xref: /linux/arch/powerpc/kernel/iommu.c (revision 2a6b6c9a226279b4f6668450ddb21ae655558087)
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 <linux/vmalloc.h>
30 #include <asm/io.h>
31 #include <asm/iommu.h>
32 #include <asm/pci-bridge.h>
33 #include <asm/machdep.h>
34 #include <asm/kdump.h>
35 #include <asm/fadump.h>
36 #include <asm/vio.h>
37 #include <asm/tce.h>
38 #include <asm/mmu_context.h>
39 #include <asm/ppc-pci.h>
40 
41 #define DBG(...)
42 
43 #ifdef CONFIG_IOMMU_DEBUGFS
44 static int iommu_debugfs_weight_get(void *data, u64 *val)
45 {
46 	struct iommu_table *tbl = data;
47 	*val = bitmap_weight(tbl->it_map, tbl->it_size);
48 	return 0;
49 }
50 DEFINE_DEBUGFS_ATTRIBUTE(iommu_debugfs_fops_weight, iommu_debugfs_weight_get, NULL, "%llu\n");
51 
52 static void iommu_debugfs_add(struct iommu_table *tbl)
53 {
54 	char name[10];
55 	struct dentry *liobn_entry;
56 
57 	sprintf(name, "%08lx", tbl->it_index);
58 	liobn_entry = debugfs_create_dir(name, iommu_debugfs_dir);
59 
60 	debugfs_create_file_unsafe("weight", 0400, liobn_entry, tbl, &iommu_debugfs_fops_weight);
61 	debugfs_create_ulong("it_size", 0400, liobn_entry, &tbl->it_size);
62 	debugfs_create_ulong("it_page_shift", 0400, liobn_entry, &tbl->it_page_shift);
63 	debugfs_create_ulong("it_reserved_start", 0400, liobn_entry, &tbl->it_reserved_start);
64 	debugfs_create_ulong("it_reserved_end", 0400, liobn_entry, &tbl->it_reserved_end);
65 	debugfs_create_ulong("it_indirect_levels", 0400, liobn_entry, &tbl->it_indirect_levels);
66 	debugfs_create_ulong("it_level_size", 0400, liobn_entry, &tbl->it_level_size);
67 }
68 
69 static void iommu_debugfs_del(struct iommu_table *tbl)
70 {
71 	char name[10];
72 
73 	sprintf(name, "%08lx", tbl->it_index);
74 	debugfs_lookup_and_remove(name, iommu_debugfs_dir);
75 }
76 #else
77 static void iommu_debugfs_add(struct iommu_table *tbl){}
78 static void iommu_debugfs_del(struct iommu_table *tbl){}
79 #endif
80 
81 static int novmerge;
82 
83 static void __iommu_free(struct iommu_table *, dma_addr_t, unsigned int);
84 
85 static 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 
96 static 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  */
105 static 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 }
114 subsys_initcall(setup_iommu_pool_hash);
115 
116 #ifdef CONFIG_FAIL_IOMMU
117 
118 static DECLARE_FAULT_ATTR(fail_iommu);
119 
120 static 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 
126 static bool should_fail_iommu(struct device *dev)
127 {
128 	return dev->archdata.fail_iommu && should_fail(&fail_iommu, 1);
129 }
130 
131 static 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 }
138 late_initcall(fail_iommu_debugfs);
139 
140 static 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 
146 static 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 
158 static DEVICE_ATTR_RW(fail_iommu);
159 
160 static 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 
176 /*
177  * PCI and VIO buses need separate notifier_block structs, since they're linked
178  * list nodes.  Sharing a notifier_block would mean that any notifiers later
179  * registered for PCI buses would also get called by VIO buses and vice versa.
180  */
181 static struct notifier_block fail_iommu_pci_bus_notifier = {
182 	.notifier_call = fail_iommu_bus_notify
183 };
184 
185 #ifdef CONFIG_IBMVIO
186 static struct notifier_block fail_iommu_vio_bus_notifier = {
187 	.notifier_call = fail_iommu_bus_notify
188 };
189 #endif
190 
191 static int __init fail_iommu_setup(void)
192 {
193 #ifdef CONFIG_PCI
194 	bus_register_notifier(&pci_bus_type, &fail_iommu_pci_bus_notifier);
195 #endif
196 #ifdef CONFIG_IBMVIO
197 	bus_register_notifier(&vio_bus_type, &fail_iommu_vio_bus_notifier);
198 #endif
199 
200 	return 0;
201 }
202 /*
203  * Must execute after PCI and VIO subsystem have initialised but before
204  * devices are probed.
205  */
206 arch_initcall(fail_iommu_setup);
207 #else
208 static inline bool should_fail_iommu(struct device *dev)
209 {
210 	return false;
211 }
212 #endif
213 
214 static unsigned long iommu_range_alloc(struct device *dev,
215 				       struct iommu_table *tbl,
216                                        unsigned long npages,
217                                        unsigned long *handle,
218                                        unsigned long mask,
219                                        unsigned int align_order)
220 {
221 	unsigned long n, end, start;
222 	unsigned long limit;
223 	int largealloc = npages > 15;
224 	int pass = 0;
225 	unsigned long align_mask;
226 	unsigned long flags;
227 	unsigned int pool_nr;
228 	struct iommu_pool *pool;
229 
230 	align_mask = (1ull << align_order) - 1;
231 
232 	/* This allocator was derived from x86_64's bit string search */
233 
234 	/* Sanity check */
235 	if (unlikely(npages == 0)) {
236 		if (printk_ratelimit())
237 			WARN_ON(1);
238 		return DMA_MAPPING_ERROR;
239 	}
240 
241 	if (should_fail_iommu(dev))
242 		return DMA_MAPPING_ERROR;
243 
244 	/*
245 	 * We don't need to disable preemption here because any CPU can
246 	 * safely use any IOMMU pool.
