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			      struct dma_attrs *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 = ppc_md.tce_build(tbl, entry, npages,
326				      (unsigned long)page &
327				      IOMMU_PAGE_MASK(tbl), direction, attrs);
328
329	/* ppc_md.tce_build() 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 (ppc_md.tce_flush)
341		ppc_md.tce_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	ppc_md.tce_free(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 (ppc_md.tce_flush)
428		ppc_md.tce_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		     struct dma_attrs *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 (printk_ratelimit())
483				dev_info(dev, "iommu_alloc failed, tbl %p "
484					 "vaddr %lx npages %lu\n", tbl, vaddr,
485					 npages);
486			goto failure;
487		}
488
489		/* Convert entry to a dma_addr_t */
490		entry += tbl->it_offset;
491		dma_addr = entry << tbl->it_page_shift;
492		dma_addr |= (s->offset & ~IOMMU_PAGE_MASK(tbl));
493
494		DBG("  - %lu pages, entry: %lx, dma_addr: %lx\n",
495			    npages, entry, dma_addr);
496
497		/* Insert into HW table */
498		build_fail = ppc_md.tce_build(tbl, entry, npages,
499					      vaddr & IOMMU_PAGE_MASK(tbl),
500					      direction, attrs);
501		if(unlikely(build_fail))
502			goto failure;
503
504		/* If we are in an open segment, try merging */
505		if (segstart != s) {
506			DBG("  - trying merge...\n");
507			/* We cannot merge if:
508			 * - allocated dma_addr isn't contiguous to previous allocation
509			 */
510			if (novmerge || (dma_addr != dma_next) ||
511			    (outs->dma_length + s->length > max_seg_size)) {
512				/* Can't merge: create a new segment */
513				segstart = s;
514				outcount++;
515				outs = sg_next(outs);
516				DBG("    can't merge, new segment.\n");
517			} else {
518				outs->dma_length += s->length;
519				DBG("    merged, new len: %ux\n", outs->dma_length);
520			}
521		}
522
523		if (segstart == s) {
524			/* This is a new segment, fill entries */
525			DBG("  - filling new segment.\n");
526			outs->dma_address = dma_addr;
527			outs->dma_length = slen;
528		}
529
530		/* Calculate next page pointer for contiguous check */
531		dma_next = dma_addr + slen;
532
533		DBG("  - dma next is: %lx\n", dma_next);
534	}
535
536	/* Flush/invalidate TLB caches if necessary */
537	if (ppc_md.tce_flush)
538		ppc_md.tce_flush(tbl);
539
540	DBG("mapped %d elements:\n", outcount);
541
542	/* For the sake of ppc_iommu_unmap_sg, we clear out the length in the
543	 * next entry of the sglist if we didn't fill the list completely
544	 */
545	if (outcount < incount) {
546		outs = sg_next(outs);
547		outs->dma_address = DMA_ERROR_CODE;
548		outs->dma_length = 0;
549	}
550
551	/* Make sure updates are seen by hardware */
552	mb();
553
554	return outcount;
555
556 failure:
557	for_each_sg(sglist, s, nelems, i) {
558		if (s->dma_length != 0) {
559			unsigned long vaddr, npages;
560
561			vaddr = s->dma_address & IOMMU_PAGE_MASK(tbl);
562			npages = iommu_num_pages(s->dma_address, s->dma_length,
563						 IOMMU_PAGE_SIZE(tbl));
564			__iommu_free(tbl, vaddr, npages);
565			s->dma_address = DMA_ERROR_CODE;
566			s->dma_length = 0;
567		}
568		if (s == outs)
569			break;
570	}
571	return 0;
572}
573
574
575void ppc_iommu_unmap_sg(struct iommu_table *tbl, struct scatterlist *sglist,
576			int nelems, enum dma_data_direction direction,
577			struct dma_attrs *attrs)
578{
579	struct scatterlist *sg;
580
581	BUG_ON(direction == DMA_NONE);
582
583	if (!tbl)
584		return;
585
586	sg = sglist;
587	while (nelems--) {
588		unsigned int npages;
589		dma_addr_t dma_handle = sg->dma_address;
590
591		if (sg->dma_length == 0)
592			break;
593		npages = iommu_num_pages(dma_handle, sg->dma_length,
594					 IOMMU_PAGE_SIZE(tbl));
595		__iommu_free(tbl, dma_handle, npages);
596		sg = sg_next(sg);
597	}
598
599	/* Flush/invalidate TLBs if necessary. As for iommu_free(), we
600	 * do not do an mb() here, the affected platforms do not need it
601	 * when freeing.
