xref: /linux/arch/sparc/kernel/iommu.c (revision 944e15f200475d530096cba489833dc6dcd8d1e1)
1 // SPDX-License-Identifier: GPL-2.0
2 /* iommu.c: Generic sparc64 IOMMU support.
3  *
4  * Copyright (C) 1999, 2007, 2008 David S. Miller (davem@davemloft.net)
5  * Copyright (C) 1999, 2000 Jakub Jelinek (jakub@redhat.com)
6  */
7 
8 #include <linux/kernel.h>
9 #include <linux/export.h>
10 #include <linux/slab.h>
11 #include <linux/delay.h>
12 #include <linux/device.h>
13 #include <linux/dma-map-ops.h>
14 #include <linux/errno.h>
15 #include <linux/iommu-helper.h>
16 #include <linux/bitmap.h>
17 #include <asm/iommu-common.h>
18 
19 #ifdef CONFIG_PCI
20 #include <linux/pci.h>
21 #endif
22 
23 #include <asm/iommu.h>
24 
25 #include "iommu_common.h"
26 #include "kernel.h"
27 
28 #define STC_CTXMATCH_ADDR(STC, CTX)	\
29 	((STC)->strbuf_ctxmatch_base + ((CTX) << 3))
30 #define STC_FLUSHFLAG_INIT(STC) \
31 	(*((STC)->strbuf_flushflag) = 0UL)
32 #define STC_FLUSHFLAG_SET(STC) \
33 	(*((STC)->strbuf_flushflag) != 0UL)
34 
35 #define iommu_read(__reg) \
36 ({	u64 __ret; \
37 	__asm__ __volatile__("ldxa [%1] %2, %0" \
38 			     : "=r" (__ret) \
39 			     : "r" (__reg), "i" (ASI_PHYS_BYPASS_EC_E) \
40 			     : "memory"); \
41 	__ret; \
42 })
43 #define iommu_write(__reg, __val) \
44 	__asm__ __volatile__("stxa %0, [%1] %2" \
45 			     : /* no outputs */ \
46 			     : "r" (__val), "r" (__reg), \
47 			       "i" (ASI_PHYS_BYPASS_EC_E))
48 
49 /* Must be invoked under the IOMMU lock. */
iommu_flushall(struct iommu_map_table * iommu_map_table)50 static void iommu_flushall(struct iommu_map_table *iommu_map_table)
51 {
52 	struct iommu *iommu = container_of(iommu_map_table, struct iommu, tbl);
53 	if (iommu->iommu_flushinv) {
54 		iommu_write(iommu->iommu_flushinv, ~(u64)0);
55 	} else {
56 		unsigned long tag;
57 		int entry;
58 
59 		tag = iommu->iommu_tags;
60 		for (entry = 0; entry < 16; entry++) {
61 			iommu_write(tag, 0);
62 			tag += 8;
63 		}
64 
65 		/* Ensure completion of previous PIO writes. */
66 		(void) iommu_read(iommu->write_complete_reg);
67 	}
68 }
69 
70 #define IOPTE_CONSISTENT(CTX) \
71 	(IOPTE_VALID | IOPTE_CACHE | \
72 	 (((CTX) << 47) & IOPTE_CONTEXT))
73 
74 #define IOPTE_STREAMING(CTX) \
75 	(IOPTE_CONSISTENT(CTX) | IOPTE_STBUF)
76 
77 /* Existing mappings are never marked invalid, instead they
78  * are pointed to a dummy page.
