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>
36 #include <asm/iommu.h>
37 #include <asm/pci-bridge.h>
38 #include <asm/machdep.h>
39 #include <asm/kdump.h>
43 #ifdef CONFIG_IOMMU_VMERGE
44 static int novmerge = 0;
46 static int novmerge = 1;
49 static int protect4gb = 1;
51 static inline unsigned long iommu_num_pages(unsigned long vaddr,
56 npages = IOMMU_PAGE_ALIGN(vaddr + slen) - (vaddr & IOMMU_PAGE_MASK);
57 npages >>= IOMMU_PAGE_SHIFT;
62 static int __init setup_protect4gb(char *str)
64 if (strcmp(str, "on") == 0)
66 else if (strcmp(str, "off") == 0)
72 static int __init setup_iommu(char *str)
74 if (!strcmp(str, "novmerge"))
76 else if (!strcmp(str, "vmerge"))
81 __setup("protect4gb=", setup_protect4gb);
82 __setup("iommu=", setup_iommu);
84 static unsigned long iommu_range_alloc(struct iommu_table *tbl,
86 unsigned long *handle,
88 unsigned int align_order)
90 unsigned long n, end, i, start;
92 int largealloc = npages > 15;
94 unsigned long align_mask;
96 align_mask = 0xffffffffffffffffl >> (64 - align_order);
98 /* This allocator was derived from x86_64's bit string search */
101 if (unlikely(npages == 0)) {
102 if (printk_ratelimit())
104 return DMA_ERROR_CODE;
107 if (handle && *handle)
110 start = largealloc ? tbl->it_largehint : tbl->it_hint;
112 /* Use only half of the table for small allocs (15 pages or less) */
113 limit = largealloc ? tbl->it_size : tbl->it_halfpoint;
115 if (largealloc && start < tbl->it_halfpoint)
116 start = tbl->it_halfpoint;
118 /* The case below can happen if we have a small segment appended
119 * to a large, or when the previous alloc was at the very end of
120 * the available space. If so, go back to the initial start.
123 start = largealloc ? tbl->it_largehint : tbl->it_hint;
127 if (limit + tbl->it_offset > mask) {
128 limit = mask - tbl->it_offset + 1;
129 /* If we're constrained on address range, first try
130 * at the masked hint to avoid O(n) search complexity,
131 * but on second pass, start at 0.
133 if ((start & mask) >= limit || pass > 0)
139 n = find_next_zero_bit(tbl->it_map, limit, start);
141 /* Align allocation */
142 n = (n + align_mask) & ~align_mask;
146 if (unlikely(end >= limit)) {
147 if (likely(pass < 2)) {
148 /* First failure, just rescan the half of the table.
149 * Second failure, rescan the other half of the table.
151 start = (largealloc ^ pass) ? tbl->it_halfpoint : 0;
152 limit = pass ? tbl->it_size : limit;
156 /* Third failure, give up */
157 return DMA_ERROR_CODE;
161 for (i = n; i < end; i++)
162 if (test_bit(i, tbl->it_map)) {
167 for (i = n; i < end; i++)
168 __set_bit(i, tbl->it_map);
170 /* Bump the hint to a new block for small allocs. */
172 /* Don't bump to new block to avoid fragmentation */
173 tbl->it_largehint = end;
175 /* Overflow will be taken care of at the next allocation */
176 tbl->it_hint = (end + tbl->it_blocksize - 1) &
177 ~(tbl->it_blocksize - 1);
180 /* Update handle for SG allocations */
187 static dma_addr_t iommu_alloc(struct iommu_table *tbl, void *page,
188 unsigned int npages, enum dma_data_direction direction,
189 unsigned long mask, unsigned int align_order)
191 unsigned long entry, flags;
192 dma_addr_t ret = DMA_ERROR_CODE;
194 spin_lock_irqsave(&(tbl->it_lock), flags);