602	 */
603	if (ppc_md.tce_flush)
604		ppc_md.tce_flush(tbl);
605}
606
607static void iommu_table_clear(struct iommu_table *tbl)
608{
609	/*
610	 * In case of firmware assisted dump system goes through clean
611	 * reboot process at the time of system crash. Hence it's safe to
612	 * clear the TCE entries if firmware assisted dump is active.
613	 */
614	if (!is_kdump_kernel() || is_fadump_active()) {
615		/* Clear the table in case firmware left allocations in it */
616		ppc_md.tce_free(tbl, tbl->it_offset, tbl->it_size);
617		return;
618	}
619
620#ifdef CONFIG_CRASH_DUMP
621	if (ppc_md.tce_get) {
622		unsigned long index, tceval, tcecount = 0;
623
624		/* Reserve the existing mappings left by the first kernel. */
625		for (index = 0; index < tbl->it_size; index++) {
626			tceval = ppc_md.tce_get(tbl, index + tbl->it_offset);
627			/*
628			 * Freed TCE entry contains 0x7fffffffffffffff on JS20
629			 */
630			if (tceval && (tceval != 0x7fffffffffffffffUL)) {
631				__set_bit(index, tbl->it_map);
632				tcecount++;
633			}
634		}
635
636		if ((tbl->it_size - tcecount) < KDUMP_MIN_TCE_ENTRIES) {
637			printk(KERN_WARNING "TCE table is full; freeing ");
638			printk(KERN_WARNING "%d entries for the kdump boot\n",
639				KDUMP_MIN_TCE_ENTRIES);
640			for (index = tbl->it_size - KDUMP_MIN_TCE_ENTRIES;
641				index < tbl->it_size; index++)
642				__clear_bit(index, tbl->it_map);
643		}
644	}
645#endif
646}
647
648/*
649 * Build a iommu_table structure.  This contains a bit map which
650 * is used to manage allocation of the tce space.
651 */
652struct iommu_table *iommu_init_table(struct iommu_table *tbl, int nid)
653{
654	unsigned long sz;
655	static int welcomed = 0;
656	struct page *page;
657	unsigned int i;
658	struct iommu_pool *p;
659
660	/* number of bytes needed for the bitmap */
661	sz = BITS_TO_LONGS(tbl->it_size) * sizeof(unsigned long);
662
663	page = alloc_pages_node(nid, GFP_KERNEL, get_order(sz));
664	if (!page)
665		panic("iommu_init_table: Can't allocate %ld bytes\n", sz);
666	tbl->it_map = page_address(page);
667	memset(tbl->it_map, 0, sz);
668
669	/*
670	 * Reserve page 0 so it will not be used for any mappings.
671	 * This avoids buggy drivers that consider page 0 to be invalid
672	 * to crash the machine or even lose data.