79  */
80 #define IOPTE_IS_DUMMY(iommu, iopte)	\
81 	((iopte_val(*iopte) & IOPTE_PAGE) == (iommu)->dummy_page_pa)
82 
iopte_make_dummy(struct iommu * iommu,iopte_t * iopte)83 static inline void iopte_make_dummy(struct iommu *iommu, iopte_t *iopte)
84 {
85 	unsigned long val = iopte_val(*iopte);
86 
87 	val &= ~IOPTE_PAGE;
88 	val |= iommu->dummy_page_pa;
89 
90 	iopte_val(*iopte) = val;
91 }
92 
iommu_table_init(struct iommu * iommu,int tsbsize,u32 dma_offset,u32 dma_addr_mask,int numa_node)93 int iommu_table_init(struct iommu *iommu, int tsbsize,
94 		     u32 dma_offset, u32 dma_addr_mask,
95 		     int numa_node)
96 {
97 	unsigned long i, order, sz, num_tsb_entries;
98 	struct page *page;
99 
100 	num_tsb_entries = tsbsize / sizeof(iopte_t);
101 
102 	/* Setup initial software IOMMU state. */
103 	spin_lock_init(&iommu->lock);
104 	iommu->ctx_lowest_free = 1;
105 	iommu->tbl.table_map_base = dma_offset;
106 	iommu->dma_addr_mask = dma_addr_mask;
107 
108 	/* Allocate and initialize the free area map.  */
109 	sz = num_tsb_entries / 8;
110 	sz = (sz + 7UL) & ~7UL;
111 	iommu->tbl.map = kzalloc_node(sz, GFP_KERNEL, numa_node);
112 	if (!iommu->tbl.map)
113 		return -ENOMEM;
114 
115 	iommu_tbl_pool_init(&iommu->tbl, num_tsb_entries, IO_PAGE_SHIFT,
116 			    (tlb_type != hypervisor ? iommu_flushall : NULL),
117 			    false, 1, false);
118 
119 	/* Allocate and initialize the dummy page which we
120 	 * set inactive IO PTEs to point to.
121 	 */
122 	page = alloc_pages_node(numa_node, GFP_KERNEL, 0);
123 	if (!page) {
124 		printk(KERN_ERR "IOMMU: Error, gfp(dummy_page) failed.\n");
125 		goto out_free_map;
126 	}
127 	iommu->dummy_page = (unsigned long) page_address(page);
128 	memset((void *)iommu->dummy_page, 0, PAGE_SIZE);
129 	iommu->dummy_page_pa = (unsigned long) __pa(iommu->dummy_page);
130 
131 	/* Now allocate and setup the IOMMU page table itself.  */
132 	order = get_order(tsbsize);
133 	page = alloc_pages_node(numa_node, GFP_KERNEL, order);
134 	if (!page) {
135 		printk(KERN_ERR "IOMMU: Error, gfp(tsb) failed.\n");
136 		goto out_free_dummy_page;
137 	}
138 	iommu->page_table = (iopte_t *)page_address(page);
139 
140 	for (i = 0; i < num_tsb_entries; i++)
141 		iopte_make_dummy(iommu, &iommu->page_table[i]);
142 
143 	return 0;
144 
145 out_free_dummy_page:
146 	free_page(iommu->dummy_page);
147 	iommu->dummy_page = 0UL;
148 
149 out_free_map:
150 	kfree(iommu->tbl.map);
151 	iommu->tbl.map = NULL;
152 
153 	return -ENOMEM;
154 }
155 
alloc_npages(struct device * dev,struct iommu * iommu,unsigned long npages)156 static inline iopte_t *alloc_npages(struct device *dev,
157 				    struct iommu *iommu,
158 				    unsigned long npages)
159 {
160 	unsigned long entry;
161 
162 	entry = iommu_tbl_range_alloc(dev, &iommu->tbl, npages, NULL,
163 				      (unsigned long)(-1), 0);
164 	if (unlikely(entry == IOMMU_ERROR_CODE))
165 		return NULL;
166 
167 	return iommu->page_table + entry;
168 }
169 
iommu_alloc_ctx(struct iommu * iommu)170 static int iommu_alloc_ctx(struct iommu *iommu)
171 {
172 	int lowest = iommu->ctx_lowest_free;
173 	int n = find_next_zero_bit(iommu->ctx_bitmap, IOMMU_NUM_CTXS, lowest);
174 
175 	if (unlikely(n == IOMMU_NUM_CTXS)) {
176 		n = find_next_zero_bit(iommu->ctx_bitmap, lowest, 1);
177 		if (unlikely(n == lowest)) {
