is flushed if the HW needs it. We do
265 * not do an mb() here on purpose, it is not needed on any of
266 * the current platforms.
268 if (ppc_md.tce_flush)
269 ppc_md.tce_flush(tbl);
271 spin_unlock_irqrestore(&(tbl->it_lock), flags);
274 int iommu_map_sg(struct iommu_table *tbl, struct scatterlist *sglist,
275 int nelems, unsigned long mask,
276 enum dma_data_direction direction)
278 dma_addr_t dma_next = 0, dma_addr;
280 struct scatterlist *s, *outs, *segstart;
281 int outcount, incount;
282 unsigned long handle;
284 BUG_ON(direction == DMA_NONE);
286 if ((nelems == 0) || !tbl)
289 outs = s = segstart = &sglist[0];
294 /* Init first segment length for backout at failure */
295 outs->dma_length = 0;
297 DBG("sg mapping %d elements:\n", nelems);
299 spin_lock_irqsave(&(tbl->it_lock), flags);
301 for (s = outs; nelems; nelems--, s++) {
302 unsigned long vaddr, npages, entry, slen;
310 /* Allocate iommu entries for that segment */
311 vaddr = (unsigned long)page_address(s->page) + s->offset;
312 npages = iommu_num_pages(vaddr, slen);
313 entry = iommu_range_alloc(tbl, npages, &handle, mask >> IOMMU_PAGE_SHIFT, 0);
315 DBG(" - vaddr: %lx, size: %lx\n", vaddr, slen);
318 if (unlikely(entry == DMA_ERROR_CODE)) {
319 if (printk_ratelimit())
320 printk(KERN_INFO "iommu_alloc failed, tbl %p vaddr %lx"
321 " npages %lx\n", tbl, vaddr, npages);
325 /* Convert entry to a dma_addr_t */
326 entry += tbl->it_offset;
327 dma_addr = entry << IOMMU_PAGE_SHIFT;
328 dma_addr |= (s->offset & ~IOMMU_PAGE_MASK);
330 DBG(" - %lu pages, entry: %lx, dma_addr: %lx\n",
331 npages, entry, dma_addr);
333 /* Insert into HW table */
334 ppc_md.tce_build(tbl, entry, npages, vaddr & IOMMU_PAGE_MASK, direction);
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 /* Can't merge: create a new segment */
346 DBG(" can't merge, new segment.\n");
348 outs->dma_length += s->length;
349 DBG(" merged, new len: %ux\n", outs->dma_length);
354 /* This is a new segment, fill entries */
355 DBG(" - filling new segment.\n");
356 outs->dma_address = dma_addr;
357 outs->dma_length = slen;
360 /* Calculate next page pointer for contiguous check */
361 dma_next = dma_addr + slen;
363 DBG(" - dma next is: %lx\n", dma_next);
366 /* Flush/invalidate TLB caches if necessary */
367 if (ppc_md.tce_flush)
368 ppc_md.tce_flush(tbl);
370 spin_unlock_irqrestore(&(tbl->it_lock), flags);
372 DBG("mapped %d elements:\n", outcount);
374 /* For the sake of iommu_unmap_sg, we clear out the length in the
375 * next entry of the sglist if we didn't fill the list completely
377 if (outcount < incount) {
379 outs->dma_address = DMA_ERROR_CODE;
380 outs->dma_length = 0;
383 /* Make sure updates are seen by hardware */
389 for (s = &sglist[0]; s <= outs; s++) {
390 if (s->dma_length != 0) {
391 unsigned long vaddr, npages;
393 vaddr = s->dma_address & IOMMU_PAGE_MASK;
394 npages = iommu_num_pages(s->dma_address, s->dma_length);
395 __iommu_free(tbl, vaddr, npages);
396 s->dma_address = DMA_ERROR_CODE;
400 spin_unlock_irqrestore(&(tbl->it_lock), flags);
405 void iommu_unmap_sg(struct iommu_table *tbl, struct scatterlist *sglist,
406 int nelems, enum dma_data_direction direction)
410 BUG_ON(direction == DMA_NONE);
415 spin_lock_irqsave(&(tbl->it_lock), flags);
419 dma_addr_t dma_handle = sglist->dma_address;
421 if (sglist->dma_length == 0)
423 npages = iommu_num_pages(dma_handle,sglist->dma_length);
424 __iommu_free(tbl, dma_handle, npages);
428 /* Flush/invalidate TLBs if necessary. As for iommu_free(), we
429 * do not do an mb() here, the affected platforms do not need it
432 if (ppc_md.tce_flush)
433 ppc_md.tce_flush(tbl);
435 spin_unlock_irqrestore(&(tbl->it_lock), flags);
439 * Build a iommu_table structure. This contains a bit map which
440 * is used to manage allocation of the tce space.
442 struct iommu_table *iommu_init_table(struct iommu_table *tbl, int nid)
445 unsigned long start_index, end_index;
446 unsigned long entries_per_4g;
448 static int welcomed = 0;
451 /* Set aside 1/4 of the table for large allocations. */
452 tbl->it_halfpoint = tbl->it_size * 3 / 4;
454 /* number of bytes needed for the bitmap */
455 sz = (tbl->it_size + 7) >> 3;
457 page = alloc_pages_node(nid, GFP_ATOMIC, get_order(sz));
459 panic("iommu_init_table: Can't allocate %ld bytes\n", sz);
460 tbl->it_map = page_address(page);
461 memset(tbl->it_map, 0, sz);
464 tbl->it_largehint = tbl->it_halfpoint;
465 spin_lock_init(&tbl->it_lock);
467 #ifdef CONFIG_CRASH_DUMP
468 if (ppc_md.tce_get) {
469 unsigned long tceval;
470 unsigned long tcecount = 0;
473 * 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);
485 if ((tbl->it_size - tcecount) < KDUMP_MIN_TCE_ENTRIES) {
486 printk(KERN_WARNING "TCE table is full; ");
487 printk(KERN_WARNING "freeing %d entries for the kdump boot\n",
488 KDUMP_MIN_TCE_ENTRIES);
489 for (index = tbl->it_size - KDUMP_MIN_TCE_ENTRIES;
490 index < tbl->it_size; index++)
491 __clear_bit(index, tbl->it_map);
495 /* Clear the hardware table in case firmware left allocations in it */
496 ppc_md.tce_free(tbl, tbl->it_offset, tbl->it_size);
500 * DMA cannot cross 4 GB boundary. Mark last entry of each 4
501 * GB chunk as reserved.