247 	 */
248 	pool_nr = raw_cpu_read(iommu_pool_hash) & (tbl->nr_pools - 1);
249 
250 	if (largealloc)
251 		pool = &(tbl->large_pool);
252 	else
253 		pool = &(tbl->pools[pool_nr]);
254 
255 	spin_lock_irqsave(&(pool->lock), flags);
256 
257 again:
258 	if ((pass == 0) && handle && *handle &&
259 	    (*handle >= pool->start) && (*handle < pool->end))
260 		start = *handle;
261 	else
262 		start = pool->hint;
263 
264 	limit = pool->end;
265 
266 	/* The case below can happen if we have a small segment appended
267 	 * to a large, or when the previous alloc was at the very end of
268 	 * the available space. If so, go back to the initial start.
269 	 */
270 	if (start >= limit)
271 		start = pool->start;
272 
273 	if (limit + tbl->it_offset > mask) {
274 		limit = mask - tbl->it_offset + 1;
275 		/* If we're constrained on address range, first try
276 		 * at the masked hint to avoid O(n) search complexity,
277 		 * but on second pass, start at 0 in pool 0.
278 		 */
279 		if ((start & mask) >= limit || pass > 0) {
280 			spin_unlock(&(pool->lock));
281 			pool = &(tbl->pools[0]);
282 			spin_lock(&(pool->lock));
283 			start = pool->start;
284 		} else {
285 			start &= mask;
286 		}
287 	}
288 
289 	n = iommu_area_alloc(tbl->it_map, limit, start, npages, tbl->it_offset,
290 			dma_get_seg_boundary_nr_pages(dev, tbl->it_page_shift),
291 			align_mask);
292 	if (n == -1) {
293 		if (likely(pass == 0)) {
294 			/* First try the pool from the start */
295 			pool->hint = pool->start;
296 			pass++;
297 			goto again;
298 
299 		} else if (pass <= tbl->nr_pools) {
300 			/* Now try scanning all the other pools */
301 			spin_unlock(&(pool->lock));
302 			pool_nr = (pool_nr + 1) & (tbl->nr_pools - 1);
303 			pool = &tbl->pools[pool_nr];
304 			spin_lock(&(pool->lock));
305 			pool->hint = pool->start;
306 			pass++;
307 			goto again;
308 
309 		} else if (pass == tbl->nr_pools + 1) {
310 			/* Last resort: try largepool */
311 			spin_unlock(&pool->lock);
312 			pool = &tbl->large_pool;
313 			spin_lock(&pool->lock);
314 			pool->hint = pool->start;
315 			pass++;
316 			goto again;
317 
318 		} else {
319 			/* Give up */
320 			spin_unlock_irqrestore(&(pool->lock), flags);
321 			return DMA_MAPPING_ERROR;
322 		}
323 	}
324 
325 	end = n + npages;
326 
327 	/* Bump the hint to a new block for small allocs. */
328 	if (largealloc) {
329 		/* Don't bump to new block to avoid fragmentation */
330 		pool->hint = end;
331 	} else {
332 		/* Overflow will be taken care of at the next allocation */
333 		pool->hint = (end + tbl->it_blocksize - 1) &
334 		                ~(tbl->it_blocksize - 1);
335 	}
336 
337 	/* Update handle for SG allocations */
338 	if (handle)
339 		*handle = end;
340 
341 	spin_unlock_irqrestore(&(pool->lock), flags);
342 
343 	return n;
344 }
345 
346 static dma_addr_t iommu_alloc(struct device *dev, struct iommu_table *tbl,
347 			      void *page, unsigned int npages,
348 			      enum dma_data_direction direction,
349 			      unsigned long mask, unsigned int align_order,
350 			      unsigned long attrs)
351 {
352 	unsigned long entry;
353 	dma_addr_t ret = DMA_MAPPING_ERROR;
354 	int build_fail;
355 
356 	entry = iommu_range_alloc(dev, tbl, npages, NULL, mask, align_order);
357 
358 	if (unlikely(entry == DMA_MAPPING_ERROR))
359 		return DMA_MAPPING_ERROR;
360 
361 	entry += tbl->it_offset;	/* Offset into real TCE table */
362 	ret = entry << tbl->it_page_shift;	/* Set the return dma address */
363 
364 	/* Put the TCEs in the HW table */
365 	build_fail = tbl->it_ops->set(tbl, entry, npages,
366 				      (unsigned long)page &
367 				      IOMMU_PAGE_MASK(tbl), direction, attrs);
368 
369 	/* tbl->it_ops->set() only returns non-zero for transient errors.
370 	 * Clean up the table bitmap in this case and return
371 	 * DMA_MAPPING_ERROR. For all other errors the functionality is
372 	 * not altered.