196 entry = iommu_range_alloc(tbl, npages, NULL, mask, align_order);
198 if (unlikely(entry == DMA_ERROR_CODE)) {
199 spin_unlock_irqrestore(&(tbl->it_lock), flags);
200 return DMA_ERROR_CODE;
203 entry += tbl->it_offset; /* Offset into real TCE table */
204 ret = entry << IOMMU_PAGE_SHIFT; /* Set the return dma address */
206 /* Put the TCEs in the HW table */
207 ppc_md.tce_build(tbl, entry, npages, (unsigned long)page & IOMMU_PAGE_MASK,
211 /* Flush/invalidate TLB caches if necessary */
212 if (ppc_md.tce_flush)
213 ppc_md.tce_flush(tbl);
215 spin_unlock_irqrestore(&(tbl->it_lock), flags);
217 /* Make sure updates are seen by hardware */
223 static void __iommu_free(struct iommu_table *tbl, dma_addr_t dma_addr,
226 unsigned long entry, free_entry;
229 entry = dma_addr >> IOMMU_PAGE_SHIFT;
230 free_entry = entry - tbl->it_offset;
232 if (((free_entry + npages) > tbl->it_size) ||
233 (entry < tbl->it_offset)) {
234 if (printk_ratelimit()) {
235 printk(KERN_INFO "iommu_free: invalid entry\n");
236 printk(KERN_INFO "\tentry = 0x%lx\n", entry);
237 printk(KERN_INFO "\tdma_addr = 0x%lx\n", (u64)dma_addr);
238 printk(KERN_INFO "\tTable = 0x%lx\n", (u64)tbl);
239 printk(KERN_INFO "\tbus# = 0x%lx\n", (u64)tbl->it_busno);
240 printk(KERN_INFO "\tsize = 0x%lx\n", (u64)tbl->it_size);
241 printk(KERN_INFO "\tstartOff = 0x%lx\n", (u64)tbl->it_offset);
242 printk(KERN_INFO "\tindex = 0x%lx\n", (u64)tbl->it_index);
248 ppc_md.tce_free(tbl, entry, npages);
250 for (i = 0; i < npages; i++)
251 __clear_bit(free_entry+i, tbl->it_map);
254 static void iommu_free(struct iommu_table *tbl, dma_addr_t dma_addr,
259 spin_lock_irqsave(&(tbl->it_lock), flags);
261 __iommu_free(tbl, dma_addr, npages);
263 /* Make sure TLB cache is flushed if the HW needs it. We do
264 * not do an mb() here on purpose, it is not needed on any of
265 * the current platforms.
267 if (ppc_md.tce_flush)
268 ppc_md.tce_flush(tbl);
270 spin_unlock_irqrestore(&(tbl->it_lock), flags);
273 int iommu_map_sg(struct iommu_table *tbl, struct scatterlist *sglist,
274 int nelems, unsigned long mask,
275 enum dma_data_direction direction)
277 dma_addr_t dma_next = 0, dma_addr;
279 struct scatterlist *s, *outs, *segstart;
280 int outcount, incount, i;
281 unsigned long handle;
283 BUG_ON(direction == DMA_NONE);
285 if ((nelems == 0) || !tbl)
288 outs = s = segstart = &sglist[0];
293 /* Init first segment length for backout at failure */
294 outs->dma_length = 0;
296 DBG("sg mapping %d elements:\n", nelems);
298 spin_lock_irqsave(&(tbl->it_lock), flags);
300 for_each_sg(sglist, s, nelems, i) {
301 unsigned long vaddr, npages, entry, slen;
309 /* Allocate iommu entries for that segment */
310 vaddr = (unsigned long)page_address(s->page) + s->offset;
311 npages = iommu_num_pages(vaddr, slen);
312 entry = iommu_range_alloc(tbl, npages, &handle, mask >> IOMMU_PAGE_SHIFT, 0);
314 DBG(" - vaddr: %lx, size: %lx\n", vaddr, slen);
317 if (unlikely(entry == DMA_ERROR_CODE)) {
318 if (printk_ratelimit())
319 printk(KERN_INFO "iommu_alloc failed, tbl %p vaddr %lx"
320 " npages %lx\n", tbl, vaddr, npages);
324 /* Convert entry to a dma_addr_t */
325 entry += tbl->it_offset;
326 dma_addr = entry << IOMMU_PAGE_SHIFT;
327 dma_addr |= (s->offset & ~IOMMU_PAGE_MASK);
