Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * Copyright (C) 2001 Mike Corrigan & Dave Engebretsen, IBM Corporation
4 *
5 * Rewrite, cleanup, new allocation schemes, virtual merging:
6 * Copyright (C) 2004 Olof Johansson, IBM Corporation
7 * and Ben. Herrenschmidt, IBM Corporation
8 *
9 * Dynamic DMA mapping support, bus-independent parts.
10 */
11
12
13#include <linux/init.h>
14#include <linux/types.h>
15#include <linux/slab.h>
16#include <linux/mm.h>
17#include <linux/spinlock.h>
18#include <linux/string.h>
19#include <linux/dma-mapping.h>
20#include <linux/bitmap.h>
21#include <linux/iommu-helper.h>
22#include <linux/crash_dump.h>
23#include <linux/hash.h>
24#include <linux/fault-inject.h>
25#include <linux/pci.h>
26#include <linux/iommu.h>
27#include <linux/sched.h>
28#include <linux/debugfs.h>
29#include <asm/io.h>
30#include <asm/iommu.h>
31#include <asm/pci-bridge.h>
32#include <asm/machdep.h>
33#include <asm/kdump.h>
34#include <asm/fadump.h>
35#include <asm/vio.h>
36#include <asm/tce.h>
37#include <asm/mmu_context.h>
38
39#define DBG(...)
40
41#ifdef CONFIG_IOMMU_DEBUGFS
42static int iommu_debugfs_weight_get(void *data, u64 *val)
43{
44 struct iommu_table *tbl = data;
45 *val = bitmap_weight(tbl->it_map, tbl->it_size);
46 return 0;
47}
48DEFINE_DEBUGFS_ATTRIBUTE(iommu_debugfs_fops_weight, iommu_debugfs_weight_get, NULL, "%llu\n");
49
50static void iommu_debugfs_add(struct iommu_table *tbl)
51{
52 char name[10];
53 struct dentry *liobn_entry;
54
55 sprintf(name, "%08lx", tbl->it_index);
56 liobn_entry = debugfs_create_dir(name, iommu_debugfs_dir);
57
58 debugfs_create_file_unsafe("weight", 0400, liobn_entry, tbl, &iommu_debugfs_fops_weight);
59 debugfs_create_ulong("it_size", 0400, liobn_entry, &tbl->it_size);
60 debugfs_create_ulong("it_page_shift", 0400, liobn_entry, &tbl->it_page_shift);
61 debugfs_create_ulong("it_reserved_start", 0400, liobn_entry, &tbl->it_reserved_start);
62 debugfs_create_ulong("it_reserved_end", 0400, liobn_entry, &tbl->it_reserved_end);
63 debugfs_create_ulong("it_indirect_levels", 0400, liobn_entry, &tbl->it_indirect_levels);
64 debugfs_create_ulong("it_level_size", 0400, liobn_entry, &tbl->it_level_size);
65}
66
67static void iommu_debugfs_del(struct iommu_table *tbl)
68{
69 char name[10];
70 struct dentry *liobn_entry;
71
72 sprintf(name, "%08lx", tbl->it_index);
73 liobn_entry = debugfs_lookup(name, iommu_debugfs_dir);
74 debugfs_remove(liobn_entry);
75}
76#else
77static void iommu_debugfs_add(struct iommu_table *tbl){}
78static void iommu_debugfs_del(struct iommu_table *tbl){}
79#endif
80
81static int novmerge;
82
83static void __iommu_free(struct iommu_table *, dma_addr_t, unsigned int);
84
85static int __init setup_iommu(char *str)
86{
87 if (!strcmp(str, "novmerge"))
88 novmerge = 1;
89 else if (!strcmp(str, "vmerge"))
90 novmerge = 0;
91 return 1;
92}
93
94__setup("iommu=", setup_iommu);
95
96static DEFINE_PER_CPU(unsigned int, iommu_pool_hash);
97
98/*
99 * We precalculate the hash to avoid doing it on every allocation.
100 *
101 * The hash is important to spread CPUs across all the pools. For example,
102 * on a POWER7 with 4 way SMT we want interrupts on the primary threads and
103 * with 4 pools all primary threads would map to the same pool.
104 */
105static int __init setup_iommu_pool_hash(void)
106{
107 unsigned int i;
108
109 for_each_possible_cpu(i)
110 per_cpu(iommu_pool_hash, i) = hash_32(i, IOMMU_POOL_HASHBITS);
111
112 return 0;
113}
114subsys_initcall(setup_iommu_pool_hash);
115
116#ifdef CONFIG_FAIL_IOMMU
117
118static DECLARE_FAULT_ATTR(fail_iommu);
119
120static int __init setup_fail_iommu(char *str)
121{
122 return setup_fault_attr(&fail_iommu, str);
123}
124__setup("fail_iommu=", setup_fail_iommu);
125
126static bool should_fail_iommu(struct device *dev)
127{
128 return dev->archdata.fail_iommu && should_fail(&fail_iommu, 1);
129}
130
131static int __init fail_iommu_debugfs(void)
132{
133 struct dentry *dir = fault_create_debugfs_attr("fail_iommu",
134 NULL, &fail_iommu);
135
136 return PTR_ERR_OR_ZERO(dir);
137}
138late_initcall(fail_iommu_debugfs);
139
140static ssize_t fail_iommu_show(struct device *dev,
141 struct device_attribute *attr, char *buf)
142{
143 return sprintf(buf, "%d\n", dev->archdata.fail_iommu);
144}
145
146static ssize_t fail_iommu_store(struct device *dev,
147 struct device_attribute *attr, const char *buf,
148 size_t count)
149{
150 int i;
151
152 if (count > 0 && sscanf(buf, "%d", &i) > 0)
153 dev->archdata.fail_iommu = (i == 0) ? 0 : 1;
154
155 return count;
156}
157
158static DEVICE_ATTR_RW(fail_iommu);
159
160static int fail_iommu_bus_notify(struct notifier_block *nb,
161 unsigned long action, void *data)
162{
163 struct device *dev = data;
164
165 if (action == BUS_NOTIFY_ADD_DEVICE) {
166 if (device_create_file(dev, &dev_attr_fail_iommu))
167 pr_warn("Unable to create IOMMU fault injection sysfs "
168 "entries\n");
169 } else if (action == BUS_NOTIFY_DEL_DEVICE) {
170 device_remove_file(dev, &dev_attr_fail_iommu);
171 }
172
173 return 0;
174}
175
176static struct notifier_block fail_iommu_bus_notifier = {
177 .notifier_call = fail_iommu_bus_notify
178};
179
180static int __init fail_iommu_setup(void)
181{
182#ifdef CONFIG_PCI
183 bus_register_notifier(&pci_bus_type, &fail_iommu_bus_notifier);
184#endif
185#ifdef CONFIG_IBMVIO
186 bus_register_notifier(&vio_bus_type, &fail_iommu_bus_notifier);
187#endif
188
189 return 0;
190}
191/*
192 * Must execute after PCI and VIO subsystem have initialised but before
193 * devices are probed.
194 */
195arch_initcall(fail_iommu_setup);
196#else
197static inline bool should_fail_iommu(struct device *dev)
198{
199 return false;
200}
201#endif
202
203static unsigned long iommu_range_alloc(struct device *dev,
204 struct iommu_table *tbl,
205 unsigned long npages,
206 unsigned long *handle,
207 unsigned long mask,
208 unsigned int align_order)
209{
210 unsigned long n, end, start;
211 unsigned long limit;
212 int largealloc = npages > 15;
213 int pass = 0;
214 unsigned long align_mask;
215 unsigned long flags;
216 unsigned int pool_nr;
217 struct iommu_pool *pool;
218
219 align_mask = (1ull << align_order) - 1;
220
221 /* This allocator was derived from x86_64's bit string search */
222
223 /* Sanity check */
224 if (unlikely(npages == 0)) {
225 if (printk_ratelimit())
226 WARN_ON(1);
227 return DMA_MAPPING_ERROR;
228 }
229
230 if (should_fail_iommu(dev))
231 return DMA_MAPPING_ERROR;
232
233 /*
234 * We don't need to disable preemption here because any CPU can
235 * safely use any IOMMU pool.
236 */
237 pool_nr = raw_cpu_read(iommu_pool_hash) & (tbl->nr_pools - 1);
238
239 if (largealloc)
240 pool = &(tbl->large_pool);
241 else
242 pool = &(tbl->pools[pool_nr]);
243
244 spin_lock_irqsave(&(pool->lock), flags);
245
246again:
247 if ((pass == 0) && handle && *handle &&
248 (*handle >= pool->start) && (*handle < pool->end))
249 start = *handle;
250 else
251 start = pool->hint;
252
253 limit = pool->end;
254
255 /* The case below can happen if we have a small segment appended
256 * to a large, or when the previous alloc was at the very end of
257 * the available space. If so, go back to the initial start.
258 */
259 if (start >= limit)
260 start = pool->start;
261
262 if (limit + tbl->it_offset > mask) {
263 limit = mask - tbl->it_offset + 1;
264 /* If we're constrained on address range, first try
265 * at the masked hint to avoid O(n) search complexity,
266 * but on second pass, start at 0 in pool 0.