373 	 */
374 	if (unlikely(build_fail)) {
375 		__iommu_free(tbl, ret, npages);
376 		return DMA_MAPPING_ERROR;
377 	}
378 
379 	/* Flush/invalidate TLB caches if necessary */
380 	if (tbl->it_ops->flush)
381 		tbl->it_ops->flush(tbl);
382 
383 	/* Make sure updates are seen by hardware */
384 	mb();
385 
386 	return ret;
387 }
388 
389 static bool iommu_free_check(struct iommu_table *tbl, dma_addr_t dma_addr,
390 			     unsigned int npages)
391 {
392 	unsigned long entry, free_entry;
393 
394 	entry = dma_addr >> tbl->it_page_shift;
395 	free_entry = entry - tbl->it_offset;
396 
397 	if (((free_entry + npages) > tbl->it_size) ||
398 	    (entry < tbl->it_offset)) {
399 		if (printk_ratelimit()) {
400 			printk(KERN_INFO "iommu_free: invalid entry\n");
401 			printk(KERN_INFO "\tentry     = 0x%lx\n", entry);
402 			printk(KERN_INFO "\tdma_addr  = 0x%llx\n", (u64)dma_addr);
403 			printk(KERN_INFO "\tTable     = 0x%llx\n", (u64)tbl);
404 			printk(KERN_INFO "\tbus#      = 0x%llx\n", (u64)tbl->it_busno);
405 			printk(KERN_INFO "\tsize      = 0x%llx\n", (u64)tbl->it_size);
406 			printk(KERN_INFO "\tstartOff  = 0x%llx\n", (u64)tbl->it_offset);
407 			printk(KERN_INFO "\tindex     = 0x%llx\n", (u64)tbl->it_index);
408 			WARN_ON(1);
409 		}
410 
411 		return false;
412 	}
413 
414 	return true;
415 }
416 
417 static struct iommu_pool *get_pool(struct iommu_table *tbl,
418 				   unsigned long entry)
419 {
420 	struct iommu_pool *p;
421 	unsigned long largepool_start = tbl->large_pool.start;
422 
423 	/* The large pool is the last pool at the top of the table */
424 	if (entry >= largepool_start) {
425 		p = &tbl->large_pool;
426 	} else {
427 		unsigned int pool_nr = entry / tbl->poolsize;
428 
429 		BUG_ON(pool_nr > tbl->nr_pools);
430 		p = &tbl->pools[pool_nr];
431 	}
432 
433 	return p;
434 }
435 
436 static void __iommu_free(struct iommu_table *tbl, dma_addr_t dma_addr,
437 			 unsigned int npages)
438 {
439 	unsigned long entry, free_entry;
440 	unsigned long flags;
441 	struct iommu_pool *pool;
442 
443 	entry = dma_addr >> tbl->it_page_shift;
444 	free_entry = entry - tbl->it_offset;
445 
446 	pool = get_pool(tbl, free_entry);
447 
448 	if (!iommu_free_check(tbl, dma_addr, npages))
449 		return;
450 
451 	tbl->it_ops->clear(tbl, entry, npages);
452 
453 	spin_lock_irqsave(&(pool->lock), flags);
454 	bitmap_clear(tbl->it_map, free_entry, npages);
455 	spin_unlock_irqrestore(&(pool->lock), flags);
456 }
457 
458 static void iommu_free(struct iommu_table *tbl, dma_addr_t dma_addr,
459 		unsigned int npages)
460 {
461 	__iommu_free(tbl, dma_addr, npages);
462 
463 	/* Make sure TLB cache is flushed if the HW needs it. We do
464 	 * not do an mb() here on purpose, it is not needed on any of
465 	 * the current platforms.
466 	 */
467 	if (tbl->it_ops->flush)
468 		tbl->it_ops->flush(tbl);
469 }
470 
471 int ppc_iommu_map_sg(struct device *dev, struct iommu_table *tbl,
472 		     struct scatterlist *sglist, int nelems,
473 		     unsigned long mask, enum dma_data_direction direction,
474 		     unsigned long attrs)
475 {
476 	dma_addr_t dma_next = 0, dma_addr;
477 	struct scatterlist *s, *outs, *segstart;
478 	int outcount, incount, i, build_fail = 0;
479 	unsigned int align;
480 	unsigned long handle;
481 	unsigned int max_seg_size;
482 
483 	BUG_ON(direction == DMA_NONE);
484 
485 	if ((nelems == 0) || !tbl)
486 		return -EINVAL;
487 
488 	outs = s = segstart = &sglist[0];
489 	outcount = 1;
490 	incount = nelems;
491 	handle = 0;
492 
493 	/* Init first segment length for backout at failure */
494 	outs->dma_length = 0;
495 
496 	DBG("sg mapping %d elements:\n", nelems);
497 
498 	max_seg_size = dma_get_max_seg_size(dev);
499 	for_each_sg(sglist, s, nelems, i) {
500 		unsigned long vaddr, npages, entry, slen;
501 
502 		slen = s->length;
503 		/* Sanity check */
504 		if (slen == 0) {
505 			dma_next = 0;
506 			continue;
507 		}
508 		/* Allocate iommu entries for that segment */
509 		vaddr = (unsigned long) sg_virt(s);
510 		npages = iommu_num_pages(vaddr, slen, IOMMU_PAGE_SIZE(tbl));
511 		align = 0;
512 		if (tbl->it_page_shift < PAGE_SHIFT && slen >= PAGE_SIZE &&
513 		    (vaddr & ~PAGE_MASK) == 0)
514 			align = PAGE_SHIFT - tbl->it_page_shift;
515 		entry = iommu_range_alloc(dev, tbl, npages, &handle,
516 					  mask >> tbl->it_page_shift, align);
517 
518 		DBG("  - vaddr: %lx, size: %lx\n", vaddr, slen);
519 
520 		/* Handle failure */
521 		if (unlikely(entry == DMA_MAPPING_ERROR)) {
522 			if (!(attrs & DMA_ATTR_NO_WARN) &&
523 			    printk_ratelimit())
524 				dev_info(dev, "iommu_alloc failed, tbl %p "
525 					 "vaddr %lx npages %lu\n", tbl, vaddr,
526 					 npages);
527 			goto failure;
528 		}
529 
530 		/* Convert entry to a dma_addr_t */
531 		entry += tbl->it_offset;
532 		dma_addr = entry << tbl->it_page_shift;
533 		dma_addr |= (vaddr & ~IOMMU_PAGE_MASK(tbl));
534 