673	 */
674	if (tbl->it_offset == 0)
675		set_bit(0, tbl->it_map);
676
677	/* We only split the IOMMU table if we have 1GB or more of space */
678	if ((tbl->it_size << tbl->it_page_shift) >= (1UL * 1024 * 1024 * 1024))
679		tbl->nr_pools = IOMMU_NR_POOLS;
680	else
681		tbl->nr_pools = 1;
682
683	/* We reserve the top 1/4 of the table for large allocations */
684	tbl->poolsize = (tbl->it_size * 3 / 4) / tbl->nr_pools;
685
686	for (i = 0; i < tbl->nr_pools; i++) {
687		p = &tbl->pools[i];
688		spin_lock_init(&(p->lock));
689		p->start = tbl->poolsize * i;
690		p->hint = p->start;
691		p->end = p->start + tbl->poolsize;
692	}
693
694	p = &tbl->large_pool;
695	spin_lock_init(&(p->lock));
696	p->start = tbl->poolsize * i;
697	p->hint = p->start;
698	p->end = tbl->it_size;
699
700	iommu_table_clear(tbl);
701
702	if (!welcomed) {
703		printk(KERN_INFO "IOMMU table initialized, virtual merging %s\n",
704		       novmerge ? "disabled" : "enabled");
705		welcomed = 1;
706	}
707
708	return tbl;
709}
710
711void iommu_free_table(struct iommu_table *tbl, const char *node_name)
712{
713	unsigned long bitmap_sz;
714	unsigned int order;
715
716	if (!tbl || !tbl->it_map) {
717		printk(KERN_ERR "%s: expected TCE map for %s\n", __func__,
718				node_name);
719		return;
720	}
721
722	/*
723	 * In case we have reserved the first bit, we should not emit
724	 * the warning below.
725	 */
726	if (tbl->it_offset == 0)
727		clear_bit(0, tbl->it_map);
728
729#ifdef CONFIG_IOMMU_API
730	if (tbl->it_group) {
731		iommu_group_put(tbl->it_group);
732		BUG_ON(tbl->it_group);
733	}
734#endif
735
736	/* verify that table contains no entries */
737	if (!bitmap_empty(tbl->it_map, tbl->it_size))
738		pr_warn("%s: Unexpected TCEs for %s\n", __func__, node_name);
739
740	/* calculate bitmap size in bytes */
741	bitmap_sz = BITS_TO_LONGS(tbl->it_size) * sizeof(unsigned long);
742
743	/* free bitmap */
744	order = get_order(bitmap_sz);
745	free_pages((unsigned long) tbl->it_map, order);
746
747	/* free table */
748	kfree(tbl);
749}
750
751/* Creates TCEs for a user provided buffer.  The user buffer must be
752 * contiguous real kernel storage (not vmalloc).  The address passed here
753 * comprises a page address and offset into that page. The dma_addr_t
754 * returned will point to the same byte within the page as was passed in.
755 */
756dma_addr_t iommu_map_page(struct device *dev, struct iommu_table *tbl,
757			  struct page *page, unsigned long offset, size_t size,
758			  unsigned long mask, enum dma_data_direction direction,
759			  struct dma_attrs *attrs)
760{
761	dma_addr_t dma_handle = DMA_ERROR_CODE;
762	void *vaddr;
763	unsigned long uaddr;
764	unsigned int npages, align;
765
766	BUG_ON(direction == DMA_NONE);
767
768	vaddr = page_address(page) + offset;
769	uaddr = (unsigned long)vaddr;
770	npages = iommu_num_pages(uaddr, size, IOMMU_PAGE_SIZE(tbl));
771
772	if (tbl) {
773		align = 0;
774		if (tbl->it_page_shift < PAGE_SHIFT && size >= PAGE_SIZE &&
775		    ((unsigned long)vaddr & ~PAGE_MASK) == 0)
776			align = PAGE_SHIFT - tbl->it_page_shift;
777
778		dma_handle = iommu_alloc(dev, tbl, vaddr, npages, direction,
779					 mask >> tbl->it_page_shift, align,
780					 attrs);
781		if (dma_handle == DMA_ERROR_CODE) {
782			if (printk_ratelimit())  {
783				dev_info(dev, "iommu_alloc failed, tbl %p "
784					 "vaddr %p npages %d\n", tbl, vaddr,
785					 npages);
786			}
787		} else
788			dma_handle |= (uaddr & ~IOMMU_PAGE_MASK(tbl));
789	}
790
791	return dma_handle;
792}
793
794void iommu_unmap_page(struct iommu_table *tbl, dma_addr_t dma_handle,
795		      size_t size, enum dma_data_direction direction,
796		      struct dma_attrs *attrs)
797{
798	unsigned int npages;
799
800	BUG_ON(direction == DMA_NONE);
801
802	if (tbl) {
803		npages = iommu_num_pages(dma_handle, size,
804					 IOMMU_PAGE_SIZE(tbl));
805		iommu_free(tbl, dma_handle, npages);
806	}
807}
808
809/* Allocates a contiguous real buffer and creates mappings over it.
