xref: /linux/arch/powerpc/kernel/iommu.c (revision f3d9478b2ce468c3115b02ecae7e975990697f15)
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/config.h>
27 #include <linux/init.h>
28 #include <linux/types.h>
29 #include <linux/slab.h>
30 #include <linux/mm.h>
31 #include <linux/spinlock.h>
32 #include <linux/string.h>
33 #include <linux/dma-mapping.h>
34 #include <linux/init.h>
35 #include <linux/bitops.h>
36 #include <asm/io.h>
37 #include <asm/prom.h>
38 #include <asm/iommu.h>
39 #include <asm/pci-bridge.h>
40 #include <asm/machdep.h>
41 
42 #define DBG(...)
43 
44 #ifdef CONFIG_IOMMU_VMERGE
45 static int novmerge = 0;
46 #else
47 static int novmerge = 1;
48 #endif
49 
50 static int __init setup_iommu(char *str)
51 {
52 	if (!strcmp(str, "novmerge"))
53 		novmerge = 1;
54 	else if (!strcmp(str, "vmerge"))
55 		novmerge = 0;
56 	return 1;
57 }
58 
59 __setup("iommu=", setup_iommu);
60 
61 static unsigned long iommu_range_alloc(struct iommu_table *tbl,
62                                        unsigned long npages,
63                                        unsigned long *handle,
64                                        unsigned long mask,
65                                        unsigned int align_order)
66 {
67 	unsigned long n, end, i, start;
68 	unsigned long limit;
69 	int largealloc = npages > 15;
70 	int pass = 0;
71 	unsigned long align_mask;
72 
73 	align_mask = 0xffffffffffffffffl >> (64 - align_order);
74 
75 	/* This allocator was derived from x86_64's bit string search */
76 
77 	/* Sanity check */
78 	if (unlikely(npages) == 0) {
79 		if (printk_ratelimit())
80 			WARN_ON(1);
81 		return DMA_ERROR_CODE;
82 	}
83 
84 	if (handle && *handle)
85 		start = *handle;
86 	else
87 		start = largealloc ? tbl->it_largehint : tbl->it_hint;
88 
89 	/* Use only half of the table for small allocs (15 pages or less) */
90 	limit = largealloc ? tbl->it_size : tbl->it_halfpoint;
91 
92 	if (largealloc && start < tbl->it_halfpoint)
93 		start = tbl->it_halfpoint;
94 
95 	/* The case below can happen if we have a small segment appended
96 	 * to a large, or when the previous alloc was at the very end of
97 	 * the available space. If so, go back to the initial start.
98 	 */
99 	if (start >= limit)
100 		start = largealloc ? tbl->it_largehint : tbl->it_hint;
101 
102  again:
103 
104 	if (limit + tbl->it_offset > mask) {
105 		limit = mask - tbl->it_offset + 1;
106 		/* If we're constrained on address range, first try
107 		 * at the masked hint to avoid O(n) search complexity,
108 		 * but on second pass, start at 0.
109 		 */
110 		if ((start & mask) >= limit || pass > 0)
111 			start = 0;
112 		else
113 			start &= mask;
114 	}
115 
116 	n = find_next_zero_bit(tbl->it_map, limit, start);
117 
118 	/* Align allocation */
119 	n = (n + align_mask) & ~align_mask;
120 
121 	end = n + npages;
122 
123 	if (unlikely(end >= limit)) {
124 		if (likely(pass < 2)) {
125 			/* First failure, just rescan the half of the table.
126 			 * Second failure, rescan the other half of the table.