267 */
268 if ((start & mask) >= limit || pass > 0) {
269 spin_unlock(&(pool->lock));
270 pool = &(tbl->pools[0]);
271 spin_lock(&(pool->lock));
272 start = pool->start;
273 } else {
274 start &= mask;
275 }
276 }
277
278 n = iommu_area_alloc(tbl->it_map, limit, start, npages, tbl->it_offset,
279 dma_get_seg_boundary_nr_pages(dev, tbl->it_page_shift),
280 align_mask);
281 if (n == -1) {
282 if (likely(pass == 0)) {
283 /* First try the pool from the start */
284 pool->hint = pool->start;
285 pass++;
286 goto again;
287
288 } else if (pass <= tbl->nr_pools) {
289 /* Now try scanning all the other pools */
290 spin_unlock(&(pool->lock));
291 pool_nr = (pool_nr + 1) & (tbl->nr_pools - 1);
292 pool = &tbl->pools[pool_nr];
293 spin_lock(&(pool->lock));
294 pool->hint = pool->start;
295 pass++;
296 goto again;
297
298 } else if (pass == tbl->nr_pools + 1) {
299 /* Last resort: try largepool */
300 spin_unlock(&pool->lock);
301 pool = &tbl->large_pool;
302 spin_lock(&pool->lock);
303 pool->hint = pool->start;
304 pass++;
305 goto again;
306
307 } else {
308 /* Give up */
309 spin_unlock_irqrestore(&(pool->lock), flags);
310 return DMA_MAPPING_ERROR;
311 }
312 }
313
314 end = n + npages;
315
316 /* Bump the hint to a new block for small allocs. */
317 if (largealloc) {
318 /* Don't bump to new block to avoid fragmentation */
319 pool->hint = end;
320 } else {
321 /* Overflow will be taken care of at the next allocation */
322 pool->hint = (end + tbl->it_blocksize - 1) &
323 ~(tbl->it_blocksize - 1);
324 }
325
326 /* Update handle for SG allocations */
327 if (handle)
328 *handle = end;
329
330 spin_unlock_irqrestore(&(pool->lock), flags);
331
332 return n;
333}
334
335static dma_addr_t iommu_alloc(struct device *dev, struct iommu_table *tbl,
336 void *page, unsigned int npages,
337 enum dma_data_direction direction,
338 unsigned long mask, unsigned int align_order,
339 unsigned long attrs)
340{
341 unsigned long entry;
342 dma_addr_t ret = DMA_MAPPING_ERROR;
343 int build_fail;
344
345 entry = iommu_range_alloc(dev, tbl, npages, NULL, mask, align_order);
346
347 if (unlikely(entry == DMA_MAPPING_ERROR))
348 return DMA_MAPPING_ERROR;
349
350 entry += tbl->it_offset; /* Offset into real TCE table */
351 ret = entry << tbl->it_page_shift; /* Set the return dma address */
352
353 /* Put the TCEs in the HW table */
354 build_fail = tbl->it_ops->set(tbl, entry, npages,
355 (unsigned long)page &
356 IOMMU_PAGE_MASK(tbl), direction, attrs);
357
358 /* tbl->it_ops->set() only returns non-zero for transient errors.
359 * Clean up the table bitmap in this case and return
360 * DMA_MAPPING_ERROR. For all other errors the functionality is
361 * not altered.
362 */
363 if (unlikely(build_fail)) {
364 __iommu_free(tbl, ret, npages);
365 return DMA_MAPPING_ERROR;
366 }
367
368 /* Flush/invalidate TLB caches if necessary */
369 if (tbl->it_ops->flush)
370 tbl->it_ops->flush(tbl);
371
372 /* Make sure updates are seen by hardware */
373 mb();
374
375 return ret;
376}
377
378static bool iommu_free_check(struct iommu_table *tbl, dma_addr_t dma_addr,
379 unsigned int npages)
380{
381 unsigned long entry, free_entry;
382
383 entry = dma_addr >> tbl->it_page_shift;
384 free_entry = entry - tbl->it_offset;
385
386 if (((free_entry + npages) > tbl->it_size) ||
387 (entry < tbl->it_offset)) {
388 if (printk_ratelimit()) {
389 printk(KERN_INFO "iommu_free: invalid entry\n");
390 printk(KERN_INFO "\tentry = 0x%lx\n", entry);
391 printk(KERN_INFO "\tdma_addr = 0x%llx\n", (u64)dma_addr);
392 printk(KERN_INFO "\tTable = 0x%llx\n", (u64)tbl);
393 printk(KERN_INFO "\tbus# = 0x%llx\n", (u64)tbl->it_busno);
394 printk(KERN_INFO "\tsize = 0x%llx\n", (u64)tbl->it_size);
395 printk(KERN_INFO "\tstartOff = 0x%llx\n", (u64)tbl->it_offset);
396 printk(KERN_INFO "\tindex = 0x%llx\n", (u64)tbl->it_index);
397 WARN_ON(1);
398 }
399
400 return false;
401 }
402
403 return true;
404}
405
406static struct iommu_pool *get_pool(struct iommu_table *tbl,
407 unsigned long entry)
408{
409 struct iommu_pool *p;
410 unsigned long largepool_start = tbl->large_pool.start;
411
412 /* The large pool is the last pool at the top of the table */
413 if (entry >= largepool_start) {
414 p = &tbl->large_pool;
415 } else {
416 unsigned int pool_nr = entry / tbl->poolsize;
417
418 BUG_ON(pool_nr > tbl->nr_pools);
419 p = &tbl->pools[pool_nr];
420 }
421
422 return p;
423}
424
425static void __iommu_free(struct iommu_table *tbl, dma_addr_t dma_addr,
426 unsigned int npages)
427{
428 unsigned long entry, free_entry;
429 unsigned long flags;
430 struct iommu_pool *pool;
431
432 entry = dma_addr >> tbl->it_page_shift;
433 free_entry = entry - tbl->it_offset;
434
435 pool = get_pool(tbl, free_entry);
436
437 if (!iommu_free_check(tbl, dma_addr, npages))
438 return;
439
440 tbl->it_ops->clear(tbl, entry, npages);
441
442 spin_lock_irqsave(&(pool->lock), flags);
443 bitmap_clear(tbl->it_map, free_entry, npages);
444 spin_unlock_irqrestore(&(pool->lock), flags);
445}
446
447static void iommu_free(struct iommu_table *tbl, dma_addr_t dma_addr,
448 unsigned int npages)
449{
450 __iommu_free(tbl, dma_addr, npages);
451
452 /* Make sure TLB cache is flushed if the HW needs it. We do
453 * not do an mb() here on purpose, it is not needed on any of
454 * the current platforms.
455 */
456 if (tbl->it_ops->flush)
457 tbl->it_ops->flush(tbl);
458}
459
460int ppc_iommu_map_sg(struct device *dev, struct iommu_table *tbl,
461 struct scatterlist *sglist, int nelems,
462 unsigned long mask, enum dma_data_direction direction,
463 unsigned long attrs)
464{
465 dma_addr_t dma_next = 0, dma_addr;
466 struct scatterlist *s, *outs, *segstart;
467 int outcount, incount, i, build_fail = 0;
468 unsigned int align;
469 unsigned long handle;
470 unsigned int max_seg_size;
471
472 BUG_ON(direction == DMA_NONE);
473
474 if ((nelems == 0) || !tbl)
475 return -EINVAL;
476
477 outs = s = segstart = &sglist[0];
478 outcount = 1;
479 incount = nelems;
480 handle = 0;
481
482 /* Init first segment length for backout at failure */
483 outs->dma_length = 0;
484
485 DBG("sg mapping %d elements:\n", nelems);
486
487 max_seg_size = dma_get_max_seg_size(dev);
488 for_each_sg(sglist, s, nelems, i) {
489 unsigned long vaddr, npages, entry, slen;
490
491 slen = s->length;
492 /* Sanity check */
493 if (slen == 0) {
494 dma_next = 0;
495 continue;
496 }
497 /* Allocate iommu entries for that segment */
498 vaddr = (unsigned long) sg_virt(s);
499 npages = iommu_num_pages(vaddr, slen, IOMMU_PAGE_SIZE(tbl));
500 align = 0;
501 if (tbl->it_page_shift < PAGE_SHIFT && slen >= PAGE_SIZE &&
502 (vaddr & ~PAGE_MASK) == 0)
503 align = PAGE_SHIFT - tbl->it_page_shift;
504 entry = iommu_range_alloc(dev, tbl, npages, &handle,
505 mask >> tbl->it_page_shift, align);
506
507 DBG(" - vaddr: %lx, size: %lx\n", vaddr, slen);
508
509 /* Handle failure */
510 if (unlikely(entry == DMA_MAPPING_ERROR)) {
511 if (!(attrs & DMA_ATTR_NO_WARN) &&
512 printk_ratelimit())
513 dev_info(dev, "iommu_alloc failed, tbl %p "
514 "vaddr %lx npages %lu\n", tbl, vaddr,
515 npages);
516 goto failure;
517 }
518
519 /* Convert entry to a dma_addr_t */
520 entry += tbl->it_offset;
521 dma_addr = entry << tbl->it_page_shift;
522 dma_addr |= (s->offset & ~IOMMU_PAGE_MASK(tbl));
523
524 DBG(" - %lu pages, entry: %lx, dma_addr: %lx\n",
525 npages, entry, dma_addr);
526
527 /* Insert into HW table */
528 build_fail = tbl->it_ops->set(tbl, entry, npages,
529 vaddr & IOMMU_PAGE_MASK(tbl),
530 direction, attrs);
531 if(unlikely(build_fail))
532 goto failure;
533
534 /* If we are in an open segment, try merging */
535 if (segstart != s) {
536 DBG(" - trying merge...\n");
537 /* We cannot merge if:
538 * - allocated dma_addr isn't contiguous to previous allocation
539 */
540 if (novmerge || (dma_addr != dma_next) ||
541 (outs->dma_length + s->length > max_seg_size)) {
542 /* Can't merge: create a new segment */
543 segstart = s;
544 outcount++;
545 outs = sg_next(outs);
546 DBG(" can't merge, new segment.\n");
547 } else {
548 outs->dma_length += s->length;
549 DBG(" merged, new len: %ux\n", outs->dma_length);
550 }
551 }
552
553 if (segstart == s) {
554 /* This is a new segment, fill entries */
555 DBG(" - filling new segment.\n");
556 outs->dma_address = dma_addr;
557 outs->dma_length = slen;
558 }
559
560 /* Calculate next page pointer for contiguous check */
561 dma_next = dma_addr + slen;
562
563 DBG(" - dma next is: %lx\n", dma_next);
564 }
565
566 /* Flush/invalidate TLB caches if necessary */
567 if (tbl->it_ops->flush)
568 tbl->it_ops->flush(tbl);
569
570 DBG("mapped %d elements:\n", outcount);
571
572 /* For the sake of ppc_iommu_unmap_sg, we clear out the length in the
573 * next entry of the sglist if we didn't fill the list completely
574 */
575 if (outcount < incount) {
576 outs = sg_next(outs);
577 outs->dma_length = 0;
578 }
579
580 /* Make sure updates are seen by hardware */
581 mb();
582
583 return outcount;
584
585 failure:
586 for_each_sg(sglist, s, nelems, i) {
587 if (s->dma_length != 0) {
588 unsigned long vaddr, npages;
589
590 vaddr = s->dma_address & IOMMU_PAGE_MASK(tbl);
591 npages = iommu_num_pages(s->dma_address, s->dma_length,
592 IOMMU_PAGE_SIZE(tbl));
593 __iommu_free(tbl, vaddr, npages);
594 s->dma_length = 0;
595 }
596 if (s == outs)
597 break;
598 }
599 return -EIO;
600}
601
602
603void ppc_iommu_unmap_sg(struct iommu_table *tbl, struct scatterlist *sglist,
604 int nelems, enum dma_data_direction direction,
605 unsigned long attrs)
606{
607 struct scatterlist *sg;
608
609 BUG_ON(direction == DMA_NONE);
610
611 if (!tbl)
612 return;
613
614 sg = sglist;
615 while (nelems--) {
616 unsigned int npages;
617 dma_addr_t dma_handle = sg->dma_address;
618
619 if (sg->dma_length == 0)
620 break;
621 npages = iommu_num_pages(dma_handle, sg->dma_length,
622 IOMMU_PAGE_SIZE(tbl));
623 __iommu_free(tbl, dma_handle, npages);
624 sg = sg_next(sg);
625 }
626
627 /* Flush/invalidate TLBs if necessary. As for iommu_free(), we
628 * do not do an mb() here, the affected platforms do not need it
629 * when freeing.