810 * Returns the virtual address of the buffer and sets dma_handle
811 * to the dma address (mapping) of the first page.
812 */
813void *iommu_alloc_coherent(struct device *dev, struct iommu_table *tbl,
814			   size_t size,	dma_addr_t *dma_handle,
815			   unsigned long mask, gfp_t flag, int node)
816{
817	void *ret = NULL;
818	dma_addr_t mapping;
819	unsigned int order;
820	unsigned int nio_pages, io_order;
821	struct page *page;
822
823	size = PAGE_ALIGN(size);
824	order = get_order(size);
825
826 	/*
827	 * Client asked for way too much space.  This is checked later
828	 * anyway.  It is easier to debug here for the drivers than in
829	 * the tce tables.
830	 */
831	if (order >= IOMAP_MAX_ORDER) {
832		dev_info(dev, "iommu_alloc_consistent size too large: 0x%lx\n",
833			 size);
834		return NULL;
835	}
836
837	if (!tbl)
838		return NULL;
839
840	/* Alloc enough pages (and possibly more) */
841	page = alloc_pages_node(node, flag, order);
842	if (!page)
843		return NULL;
844	ret = page_address(page);
845	memset(ret, 0, size);
846
847	/* Set up tces to cover the allocated range */
848	nio_pages = size >> tbl->it_page_shift;
849	io_order = get_iommu_order(size, tbl);
850	mapping = iommu_alloc(dev, tbl, ret, nio_pages, DMA_BIDIRECTIONAL,
851			      mask >> tbl->it_page_shift, io_order, NULL);
852	if (mapping == DMA_ERROR_CODE) {
853		free_pages((unsigned long)ret, order);
854		return NULL;
855	}
856	*dma_handle = mapping;
857	return ret;
858}
859
860void iommu_free_coherent(struct iommu_table *tbl, size_t size,
861			 void *vaddr, dma_addr_t dma_handle)
862{
863	if (tbl) {
864		unsigned int nio_pages;
865
866		size = PAGE_ALIGN(size);
867		nio_pages = size >> tbl->it_page_shift;
868		iommu_free(tbl, dma_handle, nio_pages);
869		size = PAGE_ALIGN(size);
870		free_pages((unsigned long)vaddr, get_order(size));
871	}
872}
873
874#ifdef CONFIG_IOMMU_API
875/*
876 * SPAPR TCE API
877 */
878static void group_release(void *iommu_data)
879{
880	struct iommu_table *tbl = iommu_data;
881	tbl->it_group = NULL;
882}
883
884void iommu_register_group(struct iommu_table *tbl,
885		int pci_domain_number, unsigned long pe_num)
886{
887	struct iommu_group *grp;
888	char *name;
889
890	grp = iommu_group_alloc();
891	if (IS_ERR(grp)) {
892		pr_warn("powerpc iommu api: cannot create new group, err=%ld\n",
893				PTR_ERR(grp));
894		return;
895	}
896	tbl->it_group = grp;
897	iommu_group_set_iommudata(grp, tbl, group_release);
898	name = kasprintf(GFP_KERNEL, "domain%d-pe%lx",
899			pci_domain_number, pe_num);
900	if (!name)
901		return;
902	iommu_group_set_name(grp, name);
903	kfree(name);
904}
905
906enum dma_data_direction iommu_tce_direction(unsigned long tce)
907{
908	if ((tce & TCE_PCI_READ) && (tce & TCE_PCI_WRITE))
909		return DMA_BIDIRECTIONAL;
