xref: /linux/drivers/gpu/drm/gma500/mmu.c (revision 64b14a184e83eb62ea0615e31a409956049d40e7)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /**************************************************************************
3  * Copyright (c) 2007, Intel Corporation.
4  *
5  **************************************************************************/
6 
7 #include <linux/highmem.h>
8 
9 #include "mmu.h"
10 #include "psb_drv.h"
11 #include "psb_reg.h"
12 
13 /*
14  * Code for the SGX MMU:
15  */
16 
17 /*
18  * clflush on one processor only:
19  * clflush should apparently flush the cache line on all processors in an
20  * SMP system.
21  */
22 
23 /*
24  * kmap atomic:
25  * The usage of the slots must be completely encapsulated within a spinlock, and
26  * no other functions that may be using the locks for other purposed may be
27  * called from within the locked region.
28  * Since the slots are per processor, this will guarantee that we are the only
29  * user.
30  */
31 
32 /*
33  * TODO: Inserting ptes from an interrupt handler:
34  * This may be desirable for some SGX functionality where the GPU can fault in
35  * needed pages. For that, we need to make an atomic insert_pages function, that
36  * may fail.
37  * If it fails, the caller need to insert the page using a workqueue function,
38  * but on average it should be fast.
39  */
40 
41 static inline uint32_t psb_mmu_pt_index(uint32_t offset)
42 {
43 	return (offset >> PSB_PTE_SHIFT) & 0x3FF;
44 }
45 
46 static inline uint32_t psb_mmu_pd_index(uint32_t offset)
47 {
48 	return offset >> PSB_PDE_SHIFT;
49 }
50 
51 static inline void psb_clflush(void *addr)
52 {
53 	__asm__ __volatile__("clflush (%0)\n" : : "r"(addr) : "memory");
54 }
55 
56 static inline void psb_mmu_clflush(struct psb_mmu_driver *driver, void *addr)
57 {
58 	if (!driver->has_clflush)
59 		return;
60 
61 	mb();
62 	psb_clflush(addr);
63 	mb();
64 }
65 
66 static void psb_mmu_flush_pd_locked(struct psb_mmu_driver *driver, int force)
67 {
68 	struct drm_device *dev = driver->dev;
69 	struct drm_psb_private *dev_priv = to_drm_psb_private(dev);
70 
71 	if (atomic_read(&driver->needs_tlbflush) || force) {
72 		uint32_t val = PSB_RSGX32(PSB_CR_BIF_CTRL);
73 		PSB_WSGX32(val | _PSB_CB_CTRL_INVALDC, PSB_CR_BIF_CTRL);
74 
75 		/* Make sure data cache is turned off before enabling it */
76 		wmb();
77 		PSB_WSGX32(val & ~_PSB_CB_CTRL_INVALDC, PSB_CR_BIF_CTRL);
78 		(void)PSB_RSGX32(PSB_CR_BIF_CTRL);
79 		if (driver->msvdx_mmu_invaldc)
80 			atomic_set(driver->msvdx_mmu_invaldc, 1);
81 	}
82 	atomic_set(&driver->needs_tlbflush, 0);
83 }
84 
85 #if 0
86 static void psb_mmu_flush_pd(struct psb_mmu_driver *driver, int force)
87 {
88 	down_write(&driver->sem);
89 	psb_mmu_flush_pd_locked(driver, force);
90 	up_write(&driver->sem);
91 }
92 #endif
93 
94 void psb_mmu_flush(struct psb_mmu_driver *driver)
95 {
96 	struct drm_device *dev = driver->dev;