630 */
631 if (tbl->it_ops->flush)
632 tbl->it_ops->flush(tbl);
633}
634
635static void iommu_table_clear(struct iommu_table *tbl)
636{
637 /*
638 * In case of firmware assisted dump system goes through clean
639 * reboot process at the time of system crash. Hence it's safe to
640 * clear the TCE entries if firmware assisted dump is active.
641 */
642 if (!is_kdump_kernel() || is_fadump_active()) {
643 /* Clear the table in case firmware left allocations in it */
644 tbl->it_ops->clear(tbl, tbl->it_offset, tbl->it_size);
645 return;
646 }
647
648#ifdef CONFIG_CRASH_DUMP
649 if (tbl->it_ops->get) {
650 unsigned long index, tceval, tcecount = 0;
651
652 /* Reserve the existing mappings left by the first kernel. */
653 for (index = 0; index < tbl->it_size; index++) {
654 tceval = tbl->it_ops->get(tbl, index + tbl->it_offset);
655 /*
656 * Freed TCE entry contains 0x7fffffffffffffff on JS20
657 */
658 if (tceval && (tceval != 0x7fffffffffffffffUL)) {
659 __set_bit(index, tbl->it_map);
660 tcecount++;
661 }
662 }
663
664 if ((tbl->it_size - tcecount) < KDUMP_MIN_TCE_ENTRIES) {
665 printk(KERN_WARNING "TCE table is full; freeing ");
666 printk(KERN_WARNING "%d entries for the kdump boot\n",
667 KDUMP_MIN_TCE_ENTRIES);
668 for (index = tbl->it_size - KDUMP_MIN_TCE_ENTRIES;
669 index < tbl->it_size; index++)
670 __clear_bit(index, tbl->it_map);
671 }
672 }
673#endif
674}
675
676static void iommu_table_reserve_pages(struct iommu_table *tbl,
677 unsigned long res_start, unsigned long res_end)
678{
679 int i;
680
681 WARN_ON_ONCE(res_end < res_start);
682 /*
683 * Reserve page 0 so it will not be used for any mappings.
684 * This avoids buggy drivers that consider page 0 to be invalid
685 * to crash the machine or even lose data.
686 */
687 if (tbl->it_offset == 0)
688 set_bit(0, tbl->it_map);
689
690 if (res_start < tbl->it_offset)
691 res_start = tbl->it_offset;
692
693 if (res_end > (tbl->it_offset + tbl->it_size))
694 res_end = tbl->it_offset + tbl->it_size;
695
696 /* Check if res_start..res_end is a valid range in the table */
697 if (res_start >= res_end) {
698 tbl->it_reserved_start = tbl->it_offset;
699 tbl->it_reserved_end = tbl->it_offset;
700 return;
701 }
702
703 tbl->it_reserved_start = res_start;
704 tbl->it_reserved_end = res_end;
705
706 for (i = tbl->it_reserved_start; i < tbl->it_reserved_end; ++i)
707 set_bit(i - tbl->it_offset, tbl->it_map);
708}
709
710/*
711 * Build a iommu_table structure. This contains a bit map which
712 * is used to manage allocation of the tce space.
713 */
714struct iommu_table *iommu_init_table(struct iommu_table *tbl, int nid,
715 unsigned long res_start, unsigned long res_end)
716{
717 unsigned long sz;
718 static int welcomed = 0;
719 unsigned int i;
720 struct iommu_pool *p;
721
722 BUG_ON(!tbl->it_ops);
723
724 /* number of bytes needed for the bitmap */
725 sz = BITS_TO_LONGS(tbl->it_size) * sizeof(unsigned long);
726
727 tbl->it_map = vzalloc_node(sz, nid);
728 if (!tbl->it_map) {
729 pr_err("%s: Can't allocate %ld bytes\n", __func__, sz);
730 return NULL;
731 }
732
733 iommu_table_reserve_pages(tbl, res_start, res_end);
734
735 /* We only split the IOMMU table if we have 1GB or more of space */
736 if ((tbl->it_size << tbl->it_page_shift) >= (1UL * 1024 * 1024 * 1024))
737 tbl->nr_pools = IOMMU_NR_POOLS;
738 else
739 tbl->nr_pools = 1;
740
741 /* We reserve the top 1/4 of the table for large allocations */
742 tbl->poolsize = (tbl->it_size * 3 / 4) / tbl->nr_pools;
743
744 for (i = 0; i < tbl->nr_pools; i++) {
745 p = &tbl->pools[i];
746 spin_lock_init(&(p->lock));
747 p->start = tbl->poolsize * i;
748 p->hint = p->start;
749 p->end = p->start + tbl->poolsize;
750 }
751
752 p = &tbl->large_pool;
753 spin_lock_init(&(p->lock));
754 p->start = tbl->poolsize * i;
755 p->hint = p->start;
756 p->end = tbl->it_size;
757
758 iommu_table_clear(tbl);
759
760 if (!welcomed) {
761 printk(KERN_INFO "IOMMU table initialized, virtual merging %s\n",
762 novmerge ? "disabled" : "enabled");
763 welcomed = 1;
764 }
765
766 iommu_debugfs_add(tbl);
767
768 return tbl;
769}
770
771bool iommu_table_in_use(struct iommu_table *tbl)
772{
773 unsigned long start = 0, end;
774
775 /* ignore reserved bit0 */
776 if (tbl->it_offset == 0)
777 start = 1;
778
779 /* Simple case with no reserved MMIO32 region */
780 if (!tbl->it_reserved_start && !tbl->it_reserved_end)
781 return find_next_bit(tbl->it_map, tbl->it_size, start) != tbl->it_size;
782
783 end = tbl->it_reserved_start - tbl->it_offset;
784 if (find_next_bit(tbl->it_map, end, start) != end)
785 return true;
786
787 start = tbl->it_reserved_end - tbl->it_offset;
788 end = tbl->it_size;
789 return find_next_bit(tbl->it_map, end, start) != end;
790}
791
792static void iommu_table_free(struct kref *kref)
793{
794 struct iommu_table *tbl;
795
796 tbl = container_of(kref, struct iommu_table, it_kref);
797
798 if (tbl->it_ops->free)
799 tbl->it_ops->free(tbl);
800
801 if (!tbl->it_map) {
802 kfree(tbl);
803 return;
804 }
805
806 iommu_debugfs_del(tbl);
807
808 /* verify that table contains no entries */
809 if (iommu_table_in_use(tbl))
810 pr_warn("%s: Unexpected TCEs\n", __func__);
811
812 /* free bitmap */
813 vfree(tbl->it_map);
814
815 /* free table */
816 kfree(tbl);
817}
818
819struct iommu_table *iommu_tce_table_get(struct iommu_table *tbl)
820{
821 if (kref_get_unless_zero(&tbl->it_kref))
822 return tbl;
823
824 return NULL;
825}
826EXPORT_SYMBOL_GPL(iommu_tce_table_get);
827
828int iommu_tce_table_put(struct iommu_table *tbl)
829{
830 if (WARN_ON(!tbl))
831 return 0;
832
833 return kref_put(&tbl->it_kref, iommu_table_free);
834}
835EXPORT_SYMBOL_GPL(iommu_tce_table_put);
836
837/* Creates TCEs for a user provided buffer. The user buffer must be
838 * contiguous real kernel storage (not vmalloc). The address passed here
839 * comprises a page address and offset into that page. The dma_addr_t
840 * returned will point to the same byte within the page as was passed in.
841 */
842dma_addr_t iommu_map_page(struct device *dev, struct iommu_table *tbl,
843 struct page *page, unsigned long offset, size_t size,
844 unsigned long mask, enum dma_data_direction direction,
845 unsigned long attrs)
846{
847 dma_addr_t dma_handle = DMA_MAPPING_ERROR;
848 void *vaddr;
849 unsigned long uaddr;
850 unsigned int npages, align;
851
852 BUG_ON(direction == DMA_NONE);
853
854 vaddr = page_address(page) + offset;
855 uaddr = (unsigned long)vaddr;
856
857 if (tbl) {
858 npages = iommu_num_pages(uaddr, size, IOMMU_PAGE_SIZE(tbl));
859 align = 0;
860 if (tbl->it_page_shift < PAGE_SHIFT && size >= PAGE_SIZE &&
861 ((unsigned long)vaddr & ~PAGE_MASK) == 0)
862 align = PAGE_SHIFT - tbl->it_page_shift;
863
864 dma_handle = iommu_alloc(dev, tbl, vaddr, npages, direction,
865 mask >> tbl->it_page_shift, align,
866 attrs);
867 if (dma_handle == DMA_MAPPING_ERROR) {
868 if (!(attrs & DMA_ATTR_NO_WARN) &&
869 printk_ratelimit()) {
870 dev_info(dev, "iommu_alloc failed, tbl %p "
871 "vaddr %p npages %d\n", tbl, vaddr,
872 npages);
873 }
874 } else
875 dma_handle |= (uaddr & ~IOMMU_PAGE_MASK(tbl));
876 }
877
878 return dma_handle;
879}
880
881void iommu_unmap_page(struct iommu_table *tbl, dma_addr_t dma_handle,
882 size_t size, enum dma_data_direction direction,
883 unsigned long attrs)
884{
885 unsigned int npages;
886
887 BUG_ON(direction == DMA_NONE);
888
889 if (tbl) {
890 npages = iommu_num_pages(dma_handle, size,
891 IOMMU_PAGE_SIZE(tbl));
892 iommu_free(tbl, dma_handle, npages);
893 }
894}
895
896/* Allocates a contiguous real buffer and creates mappings over it.