127 			 */
128 			start = (largealloc ^ pass) ? tbl->it_halfpoint : 0;
129 			limit = pass ? tbl->it_size : limit;
130 			pass++;
131 			goto again;
132 		} else {
133 			/* Third failure, give up */
134 			return DMA_ERROR_CODE;
135 		}
136 	}
137 
138 	for (i = n; i < end; i++)
139 		if (test_bit(i, tbl->it_map)) {
140 			start = i+1;
141 			goto again;
142 		}
143 
144 	for (i = n; i < end; i++)
145 		__set_bit(i, tbl->it_map);
146 
147 	/* Bump the hint to a new block for small allocs. */
148 	if (largealloc) {
149 		/* Don't bump to new block to avoid fragmentation */
150 		tbl->it_largehint = end;
151 	} else {
152 		/* Overflow will be taken care of at the next allocation */
153 		tbl->it_hint = (end + tbl->it_blocksize - 1) &
154 		                ~(tbl->it_blocksize - 1);
155 	}
156 
157 	/* Update handle for SG allocations */
158 	if (handle)
159 		*handle = end;
160 
161 	return n;
162 }
163 
164 static dma_addr_t iommu_alloc(struct iommu_table *tbl, void *page,
165 		       unsigned int npages, enum dma_data_direction direction,
166 		       unsigned long mask, unsigned int align_order)
167 {
168 	unsigned long entry, flags;
169 	dma_addr_t ret = DMA_ERROR_CODE;
170 
171 	spin_lock_irqsave(&(tbl->it_lock), flags);
172 
173 	entry = iommu_range_alloc(tbl, npages, NULL, mask, align_order);
174 
175 	if (unlikely(entry == DMA_ERROR_CODE)) {
176 		spin_unlock_irqrestore(&(tbl->it_lock), flags);
177 		return DMA_ERROR_CODE;
178 	}
179 
180 	entry += tbl->it_offset;	/* Offset into real TCE table */
181 	ret = entry << PAGE_SHIFT;	/* Set the return dma address */
182 
183 	/* Put the TCEs in the HW table */
184 	ppc_md.tce_build(tbl, entry, npages, (unsigned long)page & PAGE_MASK,
185 			 direction);
186 
187 
188 	/* Flush/invalidate TLB caches if necessary */
189 	if (ppc_md.tce_flush)
190 		ppc_md.tce_flush(tbl);
191 
192 	spin_unlock_irqrestore(&(tbl->it_lock), flags);
193 
194 	/* Make sure updates are seen by hardware */
195 	mb();
196 
197 	return ret;
198 }
199 
200 static void __iommu_free(struct iommu_table *tbl, dma_addr_t dma_addr,
201 			 unsigned int npages)
202 {
203 	unsigned long entry, free_entry;
204 	unsigned long i;
205 
206 	entry = dma_addr >> PAGE_SHIFT;
207 	free_entry = entry - tbl->it_offset;
208 
209 	if (((free_entry + npages) > tbl->it_size) ||
210 	    (entry < tbl->it_offset)) {
211 		if (printk_ratelimit()) {
212 			printk(KERN_INFO "iommu_free: invalid entry\n");
213 			printk(KERN_INFO "\tentry     = 0x%lx\n", entry);
214 			printk(KERN_INFO "\tdma_addr  = 0x%lx\n", (u64)dma_addr);
215 			printk(KERN_INFO "\tTable     = 0x%lx\n", (u64)tbl);
216 			printk(KERN_INFO "\tbus#      = 0x%lx\n", (u64)tbl->it_busno);
217 			printk(KERN_INFO "\tsize      = 0x%lx\n", (u64)tbl->it_size);
218 			printk(KERN_INFO "\tstartOff  = 0x%lx\n", (u64)tbl->it_offset);
219 			printk(KERN_INFO "\tindex     = 0x%lx\n", (u64)tbl->it_index);
220 			WARN_ON(1);
221 		}
222 		return;
223 	}
224 
225 	ppc_md.tce_free(tbl, entry, npages);
226 
227 	for (i = 0; i < npages; i++)
228 		__clear_bit(free_entry+i, tbl->it_map);
229 }
230 
231 static void iommu_free(struct iommu_table *tbl, dma_addr_t dma_addr,
232 		unsigned int npages)
233 {
234 	unsigned long flags;
235 
236 	spin_lock_irqsave(&(tbl->it_lock), flags);
237 
238 	__iommu_free(tbl, dma_addr, npages);
239 
240 	/* Make sure TLB cache is flushed if the HW needs it. We do
241 	 * not do an mb() here on purpose, it is not needed on any of
242 	 * the current platforms.