97 	struct drm_psb_private *dev_priv = to_drm_psb_private(dev);
98 	uint32_t val;
99 
100 	down_write(&driver->sem);
101 	val = PSB_RSGX32(PSB_CR_BIF_CTRL);
102 	if (atomic_read(&driver->needs_tlbflush))
103 		PSB_WSGX32(val | _PSB_CB_CTRL_INVALDC, PSB_CR_BIF_CTRL);
104 	else
105 		PSB_WSGX32(val | _PSB_CB_CTRL_FLUSH, PSB_CR_BIF_CTRL);
106 
107 	/* Make sure data cache is turned off and MMU is flushed before
108 	   restoring bank interface control register */
109 	wmb();
110 	PSB_WSGX32(val & ~(_PSB_CB_CTRL_FLUSH | _PSB_CB_CTRL_INVALDC),
111 		   PSB_CR_BIF_CTRL);
112 	(void)PSB_RSGX32(PSB_CR_BIF_CTRL);
113 
114 	atomic_set(&driver->needs_tlbflush, 0);
115 	if (driver->msvdx_mmu_invaldc)
116 		atomic_set(driver->msvdx_mmu_invaldc, 1);
117 	up_write(&driver->sem);
118 }
119 
120 void psb_mmu_set_pd_context(struct psb_mmu_pd *pd, int hw_context)
121 {
122 	struct drm_device *dev = pd->driver->dev;
123 	struct drm_psb_private *dev_priv = to_drm_psb_private(dev);
124 	uint32_t offset = (hw_context == 0) ? PSB_CR_BIF_DIR_LIST_BASE0 :
125 			  PSB_CR_BIF_DIR_LIST_BASE1 + hw_context * 4;
126 
127 	down_write(&pd->driver->sem);
128 	PSB_WSGX32(page_to_pfn(pd->p) << PAGE_SHIFT, offset);
129 	wmb();
130 	psb_mmu_flush_pd_locked(pd->driver, 1);
131 	pd->hw_context = hw_context;
132 	up_write(&pd->driver->sem);
133 
134 }
135 
136 static inline unsigned long psb_pd_addr_end(unsigned long addr,
137 					    unsigned long end)
138 {
139 	addr = (addr + PSB_PDE_MASK + 1) & ~PSB_PDE_MASK;
140 	return (addr < end) ? addr : end;
141 }
142 
143 static inline uint32_t psb_mmu_mask_pte(uint32_t pfn, int type)
144 {
145 	uint32_t mask = PSB_PTE_VALID;
146 
147 	if (type & PSB_MMU_CACHED_MEMORY)
148 		mask |= PSB_PTE_CACHED;
149 	if (type & PSB_MMU_RO_MEMORY)
150 		mask |= PSB_PTE_RO;
151 	if (type & PSB_MMU_WO_MEMORY)
152 		mask |= PSB_PTE_WO;
153 
154 	return (pfn << PAGE_SHIFT) | mask;
155 }
156 
157 struct psb_mmu_pd *psb_mmu_alloc_pd(struct psb_mmu_driver *driver,
158 				    int trap_pagefaults, int invalid_type)
159 {
160 	struct psb_mmu_pd *pd = kmalloc(sizeof(*pd), GFP_KERNEL);
161 	uint32_t *v;
162 	int i;
163 
164 	if (!pd)
165 		return NULL;
166 
167 	pd->p = alloc_page(GFP_DMA32);
168 	if (!pd->p)
169 		goto out_err1;
170 	pd->dummy_pt = alloc_page(GFP_DMA32);
171 	if (!pd->dummy_pt)
172 		goto out_err2;
173 	pd->dummy_page = alloc_page(GFP_DMA32);
174 	if (!pd->dummy_page)
175 		goto out_err3;
176 
177 	if (!trap_pagefaults) {
178 		pd->invalid_pde = psb_mmu_mask_pte(page_to_pfn(pd->dummy_pt),
179 						   invalid_type);
180 		pd->invalid_pte = psb_mmu_mask_pte(page_to_pfn(pd->dummy_page),
181 						   invalid_type);
182 	} else {
183 		pd->invalid_pde = 0;
184 		pd->invalid_pte = 0;
185 	}
186 
187 	v = kmap_local_page(pd->dummy_pt);
188 	for (i = 0; i < (PAGE_SIZE / sizeof(uint32_t)); ++i)
189 		v[i] = pd->invalid_pte;
190 
191 	kunmap_local(v);
192 
193 	v = kmap_local_page(pd->p);