897 * Returns the virtual address of the buffer and sets dma_handle
898 * to the dma address (mapping) of the first page.
899 */
900void *iommu_alloc_coherent(struct device *dev, struct iommu_table *tbl,
901 size_t size, dma_addr_t *dma_handle,
902 unsigned long mask, gfp_t flag, int node)
903{
904 void *ret = NULL;
905 dma_addr_t mapping;
906 unsigned int order;
907 unsigned int nio_pages, io_order;
908 struct page *page;
909
910 size = PAGE_ALIGN(size);
911 order = get_order(size);
912
913 /*
914 * Client asked for way too much space. This is checked later
915 * anyway. It is easier to debug here for the drivers than in
916 * the tce tables.
917 */
918 if (order >= IOMAP_MAX_ORDER) {
919 dev_info(dev, "iommu_alloc_consistent size too large: 0x%lx\n",
920 size);
921 return NULL;
922 }
923
924 if (!tbl)
925 return NULL;
926
927 /* Alloc enough pages (and possibly more) */
928 page = alloc_pages_node(node, flag, order);
929 if (!page)
930 return NULL;
931 ret = page_address(page);
932 memset(ret, 0, size);
933
934 /* Set up tces to cover the allocated range */
935 nio_pages = size >> tbl->it_page_shift;
936 io_order = get_iommu_order(size, tbl);
937 mapping = iommu_alloc(dev, tbl, ret, nio_pages, DMA_BIDIRECTIONAL,
938 mask >> tbl->it_page_shift, io_order, 0);
939 if (mapping == DMA_MAPPING_ERROR) {
940 free_pages((unsigned long)ret, order);
941 return NULL;
942 }
943 *dma_handle = mapping;
944 return ret;
945}
946
947void iommu_free_coherent(struct iommu_table *tbl, size_t size,
948 void *vaddr, dma_addr_t dma_handle)
949{
950 if (tbl) {
951 unsigned int nio_pages;
952
953 size = PAGE_ALIGN(size);
954 nio_pages = size >> tbl->it_page_shift;
955 iommu_free(tbl, dma_handle, nio_pages);
956 size = PAGE_ALIGN(size);
957 free_pages((unsigned long)vaddr, get_order(size));
958 }
959}
960
961unsigned long iommu_direction_to_tce_perm(enum dma_data_direction dir)
962{
963 switch (dir) {
964 case DMA_BIDIRECTIONAL:
965 return TCE_PCI_READ | TCE_PCI_WRITE;
966 case DMA_FROM_DEVICE:
967 return TCE_PCI_WRITE;
968 case DMA_TO_DEVICE:
969 return TCE_PCI_READ;
970 default:
971 return 0;
972 }
973}
974EXPORT_SYMBOL_GPL(iommu_direction_to_tce_perm);
975
976#ifdef CONFIG_IOMMU_API
977/*
978 * SPAPR TCE API
979 */
980static void group_release(void *iommu_data)
981{
982 struct iommu_table_group *table_group = iommu_data;
983
984 table_group->group = NULL;
985}
986
987void iommu_register_group(struct iommu_table_group *table_group,
988 int pci_domain_number, unsigned long pe_num)
989{
990 struct iommu_group *grp;
991 char *name;
992
993 grp = iommu_group_alloc();
994 if (IS_ERR(grp)) {
995 pr_warn("powerpc iommu api: cannot create new group, err=%ld\n",
996 PTR_ERR(grp));
997 return;
998 }
999 table_group->group = grp;
1000 iommu_group_set_iommudata(grp, table_group, group_release);
1001 name = kasprintf(GFP_KERNEL, "domain%d-pe%lx",
1002 pci_domain_number, pe_num);
1003 if (!name)
1004 return;
1005 iommu_group_set_name(grp, name);
1006 kfree(name);
1007}
1008
1009enum dma_data_direction iommu_tce_direction(unsigned long tce)
1010{
1011 if ((tce & TCE_PCI_READ) && (tce & TCE_PCI_WRITE))
1012 return DMA_BIDIRECTIONAL;
1013 else if (tce & TCE_PCI_READ)
1014 return DMA_TO_DEVICE;
1015 else if (tce & TCE_PCI_WRITE)
1016 return DMA_FROM_DEVICE;
1017 else
1018 return DMA_NONE;
1019}
1020EXPORT_SYMBOL_GPL(iommu_tce_direction);
1021
1022void iommu_flush_tce(struct iommu_table *tbl)
1023{
1024 /* Flush/invalidate TLB caches if necessary */
1025 if (tbl->it_ops->flush)
1026 tbl->it_ops->flush(tbl);
1027
1028 /* Make sure updates are seen by hardware */
1029 mb();
1030}
1031EXPORT_SYMBOL_GPL(iommu_flush_tce);
1032
1033int iommu_tce_check_ioba(unsigned long page_shift,
1034 unsigned long offset, unsigned long size,
1035 unsigned long ioba, unsigned long npages)
1036{
1037 unsigned long mask = (1UL << page_shift) - 1;
1038
1039 if (ioba & mask)
1040 return -EINVAL;
1041
1042 ioba >>= page_shift;
1043 if (ioba < offset)
1044 return -EINVAL;
1045
1046 if ((ioba + 1) > (offset + size))
1047 return -EINVAL;
1048
1049 return 0;
1050}
1051EXPORT_SYMBOL_GPL(iommu_tce_check_ioba);
1052
1053int iommu_tce_check_gpa(unsigned long page_shift, unsigned long gpa)
1054{
1055 unsigned long mask = (1UL << page_shift) - 1;
1056
1057 if (gpa & mask)
1058 return -EINVAL;
1059
1060 return 0;
1061}
1062EXPORT_SYMBOL_GPL(iommu_tce_check_gpa);
1063
1064extern long iommu_tce_xchg_no_kill(struct mm_struct *mm,
1065 struct iommu_table *tbl,
1066 unsigned long entry, unsigned long *hpa,
1067 enum dma_data_direction *direction)
1068{
1069 long ret;
1070 unsigned long size = 0;
1071
1072 ret = tbl->it_ops->xchg_no_kill(tbl, entry, hpa, direction);
1073 if (!ret && ((*direction == DMA_FROM_DEVICE) ||
1074 (*direction == DMA_BIDIRECTIONAL)) &&
1075 !mm_iommu_is_devmem(mm, *hpa, tbl->it_page_shift,
1076 &size))
1077 SetPageDirty(pfn_to_page(*hpa >> PAGE_SHIFT));
1078
1079 return ret;
1080}
1081EXPORT_SYMBOL_GPL(iommu_tce_xchg_no_kill);
1082
1083void iommu_tce_kill(struct iommu_table *tbl,
1084 unsigned long entry, unsigned long pages)
1085{
1086 if (tbl->it_ops->tce_kill)
1087 tbl->it_ops->tce_kill(tbl, entry, pages);
1088}
1089EXPORT_SYMBOL_GPL(iommu_tce_kill);
1090
1091int iommu_take_ownership(struct iommu_table *tbl)
1092{
1093 unsigned long flags, i, sz = (tbl->it_size + 7) >> 3;
1094 int ret = 0;
1095
1096 /*
1097 * VFIO does not control TCE entries allocation and the guest
1098 * can write new TCEs on top of existing ones so iommu_tce_build()
1099 * must be able to release old pages. This functionality
1100 * requires exchange() callback defined so if it is not
1101 * implemented, we disallow taking ownership over the table.
1102 */
1103 if (!tbl->it_ops->xchg_no_kill)
1104 return -EINVAL;
1105
1106 spin_lock_irqsave(&tbl->large_pool.lock, flags);
1107 for (i = 0; i < tbl->nr_pools; i++)
1108 spin_lock_nest_lock(&tbl->pools[i].lock, &tbl->large_pool.lock);
1109
1110 if (iommu_table_in_use(tbl)) {
1111 pr_err("iommu_tce: it_map is not empty");
1112 ret = -EBUSY;
1113 } else {
1114 memset(tbl->it_map, 0xff, sz);
1115 }
1116
1117 for (i = 0; i < tbl->nr_pools; i++)
1118 spin_unlock(&tbl->pools[i].lock);
1119 spin_unlock_irqrestore(&tbl->large_pool.lock, flags);
1120
1121 return ret;
1122}
1123EXPORT_SYMBOL_GPL(iommu_take_ownership);
1124
1125void iommu_release_ownership(struct iommu_table *tbl)
1126{
1127 unsigned long flags, i, sz = (tbl->it_size + 7) >> 3;
1128
1129 spin_lock_irqsave(&tbl->large_pool.lock, flags);
1130 for (i = 0; i < tbl->nr_pools; i++)
1131 spin_lock_nest_lock(&tbl->pools[i].lock, &tbl->large_pool.lock);
1132
1133 memset(tbl->it_map, 0, sz);
1134
1135 iommu_table_reserve_pages(tbl, tbl->it_reserved_start,
1136 tbl->it_reserved_end);
1137
1138 for (i = 0; i < tbl->nr_pools; i++)
1139 spin_unlock(&tbl->pools[i].lock);
1140 spin_unlock_irqrestore(&tbl->large_pool.lock, flags);
1141}
1142EXPORT_SYMBOL_GPL(iommu_release_ownership);
1143
1144int iommu_add_device(struct iommu_table_group *table_group, struct device *dev)
1145{
1146 /*
1147 * The sysfs entries should be populated before
1148 * binding IOMMU group. If sysfs entries isn't
1149 * ready, we simply bail.