243 	 */
244 	if (ppc_md.tce_flush)
245 		ppc_md.tce_flush(tbl);
246 
247 	spin_unlock_irqrestore(&(tbl->it_lock), flags);
248 }
249 
250 int iommu_map_sg(struct device *dev, struct iommu_table *tbl,
251 		struct scatterlist *sglist, int nelems,
252 		unsigned long mask, enum dma_data_direction direction)
253 {
254 	dma_addr_t dma_next = 0, dma_addr;
255 	unsigned long flags;
256 	struct scatterlist *s, *outs, *segstart;
257 	int outcount, incount;
258 	unsigned long handle;
259 
260 	BUG_ON(direction == DMA_NONE);
261 
262 	if ((nelems == 0) || !tbl)
263 		return 0;
264 
265 	outs = s = segstart = &sglist[0];
266 	outcount = 1;
267 	incount = nelems;
268 	handle = 0;
269 
270 	/* Init first segment length for backout at failure */
271 	outs->dma_length = 0;
272 
273 	DBG("mapping %d elements:\n", nelems);
274 
275 	spin_lock_irqsave(&(tbl->it_lock), flags);
276 
277 	for (s = outs; nelems; nelems--, s++) {
278 		unsigned long vaddr, npages, entry, slen;
279 
280 		slen = s->length;
281 		/* Sanity check */
282 		if (slen == 0) {
283 			dma_next = 0;
284 			continue;
285 		}
286 		/* Allocate iommu entries for that segment */
287 		vaddr = (unsigned long)page_address(s->page) + s->offset;
288 		npages = PAGE_ALIGN(vaddr + slen) - (vaddr & PAGE_MASK);
289 		npages >>= PAGE_SHIFT;
290 		entry = iommu_range_alloc(tbl, npages, &handle, mask >> PAGE_SHIFT, 0);
291 
292 		DBG("  - vaddr: %lx, size: %lx\n", vaddr, slen);
293 
294 		/* Handle failure */
295 		if (unlikely(entry == DMA_ERROR_CODE)) {
296 			if (printk_ratelimit())
297 				printk(KERN_INFO "iommu_alloc failed, tbl %p vaddr %lx"
298 				       " npages %lx\n", tbl, vaddr, npages);
299 			goto failure;
300 		}
301 
302 		/* Convert entry to a dma_addr_t */
303 		entry += tbl->it_offset;
304 		dma_addr = entry << PAGE_SHIFT;
305 		dma_addr |= s->offset;
306 
307 		DBG("  - %lx pages, entry: %lx, dma_addr: %lx\n",
308 			    npages, entry, dma_addr);
309 
310 		/* Insert into HW table */
311 		ppc_md.tce_build(tbl, entry, npages, vaddr & PAGE_MASK, direction);
312 
313 		/* If we are in an open segment, try merging */
314 		if (segstart != s) {
315 			DBG("  - trying merge...\n");
316 			/* We cannot merge if:
317 			 * - allocated dma_addr isn't contiguous to previous allocation
318 			 */
319 			if (novmerge || (dma_addr != dma_next)) {
320 				/* Can't merge: create a new segment */
321 				segstart = s;
322 				outcount++; outs++;
323 				DBG("    can't merge, new segment.\n");
324 			} else {
325 				outs->dma_length += s->length;
326 				DBG("    merged, new len: %lx\n", outs->dma_length);
327 			}
328 		}
329 
330 		if (segstart == s) {
331 			/* This is a new segment, fill entries */
332 			DBG("  - filling new segment.\n");
333 			outs->dma_address = dma_addr;
334 			outs->dma_length = slen;
335 		}
336 
337 		/* Calculate next page pointer for contiguous check */
338 		dma_next = dma_addr + slen;
339 
340 		DBG("  - dma next is: %lx\n", dma_next);
341 	}
342 
343 	/* Flush/invalidate TLB caches if necessary */
344 	if (ppc_md.tce_flush)
345 		ppc_md.tce_flush(tbl);
346 
347 	spin_unlock_irqrestore(&(tbl->it_lock), flags);
348 
349 	DBG("mapped %d elements:\n", outcount);
350 
351 	/* For the sake of iommu_unmap_sg, we clear out the length in the
352 	 * next entry of the sglist if we didn't fill the list completely
353 	 */
354 	if (outcount < incount) {
355 		outs++;
356 		outs->dma_address = DMA_ERROR_CODE;
357 		outs->dma_length = 0;
358 	}
359 
360 	/* Make sure updates are seen by hardware */
361 	mb();
362 
363 	return outcount;
364 
365  failure:
366 	for (s = &sglist[0]; s <= outs; s++) {
367 		if (s->dma_length != 0) {
368 			unsigned long vaddr, npages;