194 	for (i = 0; i < (PAGE_SIZE / sizeof(uint32_t)); ++i)
195 		v[i] = pd->invalid_pde;
196 
197 	kunmap_local(v);
198 
199 	clear_page(kmap(pd->dummy_page));
200 	kunmap(pd->dummy_page);
201 
202 	pd->tables = vmalloc_user(sizeof(struct psb_mmu_pt *) * 1024);
203 	if (!pd->tables)
204 		goto out_err4;
205 
206 	pd->hw_context = -1;
207 	pd->pd_mask = PSB_PTE_VALID;
208 	pd->driver = driver;
209 
210 	return pd;
211 
212 out_err4:
213 	__free_page(pd->dummy_page);
214 out_err3:
215 	__free_page(pd->dummy_pt);
216 out_err2:
217 	__free_page(pd->p);
218 out_err1:
219 	kfree(pd);
220 	return NULL;
221 }
222 
223 static void psb_mmu_free_pt(struct psb_mmu_pt *pt)
224 {
225 	__free_page(pt->p);
226 	kfree(pt);
227 }
228 
229 void psb_mmu_free_pagedir(struct psb_mmu_pd *pd)
230 {
231 	struct psb_mmu_driver *driver = pd->driver;
232 	struct drm_device *dev = driver->dev;
233 	struct drm_psb_private *dev_priv = to_drm_psb_private(dev);
234 	struct psb_mmu_pt *pt;
235 	int i;
236 
237 	down_write(&driver->sem);
238 	if (pd->hw_context != -1) {
239 		PSB_WSGX32(0, PSB_CR_BIF_DIR_LIST_BASE0 + pd->hw_context * 4);
240 		psb_mmu_flush_pd_locked(driver, 1);
241 	}
242 
243 	/* Should take the spinlock here, but we don't need to do that
244 	   since we have the semaphore in write mode. */
245 
246 	for (i = 0; i < 1024; ++i) {
247 		pt = pd->tables[i];
248 		if (pt)
249 			psb_mmu_free_pt(pt);
250 	}
251 
252 	vfree(pd->tables);
253 	__free_page(pd->dummy_page);
254 	__free_page(pd->dummy_pt);
255 	__free_page(pd->p);
256 	kfree(pd);
257 	up_write(&driver->sem);
258 }
259 
260 static struct psb_mmu_pt *psb_mmu_alloc_pt(struct psb_mmu_pd *pd)
261 {
262 	struct psb_mmu_pt *pt = kmalloc(sizeof(*pt), GFP_KERNEL);
263 	void *v;
264 	uint32_t clflush_add = pd->driver->clflush_add >> PAGE_SHIFT;
265 	uint32_t clflush_count = PAGE_SIZE / clflush_add;
266 	spinlock_t *lock = &pd->driver->lock;
267 	uint8_t *clf;
268 	uint32_t *ptes;
269 	int i;
270 
271 	if (!pt)
272 		return NULL;
273 
274 	pt->p = alloc_page(GFP_DMA32);
275 	if (!pt->p) {
276 		kfree(pt);
277 		return NULL;
278 	}
279 
280 	spin_lock(lock);
281 
282 	v = kmap_atomic(pt->p);
283 	clf = (uint8_t *) v;
284 	ptes = (uint32_t *) v;
285 	for (i = 0; i < (PAGE_SIZE / sizeof(uint32_t)); ++i)
286 		*ptes++ = pd->invalid_pte;
287 
288 	if (pd->driver->has_clflush && pd->hw_context != -1) {
289 		mb();
290 		for (i = 0; i < clflush_count; ++i) {
291 			psb_clflush(clf);
292 			clf += clflush_add;
293 		}
294 		mb();
295 	}
296 	kunmap_atomic(v);
297 	spin_unlock(lock);
298 
299 	pt->count = 0;
300 	pt->pd = pd;
301 	pt->index = 0;
302 
303 	return pt;
304 }
305 
306 static struct psb_mmu_pt *psb_mmu_pt_alloc_map_lock(struct psb_mmu_pd *pd,
307 						    unsigned long addr)
308 {
309 	uint32_t index = psb_mmu_pd_index(addr);
310 	struct psb_mmu_pt *pt;
311 	uint32_t *v;
312 	spinlock_t *lock = &pd->driver->lock;
313 
314 	spin_lock(lock);
315 	pt = pd->tables[index];