910	else if (tce & TCE_PCI_READ)
911		return DMA_TO_DEVICE;
912	else if (tce & TCE_PCI_WRITE)
913		return DMA_FROM_DEVICE;
914	else
915		return DMA_NONE;
916}
917EXPORT_SYMBOL_GPL(iommu_tce_direction);
918
919void iommu_flush_tce(struct iommu_table *tbl)
920{
921	/* Flush/invalidate TLB caches if necessary */
922	if (ppc_md.tce_flush)
923		ppc_md.tce_flush(tbl);
924
925	/* Make sure updates are seen by hardware */
926	mb();
927}
928EXPORT_SYMBOL_GPL(iommu_flush_tce);
929
930int iommu_tce_clear_param_check(struct iommu_table *tbl,
931		unsigned long ioba, unsigned long tce_value,
932		unsigned long npages)
933{
934	/* ppc_md.tce_free() does not support any value but 0 */
935	if (tce_value)
936		return -EINVAL;
937
938	if (ioba & ~IOMMU_PAGE_MASK(tbl))
939		return -EINVAL;
940
941	ioba >>= tbl->it_page_shift;
942	if (ioba < tbl->it_offset)
943		return -EINVAL;
944
945	if ((ioba + npages) > (tbl->it_offset + tbl->it_size))
946		return -EINVAL;
947
948	return 0;
949}
950EXPORT_SYMBOL_GPL(iommu_tce_clear_param_check);
951
952int iommu_tce_put_param_check(struct iommu_table *tbl,
953		unsigned long ioba, unsigned long tce)
954{
955	if (!(tce & (TCE_PCI_WRITE | TCE_PCI_READ)))
956		return -EINVAL;
957
958	if (tce & ~(IOMMU_PAGE_MASK(tbl) | TCE_PCI_WRITE | TCE_PCI_READ))
959		return -EINVAL;
960
961	if (ioba & ~IOMMU_PAGE_MASK(tbl))
962		return -EINVAL;
963
964	ioba >>= tbl->it_page_shift;
965	if (ioba < tbl->it_offset)
966		return -EINVAL;
967
968	if ((ioba + 1) > (tbl->it_offset + tbl->it_size))
969		return -EINVAL;
970
971	return 0;
972}
973EXPORT_SYMBOL_GPL(iommu_tce_put_param_check);
974
975unsigned long iommu_clear_tce(struct iommu_table *tbl, unsigned long entry)
976{
977	unsigned long oldtce;
978	struct iommu_pool *pool = get_pool(tbl, entry);
979
980	spin_lock(&(pool->lock));
981
982	oldtce = ppc_md.tce_get(tbl, entry);
983	if (oldtce & (TCE_PCI_WRITE | TCE_PCI_READ))
984		ppc_md.tce_free(tbl, entry, 1);
985	else
986		oldtce = 0;
987
988	spin_unlock(&(pool->lock));
989
990	return oldtce;
991}
992EXPORT_SYMBOL_GPL(iommu_clear_tce);
993
994int iommu_clear_tces_and_put_pages(struct iommu_table *tbl,
995		unsigned long entry, unsigned long pages)
996{
997	unsigned long oldtce;
998	struct page *page;
999
1000	for ( ; pages; --pages, ++entry) {
1001		oldtce = iommu_clear_tce(tbl, entry);
1002		if (!oldtce)
1003			continue;
1004
1005		page = pfn_to_page(oldtce >> PAGE_SHIFT);
1006		WARN_ON(!page);
1007		if (page) {
1008			if (oldtce & TCE_PCI_WRITE)
1009				SetPageDirty(page);
1010			put_page(page);
1011		}
1012	}
1013
1014	return 0;
1015}
1016EXPORT_SYMBOL_GPL(iommu_clear_tces_and_put_pages);
1017
1018/*
1019 * hwaddr is a kernel virtual address here (0xc... bazillion),
1020 * tce_build converts it to a physical address.