329 DBG(" - %lu pages, entry: %lx, dma_addr: %lx\n",
330 npages, entry, dma_addr);
332 /* Insert into HW table */
333 ppc_md.tce_build(tbl, entry, npages, vaddr & IOMMU_PAGE_MASK, direction);
335 /* If we are in an open segment, try merging */
337 DBG(" - trying merge...\n");
338 /* We cannot merge if:
339 * - allocated dma_addr isn't contiguous to previous allocation
341 if (novmerge || (dma_addr != dma_next)) {
342 /* Can't merge: create a new segment */
345 outs = sg_next(outs);
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) {
378 outs = sg_next(outs);
379 outs->dma_address = DMA_ERROR_CODE;
380 outs->dma_length = 0;
383 /* Make sure updates are seen by hardware */
389 for_each_sg(sglist, s, nelems, i) {
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;
402 spin_unlock_irqrestore(&(tbl->it_lock), flags);
407 void iommu_unmap_sg(struct iommu_table *tbl, struct scatterlist *sglist,
408 int nelems, enum dma_data_direction direction)
410 struct scatterlist *sg;
413 BUG_ON(direction == DMA_NONE);
418 spin_lock_irqsave(&(tbl->it_lock), flags);
423 dma_addr_t dma_handle = sg->dma_address;
425 if (sg->dma_length == 0)
427 npages = iommu_num_pages(dma_handle, sg->dma_length);
428 __iommu_free(tbl, dma_handle, npages);
432 /* Flush/invalidate TLBs if necessary. As for iommu_free(), we
433 * do not do an mb() here, the affected platforms do not need it
436 if (ppc_md.tce_flush)
437 ppc_md.tce_flush(tbl);
439 spin_unlock_irqrestore(&(tbl->it_lock), flags);
443 * Build a iommu_table structure. This contains a bit map which
444 * is used to manage allocation of the tce space.
446 struct iommu_table *iommu_init_table(struct iommu_table *tbl, int nid)
449 unsigned long start_index, end_index;
450 unsigned long entries_per_4g;
452 static int welcomed = 0;
455 /* Set aside 1/4 of the table for large allocations. */
456 tbl->it_halfpoint = tbl->it_size * 3 / 4;
458 /* number of bytes needed for the bitmap */
459 sz = (tbl->it_size + 7) >> 3;
461 page = alloc_pages_node(nid, GFP_ATOMIC, get_order(sz));
463 panic("iommu_init_table: Can't allocate %ld bytes\n", sz);
464 tbl->it_map = page_address(page);
465 memset(tbl->it_map, 0, sz);
468 tbl->it_largehint = tbl->it_halfpoint;
469 spin_lock_init(&tbl->it_lock);
471 #ifdef CONFIG_CRASH_DUMP
472 if (ppc_md.tce_get) {
473 unsigned long tceval;
474 unsigned long tcecount = 0;
477 * Reserve the existing mappings left by the first kernel.
479 for (index = 0; index < tbl->it_size; index++) {
480 tceval = ppc_md.tce_get(tbl, index + tbl->it_offset);
482 * Freed TCE entry contains 0x7fffffffffffffff on JS20
484 if (tceval && (tceval != 0x7fffffffffffffffUL)) {
485 __set_bit(index, tbl->it_map);
489 if ((tbl->it_size - tcecount) < KDUMP_MIN_TCE_ENTRIES) {
490 printk(KERN_WARNING "TCE table is full; ");
491 printk(KERN_WARNING "freeing %d entries for the kdump boot\n",
492 KDUMP_MIN_TCE_ENTRIES);
493 for (index = tbl->it_size - KDUMP_MIN_TCE_ENTRIES;
494 index < tbl->it_size; index++)
495 __clear_bit(index, tbl->it_map);
499 /* Clear the hardware table in case firmware left allocations in it */
500 ppc_md.tce_free(tbl, tbl->it_offset, tbl->it_size);
504 * DMA cannot cross 4 GB boundary. Mark last entry of each 4
505 * GB chunk as reserved.