178 			printk(KERN_WARNING "IOMMU: Ran out of contexts.\n");
179 			n = 0;
180 		}
181 	}
182 	if (n)
183 		__set_bit(n, iommu->ctx_bitmap);
184 
185 	return n;
186 }
187 
iommu_free_ctx(struct iommu * iommu,int ctx)188 static inline void iommu_free_ctx(struct iommu *iommu, int ctx)
189 {
190 	if (likely(ctx)) {
191 		__clear_bit(ctx, iommu->ctx_bitmap);
192 		if (ctx < iommu->ctx_lowest_free)
193 			iommu->ctx_lowest_free = ctx;
194 	}
195 }
196 
dma_4u_alloc_coherent(struct device * dev,size_t size,dma_addr_t * dma_addrp,gfp_t gfp,unsigned long attrs)197 static void *dma_4u_alloc_coherent(struct device *dev, size_t size,
198 				   dma_addr_t *dma_addrp, gfp_t gfp,
199 				   unsigned long attrs)
200 {
201 	unsigned long order, first_page;
202 	struct iommu *iommu;
203 	struct page *page;
204 	int npages, nid;
205 	iopte_t *iopte;
206 	void *ret;
207 
208 	size = IO_PAGE_ALIGN(size);
209 	order = get_order(size);
210 	if (order >= 10)
211 		return NULL;
212 
213 	nid = dev->archdata.numa_node;
214 	page = alloc_pages_node(nid, gfp, order);
215 	if (unlikely(!page))
216 		return NULL;
217 
218 	first_page = (unsigned long) page_address(page);
219 	memset((char *)first_page, 0, PAGE_SIZE << order);
220 
221 	iommu = dev->archdata.iommu;
222 
223 	iopte = alloc_npages(dev, iommu, size >> IO_PAGE_SHIFT);
224 
225 	if (unlikely(iopte == NULL)) {
226 		free_pages(first_page, order);
227 		return NULL;
228 	}
229 
230 	*dma_addrp = (iommu->tbl.table_map_base +
231 		      ((iopte - iommu->page_table) << IO_PAGE_SHIFT));
232 	ret = (void *) first_page;
233 	npages = size >> IO_PAGE_SHIFT;
234 	first_page = __pa(first_page);
235 	while (npages--) {
236 		iopte_val(*iopte) = (IOPTE_CONSISTENT(0UL) |
237 				     IOPTE_WRITE |
238 				     (first_page & IOPTE_PAGE));
239 		iopte++;
240 		first_page += IO_PAGE_SIZE;
241 	}
242 
243 	return ret;
244 }
245 
dma_4u_free_coherent(struct device * dev,size_t size,void * cpu,dma_addr_t dvma,unsigned long attrs)246 static void dma_4u_free_coherent(struct device *dev, size_t size,
247 				 void *cpu, dma_addr_t dvma,
248 				 unsigned long attrs)
249 {
250 	struct iommu *iommu;
251 	unsigned long order, npages;
252 
253 	npages = IO_PAGE_ALIGN(size) >> IO_PAGE_SHIFT;
254 	iommu = dev->archdata.iommu;
255 
256 	iommu_tbl_range_free(&iommu->tbl, dvma, npages, IOMMU_ERROR_CODE);
257 
258 	order = get_order(size);
259 	if (order < 10)
260 		free_pages((unsigned long)cpu, order);
261 }
262 
dma_4u_map_phys(struct device * dev,phys_addr_t phys,size_t sz,enum dma_data_direction direction,unsigned long attrs)263 static dma_addr_t dma_4u_map_phys(struct device *dev, phys_addr_t phys,
264 				  size_t sz, enum dma_data_direction direction,
265 				  unsigned long attrs)
266 {
267 	struct iommu *iommu;
268 	struct strbuf *strbuf;
269 	iopte_t *base;
270 	unsigned long flags, npages, oaddr;
271 	unsigned long i, ctx;
272 	u32 bus_addr, ret;
273 	unsigned long iopte_protection;
274 
275 	if (unlikely(attrs & DMA_ATTR_MMIO))
276 		/*
277 		 * This check is included because older versions of the code
278 		 * lacked MMIO path support, and my ability to test this path
279 		 * is limited. However, from a software technical standpoint,
280 		 * there is no restriction, as the following code operates
281 		 * solely on physical addresses.