369 
370 			vaddr = s->dma_address & PAGE_MASK;
371 			npages = (PAGE_ALIGN(s->dma_address + s->dma_length) - vaddr)
372 				>> PAGE_SHIFT;
373 			__iommu_free(tbl, vaddr, npages);
374 			s->dma_address = DMA_ERROR_CODE;
375 			s->dma_length = 0;
376 		}
377 	}
378 	spin_unlock_irqrestore(&(tbl->it_lock), flags);
379 	return 0;
380 }
381 
382 
383 void iommu_unmap_sg(struct iommu_table *tbl, struct scatterlist *sglist,
384 		int nelems, enum dma_data_direction direction)
385 {
386 	unsigned long flags;
387 
388 	BUG_ON(direction == DMA_NONE);
389 
390 	if (!tbl)
391 		return;
392 
393 	spin_lock_irqsave(&(tbl->it_lock), flags);
394 
395 	while (nelems--) {
396 		unsigned int npages;
397 		dma_addr_t dma_handle = sglist->dma_address;
398 
399 		if (sglist->dma_length == 0)
400 			break;
401 		npages = (PAGE_ALIGN(dma_handle + sglist->dma_length)
402 			  - (dma_handle & PAGE_MASK)) >> PAGE_SHIFT;
403 		__iommu_free(tbl, dma_handle, npages);
404 		sglist++;
405 	}
406 
407 	/* Flush/invalidate TLBs if necessary. As for iommu_free(), we
408 	 * do not do an mb() here, the affected platforms do not need it
409 	 * when freeing.
410 	 */
411 	if (ppc_md.tce_flush)
412 		ppc_md.tce_flush(tbl);
413 
414 	spin_unlock_irqrestore(&(tbl->it_lock), flags);
415 }
416 
417 /*
418  * Build a iommu_table structure.  This contains a bit map which
419  * is used to manage allocation of the tce space.
420  */
421 struct iommu_table *iommu_init_table(struct iommu_table *tbl)
422 {
423 	unsigned long sz;
424 	static int welcomed = 0;
425 
426 	/* Set aside 1/4 of the table for large allocations. */
427 	tbl->it_halfpoint = tbl->it_size * 3 / 4;
428 
429 	/* number of bytes needed for the bitmap */
430 	sz = (tbl->it_size + 7) >> 3;
431 
432 	tbl->it_map = (unsigned long *)__get_free_pages(GFP_ATOMIC, get_order(sz));
433 	if (!tbl->it_map)
434 		panic("iommu_init_table: Can't allocate %ld bytes\n", sz);
435 
436 	memset(tbl->it_map, 0, sz);
437 
438 	tbl->it_hint = 0;
439 	tbl->it_largehint = tbl->it_halfpoint;
440 	spin_lock_init(&tbl->it_lock);
441 
442 	/* Clear the hardware table in case firmware left allocations in it */
443 	ppc_md.tce_free(tbl, tbl->it_offset, tbl->it_size);
444 
445 	if (!welcomed) {
446 		printk(KERN_INFO "IOMMU table initialized, virtual merging %s\n",
447 		       novmerge ? "disabled" : "enabled");
448 		welcomed = 1;
449 	}
450 
451 	return tbl;
452 }
453 
454 void iommu_free_table(struct device_node *dn)
455 {
456 	struct pci_dn *pdn = dn->data;
457 	struct iommu_table *tbl = pdn->iommu_table;
458 	unsigned long bitmap_sz, i;
459 	unsigned int order;
460 
461 	if (!tbl || !tbl->it_map) {
462 		printk(KERN_ERR "%s: expected TCE map for %s\n", __FUNCTION__,
463 				dn->full_name);
464 		return;
465 	}
466 
467 	/* verify that table contains no entries */
468 	/* it_size is in entries, and we're examining 64 at a time */
469 	for (i = 0; i < (tbl->it_size/64); i++) {
470 		if (tbl->it_map[i] != 0) {
471 			printk(KERN_WARNING "%s: Unexpected TCEs for %s\n",
472 				__FUNCTION__, dn->full_name);
473 			break;
474 		}
475 	}
476 
477 	/* calculate bitmap size in bytes */
478 	bitmap_sz = (tbl->it_size + 7) / 8;
479 
480 	/* free bitmap */
481 	order = get_order(bitmap_sz);
482 	free_pages((unsigned long) tbl->it_map, order);
483 
484 	/* free table */
485 	kfree(tbl);
486 }
487 
488 /* Creates TCEs for a user provided buffer.  The user buffer must be
489  * contiguous real kernel storage (not vmalloc).  The address of the buffer
490  * passed here is the kernel (virtual) address of the buffer.  The buffer
491  * need not be page aligned, the dma_addr_t returned will point to the same
492  * byte within the page as vaddr.