1150 */
1151 if (!device_is_registered(dev))
1152 return -ENOENT;
1153
1154 if (device_iommu_mapped(dev)) {
1155 pr_debug("%s: Skipping device %s with iommu group %d\n",
1156 __func__, dev_name(dev),
1157 iommu_group_id(dev->iommu_group));
1158 return -EBUSY;
1159 }
1160
1161 pr_debug("%s: Adding %s to iommu group %d\n",
1162 __func__, dev_name(dev), iommu_group_id(table_group->group));
1163
1164 return iommu_group_add_device(table_group->group, dev);
1165}
1166EXPORT_SYMBOL_GPL(iommu_add_device);
1167
1168void iommu_del_device(struct device *dev)
1169{
1170 /*
1171 * Some devices might not have IOMMU table and group
1172 * and we needn't detach them from the associated
1173 * IOMMU groups
1174 */
1175 if (!device_iommu_mapped(dev)) {
1176 pr_debug("iommu_tce: skipping device %s with no tbl\n",
1177 dev_name(dev));
1178 return;
1179 }
1180
1181 iommu_group_remove_device(dev);
1182}
1183EXPORT_SYMBOL_GPL(iommu_del_device);
1184#endif /* CONFIG_IOMMU_API */
1/*
2 * Copyright (C) 2001 Mike Corrigan & Dave Engebretsen, IBM Corporation
3 *
4 * Rewrite, cleanup, new allocation schemes, virtual merging:
5 * Copyright (C) 2004 Olof Johansson, IBM Corporation
6 * and Ben. Herrenschmidt, IBM Corporation
7 *
8 * Dynamic DMA mapping support, bus-independent parts.
9 *
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.
14 *
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.
19 *
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
23 */
24
25
26#include <linux/init.h>
27#include <linux/types.h>
28#include <linux/slab.h>
29#include <linux/mm.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>
36#include <linux/hash.h>
37#include <linux/fault-inject.h>
38#include <linux/pci.h>
39#include <linux/iommu.h>
40#include <linux/sched.h>
41#include <asm/io.h>
42#include <asm/prom.h>
43#include <asm/iommu.h>
44#include <asm/pci-bridge.h>
45#include <asm/machdep.h>
46#include <asm/kdump.h>
47#include <asm/fadump.h>
48#include <asm/vio.h>
49#include <asm/tce.h>
50
51#define DBG(...)
52
53static int novmerge;
54
55static void __iommu_free(struct iommu_table *, dma_addr_t, unsigned int);
56
57static int __init setup_iommu(char *str)
58{
59 if (!strcmp(str, "novmerge"))
60 novmerge = 1;
61 else if (!strcmp(str, "vmerge"))
62 novmerge = 0;
63 return 1;
64}
65
66__setup("iommu=", setup_iommu);
67
68static DEFINE_PER_CPU(unsigned int, iommu_pool_hash);
69
70/*
71 * We precalculate the hash to avoid doing it on every allocation.
72 *
73 * The hash is important to spread CPUs across all the pools. For example,
74 * on a POWER7 with 4 way SMT we want interrupts on the primary threads and
75 * with 4 pools all primary threads would map to the same pool.
76 */
77static int __init setup_iommu_pool_hash(void)
78{
79 unsigned int i;
80
81 for_each_possible_cpu(i)
82 per_cpu(iommu_pool_hash, i) = hash_32(i, IOMMU_POOL_HASHBITS);
83
84 return 0;
85}
86subsys_initcall(setup_iommu_pool_hash);
87
88#ifdef CONFIG_FAIL_IOMMU
89
90static DECLARE_FAULT_ATTR(fail_iommu);
91
92static int __init setup_fail_iommu(char *str)
93{
94 return setup_fault_attr(&fail_iommu, str);
95}
96__setup("fail_iommu=", setup_fail_iommu);
97
98static bool should_fail_iommu(struct device *dev)
99{
100 return dev->archdata.fail_iommu && should_fail(&fail_iommu, 1);
101}
102
103static int __init fail_iommu_debugfs(void)
104{
105 struct dentry *dir = fault_create_debugfs_attr("fail_iommu",
106 NULL, &fail_iommu);
107
108 return PTR_ERR_OR_ZERO(dir);
109}
110late_initcall(fail_iommu_debugfs);
111
112static ssize_t fail_iommu_show(struct device *dev,
113 struct device_attribute *attr, char *buf)
114{
115 return sprintf(buf, "%d\n", dev->archdata.fail_iommu);
116}
117
118static ssize_t fail_iommu_store(struct device *dev,
119 struct device_attribute *attr, const char *buf,
120 size_t count)
121{
122 int i;
123
124 if (count > 0 && sscanf(buf, "%d", &i) > 0)
125 dev->archdata.fail_iommu = (i == 0) ? 0 : 1;
126
127 return count;
128}
129
130static DEVICE_ATTR(fail_iommu, S_IRUGO|S_IWUSR, fail_iommu_show,
131 fail_iommu_store);
132
133static int fail_iommu_bus_notify(struct notifier_block *nb,
134 unsigned long action, void *data)
135{
136 struct device *dev = data;
137
138 if (action == BUS_NOTIFY_ADD_DEVICE) {
139 if (device_create_file(dev, &dev_attr_fail_iommu))
140 pr_warn("Unable to create IOMMU fault injection sysfs "
141 "entries\n");
142 } else if (action == BUS_NOTIFY_DEL_DEVICE) {
143 device_remove_file(dev, &dev_attr_fail_iommu);
144 }
145
146 return 0;
147}
148
149static struct notifier_block fail_iommu_bus_notifier = {
150 .notifier_call = fail_iommu_bus_notify
151};
152
153static int __init fail_iommu_setup(void)
154{
155#ifdef CONFIG_PCI
156 bus_register_notifier(&pci_bus_type, &fail_iommu_bus_notifier);
157#endif
158#ifdef CONFIG_IBMVIO
159 bus_register_notifier(&vio_bus_type, &fail_iommu_bus_notifier);
160#endif
161
162 return 0;
163}
164/*
165 * Must execute after PCI and VIO subsystem have initialised but before
166 * devices are probed.
167 */
168arch_initcall(fail_iommu_setup);
169#else
170static inline bool should_fail_iommu(struct device *dev)
171{
172 return false;
173}
174#endif
175
176static unsigned long iommu_range_alloc(struct device *dev,
177 struct iommu_table *tbl,
178 unsigned long npages,
179 unsigned long *handle,
180 unsigned long mask,
181 unsigned int align_order)
182{
183 unsigned long n, end, start;
184 unsigned long limit;
185 int largealloc = npages > 15;
186 int pass = 0;
187 unsigned long align_mask;
188 unsigned long boundary_size;
189 unsigned long flags;
190 unsigned int pool_nr;
191 struct iommu_pool *pool;
192
193 align_mask = 0xffffffffffffffffl >> (64 - align_order);
194
195 /* This allocator was derived from x86_64's bit string search */
196
197 /* Sanity check */
198 if (unlikely(npages == 0)) {
199 if (printk_ratelimit())
200 WARN_ON(1);
201 return DMA_ERROR_CODE;
202 }
203
204 if (should_fail_iommu(dev))
205 return DMA_ERROR_CODE;
206
207 /*
208 * We don't need to disable preemption here because any CPU can
209 * safely use any IOMMU pool.
210 */
211 pool_nr = __this_cpu_read(iommu_pool_hash) & (tbl->nr_pools - 1);
212
213 if (largealloc)
214 pool = &(tbl->large_pool);
215 else
216 pool = &(tbl->pools[pool_nr]);
217
218 spin_lock_irqsave(&(pool->lock), flags);
219
220again:
221 if ((pass == 0) && handle && *handle &&
222 (*handle >= pool->start) && (*handle < pool->end))
223 start = *handle;
224 else
225 start = pool->hint;
226
227 limit = pool->end;
228
229 /* The case below can happen if we have a small segment appended
230 * to a large, or when the previous alloc was at the very end of
231 * the available space. If so, go back to the initial start.
232 */
233 if (start >= limit)
234 start = pool->start;
235
236 if (limit + tbl->it_offset > mask) {
237 limit = mask - tbl->it_offset + 1;
238 /* If we're constrained on address range, first try
239 * at the masked hint to avoid O(n) search complexity,
240 * but on second pass, start at 0 in pool 0.
241 */
242 if ((start & mask) >= limit || pass > 0) {
243 spin_unlock(&(pool->lock));
244 pool = &(tbl->pools[0]);
245 spin_lock(&(pool->lock));
246 start = pool->start;
247 } else {
248 start &= mask;
249 }
250 }
251
252 if (dev)
253 boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
254 1 << tbl->it_page_shift);
255 else
256 boundary_size = ALIGN(1UL << 32, 1 << tbl->it_page_shift);
257 /* 4GB boundary for iseries_hv_alloc and iseries_hv_map */
258
259 n = iommu_area_alloc(tbl->it_map, limit, start, npages, tbl->it_offset,
260 boundary_size >> tbl->it_page_shift, align_mask);
261 if (n == -1) {
262 if (likely(pass == 0)) {
263 /* First try the pool from the start */
264 pool->hint = pool->start;
265 pass++;
266 goto again;
267
268 } else if (pass <= tbl->nr_pools) {
269 /* Now try scanning all the other pools */
270 spin_unlock(&(pool->lock));
271 pool_nr = (pool_nr + 1) & (tbl->nr_pools - 1);
272 pool = &tbl->pools[pool_nr];
273 spin_lock(&(pool->lock));
274 pool->hint = pool->start;
275 pass++;
276 goto again;
277
278 } else {
279 /* Give up */
280 spin_unlock_irqrestore(&(pool->lock), flags);
281 return DMA_ERROR_CODE;
282 }
283 }
284
285 end = n + npages;
286
287 /* Bump the hint to a new block for small allocs. */
288 if (largealloc) {
289 /* Don't bump to new block to avoid fragmentation */
290 pool->hint = end;
291 } else {
292 /* Overflow will be taken care of at the next allocation */
293 pool->hint = (end + tbl->it_blocksize - 1) &
294 ~(tbl->it_blocksize - 1);
295 }
296
297 /* Update handle for SG allocations */
298 if (handle)
299 *handle = end;
300
301 spin_unlock_irqrestore(&(pool->lock), flags);
302
303 return n;
304}
305
306static dma_addr_t iommu_alloc(struct device *dev, struct iommu_table *tbl,
307 void *page, unsigned int npages,
308 enum dma_data_direction direction,
309 unsigned long mask, unsigned int align_order,
310 unsigned long attrs)
311{
312 unsigned long entry;
313 dma_addr_t ret = DMA_ERROR_CODE;
314 int build_fail;
315
316 entry = iommu_range_alloc(dev, tbl, npages, NULL, mask, align_order);
317
318 if (unlikely(entry == DMA_ERROR_CODE))
319 return DMA_ERROR_CODE;
320
321 entry += tbl->it_offset; /* Offset into real TCE table */
322 ret = entry << tbl->it_page_shift; /* Set the return dma address */
323
324 /* Put the TCEs in the HW table */
325 build_fail = tbl->it_ops->set(tbl, entry, npages,
326 (unsigned long)page &
327 IOMMU_PAGE_MASK(tbl), direction, attrs);
328
329 /* tbl->it_ops->set() only returns non-zero for transient errors.