282 		 */
283 		goto bad_no_ctx;
284 
285 	iommu = dev->archdata.iommu;
286 	strbuf = dev->archdata.stc;
287 
288 	if (unlikely(direction == DMA_NONE))
289 		goto bad_no_ctx;
290 
291 	oaddr = (unsigned long)(phys_to_virt(phys));
292 	npages = IO_PAGE_ALIGN(oaddr + sz) - (oaddr & IO_PAGE_MASK);
293 	npages >>= IO_PAGE_SHIFT;
294 
295 	base = alloc_npages(dev, iommu, npages);
296 	spin_lock_irqsave(&iommu->lock, flags);
297 	ctx = 0;
298 	if (iommu->iommu_ctxflush)
299 		ctx = iommu_alloc_ctx(iommu);
300 	spin_unlock_irqrestore(&iommu->lock, flags);
301 
302 	if (unlikely(!base))
303 		goto bad;
304 
305 	bus_addr = (iommu->tbl.table_map_base +
306 		    ((base - iommu->page_table) << IO_PAGE_SHIFT));
307 	ret = bus_addr | (oaddr & ~IO_PAGE_MASK);
308 	if (strbuf->strbuf_enabled)
309 		iopte_protection = IOPTE_STREAMING(ctx);
310 	else
311 		iopte_protection = IOPTE_CONSISTENT(ctx);
312 	if (direction != DMA_TO_DEVICE)
313 		iopte_protection |= IOPTE_WRITE;
314 
315 	phys &= IO_PAGE_MASK;
316 
317 	for (i = 0; i < npages; i++, base++, phys += IO_PAGE_SIZE)
318 		iopte_val(*base) = iopte_protection | phys;
319 
320 	return ret;
321 
322 bad:
323 	iommu_free_ctx(iommu, ctx);
324 bad_no_ctx:
325 	if (printk_ratelimit())
326 		WARN_ON(1);
327 	return DMA_MAPPING_ERROR;
328 }
329 
strbuf_flush(struct strbuf * strbuf,struct iommu * iommu,u32 vaddr,unsigned long ctx,unsigned long npages,enum dma_data_direction direction)330 static void strbuf_flush(struct strbuf *strbuf, struct iommu *iommu,
331 			 u32 vaddr, unsigned long ctx, unsigned long npages,
332 			 enum dma_data_direction direction)
333 {
334 	int limit;
335 
336 	if (strbuf->strbuf_ctxflush &&
337 	    iommu->iommu_ctxflush) {
338 		unsigned long matchreg, flushreg;
339 		u64 val;
340 
341 		flushreg = strbuf->strbuf_ctxflush;
342 		matchreg = STC_CTXMATCH_ADDR(strbuf, ctx);
343 
344 		iommu_write(flushreg, ctx);
345 		val = iommu_read(matchreg);
346 		val &= 0xffff;
347 		if (!val)
348 			goto do_flush_sync;
349 
350 		while (val) {
351 			if (val & 0x1)
352 				iommu_write(flushreg, ctx);
353 			val >>= 1;
354 		}
355 		val = iommu_read(matchreg);
356 		if (unlikely(val)) {
357 			printk(KERN_WARNING "strbuf_flush: ctx flush "
358 			       "timeout matchreg[%llx] ctx[%lx]\n",
359 			       val, ctx);
360 			goto do_page_flush;
361 		}
362 	} else {
363 		unsigned long i;
364 
365 	do_page_flush:
366 		for (i = 0; i < npages; i++, vaddr += IO_PAGE_SIZE)
367 			iommu_write(strbuf->strbuf_pflush, vaddr);
368 	}
369 
370 do_flush_sync:
371 	/* If the device could not have possibly put dirty data into
372 	 * the streaming cache, no flush-flag synchronization needs
373 	 * to be performed.