316 	while (!pt) {
317 		spin_unlock(lock);
318 		pt = psb_mmu_alloc_pt(pd);
319 		if (!pt)
320 			return NULL;
321 		spin_lock(lock);
322 
323 		if (pd->tables[index]) {
324 			spin_unlock(lock);
325 			psb_mmu_free_pt(pt);
326 			spin_lock(lock);
327 			pt = pd->tables[index];
328 			continue;
329 		}
330 
331 		v = kmap_atomic(pd->p);
332 		pd->tables[index] = pt;
333 		v[index] = (page_to_pfn(pt->p) << 12) | pd->pd_mask;
334 		pt->index = index;
335 		kunmap_atomic((void *) v);
336 
337 		if (pd->hw_context != -1) {
338 			psb_mmu_clflush(pd->driver, (void *)&v[index]);
339 			atomic_set(&pd->driver->needs_tlbflush, 1);
340 		}
341 	}
342 	pt->v = kmap_atomic(pt->p);
343 	return pt;
344 }
345 
346 static struct psb_mmu_pt *psb_mmu_pt_map_lock(struct psb_mmu_pd *pd,
347 					      unsigned long addr)
348 {
349 	uint32_t index = psb_mmu_pd_index(addr);
350 	struct psb_mmu_pt *pt;
351 	spinlock_t *lock = &pd->driver->lock;
352 
353 	spin_lock(lock);
354 	pt = pd->tables[index];
355 	if (!pt) {
356 		spin_unlock(lock);
357 		return NULL;
358 	}
359 	pt->v = kmap_atomic(pt->p);
360 	return pt;
361 }
362 
363 static void psb_mmu_pt_unmap_unlock(struct psb_mmu_pt *pt)
364 {
365 	struct psb_mmu_pd *pd = pt->pd;
366 	uint32_t *v;
367 
368 	kunmap_atomic(pt->v);
369 	if (pt->count == 0) {
370 		v = kmap_atomic(pd->p);
371 		v[pt->index] = pd->invalid_pde;
372 		pd->tables[pt->index] = NULL;
373 
374 		if (pd->hw_context != -1) {
375 			psb_mmu_clflush(pd->driver, (void *)&v[pt->index]);
376 			atomic_set(&pd->driver->needs_tlbflush, 1);
377 		}
378 		kunmap_atomic(v);
379 		spin_unlock(&pd->driver->lock);
380 		psb_mmu_free_pt(pt);
381 		return;
382 	}
383 	spin_unlock(&pd->driver->lock);
384 }
385 
386 static inline void psb_mmu_set_pte(struct psb_mmu_pt *pt, unsigned long addr,
387 				   uint32_t pte)
388 {
389 	pt->v[psb_mmu_pt_index(addr)] = pte;
390 }
391 
392 static inline void psb_mmu_invalidate_pte(struct psb_mmu_pt *pt,
393 					  unsigned long addr)
394 {
395 	pt->v[psb_mmu_pt_index(addr)] = pt->pd->invalid_pte;
396 }
397 
398 struct psb_mmu_pd *psb_mmu_get_default_pd(struct psb_mmu_driver *driver)
399 {
400 	struct psb_mmu_pd *pd;
401 
402 	down_read(&driver->sem);
403 	pd = driver->default_pd;
404 	up_read(&driver->sem);
405 
406 	return pd;
407 }
408 
409 void psb_mmu_driver_takedown(struct psb_mmu_driver *driver)
410 {
411 	struct drm_device *dev = driver->dev;
412 	struct drm_psb_private *dev_priv = to_drm_psb_private(dev);
413 
414 	PSB_WSGX32(driver->bif_ctrl, PSB_CR_BIF_CTRL);
415 	psb_mmu_free_pagedir(driver->default_pd);
416 	kfree(driver);
417 }
418 
419 struct psb_mmu_driver *psb_mmu_driver_init(struct drm_device *dev,
420 					   int trap_pagefaults,
421 					   int invalid_type,
422 					   atomic_t *msvdx_mmu_invaldc)
423 {
424 	struct psb_mmu_driver *driver;
425 	struct drm_psb_private *dev_priv = to_drm_psb_private(dev);
426 
427 	driver = kmalloc(sizeof(*driver), GFP_KERNEL);