535 		DBG("  - %lu pages, entry: %lx, dma_addr: %lx\n",
536 			    npages, entry, dma_addr);
537 
538 		/* Insert into HW table */
539 		build_fail = tbl->it_ops->set(tbl, entry, npages,
540 					      vaddr & IOMMU_PAGE_MASK(tbl),
541 					      direction, attrs);
542 		if(unlikely(build_fail))
543 			goto failure;
544 
545 		/* If we are in an open segment, try merging */
546 		if (segstart != s) {
547 			DBG("  - trying merge...\n");
548 			/* We cannot merge if:
549 			 * - allocated dma_addr isn't contiguous to previous allocation
550 			 */
551 			if (novmerge || (dma_addr != dma_next) ||
552 			    (outs->dma_length + s->length > max_seg_size)) {
553 				/* Can't merge: create a new segment */
554 				segstart = s;
555 				outcount++;
556 				outs = sg_next(outs);
557 				DBG("    can't merge, new segment.\n");
558 			} else {
559 				outs->dma_length += s->length;
560 				DBG("    merged, new len: %ux\n", outs->dma_length);
561 			}
562 		}
563 
564 		if (segstart == s) {
565 			/* This is a new segment, fill entries */
566 			DBG("  - filling new segment.\n");
567 			outs->dma_address = dma_addr;
568 			outs->dma_length = slen;
569 		}
570 
571 		/* Calculate next page pointer for contiguous check */
572 		dma_next = dma_addr + slen;
573 
574 		DBG("  - dma next is: %lx\n", dma_next);
575 	}
576 
577 	/* Flush/invalidate TLB caches if necessary */
578 	if (tbl->it_ops->flush)
579 		tbl->it_ops->flush(tbl);
580 
581 	DBG("mapped %d elements:\n", outcount);
582 
583 	/* For the sake of ppc_iommu_unmap_sg, we clear out the length in the
584 	 * next entry of the sglist if we didn't fill the list completely
585 	 */
586 	if (outcount < incount) {
587 		outs = sg_next(outs);
588 		outs->dma_length = 0;
589 	}
590 
591 	/* Make sure updates are seen by hardware */
592 	mb();
593 
594 	return outcount;
595 
596  failure:
597 	for_each_sg(sglist, s, nelems, i) {
598 		if (s->dma_length != 0) {
599 			unsigned long vaddr, npages;
600 
601 			vaddr = s->dma_address & IOMMU_PAGE_MASK(tbl);
602 			npages = iommu_num_pages(s->dma_address, s->dma_length,
603 						 IOMMU_PAGE_SIZE(tbl));
604 			__iommu_free(tbl, vaddr, npages);
605 			s->dma_length = 0;
606 		}
607 		if (s == outs)
608 			break;
609 	}
610 	return -EIO;
611 }
612 
613 
614 void ppc_iommu_unmap_sg(struct iommu_table *tbl, struct scatterlist *sglist,
615 			int nelems, enum dma_data_direction direction,
616 			unsigned long attrs)
617 {
618 	struct scatterlist *sg;
619 
620 	BUG_ON(direction == DMA_NONE);
621 
622 	if (!tbl)
623 		return;
624 
625 	sg = sglist;
626 	while (nelems--) {
627 		unsigned int npages;
628 		dma_addr_t dma_handle = sg->dma_address;
629 
630 		if (sg->dma_length == 0)
631 			break;
632 		npages = iommu_num_pages(dma_handle, sg->dma_length,
633 					 IOMMU_PAGE_SIZE(tbl));
634 		__iommu_free(tbl, dma_handle, npages);
635 		sg = sg_next(sg);
636 	}
637 
638 	/* Flush/invalidate TLBs if necessary. As for iommu_free(), we
639 	 * do not do an mb() here, the affected platforms do not need it
640 	 * when freeing.
641 	 */
642 	if (tbl->it_ops->flush)
643 		tbl->it_ops->flush(tbl);
644 }
645 
646 void iommu_table_clear(struct iommu_table *tbl)
647 {
648 	/*
649 	 * In case of firmware assisted dump system goes through clean
650 	 * reboot process at the time of system crash. Hence it's safe to
651 	 * clear the TCE entries if firmware assisted dump is active.
652 	 */
653 	if (!is_kdump_kernel() || is_fadump_active()) {
654 		/* Clear the table in case firmware left allocations in it */
655 		tbl->it_ops->clear(tbl, tbl->it_offset, tbl->it_size);
656 		return;
657 	}
658 
659 #ifdef CONFIG_CRASH_DUMP
660 	if (tbl->it_ops->get) {
661 		unsigned long index, tceval, tcecount = 0;
662 
663 		/* Reserve the existing mappings left by the first kernel. */
664 		for (index = 0; index < tbl->it_size; index++) {
665 			tceval = tbl->it_ops->get(tbl, index + tbl->it_offset);
666 			/*
667 			 * Freed TCE entry contains 0x7fffffffffffffff on JS20
668 			 */
669 			if (tceval && (tceval != 0x7fffffffffffffffUL)) {
670 				__set_bit(index, tbl->it_map);
671 				tcecount++;
672 			}
673 		}
674 
675 		if ((tbl->it_size - tcecount) < KDUMP_MIN_TCE_ENTRIES) {
676 			printk(KERN_WARNING "TCE table is full; freeing ");
677 			printk(KERN_WARNING "%d entries for the kdump boot\n",
678 				KDUMP_MIN_TCE_ENTRIES);
679 			for (index = tbl->it_size - KDUMP_MIN_TCE_ENTRIES;
680 				index < tbl->it_size; index++)
681 				__clear_bit(index, tbl->it_map);
682 		}
683 	}
684 #endif
685 }
686 
687 void iommu_table_reserve_pages(struct iommu_table *tbl,
688 		unsigned long res_start, unsigned long res_end)
689 {
690 	int i;
691 
692 	WARN_ON_ONCE(res_end < res_start);
693 	/*
694 	 * Reserve page 0 so it will not be used for any mappings.
695 	 * This avoids buggy drivers that consider page 0 to be invalid
696 	 * to crash the machine or even lose data.