374 	 */
375 	if (direction == DMA_TO_DEVICE)
376 		return;
377 
378 	STC_FLUSHFLAG_INIT(strbuf);
379 	iommu_write(strbuf->strbuf_fsync, strbuf->strbuf_flushflag_pa);
380 	(void) iommu_read(iommu->write_complete_reg);
381 
382 	limit = 100000;
383 	while (!STC_FLUSHFLAG_SET(strbuf)) {
384 		limit--;
385 		if (!limit)
386 			break;
387 		udelay(1);
388 		rmb();
389 	}
390 	if (!limit)
391 		printk(KERN_WARNING "strbuf_flush: flushflag timeout "
392 		       "vaddr[%08x] ctx[%lx] npages[%ld]\n",
393 		       vaddr, ctx, npages);
394 }
395 
dma_4u_unmap_phys(struct device * dev,dma_addr_t bus_addr,size_t sz,enum dma_data_direction direction,unsigned long attrs)396 static void dma_4u_unmap_phys(struct device *dev, dma_addr_t bus_addr,
397 			      size_t sz, enum dma_data_direction direction,
398 			      unsigned long attrs)
399 {
400 	struct iommu *iommu;
401 	struct strbuf *strbuf;
402 	iopte_t *base;
403 	unsigned long flags, npages, ctx, i;
404 
405 	if (unlikely(direction == DMA_NONE)) {
406 		if (printk_ratelimit())
407 			WARN_ON(1);
408 		return;
409 	}
410 
411 	iommu = dev->archdata.iommu;
412 	strbuf = dev->archdata.stc;
413 
414 	npages = IO_PAGE_ALIGN(bus_addr + sz) - (bus_addr & IO_PAGE_MASK);
415 	npages >>= IO_PAGE_SHIFT;
416 	base = iommu->page_table +
417 		((bus_addr - iommu->tbl.table_map_base) >> IO_PAGE_SHIFT);
418 	bus_addr &= IO_PAGE_MASK;
419 
420 	spin_lock_irqsave(&iommu->lock, flags);
421 
422 	/* Record the context, if any. */
423 	ctx = 0;
424 	if (iommu->iommu_ctxflush)
425 		ctx = (iopte_val(*base) & IOPTE_CONTEXT) >> 47UL;
426 
427 	/* Step 1: Kick data out of streaming buffers if necessary. */
428 	if (strbuf->strbuf_enabled && !(attrs & DMA_ATTR_SKIP_CPU_SYNC))
429 		strbuf_flush(strbuf, iommu, bus_addr, ctx,
430 			     npages, direction);
431 
432 	/* Step 2: Clear out TSB entries. */
433 	for (i = 0; i < npages; i++)
434 		iopte_make_dummy(iommu, base + i);
435 
436 	iommu_free_ctx(iommu, ctx);
437 	spin_unlock_irqrestore(&iommu->lock, flags);
438 
439 	iommu_tbl_range_free(&iommu->tbl, bus_addr, npages, IOMMU_ERROR_CODE);
440 }
441 
dma_4u_map_sg(struct device * dev,struct scatterlist * sglist,int nelems,enum dma_data_direction direction,unsigned long attrs)442 static int dma_4u_map_sg(struct device *dev, struct scatterlist *sglist,
443 			 int nelems, enum dma_data_direction direction,
444 			 unsigned long attrs)
445 {
446 	struct scatterlist *s, *outs, *segstart;
447 	unsigned long flags, handle, prot, ctx;
448 	dma_addr_t dma_next = 0, dma_addr;
449 	unsigned int max_seg_size;
450 	unsigned long seg_boundary_size;
451 	int outcount, incount, i;
452 	struct strbuf *strbuf;
453 	struct iommu *iommu;
454 	unsigned long base_shift;
455 
456 	BUG_ON(direction == DMA_NONE);
457 
458 	iommu = dev->archdata.iommu;
459 	strbuf = dev->archdata.stc;
460 	if (nelems == 0 || !iommu)
461 		return -EINVAL;
462 
463 	spin_lock_irqsave(&iommu->lock, flags);
464 
465 	ctx = 0;
466 	if (iommu->iommu_ctxflush)
467 		ctx = iommu_alloc_ctx(iommu);
468 
469 	if (strbuf->strbuf_enabled)
470 		prot = IOPTE_STREAMING(ctx);
471 	else
472 		prot = IOPTE_CONSISTENT(ctx);
473 	if (direction != DMA_TO_DEVICE)
474 		prot |= IOPTE_WRITE;
475 
476 	outs = s = segstart = &sglist[0];
477 	outcount = 1;
478 	incount = nelems;
479 	handle = 0;