428 
429 	if (!driver)
430 		return NULL;
431 
432 	driver->dev = dev;
433 	driver->default_pd = psb_mmu_alloc_pd(driver, trap_pagefaults,
434 					      invalid_type);
435 	if (!driver->default_pd)
436 		goto out_err1;
437 
438 	spin_lock_init(&driver->lock);
439 	init_rwsem(&driver->sem);
440 	down_write(&driver->sem);
441 	atomic_set(&driver->needs_tlbflush, 1);
442 	driver->msvdx_mmu_invaldc = msvdx_mmu_invaldc;
443 
444 	driver->bif_ctrl = PSB_RSGX32(PSB_CR_BIF_CTRL);
445 	PSB_WSGX32(driver->bif_ctrl | _PSB_CB_CTRL_CLEAR_FAULT,
446 		   PSB_CR_BIF_CTRL);
447 	PSB_WSGX32(driver->bif_ctrl & ~_PSB_CB_CTRL_CLEAR_FAULT,
448 		   PSB_CR_BIF_CTRL);
449 
450 	driver->has_clflush = 0;
451 
452 	if (boot_cpu_has(X86_FEATURE_CLFLUSH)) {
453 		uint32_t tfms, misc, cap0, cap4, clflush_size;
454 
455 		/*
456 		 * clflush size is determined at kernel setup for x86_64 but not
457 		 * for i386. We have to do it here.
458 		 */
459 
460 		cpuid(0x00000001, &tfms, &misc, &cap0, &cap4);
461 		clflush_size = ((misc >> 8) & 0xff) * 8;
462 		driver->has_clflush = 1;
463 		driver->clflush_add =
464 		    PAGE_SIZE * clflush_size / sizeof(uint32_t);
465 		driver->clflush_mask = driver->clflush_add - 1;
466 		driver->clflush_mask = ~driver->clflush_mask;
467 	}
468 
469 	up_write(&driver->sem);
470 	return driver;
471 
472 out_err1:
473 	kfree(driver);
474 	return NULL;
475 }
476 
477 static void psb_mmu_flush_ptes(struct psb_mmu_pd *pd, unsigned long address,
478 			       uint32_t num_pages, uint32_t desired_tile_stride,
479 			       uint32_t hw_tile_stride)
480 {
481 	struct psb_mmu_pt *pt;
482 	uint32_t rows = 1;
483 	uint32_t i;
484 	unsigned long addr;
485 	unsigned long end;
486 	unsigned long next;
487 	unsigned long add;
488 	unsigned long row_add;
489 	unsigned long clflush_add = pd->driver->clflush_add;
490 	unsigned long clflush_mask = pd->driver->clflush_mask;
491 
492 	if (!pd->driver->has_clflush)
493 		return;
494 
495 	if (hw_tile_stride)
496 		rows = num_pages / desired_tile_stride;
497 	else
498 		desired_tile_stride = num_pages;
499 
500 	add = desired_tile_stride << PAGE_SHIFT;
501 	row_add = hw_tile_stride << PAGE_SHIFT;
502 	mb();
503 	for (i = 0; i < rows; ++i) {
504 
505 		addr = address;
506 		end = addr + add;
507 
508 		do {
509 			next = psb_pd_addr_end(addr, end);
510 			pt = psb_mmu_pt_map_lock(pd, addr);
511 			if (!pt)
512 				continue;
513 			do {
514 				psb_clflush(&pt->v[psb_mmu_pt_index(addr)]);
515 			} while (addr += clflush_add,
516 				 (addr & clflush_mask) < next);
517 
518 			psb_mmu_pt_unmap_unlock(pt);
519 		} while (addr = next, next != end);
520 		address += row_add;
521 	}
522 	mb();
523 }
524 
525 void psb_mmu_remove_pfn_sequence(struct psb_mmu_pd *pd,
526 				 unsigned long address, uint32_t num_pages)
527 {
528 	struct psb_mmu_pt *pt;
529 	unsigned long addr;
530 	unsigned long end;
531 	unsigned long next;