697 	 */
698 	if (tbl->it_offset == 0)
699 		set_bit(0, tbl->it_map);
700 
701 	if (res_start < tbl->it_offset)
702 		res_start = tbl->it_offset;
703 
704 	if (res_end > (tbl->it_offset + tbl->it_size))
705 		res_end = tbl->it_offset + tbl->it_size;
706 
707 	/* Check if res_start..res_end is a valid range in the table */
708 	if (res_start >= res_end) {
709 		tbl->it_reserved_start = tbl->it_offset;
710 		tbl->it_reserved_end = tbl->it_offset;
711 		return;
712 	}
713 
714 	tbl->it_reserved_start = res_start;
715 	tbl->it_reserved_end = res_end;
716 
717 	for (i = tbl->it_reserved_start; i < tbl->it_reserved_end; ++i)
718 		set_bit(i - tbl->it_offset, tbl->it_map);
719 }
720 
721 /*
722  * Build a iommu_table structure.  This contains a bit map which
723  * is used to manage allocation of the tce space.
724  */
725 struct iommu_table *iommu_init_table(struct iommu_table *tbl, int nid,
726 		unsigned long res_start, unsigned long res_end)
727 {
728 	unsigned long sz;
729 	static int welcomed = 0;
730 	unsigned int i;
731 	struct iommu_pool *p;
732 
733 	BUG_ON(!tbl->it_ops);
734 
735 	/* number of bytes needed for the bitmap */
736 	sz = BITS_TO_LONGS(tbl->it_size) * sizeof(unsigned long);
737 
738 	tbl->it_map = vzalloc_node(sz, nid);
739 	if (!tbl->it_map) {
740 		pr_err("%s: Can't allocate %ld bytes\n", __func__, sz);
741 		return NULL;
742 	}
743 
744 	iommu_table_reserve_pages(tbl, res_start, res_end);
745 
746 	/* We only split the IOMMU table if we have 1GB or more of space */
747 	if ((tbl->it_size << tbl->it_page_shift) >= (1UL * 1024 * 1024 * 1024))
748 		tbl->nr_pools = IOMMU_NR_POOLS;
749 	else
750 		tbl->nr_pools = 1;
751 
752 	/* We reserve the top 1/4 of the table for large allocations */
753 	tbl->poolsize = (tbl->it_size * 3 / 4) / tbl->nr_pools;
754 
755 	for (i = 0; i < tbl->nr_pools; i++) {
756 		p = &tbl->pools[i];
757 		spin_lock_init(&(p->lock));
758 		p->start = tbl->poolsize * i;
759 		p->hint = p->start;
760 		p->end = p->start + tbl->poolsize;
761 	}
762 
763 	p = &tbl->large_pool;
764 	spin_lock_init(&(p->lock));
765 	p->start = tbl->poolsize * i;
766 	p->hint = p->start;
767 	p->end = tbl->it_size;
768 
769 	iommu_table_clear(tbl);
770 
771 	if (!welcomed) {
772 		printk(KERN_INFO "IOMMU table initialized, virtual merging %s\n",
773 		       novmerge ? "disabled" : "enabled");
774 		welcomed = 1;
775 	}
776 
777 	iommu_debugfs_add(tbl);
778 
779 	return tbl;
780 }
781 
782 bool iommu_table_in_use(struct iommu_table *tbl)
783 {
784 	unsigned long start = 0, end;
785 
786 	/* ignore reserved bit0 */
787 	if (tbl->it_offset == 0)
788 		start = 1;
789 
790 	/* Simple case with no reserved MMIO32 region */
791 	if (!tbl->it_reserved_start && !tbl->it_reserved_end)
792 		return find_next_bit(tbl->it_map, tbl->it_size, start) != tbl->it_size;
793 
794 	end = tbl->it_reserved_start - tbl->it_offset;
795 	if (find_next_bit(tbl->it_map, end, start) != end)
796 		return true;
797 
798 	start = tbl->it_reserved_end - tbl->it_offset;
799 	end = tbl->it_size;
800 	return find_next_bit(tbl->it_map, end, start) != end;
801 }
802 
803 static void iommu_table_free(struct kref *kref)
804 {
805 	struct iommu_table *tbl;
806 
807 	tbl = container_of(kref, struct iommu_table, it_kref);
808 
809 	if (tbl->it_ops->free)
810 		tbl->it_ops->free(tbl);
811 
812 	if (!tbl->it_map) {
813 		kfree(tbl);
814 		return;
815 	}
816 
817 	iommu_debugfs_del(tbl);
818 
819 	/* verify that table contains no entries */
820 	if (iommu_table_in_use(tbl))
821 		pr_warn("%s: Unexpected TCEs\n", __func__);
822 
823 	/* free bitmap */
824 	vfree(tbl->it_map);
825 
826 	/* free table */
827 	kfree(tbl);
828 }
829 
830 struct iommu_table *iommu_tce_table_get(struct iommu_table *tbl)
831 {
832 	if (kref_get_unless_zero(&tbl->it_kref))
833 		return tbl;
834 
835 	return NULL;
836 }
837 EXPORT_SYMBOL_GPL(iommu_tce_table_get);
838 
839 int iommu_tce_table_put(struct iommu_table *tbl)
840 {
841 	if (WARN_ON(!tbl))
842 		return 0;
843 
844 	return kref_put(&tbl->it_kref, iommu_table_free);
845 }
846 EXPORT_SYMBOL_GPL(iommu_tce_table_put);
847 
848 /* Creates TCEs for a user provided buffer.  The user buffer must be
849  * contiguous real kernel storage (not vmalloc).  The address passed here
850  * comprises a page address and offset into that page. The dma_addr_t
851  * returned will point to the same byte within the page as was passed in.