480 
481 	/* Init first segment length for backout at failure */
482 	outs->dma_length = 0;
483 
484 	max_seg_size = dma_get_max_seg_size(dev);
485 	seg_boundary_size = dma_get_seg_boundary_nr_pages(dev, IO_PAGE_SHIFT);
486 	base_shift = iommu->tbl.table_map_base >> IO_PAGE_SHIFT;
487 	for_each_sg(sglist, s, nelems, i) {
488 		unsigned long paddr, npages, entry, out_entry = 0, slen;
489 		iopte_t *base;
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 		paddr = (unsigned long) SG_ENT_PHYS_ADDRESS(s);
499 		npages = iommu_num_pages(paddr, slen, IO_PAGE_SIZE);
500 		entry = iommu_tbl_range_alloc(dev, &iommu->tbl, npages,
501 					      &handle, (unsigned long)(-1), 0);
502 
503 		/* Handle failure */
504 		if (unlikely(entry == IOMMU_ERROR_CODE)) {
505 			if (printk_ratelimit())
506 				printk(KERN_INFO "iommu_alloc failed, iommu %p paddr %lx"
507 				       " npages %lx\n", iommu, paddr, npages);
508 			goto iommu_map_failed;
509 		}
510 
511 		base = iommu->page_table + entry;
512 
513 		/* Convert entry to a dma_addr_t */
514 		dma_addr = iommu->tbl.table_map_base +
515 			(entry << IO_PAGE_SHIFT);
516 		dma_addr |= (s->offset & ~IO_PAGE_MASK);
517 
518 		/* Insert into HW table */
519 		paddr &= IO_PAGE_MASK;
520 		while (npages--) {
521 			iopte_val(*base) = prot | paddr;
522 			base++;
523 			paddr += IO_PAGE_SIZE;
524 		}
525 
526 		/* If we are in an open segment, try merging */
527 		if (segstart != s) {
528 			/* We cannot merge if:
529 			 * - allocated dma_addr isn't contiguous to previous allocation
530 			 */
531 			if ((dma_addr != dma_next) ||
532 			    (outs->dma_length + s->length > max_seg_size) ||
533 			    (is_span_boundary(out_entry, base_shift,
534 					      seg_boundary_size, outs, s))) {
535 				/* Can't merge: create a new segment */
536 				segstart = s;
537 				outcount++;
538 				outs = sg_next(outs);
539 			} else {
540 				outs->dma_length += s->length;
541 			}
542 		}
543 
544 		if (segstart == s) {
545 			/* This is a new segment, fill entries */
546 			outs->dma_address = dma_addr;
547 			outs->dma_length = slen;
548 			out_entry = entry;
549 		}
550 
551 		/* Calculate next page pointer for contiguous check */
552 		dma_next = dma_addr + slen;
553 	}
554 
555 	spin_unlock_irqrestore(&iommu->lock, flags);
556 
557 	if (outcount < incount) {
558 		outs = sg_next(outs);
559 		outs->dma_length = 0;
560 	}
561 
562 	return outcount;
563 
564 iommu_map_failed:
565 	for_each_sg(sglist, s, nelems, i) {
566 		if (s->dma_length != 0) {
567 			unsigned long vaddr, npages, entry, j;
568 			iopte_t *base;
569 
570 			vaddr = s->dma_address & IO_PAGE_MASK;
571 			npages = iommu_num_pages(s->dma_address, s->dma_length,
572 						 IO_PAGE_SIZE);
573 
574 			entry = (vaddr - iommu->tbl.table_map_base)
575 				>> IO_PAGE_SHIFT;
576 			base = iommu->page_table + entry;
577 
578 			for (j = 0; j < npages; j++)
579 				iopte_make_dummy(iommu, base + j);
580 
581 			iommu_tbl_range_free(&iommu->tbl, vaddr, npages,
582 					     IOMMU_ERROR_CODE);
583 
584 			s->dma_length = 0;
585 		}
586 		if (s == outs)
587 			break;
588 	}
589 	spin_unlock_irqrestore(&iommu->lock, flags);
590 
591 	return -EINVAL;
592 }
593 
594 /* If contexts are being used, they are the same in all of the mappings
595  * we make for a particular SG.