508 entries_per_4g = 0x100000000l >> IOMMU_PAGE_SHIFT;
510 /* Mark the last bit before a 4GB boundary as used */
511 start_index = tbl->it_offset | (entries_per_4g - 1);
512 start_index -= tbl->it_offset;
514 end_index = tbl->it_size;
516 for (index = start_index; index < end_index - 1; index += entries_per_4g)
517 __set_bit(index, tbl->it_map);
521 printk(KERN_INFO "IOMMU table initialized, virtual merging %s\n",
522 novmerge ? "disabled" : "enabled");
529 void iommu_free_table(struct device_node *dn)
531 struct pci_dn *pdn = dn->data;
532 struct iommu_table *tbl = pdn->iommu_table;
533 unsigned long bitmap_sz, i;
536 if (!tbl || !tbl->it_map) {
537 printk(KERN_ERR "%s: expected TCE map for %s\n", __FUNCTION__,
542 /* verify that table contains no entries */
543 /* it_size is in entries, and we're examining 64 at a time */
544 for (i = 0; i < (tbl->it_size/64); i++) {
545 if (tbl->it_map[i] != 0) {
546 printk(KERN_WARNING "%s: Unexpected TCEs for %s\n",
547 __FUNCTION__, dn->full_name);
552 /* calculate bitmap size in bytes */
553 bitmap_sz = (tbl->it_size + 7) / 8;
556 order = get_order(bitmap_sz);
557 free_pages((unsigned long) tbl->it_map, order);
563 /* Creates TCEs for a user provided buffer. The user buffer must be
564 * contiguous real kernel storage (not vmalloc). The address of the buffer
565 * passed here is the kernel (virtual) address of the buffer. The buffer
566 * need not be page aligned, the dma_addr_t returned will point to the same
567 * byte within the page as vaddr.
569 dma_addr_t iommu_map_single(struct iommu_table *tbl, void *vaddr,
570 size_t size, unsigned long mask,
571 enum dma_data_direction direction)
573 dma_addr_t dma_handle = DMA_ERROR_CODE;
577 BUG_ON(direction == DMA_NONE);
579 uaddr = (unsigned long)vaddr;
580 npages = iommu_num_pages(uaddr, size);
583 dma_handle = iommu_alloc(tbl, vaddr, npages, direction,
584 mask >> IOMMU_PAGE_SHIFT, 0);
585 if (dma_handle == DMA_ERROR_CODE) {
586 if (printk_ratelimit()) {
587 printk(KERN_INFO "iommu_alloc failed, "
588 "tbl %p vaddr %p npages %d\n",
592 dma_handle |= (uaddr & ~IOMMU_PAGE_MASK);
598 void iommu_unmap_single(struct iommu_table *tbl, dma_addr_t dma_handle,
599 size_t size, enum dma_data_direction direction)
603 BUG_ON(direction == DMA_NONE);
606 npages = iommu_num_pages(dma_handle, size);
607 iommu_free(tbl, dma_handle, npages);
611 /* Allocates a contiguous real buffer and creates mappings over it.
612 * Returns the virtual address of the buffer and sets dma_handle
613 * to the dma address (mapping) of the first page.
615 void *iommu_alloc_coherent(struct iommu_table *tbl, size_t size,
616 dma_addr_t *dma_handle, unsigned long mask, gfp_t flag, int node)
621 unsigned int nio_pages, io_order;
624 size = PAGE_ALIGN(size);
625 order = get_order(size);
628 * Client asked for way too much space. This is checked later
629 * anyway. It is easier to debug here for the drivers than in
632 if (order >= IOMAP_MAX_ORDER) {
633 printk("iommu_alloc_consistent size too large: 0x%lx\n", size);
640 /* Alloc enough pages (and possibly more) */
641 page = alloc_pages_node(node, flag, order);
644 ret = page_address(page);
645 memset(ret, 0, size);
647 /* Set up tces to cover the allocated range */
648 nio_pages = size >> IOMMU_PAGE_SHIFT;
649 io_order = get_iommu_order(size);
650 mapping = iommu_alloc(tbl, ret, nio_pages, DMA_BIDIRECTIONAL,
651 mask >> IOMMU_PAGE_SHIFT, io_order);
652 if (mapping == DMA_ERROR_CODE) {
653 free_pages((unsigned long)ret, order);
656 *dma_handle = mapping;
660 void iommu_free_coherent(struct iommu_table *tbl, size_t size,
661 void *vaddr, dma_addr_t dma_handle)
664 unsigned int nio_pages;
666 size = PAGE_ALIGN(size);
667 nio_pages = size >> IOMMU_PAGE_SHIFT;
668 iommu_free(tbl, dma_handle, nio_pages);
669 size = PAGE_ALIGN(size);
670 free_pages((unsigned long)vaddr, get_order(size));