493  */
494 dma_addr_t iommu_map_single(struct iommu_table *tbl, void *vaddr,
495 		size_t size, unsigned long mask,
496 		enum dma_data_direction direction)
497 {
498 	dma_addr_t dma_handle = DMA_ERROR_CODE;
499 	unsigned long uaddr;
500 	unsigned int npages;
501 
502 	BUG_ON(direction == DMA_NONE);
503 
504 	uaddr = (unsigned long)vaddr;
505 	npages = PAGE_ALIGN(uaddr + size) - (uaddr & PAGE_MASK);
506 	npages >>= PAGE_SHIFT;
507 
508 	if (tbl) {
509 		dma_handle = iommu_alloc(tbl, vaddr, npages, direction,
510 					 mask >> PAGE_SHIFT, 0);
511 		if (dma_handle == DMA_ERROR_CODE) {
512 			if (printk_ratelimit())  {
513 				printk(KERN_INFO "iommu_alloc failed, "
514 						"tbl %p vaddr %p npages %d\n",
515 						tbl, vaddr, npages);
516 			}
517 		} else
518 			dma_handle |= (uaddr & ~PAGE_MASK);
519 	}
520 
521 	return dma_handle;
522 }
523 
524 void iommu_unmap_single(struct iommu_table *tbl, dma_addr_t dma_handle,
525 		size_t size, enum dma_data_direction direction)
526 {
527 	BUG_ON(direction == DMA_NONE);
528 
529 	if (tbl)
530 		iommu_free(tbl, dma_handle, (PAGE_ALIGN(dma_handle + size) -
531 					(dma_handle & PAGE_MASK)) >> PAGE_SHIFT);
532 }
533 
534 /* Allocates a contiguous real buffer and creates mappings over it.
535  * Returns the virtual address of the buffer and sets dma_handle
536  * to the dma address (mapping) of the first page.
537  */
538 void *iommu_alloc_coherent(struct iommu_table *tbl, size_t size,
539 		dma_addr_t *dma_handle, unsigned long mask, gfp_t flag)
540 {
541 	void *ret = NULL;
542 	dma_addr_t mapping;
543 	unsigned int npages, order;
544 
545 	size = PAGE_ALIGN(size);
546 	npages = size >> PAGE_SHIFT;
547 	order = get_order(size);
548 
549  	/*
550 	 * Client asked for way too much space.  This is checked later
551 	 * anyway.  It is easier to debug here for the drivers than in
552 	 * the tce tables.
553 	 */
554 	if (order >= IOMAP_MAX_ORDER) {
555 		printk("iommu_alloc_consistent size too large: 0x%lx\n", size);
556 		return NULL;
557 	}
558 
559 	if (!tbl)
560 		return NULL;
561 
562 	/* Alloc enough pages (and possibly more) */
563 	ret = (void *)__get_free_pages(flag, order);
564 	if (!ret)
565 		return NULL;
566 	memset(ret, 0, size);
567 
568 	/* Set up tces to cover the allocated range */
569 	mapping = iommu_alloc(tbl, ret, npages, DMA_BIDIRECTIONAL,
570 			      mask >> PAGE_SHIFT, order);
571 	if (mapping == DMA_ERROR_CODE) {
572 		free_pages((unsigned long)ret, order);
573 		ret = NULL;
574 	} else
575 		*dma_handle = mapping;
576 	return ret;
577 }
578 
579 void iommu_free_coherent(struct iommu_table *tbl, size_t size,
580 			 void *vaddr, dma_addr_t dma_handle)
581 {
582 	unsigned int npages;
583 
584 	if (tbl) {
585 		size = PAGE_ALIGN(size);
586 		npages = size >> PAGE_SHIFT;
587 		iommu_free(tbl, dma_handle, npages);
588 		free_pages((unsigned long)vaddr, get_order(size));
589 	}
590 }
591