1021 */
1022int iommu_tce_build(struct iommu_table *tbl, unsigned long entry,
1023		unsigned long hwaddr, enum dma_data_direction direction)
1024{
1025	int ret = -EBUSY;
1026	unsigned long oldtce;
1027	struct iommu_pool *pool = get_pool(tbl, entry);
1028
1029	spin_lock(&(pool->lock));
1030
1031	oldtce = ppc_md.tce_get(tbl, entry);
1032	/* Add new entry if it is not busy */
1033	if (!(oldtce & (TCE_PCI_WRITE | TCE_PCI_READ)))
1034		ret = ppc_md.tce_build(tbl, entry, 1, hwaddr, direction, NULL);
1035
1036	spin_unlock(&(pool->lock));
1037
1038	/* if (unlikely(ret))
1039		pr_err("iommu_tce: %s failed on hwaddr=%lx ioba=%lx kva=%lx ret=%d\n",
1040			__func__, hwaddr, entry << tbl->it_page_shift,
1041				hwaddr, ret); */
1042
1043	return ret;
1044}
1045EXPORT_SYMBOL_GPL(iommu_tce_build);
1046
1047int iommu_put_tce_user_mode(struct iommu_table *tbl, unsigned long entry,
1048		unsigned long tce)
1049{
1050	int ret;
1051	struct page *page = NULL;
1052	unsigned long hwaddr, offset = tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
1053	enum dma_data_direction direction = iommu_tce_direction(tce);
1054
1055	ret = get_user_pages_fast(tce & PAGE_MASK, 1,
1056			direction != DMA_TO_DEVICE, &page);
1057	if (unlikely(ret != 1)) {
1058		/* pr_err("iommu_tce: get_user_pages_fast failed tce=%lx ioba=%lx ret=%d\n",
1059				tce, entry << tbl->it_page_shift, ret); */
1060		return -EFAULT;
1061	}
1062	hwaddr = (unsigned long) page_address(page) + offset;
1063
1064	ret = iommu_tce_build(tbl, entry, hwaddr, direction);
1065	if (ret)
1066		put_page(page);
1067
1068	if (ret < 0)
1069		pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%d\n",
1070			__func__, entry << tbl->it_page_shift, tce, ret);
1071
1072	return ret;
1073}
1074EXPORT_SYMBOL_GPL(iommu_put_tce_user_mode);
1075
1076int iommu_take_ownership(struct iommu_table *tbl)
1077{
1078	unsigned long sz = (tbl->it_size + 7) >> 3;
1079
1080	if (tbl->it_offset == 0)
1081		clear_bit(0, tbl->it_map);
1082
1083	if (!bitmap_empty(tbl->it_map, tbl->it_size)) {
1084		pr_err("iommu_tce: it_map is not empty");
1085		return -EBUSY;
1086	}
1087
1088	memset(tbl->it_map, 0xff, sz);
1089	iommu_clear_tces_and_put_pages(tbl, tbl->it_offset, tbl->it_size);
1090
1091	/*
1092	 * Disable iommu bypass, otherwise the user can DMA to all of
1093	 * our physical memory via the bypass window instead of just
1094	 * the pages that has been explicitly mapped into the iommu
1095	 */
1096	if (tbl->set_bypass)
1097		tbl->set_bypass(tbl, false);
1098
1099	return 0;
1100}
1101EXPORT_SYMBOL_GPL(iommu_take_ownership);
1102
1103void iommu_release_ownership(struct iommu_table *tbl)
1104{
1105	unsigned long sz = (tbl->it_size + 7) >> 3;
1106
1107	iommu_clear_tces_and_put_pages(tbl, tbl->it_offset, tbl->it_size);
1108	memset(tbl->it_map, 0, sz);
1109
1110	/* Restore bit#0 set by iommu_init_table() */
1111	if (tbl->it_offset == 0)
1112		set_bit(0, tbl->it_map);
1113
1114	/* The kernel owns the device now, we can restore the iommu bypass */
1115	if (tbl->set_bypass)
1116		tbl->set_bypass(tbl, true);
1117}
1118EXPORT_SYMBOL_GPL(iommu_release_ownership);
1119
1120int iommu_add_device(struct device *dev)
1121{
1122	struct iommu_table *tbl;
1123
1124	/*
1125	 * The sysfs entries should be populated before
1126	 * binding IOMMU group. If sysfs entries isn't
1127	 * ready, we simply bail.