330 * Clean up the table bitmap in this case and return
331 * DMA_ERROR_CODE. For all other errors the functionality is
332 * not altered.
333 */
334 if (unlikely(build_fail)) {
335 __iommu_free(tbl, ret, npages);
336 return DMA_ERROR_CODE;
337 }
338
339 /* Flush/invalidate TLB caches if necessary */
340 if (tbl->it_ops->flush)
341 tbl->it_ops->flush(tbl);
342
343 /* Make sure updates are seen by hardware */
344 mb();
345
346 return ret;
347}
348
349static bool iommu_free_check(struct iommu_table *tbl, dma_addr_t dma_addr,
350 unsigned int npages)
351{
352 unsigned long entry, free_entry;
353
354 entry = dma_addr >> tbl->it_page_shift;
355 free_entry = entry - tbl->it_offset;
356
357 if (((free_entry + npages) > tbl->it_size) ||
358 (entry < tbl->it_offset)) {
359 if (printk_ratelimit()) {
360 printk(KERN_INFO "iommu_free: invalid entry\n");
361 printk(KERN_INFO "\tentry = 0x%lx\n", entry);
362 printk(KERN_INFO "\tdma_addr = 0x%llx\n", (u64)dma_addr);
363 printk(KERN_INFO "\tTable = 0x%llx\n", (u64)tbl);
364 printk(KERN_INFO "\tbus# = 0x%llx\n", (u64)tbl->it_busno);
365 printk(KERN_INFO "\tsize = 0x%llx\n", (u64)tbl->it_size);
366 printk(KERN_INFO "\tstartOff = 0x%llx\n", (u64)tbl->it_offset);
367 printk(KERN_INFO "\tindex = 0x%llx\n", (u64)tbl->it_index);
368 WARN_ON(1);
369 }
370
371 return false;
372 }
373
374 return true;
375}
376
377static struct iommu_pool *get_pool(struct iommu_table *tbl,
378 unsigned long entry)
379{
380 struct iommu_pool *p;
381 unsigned long largepool_start = tbl->large_pool.start;
382
383 /* The large pool is the last pool at the top of the table */
384 if (entry >= largepool_start) {
385 p = &tbl->large_pool;
386 } else {
387 unsigned int pool_nr = entry / tbl->poolsize;
388
389 BUG_ON(pool_nr > tbl->nr_pools);
390 p = &tbl->pools[pool_nr];
391 }
392
393 return p;
394}
395
396static void __iommu_free(struct iommu_table *tbl, dma_addr_t dma_addr,
397 unsigned int npages)
398{
399 unsigned long entry, free_entry;
400 unsigned long flags;
401 struct iommu_pool *pool;
402
403 entry = dma_addr >> tbl->it_page_shift;
404 free_entry = entry - tbl->it_offset;
405
406 pool = get_pool(tbl, free_entry);
407
408 if (!iommu_free_check(tbl, dma_addr, npages))
409 return;
410
411 tbl->it_ops->clear(tbl, entry, npages);
412
413 spin_lock_irqsave(&(pool->lock), flags);
414 bitmap_clear(tbl->it_map, free_entry, npages);
415 spin_unlock_irqrestore(&(pool->lock), flags);
416}
417
418static void iommu_free(struct iommu_table *tbl, dma_addr_t dma_addr,
419 unsigned int npages)
420{
421 __iommu_free(tbl, dma_addr, npages);
422
423 /* Make sure TLB cache is flushed if the HW needs it. We do
424 * not do an mb() here on purpose, it is not needed on any of
425 * the current platforms.
426 */
427 if (tbl->it_ops->flush)
428 tbl->it_ops->flush(tbl);
429}
430
431int ppc_iommu_map_sg(struct device *dev, struct iommu_table *tbl,
432 struct scatterlist *sglist, int nelems,
433 unsigned long mask, enum dma_data_direction direction,
434 unsigned long attrs)
435{
436 dma_addr_t dma_next = 0, dma_addr;
437 struct scatterlist *s, *outs, *segstart;
438 int outcount, incount, i, build_fail = 0;
439 unsigned int align;
440 unsigned long handle;
441 unsigned int max_seg_size;
442
443 BUG_ON(direction == DMA_NONE);
444
445 if ((nelems == 0) || !tbl)
446 return 0;
447
448 outs = s = segstart = &sglist[0];
449 outcount = 1;
450 incount = nelems;
451 handle = 0;
452
453 /* Init first segment length for backout at failure */
454 outs->dma_length = 0;
455
456 DBG("sg mapping %d elements:\n", nelems);
457
458 max_seg_size = dma_get_max_seg_size(dev);
459 for_each_sg(sglist, s, nelems, i) {
460 unsigned long vaddr, npages, entry, slen;
461
462 slen = s->length;
463 /* Sanity check */
464 if (slen == 0) {
465 dma_next = 0;
466 continue;
467 }
468 /* Allocate iommu entries for that segment */
469 vaddr = (unsigned long) sg_virt(s);
470 npages = iommu_num_pages(vaddr, slen, IOMMU_PAGE_SIZE(tbl));
471 align = 0;
472 if (tbl->it_page_shift < PAGE_SHIFT && slen >= PAGE_SIZE &&
473 (vaddr & ~PAGE_MASK) == 0)
474 align = PAGE_SHIFT - tbl->it_page_shift;
475 entry = iommu_range_alloc(dev, tbl, npages, &handle,
476 mask >> tbl->it_page_shift, align);
477
478 DBG(" - vaddr: %lx, size: %lx\n", vaddr, slen);
479
480 /* Handle failure */
481 if (unlikely(entry == DMA_ERROR_CODE)) {
482 if (!(attrs & DMA_ATTR_NO_WARN) &&
483 printk_ratelimit())
484 dev_info(dev, "iommu_alloc failed, tbl %p "
485 "vaddr %lx npages %lu\n", tbl, vaddr,
486 npages);
487 goto failure;
488 }
489
490 /* Convert entry to a dma_addr_t */
491 entry += tbl->it_offset;
492 dma_addr = entry << tbl->it_page_shift;
493 dma_addr |= (s->offset & ~IOMMU_PAGE_MASK(tbl));
494
495 DBG(" - %lu pages, entry: %lx, dma_addr: %lx\n",
496 npages, entry, dma_addr);
497
498 /* Insert into HW table */
499 build_fail = tbl->it_ops->set(tbl, entry, npages,
500 vaddr & IOMMU_PAGE_MASK(tbl),
501 direction, attrs);
502 if(unlikely(build_fail))
503 goto failure;
504
505 /* If we are in an open segment, try merging */
506 if (segstart != s) {
507 DBG(" - trying merge...\n");
508 /* We cannot merge if:
509 * - allocated dma_addr isn't contiguous to previous allocation
510 */
511 if (novmerge || (dma_addr != dma_next) ||
512 (outs->dma_length + s->length > max_seg_size)) {
513 /* Can't merge: create a new segment */
514 segstart = s;
515 outcount++;
516 outs = sg_next(outs);
517 DBG(" can't merge, new segment.\n");
518 } else {
519 outs->dma_length += s->length;
520 DBG(" merged, new len: %ux\n", outs->dma_length);
521 }
522 }
523
524 if (segstart == s) {
525 /* This is a new segment, fill entries */
526 DBG(" - filling new segment.\n");
527 outs->dma_address = dma_addr;
528 outs->dma_length = slen;
529 }
530
531 /* Calculate next page pointer for contiguous check */
532 dma_next = dma_addr + slen;
533
534 DBG(" - dma next is: %lx\n", dma_next);
535 }
536
537 /* Flush/invalidate TLB caches if necessary */
538 if (tbl->it_ops->flush)
539 tbl->it_ops->flush(tbl);
540
541 DBG("mapped %d elements:\n", outcount);
542
543 /* For the sake of ppc_iommu_unmap_sg, we clear out the length in the
544 * next entry of the sglist if we didn't fill the list completely
545 */
546 if (outcount < incount) {
547 outs = sg_next(outs);
548 outs->dma_address = DMA_ERROR_CODE;
549 outs->dma_length = 0;
550 }
551
552 /* Make sure updates are seen by hardware */
553 mb();
554
555 return outcount;
556
557 failure:
558 for_each_sg(sglist, s, nelems, i) {
559 if (s->dma_length != 0) {
560 unsigned long vaddr, npages;
561
562 vaddr = s->dma_address & IOMMU_PAGE_MASK(tbl);
563 npages = iommu_num_pages(s->dma_address, s->dma_length,
564 IOMMU_PAGE_SIZE(tbl));
565 __iommu_free(tbl, vaddr, npages);
566 s->dma_address = DMA_ERROR_CODE;
567 s->dma_length = 0;
568 }
569 if (s == outs)
570 break;
571 }
572 return 0;
573}
574
575
576void ppc_iommu_unmap_sg(struct iommu_table *tbl, struct scatterlist *sglist,
577 int nelems, enum dma_data_direction direction,
578 unsigned long attrs)
579{
580 struct scatterlist *sg;
581
582 BUG_ON(direction == DMA_NONE);
583
584 if (!tbl)
585 return;
586
587 sg = sglist;
588 while (nelems--) {
589 unsigned int npages;
590 dma_addr_t dma_handle = sg->dma_address;
591
592 if (sg->dma_length == 0)
593 break;
594 npages = iommu_num_pages(dma_handle, sg->dma_length,
595 IOMMU_PAGE_SIZE(tbl));
596 __iommu_free(tbl, dma_handle, npages);
597 sg = sg_next(sg);
598 }
599
600 /* Flush/invalidate TLBs if necessary. As for iommu_free(), we
601 * do not do an mb() here, the affected platforms do not need it
602 * when freeing.