596  */
fetch_sg_ctx(struct iommu * iommu,struct scatterlist * sg)597 static unsigned long fetch_sg_ctx(struct iommu *iommu, struct scatterlist *sg)
598 {
599 	unsigned long ctx = 0;
600 
601 	if (iommu->iommu_ctxflush) {
602 		iopte_t *base;
603 		u32 bus_addr;
604 		struct iommu_map_table *tbl = &iommu->tbl;
605 
606 		bus_addr = sg->dma_address & IO_PAGE_MASK;
607 		base = iommu->page_table +
608 			((bus_addr - tbl->table_map_base) >> IO_PAGE_SHIFT);
609 
610 		ctx = (iopte_val(*base) & IOPTE_CONTEXT) >> 47UL;
611 	}
612 	return ctx;
613 }
614 
dma_4u_unmap_sg(struct device * dev,struct scatterlist * sglist,int nelems,enum dma_data_direction direction,unsigned long attrs)615 static void dma_4u_unmap_sg(struct device *dev, struct scatterlist *sglist,
616 			    int nelems, enum dma_data_direction direction,
617 			    unsigned long attrs)
618 {
619 	unsigned long flags, ctx;
620 	struct scatterlist *sg;
621 	struct strbuf *strbuf;
622 	struct iommu *iommu;
623 
624 	BUG_ON(direction == DMA_NONE);
625 
626 	iommu = dev->archdata.iommu;
627 	strbuf = dev->archdata.stc;
628 
629 	ctx = fetch_sg_ctx(iommu, sglist);
630 
631 	spin_lock_irqsave(&iommu->lock, flags);
632 
633 	sg = sglist;
634 	while (nelems--) {
635 		dma_addr_t dma_handle = sg->dma_address;
636 		unsigned int len = sg->dma_length;
637 		unsigned long npages, entry;
638 		iopte_t *base;
639 		int i;
640 
641 		if (!len)
642 			break;
643 		npages = iommu_num_pages(dma_handle, len, IO_PAGE_SIZE);
644 
645 		entry = ((dma_handle - iommu->tbl.table_map_base)
646 			 >> IO_PAGE_SHIFT);
647 		base = iommu->page_table + entry;
648 
649 		dma_handle &= IO_PAGE_MASK;
650 		if (strbuf->strbuf_enabled && !(attrs & DMA_ATTR_SKIP_CPU_SYNC))
651 			strbuf_flush(strbuf, iommu, dma_handle, ctx,
652 				     npages, direction);
653 
654 		for (i = 0; i < npages; i++)
655 			iopte_make_dummy(iommu, base + i);
656 
657 		iommu_tbl_range_free(&iommu->tbl, dma_handle, npages,
658 				     IOMMU_ERROR_CODE);
659 		sg = sg_next(sg);
660 	}
661 
662 	iommu_free_ctx(iommu, ctx);
663 
664 	spin_unlock_irqrestore(&iommu->lock, flags);
665 }
666 
dma_4u_sync_single_for_cpu(struct device * dev,dma_addr_t bus_addr,size_t sz,enum dma_data_direction direction)667 static void dma_4u_sync_single_for_cpu(struct device *dev,
668 				       dma_addr_t bus_addr, size_t sz,
669 				       enum dma_data_direction direction)
670 {
671 	struct iommu *iommu;
672 	struct strbuf *strbuf;
673 	unsigned long flags, ctx, npages;
674 
675 	iommu = dev->archdata.iommu;
676 	strbuf = dev->archdata.stc;
677 
678 	if (!strbuf->strbuf_enabled)
679 		return;
680 
681 	spin_lock_irqsave(&iommu->lock, flags);
682 
683 	npages = IO_PAGE_ALIGN(bus_addr + sz) - (bus_addr & IO_PAGE_MASK);
684 	npages >>= IO_PAGE_SHIFT;
685 	bus_addr &= IO_PAGE_MASK;
686 
687 	/* Step 1: Record the context, if any. */