1128	 */
1129	if (!device_is_registered(dev))
1130		return -ENOENT;
1131
1132	if (dev->iommu_group) {
1133		pr_debug("%s: Skipping device %s with iommu group %d\n",
1134			 __func__, dev_name(dev),
1135			 iommu_group_id(dev->iommu_group));
1136		return -EBUSY;
1137	}
1138
1139	tbl = get_iommu_table_base(dev);
1140	if (!tbl || !tbl->it_group) {
1141		pr_debug("%s: Skipping device %s with no tbl\n",
1142			 __func__, dev_name(dev));
1143		return 0;
1144	}
1145
1146	pr_debug("%s: Adding %s to iommu group %d\n",
1147		 __func__, dev_name(dev),
1148		 iommu_group_id(tbl->it_group));
1149
1150	if (PAGE_SIZE < IOMMU_PAGE_SIZE(tbl)) {
1151		pr_err("%s: Invalid IOMMU page size %lx (%lx) on %s\n",
1152		       __func__, IOMMU_PAGE_SIZE(tbl),
1153		       PAGE_SIZE, dev_name(dev));
1154		return -EINVAL;
1155	}
1156
1157	return iommu_group_add_device(tbl->it_group, dev);
1158}
1159EXPORT_SYMBOL_GPL(iommu_add_device);
1160
1161void iommu_del_device(struct device *dev)
1162{
1163	/*
1164	 * Some devices might not have IOMMU table and group
1165	 * and we needn't detach them from the associated
1166	 * IOMMU groups
1167	 */
1168	if (!dev->iommu_group) {
1169		pr_debug("iommu_tce: skipping device %s with no tbl\n",
1170			 dev_name(dev));
1171		return;
1172	}
1173
1174	iommu_group_remove_device(dev);
1175}
1176EXPORT_SYMBOL_GPL(iommu_del_device);
1177
1178static int tce_iommu_bus_notifier(struct notifier_block *nb,
1179                unsigned long action, void *data)
1180{
1181        struct device *dev = data;
1182
1183        switch (action) {
1184        case BUS_NOTIFY_ADD_DEVICE:
1185                return iommu_add_device(dev);
1186        case BUS_NOTIFY_DEL_DEVICE:
1187                if (dev->iommu_group)
1188                        iommu_del_device(dev);
1189                return 0;
1190        default:
1191                return 0;
1192        }
1193}
1194
1195static struct notifier_block tce_iommu_bus_nb = {
1196        .notifier_call = tce_iommu_bus_notifier,
1197};
1198
1199int __init tce_iommu_bus_notifier_init(void)
1200{
1201        bus_register_notifier(&pci_bus_type, &tce_iommu_bus_nb);
1202        return 0;
1203}
1204#endif /* CONFIG_IOMMU_API */
1205