852  */
853 dma_addr_t iommu_map_page(struct device *dev, struct iommu_table *tbl,
854 			  struct page *page, unsigned long offset, size_t size,
855 			  unsigned long mask, enum dma_data_direction direction,
856 			  unsigned long attrs)
857 {
858 	dma_addr_t dma_handle = DMA_MAPPING_ERROR;
859 	void *vaddr;
860 	unsigned long uaddr;
861 	unsigned int npages, align;
862 
863 	BUG_ON(direction == DMA_NONE);
864 
865 	vaddr = page_address(page) + offset;
866 	uaddr = (unsigned long)vaddr;
867 
868 	if (tbl) {
869 		npages = iommu_num_pages(uaddr, size, IOMMU_PAGE_SIZE(tbl));
870 		align = 0;
871 		if (tbl->it_page_shift < PAGE_SHIFT && size >= PAGE_SIZE &&
872 		    ((unsigned long)vaddr & ~PAGE_MASK) == 0)
873 			align = PAGE_SHIFT - tbl->it_page_shift;
874 
875 		dma_handle = iommu_alloc(dev, tbl, vaddr, npages, direction,
876 					 mask >> tbl->it_page_shift, align,
877 					 attrs);
878 		if (dma_handle == DMA_MAPPING_ERROR) {
879 			if (!(attrs & DMA_ATTR_NO_WARN) &&
880 			    printk_ratelimit())  {
881 				dev_info(dev, "iommu_alloc failed, tbl %p "
882 					 "vaddr %p npages %d\n", tbl, vaddr,
883 					 npages);
884 			}
885 		} else
886 			dma_handle |= (uaddr & ~IOMMU_PAGE_MASK(tbl));
887 	}
888 
889 	return dma_handle;
890 }
891 
892 void iommu_unmap_page(struct iommu_table *tbl, dma_addr_t dma_handle,
893 		      size_t size, enum dma_data_direction direction,
894 		      unsigned long attrs)
895 {
896 	unsigned int npages;
897 
898 	BUG_ON(direction == DMA_NONE);
899 
900 	if (tbl) {
901 		npages = iommu_num_pages(dma_handle, size,
902 					 IOMMU_PAGE_SIZE(tbl));
903 		iommu_free(tbl, dma_handle, npages);
904 	}
905 }
906 
907 /* Allocates a contiguous real buffer and creates mappings over it.
908  * Returns the virtual address of the buffer and sets dma_handle
909  * to the dma address (mapping) of the first page.
910  */
911 void *iommu_alloc_coherent(struct device *dev, struct iommu_table *tbl,
912 			   size_t size,	dma_addr_t *dma_handle,
913 			   unsigned long mask, gfp_t flag, int node)
914 {
915 	void *ret = NULL;
916 	dma_addr_t mapping;
917 	unsigned int order;
918 	unsigned int nio_pages, io_order;
919 	struct page *page;
920 	int tcesize = (1 << tbl->it_page_shift);
921 
922 	size = PAGE_ALIGN(size);
923 	order = get_order(size);
924 
925  	/*
926 	 * Client asked for way too much space.  This is checked later
927 	 * anyway.  It is easier to debug here for the drivers than in
928 	 * the tce tables.
929 	 */
930 	if (order >= IOMAP_MAX_ORDER) {
931 		dev_info(dev, "iommu_alloc_consistent size too large: 0x%lx\n",
932 			 size);
933 		return NULL;
934 	}
935 
936 	if (!tbl)
937 		return NULL;
938 
939 	/* Alloc enough pages (and possibly more) */
940 	page = alloc_pages_node(node, flag, order);
941 	if (!page)
942 		return NULL;
943 	ret = page_address(page);
944 	memset(ret, 0, size);
945 
946 	/* Set up tces to cover the allocated range */
947 	nio_pages = IOMMU_PAGE_ALIGN(size, tbl) >> tbl->it_page_shift;
948 
949 	io_order = get_iommu_order(size, tbl);
950 	mapping = iommu_alloc(dev, tbl, ret, nio_pages, DMA_BIDIRECTIONAL,
951 			      mask >> tbl->it_page_shift, io_order, 0);
952 	if (mapping == DMA_MAPPING_ERROR) {
953 		free_pages((unsigned long)ret, order);
954 		return NULL;
955 	}
956 
957 	*dma_handle = mapping | ((u64)ret & (tcesize - 1));
958 	return ret;
959 }
960 
961 void iommu_free_coherent(struct iommu_table *tbl, size_t size,
962 			 void *vaddr, dma_addr_t dma_handle)
963 {
964 	if (tbl) {
965 		unsigned int nio_pages;
966 
967 		size = PAGE_ALIGN(size);
968 		nio_pages = IOMMU_PAGE_ALIGN(size, tbl) >> tbl->it_page_shift;
969 		iommu_free(tbl, dma_handle, nio_pages);
970 		size = PAGE_ALIGN(size);
971 		free_pages((unsigned long)vaddr, get_order(size));
972 	}
973 }
974 
975 unsigned long iommu_direction_to_tce_perm(enum dma_data_direction dir)
976 {
977 	switch (dir) {
978 	case DMA_BIDIRECTIONAL:
979 		return TCE_PCI_READ | TCE_PCI_WRITE;
980 	case DMA_FROM_DEVICE:
981 		return TCE_PCI_WRITE;
982 	case DMA_TO_DEVICE:
983 		return TCE_PCI_READ;
984 	default:
985 		return 0;
986 	}
987 }
988 EXPORT_SYMBOL_GPL(iommu_direction_to_tce_perm);
989 
990 #ifdef CONFIG_IOMMU_API
991 
992 int dev_has_iommu_table(struct device *dev, void *data)
993 {
994 	struct pci_dev *pdev = to_pci_dev(dev);
995 	struct pci_dev **ppdev = data;
996 
997 	if (!dev)
998 		return 0;
999 
1000 	if (device_iommu_mapped(dev)) {
1001 		*ppdev = pdev;
1002 		return 1;
1003 	}
1004 
1005 	return 0;
1006 }
1007 
1008 /*
1009  * SPAPR TCE API
1010  */
1011 static void group_release(void *iommu_data)
1012 {