603 */
604 if (tbl->it_ops->flush)
605 tbl->it_ops->flush(tbl);
606}
607
608static void iommu_table_clear(struct iommu_table *tbl)
609{
610 /*
611 * In case of firmware assisted dump system goes through clean
612 * reboot process at the time of system crash. Hence it's safe to
613 * clear the TCE entries if firmware assisted dump is active.
614 */
615 if (!is_kdump_kernel() || is_fadump_active()) {
616 /* Clear the table in case firmware left allocations in it */
617 tbl->it_ops->clear(tbl, tbl->it_offset, tbl->it_size);
618 return;
619 }
620
621#ifdef CONFIG_CRASH_DUMP
622 if (tbl->it_ops->get) {
623 unsigned long index, tceval, tcecount = 0;
624
625 /* Reserve the existing mappings left by the first kernel. */
626 for (index = 0; index < tbl->it_size; index++) {
627 tceval = tbl->it_ops->get(tbl, index + tbl->it_offset);
628 /*
629 * Freed TCE entry contains 0x7fffffffffffffff on JS20
630 */
631 if (tceval && (tceval != 0x7fffffffffffffffUL)) {
632 __set_bit(index, tbl->it_map);
633 tcecount++;
634 }
635 }
636
637 if ((tbl->it_size - tcecount) < KDUMP_MIN_TCE_ENTRIES) {
638 printk(KERN_WARNING "TCE table is full; freeing ");
639 printk(KERN_WARNING "%d entries for the kdump boot\n",
640 KDUMP_MIN_TCE_ENTRIES);
641 for (index = tbl->it_size - KDUMP_MIN_TCE_ENTRIES;
642 index < tbl->it_size; index++)
643 __clear_bit(index, tbl->it_map);
644 }
645 }
646#endif
647}
648
649/*
650 * Build a iommu_table structure. This contains a bit map which
651 * is used to manage allocation of the tce space.
652 */
653struct iommu_table *iommu_init_table(struct iommu_table *tbl, int nid)
654{
655 unsigned long sz;
656 static int welcomed = 0;
657 struct page *page;
658 unsigned int i;
659 struct iommu_pool *p;
660
661 BUG_ON(!tbl->it_ops);
662
663 /* number of bytes needed for the bitmap */
664 sz = BITS_TO_LONGS(tbl->it_size) * sizeof(unsigned long);
665
666 page = alloc_pages_node(nid, GFP_KERNEL, get_order(sz));
667 if (!page)
668 panic("iommu_init_table: Can't allocate %ld bytes\n", sz);
669 tbl->it_map = page_address(page);
670 memset(tbl->it_map, 0, sz);
671
672 /*
673 * Reserve page 0 so it will not be used for any mappings.
674 * This avoids buggy drivers that consider page 0 to be invalid
675 * to crash the machine or even lose data.
676 */
677 if (tbl->it_offset == 0)
678 set_bit(0, tbl->it_map);
679
680 /* We only split the IOMMU table if we have 1GB or more of space */
681 if ((tbl->it_size << tbl->it_page_shift) >= (1UL * 1024 * 1024 * 1024))
682 tbl->nr_pools = IOMMU_NR_POOLS;
683 else
684 tbl->nr_pools = 1;
685
686 /* We reserve the top 1/4 of the table for large allocations */
687 tbl->poolsize = (tbl->it_size * 3 / 4) / tbl->nr_pools;
688
689 for (i = 0; i < tbl->nr_pools; i++) {
690 p = &tbl->pools[i];
691 spin_lock_init(&(p->lock));
692 p->start = tbl->poolsize * i;
693 p->hint = p->start;
694 p->end = p->start + tbl->poolsize;
695 }
696
697 p = &tbl->large_pool;
698 spin_lock_init(&(p->lock));
699 p->start = tbl->poolsize * i;
700 p->hint = p->start;
701 p->end = tbl->it_size;
702
703 iommu_table_clear(tbl);
704
705 if (!welcomed) {
706 printk(KERN_INFO "IOMMU table initialized, virtual merging %s\n",
707 novmerge ? "disabled" : "enabled");
708 welcomed = 1;
709 }
710
711 return tbl;
712}
713
714void iommu_free_table(struct iommu_table *tbl, const char *node_name)
715{
716 unsigned long bitmap_sz;
717 unsigned int order;
718
719 if (!tbl)
720 return;
721
722 if (!tbl->it_map) {
723 kfree(tbl);
724 return;
725 }
726
727 /*
728 * In case we have reserved the first bit, we should not emit
729 * the warning below.
730 */
731 if (tbl->it_offset == 0)
732 clear_bit(0, tbl->it_map);
733
734 /* verify that table contains no entries */
735 if (!bitmap_empty(tbl->it_map, tbl->it_size))
736 pr_warn("%s: Unexpected TCEs for %s\n", __func__, node_name);
737
738 /* calculate bitmap size in bytes */
739 bitmap_sz = BITS_TO_LONGS(tbl->it_size) * sizeof(unsigned long);
740
741 /* free bitmap */
742 order = get_order(bitmap_sz);
743 free_pages((unsigned long) tbl->it_map, order);
744
745 /* free table */
746 kfree(tbl);
747}
748
749/* Creates TCEs for a user provided buffer. The user buffer must be
750 * contiguous real kernel storage (not vmalloc). The address passed here
751 * comprises a page address and offset into that page. The dma_addr_t
752 * returned will point to the same byte within the page as was passed in.
753 */
754dma_addr_t iommu_map_page(struct device *dev, struct iommu_table *tbl,
755 struct page *page, unsigned long offset, size_t size,
756 unsigned long mask, enum dma_data_direction direction,
757 unsigned long attrs)
758{
759 dma_addr_t dma_handle = DMA_ERROR_CODE;
760 void *vaddr;
761 unsigned long uaddr;
762 unsigned int npages, align;
763
764 BUG_ON(direction == DMA_NONE);
765
766 vaddr = page_address(page) + offset;
767 uaddr = (unsigned long)vaddr;
768 npages = iommu_num_pages(uaddr, size, IOMMU_PAGE_SIZE(tbl));
769
770 if (tbl) {
771 align = 0;
772 if (tbl->it_page_shift < PAGE_SHIFT && size >= PAGE_SIZE &&
773 ((unsigned long)vaddr & ~PAGE_MASK) == 0)
774 align = PAGE_SHIFT - tbl->it_page_shift;
775
776 dma_handle = iommu_alloc(dev, tbl, vaddr, npages, direction,
777 mask >> tbl->it_page_shift, align,
778 attrs);
779 if (dma_handle == DMA_ERROR_CODE) {
780 if (!(attrs & DMA_ATTR_NO_WARN) &&
781 printk_ratelimit()) {
782 dev_info(dev, "iommu_alloc failed, tbl %p "
783 "vaddr %p npages %d\n", tbl, vaddr,
784 npages);
785 }
786 } else
787 dma_handle |= (uaddr & ~IOMMU_PAGE_MASK(tbl));
788 }
789
790 return dma_handle;
791}
792
793void iommu_unmap_page(struct iommu_table *tbl, dma_addr_t dma_handle,
794 size_t size, enum dma_data_direction direction,
795 unsigned long attrs)
796{
797 unsigned int npages;
798
799 BUG_ON(direction == DMA_NONE);
800
801 if (tbl) {
802 npages = iommu_num_pages(dma_handle, size,
803 IOMMU_PAGE_SIZE(tbl));
804 iommu_free(tbl, dma_handle, npages);
805 }
806}
807
808/* Allocates a contiguous real buffer and creates mappings over it.
809 * Returns the virtual address of the buffer and sets dma_handle
810 * to the dma address (mapping) of the first page.
811 */
812void *iommu_alloc_coherent(struct device *dev, struct iommu_table *tbl,
813 size_t size, dma_addr_t *dma_handle,
814 unsigned long mask, gfp_t flag, int node)
815{
816 void *ret = NULL;
817 dma_addr_t mapping;
818 unsigned int order;
819 unsigned int nio_pages, io_order;
820 struct page *page;
821
822 size = PAGE_ALIGN(size);
823 order = get_order(size);
824
825 /*
826 * Client asked for way too much space. This is checked later
827 * anyway. It is easier to debug here for the drivers than in
828 * the tce tables.