688 	ctx = 0;
689 	if (iommu->iommu_ctxflush &&
690 	    strbuf->strbuf_ctxflush) {
691 		iopte_t *iopte;
692 		struct iommu_map_table *tbl = &iommu->tbl;
693 
694 		iopte = iommu->page_table +
695 			((bus_addr - tbl->table_map_base)>>IO_PAGE_SHIFT);
696 		ctx = (iopte_val(*iopte) & IOPTE_CONTEXT) >> 47UL;
697 	}
698 
699 	/* Step 2: Kick data out of streaming buffers. */
700 	strbuf_flush(strbuf, iommu, bus_addr, ctx, npages, direction);
701 
702 	spin_unlock_irqrestore(&iommu->lock, flags);
703 }
704 
dma_4u_sync_sg_for_cpu(struct device * dev,struct scatterlist * sglist,int nelems,enum dma_data_direction direction)705 static void dma_4u_sync_sg_for_cpu(struct device *dev,
706 				   struct scatterlist *sglist, int nelems,
707 				   enum dma_data_direction direction)
708 {
709 	struct iommu *iommu;
710 	struct strbuf *strbuf;
711 	unsigned long flags, ctx, npages, i;
712 	struct scatterlist *sg, *sgprv;
713 	u32 bus_addr;
714 
715 	iommu = dev->archdata.iommu;
716 	strbuf = dev->archdata.stc;
717 
718 	if (!strbuf->strbuf_enabled)
719 		return;
720 
721 	spin_lock_irqsave(&iommu->lock, flags);
722 
723 	/* Step 1: Record the context, if any. */
724 	ctx = 0;
725 	if (iommu->iommu_ctxflush &&
726 	    strbuf->strbuf_ctxflush) {
727 		iopte_t *iopte;
728 		struct iommu_map_table *tbl = &iommu->tbl;
729 
730 		iopte = iommu->page_table + ((sglist[0].dma_address -
731 			tbl->table_map_base) >> IO_PAGE_SHIFT);
732 		ctx = (iopte_val(*iopte) & IOPTE_CONTEXT) >> 47UL;
733 	}
734 
735 	/* Step 2: Kick data out of streaming buffers. */
736 	bus_addr = sglist[0].dma_address & IO_PAGE_MASK;
737 	sgprv = NULL;
738 	for_each_sg(sglist, sg, nelems, i) {
739 		if (sg->dma_length == 0)
740 			break;
741 		sgprv = sg;
742 	}
743 
744 	npages = (IO_PAGE_ALIGN(sgprv->dma_address + sgprv->dma_length)
745 		  - bus_addr) >> IO_PAGE_SHIFT;
746 	strbuf_flush(strbuf, iommu, bus_addr, ctx, npages, direction);
747 
748 	spin_unlock_irqrestore(&iommu->lock, flags);
749 }
750 
dma_4u_supported(struct device * dev,u64 device_mask)751 static int dma_4u_supported(struct device *dev, u64 device_mask)
752 {
753 	struct iommu *iommu = dev->archdata.iommu;
754 
755 	if (ali_sound_dma_hack(dev, device_mask))
756 		return 1;
757 
758 	if (device_mask < iommu->dma_addr_mask)
759 		return 0;
760 	return 1;
761 }
762 
763 static const struct dma_map_ops sun4u_dma_ops = {
764 	.alloc			= dma_4u_alloc_coherent,
765 	.free			= dma_4u_free_coherent,
766 	.map_phys		= dma_4u_map_phys,
767 	.unmap_phys		= dma_4u_unmap_phys,
768 	.map_sg			= dma_4u_map_sg,
769 	.unmap_sg		= dma_4u_unmap_sg,
770 	.sync_single_for_cpu	= dma_4u_sync_single_for_cpu,
771 	.sync_sg_for_cpu	= dma_4u_sync_sg_for_cpu,
772 	.dma_supported		= dma_4u_supported,
773 };
774 
775 const struct dma_map_ops *dma_ops = &sun4u_dma_ops;
776 EXPORT_SYMBOL(dma_ops);
777