1013 	struct iommu_table_group *table_group = iommu_data;
1014 
1015 	table_group->group = NULL;
1016 }
1017 
1018 void iommu_register_group(struct iommu_table_group *table_group,
1019 		int pci_domain_number, unsigned long pe_num)
1020 {
1021 	struct iommu_group *grp;
1022 	char *name;
1023 
1024 	grp = iommu_group_alloc();
1025 	if (IS_ERR(grp)) {
1026 		pr_warn("powerpc iommu api: cannot create new group, err=%ld\n",
1027 				PTR_ERR(grp));
1028 		return;
1029 	}
1030 	table_group->group = grp;
1031 	iommu_group_set_iommudata(grp, table_group, group_release);
1032 	name = kasprintf(GFP_KERNEL, "domain%d-pe%lx",
1033 			pci_domain_number, pe_num);
1034 	if (!name)
1035 		return;
1036 	iommu_group_set_name(grp, name);
1037 	kfree(name);
1038 }
1039 
1040 enum dma_data_direction iommu_tce_direction(unsigned long tce)
1041 {
1042 	if ((tce & TCE_PCI_READ) && (tce & TCE_PCI_WRITE))
1043 		return DMA_BIDIRECTIONAL;
1044 	else if (tce & TCE_PCI_READ)
1045 		return DMA_TO_DEVICE;
1046 	else if (tce & TCE_PCI_WRITE)
1047 		return DMA_FROM_DEVICE;
1048 	else
1049 		return DMA_NONE;
1050 }
1051 EXPORT_SYMBOL_GPL(iommu_tce_direction);
1052 
1053 void iommu_flush_tce(struct iommu_table *tbl)
1054 {
1055 	/* Flush/invalidate TLB caches if necessary */
1056 	if (tbl->it_ops->flush)
1057 		tbl->it_ops->flush(tbl);
1058 
1059 	/* Make sure updates are seen by hardware */
1060 	mb();
1061 }
1062 EXPORT_SYMBOL_GPL(iommu_flush_tce);
1063 
1064 int iommu_tce_check_ioba(unsigned long page_shift,
1065 		unsigned long offset, unsigned long size,
1066 		unsigned long ioba, unsigned long npages)
1067 {
1068 	unsigned long mask = (1UL << page_shift) - 1;
1069 
1070 	if (ioba & mask)
1071 		return -EINVAL;
1072 
1073 	ioba >>= page_shift;
1074 	if (ioba < offset)
1075 		return -EINVAL;
1076 
1077 	if ((ioba + 1) > (offset + size))
1078 		return -EINVAL;
1079 
1080 	return 0;
1081 }
1082 EXPORT_SYMBOL_GPL(iommu_tce_check_ioba);
1083 
1084 int iommu_tce_check_gpa(unsigned long page_shift, unsigned long gpa)
1085 {
1086 	unsigned long mask = (1UL << page_shift) - 1;
1087 
1088 	if (gpa & mask)
1089 		return -EINVAL;
1090 
1091 	return 0;
1092 }
1093 EXPORT_SYMBOL_GPL(iommu_tce_check_gpa);
1094 
1095 long iommu_tce_xchg_no_kill(struct mm_struct *mm,
1096 			    struct iommu_table *tbl,
1097 			    unsigned long entry, unsigned long *hpa,
1098 			    enum dma_data_direction *direction)
1099 {
1100 	long ret;
1101 	unsigned long size = 0;
1102 
1103 	ret = tbl->it_ops->xchg_no_kill(tbl, entry, hpa, direction);
1104 	if (!ret && ((*direction == DMA_FROM_DEVICE) ||
1105 			(*direction == DMA_BIDIRECTIONAL)) &&
1106 			!mm_iommu_is_devmem(mm, *hpa, tbl->it_page_shift,
1107 					&size))
1108 		SetPageDirty(pfn_to_page(*hpa >> PAGE_SHIFT));
1109 
1110 	return ret;
1111 }
1112 EXPORT_SYMBOL_GPL(iommu_tce_xchg_no_kill);
1113 
1114 void iommu_tce_kill(struct iommu_table *tbl,
1115 		unsigned long entry, unsigned long pages)
1116 {
1117 	if (tbl->it_ops->tce_kill)
1118 		tbl->it_ops->tce_kill(tbl, entry, pages);
1119 }
1120 EXPORT_SYMBOL_GPL(iommu_tce_kill);
1121 
1122 int iommu_add_device(struct iommu_table_group *table_group, struct device *dev)
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 (device_iommu_mapped(dev)) {
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 	pr_debug("%s: Adding %s to iommu group %d\n",
1140 		 __func__, dev_name(dev),  iommu_group_id(table_group->group));
1141 	/*
1142 	 * This is still not adding devices via the IOMMU bus notifier because
1143 	 * of pcibios_init() from arch/powerpc/kernel/pci_64.c which calls
1144 	 * pcibios_scan_phb() first (and this guy adds devices and triggers
1145 	 * the notifier) and only then it calls pci_bus_add_devices() which
1146 	 * configures DMA for buses which also creates PEs and IOMMU groups.
1147 	 */
1148 	return iommu_probe_device(dev);
1149 }
1150 EXPORT_SYMBOL_GPL(iommu_add_device);
1151 
1152 #if defined(CONFIG_PPC_PSERIES) || defined(CONFIG_PPC_POWERNV)
1153 /*
1154  * A simple iommu_ops to allow less cruft in generic VFIO code.
1155  */
1156 static int
1157 spapr_tce_platform_iommu_attach_dev(struct iommu_domain *platform_domain,
1158 				    struct device *dev)
1159 {
1160 	struct iommu_domain *domain = iommu_get_domain_for_dev(dev);
1161 	struct iommu_table_group *table_group;
1162 	struct iommu_group *grp;
1163 
1164 	/* At first attach the ownership is already set */
1165 	if (!domain)
1166 		return 0;
1167 
1168 	grp = iommu_group_get(dev);
1169 	table_group = iommu_group_get_iommudata(grp);
1170 	/*
1171 	 * The domain being set to PLATFORM from earlier
1172 	 * BLOCKED. The table_group ownership has to be released.