829 */
830 if (order >= IOMAP_MAX_ORDER) {
831 dev_info(dev, "iommu_alloc_consistent size too large: 0x%lx\n",
832 size);
833 return NULL;
834 }
835
836 if (!tbl)
837 return NULL;
838
839 /* Alloc enough pages (and possibly more) */
840 page = alloc_pages_node(node, flag, order);
841 if (!page)
842 return NULL;
843 ret = page_address(page);
844 memset(ret, 0, size);
845
846 /* Set up tces to cover the allocated range */
847 nio_pages = size >> tbl->it_page_shift;
848 io_order = get_iommu_order(size, tbl);
849 mapping = iommu_alloc(dev, tbl, ret, nio_pages, DMA_BIDIRECTIONAL,
850 mask >> tbl->it_page_shift, io_order, 0);
851 if (mapping == DMA_ERROR_CODE) {
852 free_pages((unsigned long)ret, order);
853 return NULL;
854 }
855 *dma_handle = mapping;
856 return ret;
857}
858
859void iommu_free_coherent(struct iommu_table *tbl, size_t size,
860 void *vaddr, dma_addr_t dma_handle)
861{
862 if (tbl) {
863 unsigned int nio_pages;
864
865 size = PAGE_ALIGN(size);
866 nio_pages = size >> tbl->it_page_shift;
867 iommu_free(tbl, dma_handle, nio_pages);
868 size = PAGE_ALIGN(size);
869 free_pages((unsigned long)vaddr, get_order(size));
870 }
871}
872
873unsigned long iommu_direction_to_tce_perm(enum dma_data_direction dir)
874{
875 switch (dir) {
876 case DMA_BIDIRECTIONAL:
877 return TCE_PCI_READ | TCE_PCI_WRITE;
878 case DMA_FROM_DEVICE:
879 return TCE_PCI_WRITE;
880 case DMA_TO_DEVICE:
881 return TCE_PCI_READ;
882 default:
883 return 0;
884 }
885}
886EXPORT_SYMBOL_GPL(iommu_direction_to_tce_perm);
887
888#ifdef CONFIG_IOMMU_API
889/*
890 * SPAPR TCE API
891 */
892static void group_release(void *iommu_data)
893{
894 struct iommu_table_group *table_group = iommu_data;
895
896 table_group->group = NULL;
897}
898
899void iommu_register_group(struct iommu_table_group *table_group,
900 int pci_domain_number, unsigned long pe_num)
901{
902 struct iommu_group *grp;
903 char *name;
904
905 grp = iommu_group_alloc();
906 if (IS_ERR(grp)) {
907 pr_warn("powerpc iommu api: cannot create new group, err=%ld\n",
908 PTR_ERR(grp));
909 return;
910 }
911 table_group->group = grp;
912 iommu_group_set_iommudata(grp, table_group, group_release);
913 name = kasprintf(GFP_KERNEL, "domain%d-pe%lx",
914 pci_domain_number, pe_num);
915 if (!name)
916 return;
917 iommu_group_set_name(grp, name);
918 kfree(name);
919}
920
921enum dma_data_direction iommu_tce_direction(unsigned long tce)
922{
923 if ((tce & TCE_PCI_READ) && (tce & TCE_PCI_WRITE))
924 return DMA_BIDIRECTIONAL;
925 else if (tce & TCE_PCI_READ)
926 return DMA_TO_DEVICE;
927 else if (tce & TCE_PCI_WRITE)
928 return DMA_FROM_DEVICE;
929 else
930 return DMA_NONE;
931}
932EXPORT_SYMBOL_GPL(iommu_tce_direction);
933
934void iommu_flush_tce(struct iommu_table *tbl)
935{
936 /* Flush/invalidate TLB caches if necessary */
937 if (tbl->it_ops->flush)
938 tbl->it_ops->flush(tbl);
939
940 /* Make sure updates are seen by hardware */
941 mb();
942}
943EXPORT_SYMBOL_GPL(iommu_flush_tce);
944
945int iommu_tce_clear_param_check(struct iommu_table *tbl,
946 unsigned long ioba, unsigned long tce_value,
947 unsigned long npages)
948{
949 /* tbl->it_ops->clear() does not support any value but 0 */
950 if (tce_value)
951 return -EINVAL;
952
953 if (ioba & ~IOMMU_PAGE_MASK(tbl))
954 return -EINVAL;
955
956 ioba >>= tbl->it_page_shift;
957 if (ioba < tbl->it_offset)
958 return -EINVAL;
959
960 if ((ioba + npages) > (tbl->it_offset + tbl->it_size))
961 return -EINVAL;
962
963 return 0;
964}
965EXPORT_SYMBOL_GPL(iommu_tce_clear_param_check);
966
967int iommu_tce_put_param_check(struct iommu_table *tbl,
968 unsigned long ioba, unsigned long tce)
969{
970 if (tce & ~IOMMU_PAGE_MASK(tbl))
971 return -EINVAL;
972
973 if (ioba & ~IOMMU_PAGE_MASK(tbl))
974 return -EINVAL;
975
976 ioba >>= tbl->it_page_shift;
977 if (ioba < tbl->it_offset)
978 return -EINVAL;
979
980 if ((ioba + 1) > (tbl->it_offset + tbl->it_size))
981 return -EINVAL;
982
983 return 0;
984}
985EXPORT_SYMBOL_GPL(iommu_tce_put_param_check);
986
987long iommu_tce_xchg(struct iommu_table *tbl, unsigned long entry,
988 unsigned long *hpa, enum dma_data_direction *direction)
989{
990 long ret;
991
992 ret = tbl->it_ops->exchange(tbl, entry, hpa, direction);
993
994 if (!ret && ((*direction == DMA_FROM_DEVICE) ||
995 (*direction == DMA_BIDIRECTIONAL)))
996 SetPageDirty(pfn_to_page(*hpa >> PAGE_SHIFT));
997
998 /* if (unlikely(ret))
999 pr_err("iommu_tce: %s failed on hwaddr=%lx ioba=%lx kva=%lx ret=%d\n",
1000 __func__, hwaddr, entry << tbl->it_page_shift,
1001 hwaddr, ret); */
1002
1003 return ret;
1004}
1005EXPORT_SYMBOL_GPL(iommu_tce_xchg);
1006
1007int iommu_take_ownership(struct iommu_table *tbl)
1008{
1009 unsigned long flags, i, sz = (tbl->it_size + 7) >> 3;
1010 int ret = 0;
1011
1012 /*
1013 * VFIO does not control TCE entries allocation and the guest
1014 * can write new TCEs on top of existing ones so iommu_tce_build()
1015 * must be able to release old pages. This functionality
1016 * requires exchange() callback defined so if it is not
1017 * implemented, we disallow taking ownership over the table.
1018 */
1019 if (!tbl->it_ops->exchange)
1020 return -EINVAL;
1021
1022 spin_lock_irqsave(&tbl->large_pool.lock, flags);
1023 for (i = 0; i < tbl->nr_pools; i++)
1024 spin_lock(&tbl->pools[i].lock);
1025
1026 if (tbl->it_offset == 0)
1027 clear_bit(0, tbl->it_map);
1028
1029 if (!bitmap_empty(tbl->it_map, tbl->it_size)) {
1030 pr_err("iommu_tce: it_map is not empty");
1031 ret = -EBUSY;
1032 /* Restore bit#0 set by iommu_init_table() */
1033 if (tbl->it_offset == 0)
1034 set_bit(0, tbl->it_map);
1035 } else {
1036 memset(tbl->it_map, 0xff, sz);
1037 }
1038
1039 for (i = 0; i < tbl->nr_pools; i++)
1040 spin_unlock(&tbl->pools[i].lock);
1041 spin_unlock_irqrestore(&tbl->large_pool.lock, flags);
1042
1043 return ret;
1044}
1045EXPORT_SYMBOL_GPL(iommu_take_ownership);
1046
1047void iommu_release_ownership(struct iommu_table *tbl)
1048{
1049 unsigned long flags, i, sz = (tbl->it_size + 7) >> 3;
1050
1051 spin_lock_irqsave(&tbl->large_pool.lock, flags);
1052 for (i = 0; i < tbl->nr_pools; i++)
1053 spin_lock(&tbl->pools[i].lock);
1054
1055 memset(tbl->it_map, 0, sz);
1056
1057 /* Restore bit#0 set by iommu_init_table() */
1058 if (tbl->it_offset == 0)
1059 set_bit(0, tbl->it_map);
1060
1061 for (i = 0; i < tbl->nr_pools; i++)
1062 spin_unlock(&tbl->pools[i].lock);
1063 spin_unlock_irqrestore(&tbl->large_pool.lock, flags);
1064}
1065EXPORT_SYMBOL_GPL(iommu_release_ownership);
1066
1067int iommu_add_device(struct device *dev)
1068{
1069 struct iommu_table *tbl;
1070 struct iommu_table_group_link *tgl;
1071
1072 /*
1073 * The sysfs entries should be populated before
1074 * binding IOMMU group. If sysfs entries isn't
1075 * ready, we simply bail.
1076 */
1077 if (!device_is_registered(dev))
1078 return -ENOENT;
1079
1080 if (dev->iommu_group) {
1081 pr_debug("%s: Skipping device %s with iommu group %d\n",
1082 __func__, dev_name(dev),
1083 iommu_group_id(dev->iommu_group));
1084 return -EBUSY;
1085 }
1086
1087 tbl = get_iommu_table_base(dev);
1088 if (!tbl) {
1089 pr_debug("%s: Skipping device %s with no tbl\n",
1090 __func__, dev_name(dev));
1091 return 0;
1092 }
1093
1094 tgl = list_first_entry_or_null(&tbl->it_group_list,
1095 struct iommu_table_group_link, next);
1096 if (!tgl) {
1097 pr_debug("%s: Skipping device %s with no group\n",
1098 __func__, dev_name(dev));
1099 return 0;
1100 }
1101 pr_debug("%s: Adding %s to iommu group %d\n",
1102 __func__, dev_name(dev),
1103 iommu_group_id(tgl->table_group->group));
1104
1105 if (PAGE_SIZE < IOMMU_PAGE_SIZE(tbl)) {
1106 pr_err("%s: Invalid IOMMU page size %lx (%lx) on %s\n",
1107 __func__, IOMMU_PAGE_SIZE(tbl),
1108 PAGE_SIZE, dev_name(dev));
1109 return -EINVAL;
1110 }
1111
1112 return iommu_group_add_device(tgl->table_group->group, dev);
1113}
1114EXPORT_SYMBOL_GPL(iommu_add_device);
1115
1116void iommu_del_device(struct device *dev)
1117{
1118 /*
1119 * Some devices might not have IOMMU table and group
1120 * and we needn't detach them from the associated
1121 * IOMMU groups
1122 */
1123 if (!dev->iommu_group) {
1124 pr_debug("iommu_tce: skipping device %s with no tbl\n",
1125 dev_name(dev));
1126 return;
1127 }
1128
1129 iommu_group_remove_device(dev);
1130}
1131EXPORT_SYMBOL_GPL(iommu_del_device);
1132
1133static int tce_iommu_bus_notifier(struct notifier_block *nb,
1134 unsigned long action, void *data)
1135{
1136 struct device *dev = data;
1137
1138 switch (action) {
1139 case BUS_NOTIFY_ADD_DEVICE:
1140 return iommu_add_device(dev);
1141 case BUS_NOTIFY_DEL_DEVICE:
1142 if (dev->iommu_group)
1143 iommu_del_device(dev);
1144 return 0;
1145 default:
1146 return 0;
1147 }
1148}
1149
1150static struct notifier_block tce_iommu_bus_nb = {
1151 .notifier_call = tce_iommu_bus_notifier,
1152};
1153
1154int __init tce_iommu_bus_notifier_init(void)
1155{
1156 bus_register_notifier(&pci_bus_type, &tce_iommu_bus_nb);
1157 return 0;
1158}
1159#endif /* CONFIG_IOMMU_API */