504 entries_per_4g = 0x100000000l >> IOMMU_PAGE_SHIFT;
506 /* Mark the last bit before a 4GB boundary as used */
507 start_index = tbl->it_offset | (entries_per_4g - 1);
508 start_index -= tbl->it_offset;
510 end_index = tbl->it_size;
512 for (index = start_index; index < end_index - 1; index += entries_per_4g)
513 __set_bit(index, tbl->it_map);
517 printk(KERN_INFO "IOMMU table initialized, virtual merging %s\n",
518 novmerge ? "disabled" : "enabled");
525 void iommu_free_table(struct device_node *dn)
527 struct pci_dn *pdn = dn->data;
528 struct iommu_table *tbl = pdn->iommu_table;
529 unsigned long bitmap_sz, i;
532 if (!tbl || !tbl->it_map) {
533 printk(KERN_ERR "%s: expected TCE map for %s\n", __FUNCTION__,
538 /* verify that table contains no entries */
539 /* it_size is in entries, and we're examining 64 at a time */
540 for (i = 0; i < (tbl->it_size/64); i++) {
541 if (tbl->it_map[i] != 0) {
542 printk(KERN_WARNING "%s: Unexpected TCEs for %s\n",
543 __FUNCTION__, dn->full_name);
548 /* calculate bitmap size in bytes */
549 bitmap_sz = (tbl->it_size + 7) / 8;
552 order = get_order(bitmap_sz);
553 free_pages((unsigned long) tbl->it_map, order);
559 /* Creates TCEs for a user provided buffer. The user buffer must be
560 * contiguous real kernel storage (not vmalloc). The address of the buffer
561 * passed here is the kernel (virtual) address of the buffer. The buffer
562 * need not be page aligned, the dma_addr_t returned will point to the same
563 * byte within the page as vaddr.
565 dma_addr_t iommu_map_single(struct iommu_table *tbl, void *vaddr,
566 size_t size, unsigned long mask,
567 enum dma_data_direction direction)
569 dma_addr_t dma_handle = DMA_ERROR_CODE;
573 BUG_ON(direction == DMA_NONE);
575 uaddr = (unsigned long)vaddr;
576 npages = iommu_num_pages(uaddr, size);
579 dma_handle = iommu_alloc(tbl, vaddr, npages, direction,
580 mask >> IOMMU_PAGE_SHIFT, 0);
581 if (dma_handle == DMA_ERROR_CODE) {
582 if (printk_ratelimit()) {
583 printk(KERN_INFO "iommu_alloc failed, "
584 "tbl %p vaddr %p npages %d\n",
588 dma_handle |= (uaddr & ~IOMMU_PAGE_MASK);
594 void iommu_unmap_single(struct iommu_table *tbl, dma_addr_t dma_handle,
595 size_t size, enum dma_data_direction direction)
599 BUG_ON(direction == DMA_NONE);
602 npages = iommu_num_pages(dma_handle, size);
603 iommu_free(tbl, dma_handle, npages);
607 /* Allocates a contiguous real buffer and creates mappings over it.
608 * Returns the virtual address of the buffer and sets dma_handle
609 * to the dma address (mapping) of the first page.
611 void *iommu_alloc_coherent(struct iommu_table *tbl, size_t size,
612 dma_addr_t *dma_handle, unsigned long mask, gfp_t flag, int node)
617 unsigned int nio_pages, io_order;
620 size = PAGE_ALIGN(size);
621 order = get_order(size);
624 * Client asked for way too much space. This is checked later
625 * anyway. It is easier to debug here for the drivers than in
628 if (order >= IOMAP_MAX_ORDER) {
629 printk("iommu_alloc_consistent size too large: 0x%lx\n", size);
636 /* Alloc enough pages (and possibly more) */
637 page = alloc_pages_node(node, flag, order);
640 ret = page_address(page);
641 memset(ret, 0, size);
643 /* Set up tces to cover the allocated range */
644 nio_pages = size >> IOMMU_PAGE_SHIFT;
645 io_order = get_iommu_order(size);
646 mapping = iommu_alloc(tbl, ret, nio_pages, DMA_BIDIRECTIONAL,
647 mask >> IOMMU_PAGE_SHIFT, io_order);
648 if (mapping == DMA_ERROR_CODE) {
649 free_pages((unsigned long)ret, order);
652 *dma_handle = mapping;
656 void iommu_free_coherent(struct iommu_table *tbl, size_t size,
657 void *vaddr, dma_addr_t dma_handle)
660 unsigned int nio_pages;
662 size = PAGE_ALIGN(size);
663 nio_pages = size >> IOMMU_PAGE_SHIFT;
664 iommu_free(tbl, dma_handle, nio_pages);
665 size = PAGE_ALIGN(size);
666 free_pages((unsigned long)vaddr, get_order(size));