1173 	 */
1174 	table_group->ops->release_ownership(table_group, dev);
1175 	iommu_group_put(grp);
1176 
1177 	return 0;
1178 }
1179 
1180 static const struct iommu_domain_ops spapr_tce_platform_domain_ops = {
1181 	.attach_dev = spapr_tce_platform_iommu_attach_dev,
1182 };
1183 
1184 static struct iommu_domain spapr_tce_platform_domain = {
1185 	.type = IOMMU_DOMAIN_PLATFORM,
1186 	.ops = &spapr_tce_platform_domain_ops,
1187 };
1188 
1189 static int
1190 spapr_tce_blocked_iommu_attach_dev(struct iommu_domain *platform_domain,
1191 				     struct device *dev)
1192 {
1193 	struct iommu_group *grp = iommu_group_get(dev);
1194 	struct iommu_table_group *table_group;
1195 	int ret = -EINVAL;
1196 
1197 	/*
1198 	 * FIXME: SPAPR mixes blocked and platform behaviors, the blocked domain
1199 	 * also sets the dma_api ops
1200 	 */
1201 	table_group = iommu_group_get_iommudata(grp);
1202 	ret = table_group->ops->take_ownership(table_group, dev);
1203 	iommu_group_put(grp);
1204 
1205 	return ret;
1206 }
1207 
1208 static const struct iommu_domain_ops spapr_tce_blocked_domain_ops = {
1209 	.attach_dev = spapr_tce_blocked_iommu_attach_dev,
1210 };
1211 
1212 static struct iommu_domain spapr_tce_blocked_domain = {
1213 	.type = IOMMU_DOMAIN_BLOCKED,
1214 	.ops = &spapr_tce_blocked_domain_ops,
1215 };
1216 
1217 static bool spapr_tce_iommu_capable(struct device *dev, enum iommu_cap cap)
1218 {
1219 	switch (cap) {
1220 	case IOMMU_CAP_CACHE_COHERENCY:
1221 		return true;
1222 	default:
1223 		break;
1224 	}
1225 
1226 	return false;
1227 }
1228 
1229 static struct iommu_device *spapr_tce_iommu_probe_device(struct device *dev)
1230 {
1231 	struct pci_dev *pdev;
1232 	struct pci_controller *hose;
1233 
1234 	if (!dev_is_pci(dev))
1235 		return ERR_PTR(-ENODEV);
1236 
1237 	pdev = to_pci_dev(dev);
1238 	hose = pdev->bus->sysdata;
1239 
1240 	return &hose->iommu;
1241 }
1242 
1243 static void spapr_tce_iommu_release_device(struct device *dev)
1244 {
1245 }
1246 
1247 static struct iommu_group *spapr_tce_iommu_device_group(struct device *dev)
1248 {
1249 	struct pci_controller *hose;
1250 	struct pci_dev *pdev;
1251 
1252 	pdev = to_pci_dev(dev);
1253 	hose = pdev->bus->sysdata;
1254 
1255 	if (!hose->controller_ops.device_group)
1256 		return ERR_PTR(-ENOENT);
1257 
1258 	return hose->controller_ops.device_group(hose, pdev);
1259 }
1260 
1261 static const struct iommu_ops spapr_tce_iommu_ops = {
1262 	.default_domain = &spapr_tce_platform_domain,
1263 	.blocked_domain = &spapr_tce_blocked_domain,
1264 	.capable = spapr_tce_iommu_capable,
1265 	.probe_device = spapr_tce_iommu_probe_device,
1266 	.release_device = spapr_tce_iommu_release_device,
1267 	.device_group = spapr_tce_iommu_device_group,
1268 };
1269 
1270 static struct attribute *spapr_tce_iommu_attrs[] = {
1271 	NULL,
1272 };
1273 
1274 static struct attribute_group spapr_tce_iommu_group = {
1275 	.name = "spapr-tce-iommu",
1276 	.attrs = spapr_tce_iommu_attrs,
1277 };
1278 
1279 static const struct attribute_group *spapr_tce_iommu_groups[] = {
1280 	&spapr_tce_iommu_group,
1281 	NULL,
1282 };
1283 
1284 void ppc_iommu_register_device(struct pci_controller *phb)
1285 {
1286 	iommu_device_sysfs_add(&phb->iommu, phb->parent,
1287 				spapr_tce_iommu_groups, "iommu-phb%04x",
1288 				phb->global_number);
1289 	iommu_device_register(&phb->iommu, &spapr_tce_iommu_ops,
1290 				phb->parent);
1291 }
1292 
1293 void ppc_iommu_unregister_device(struct pci_controller *phb)
1294 {
1295 	iommu_device_unregister(&phb->iommu);
1296 	iommu_device_sysfs_remove(&phb->iommu);
1297 }
1298 
1299 /*
1300  * This registers IOMMU devices of PHBs. This needs to happen
1301  * after core_initcall(iommu_init) + postcore_initcall(pci_driver_init) and
1302  * before subsys_initcall(iommu_subsys_init).
1303  */
1304 static int __init spapr_tce_setup_phb_iommus_initcall(void)
1305 {
1306 	struct pci_controller *hose;
1307 
1308 	list_for_each_entry(hose, &hose_list, list_node) {
1309 		ppc_iommu_register_device(hose);
1310 	}
1311 	return 0;
1312 }
1313 postcore_initcall_sync(spapr_tce_setup_phb_iommus_initcall);
1314 #endif
1315 
1316 #endif /* CONFIG_IOMMU_API */
1317