532 	unsigned long f_address = address;
533 
534 	down_read(&pd->driver->sem);
535 
536 	addr = address;
537 	end = addr + (num_pages << PAGE_SHIFT);
538 
539 	do {
540 		next = psb_pd_addr_end(addr, end);
541 		pt = psb_mmu_pt_alloc_map_lock(pd, addr);
542 		if (!pt)
543 			goto out;
544 		do {
545 			psb_mmu_invalidate_pte(pt, addr);
546 			--pt->count;
547 		} while (addr += PAGE_SIZE, addr < next);
548 		psb_mmu_pt_unmap_unlock(pt);
549 
550 	} while (addr = next, next != end);
551 
552 out:
553 	if (pd->hw_context != -1)
554 		psb_mmu_flush_ptes(pd, f_address, num_pages, 1, 1);
555 
556 	up_read(&pd->driver->sem);
557 
558 	if (pd->hw_context != -1)
559 		psb_mmu_flush(pd->driver);
560 
561 	return;
562 }
563 
564 void psb_mmu_remove_pages(struct psb_mmu_pd *pd, unsigned long address,
565 			  uint32_t num_pages, uint32_t desired_tile_stride,
566 			  uint32_t hw_tile_stride)
567 {
568 	struct psb_mmu_pt *pt;
569 	uint32_t rows = 1;
570 	uint32_t i;
571 	unsigned long addr;
572 	unsigned long end;
573 	unsigned long next;
574 	unsigned long add;
575 	unsigned long row_add;
576 	unsigned long f_address = address;
577 
578 	if (hw_tile_stride)
579 		rows = num_pages / desired_tile_stride;
580 	else
581 		desired_tile_stride = num_pages;
582 
583 	add = desired_tile_stride << PAGE_SHIFT;
584 	row_add = hw_tile_stride << PAGE_SHIFT;
585 
586 	down_read(&pd->driver->sem);
587 
588 	/* Make sure we only need to flush this processor's cache */
589 
590 	for (i = 0; i < rows; ++i) {
591 
592 		addr = address;
593 		end = addr + add;
594 
595 		do {
596 			next = psb_pd_addr_end(addr, end);
597 			pt = psb_mmu_pt_map_lock(pd, addr);
598 			if (!pt)
599 				continue;
600 			do {
601 				psb_mmu_invalidate_pte(pt, addr);
602 				--pt->count;
603 
604 			} while (addr += PAGE_SIZE, addr < next);
605 			psb_mmu_pt_unmap_unlock(pt);
606 
607 		} while (addr = next, next != end);
608 		address += row_add;
609 	}
610 	if (pd->hw_context != -1)
611 		psb_mmu_flush_ptes(pd, f_address, num_pages,
612 				   desired_tile_stride, hw_tile_stride);
613 
614 	up_read(&pd->driver->sem);
615 
616 	if (pd->hw_context != -1)
617 		psb_mmu_flush(pd->driver);
618 }
619 
620 int psb_mmu_insert_pfn_sequence(struct psb_mmu_pd *pd, uint32_t start_pfn,
621 				unsigned long address, uint32_t num_pages,
622 				int type)
623 {
624 	struct psb_mmu_pt *pt;
625 	uint32_t pte;
626 	unsigned long addr;
627 	unsigned long end;
628 	unsigned long next;
629 	unsigned long f_address = address;
630 	int ret = -ENOMEM;
631 
632 	down_read(&pd->driver->sem);
633 
634 	addr = address;
635 	end = addr + (num_pages << PAGE_SHIFT);
636 
637 	do {
638 		next = psb_pd_addr_end(addr, end);
639 		pt = psb_mmu_pt_alloc_map_lock(pd, addr);
640 		if (!pt) {
641 			ret = -ENOMEM;
642 			goto out;
643 		}
644 		do {
645 			pte = psb_mmu_mask_pte(start_pfn++, type);
646 			psb_mmu_set_pte(pt, addr, pte);
647 			pt->count++;
648 		} while (addr += PAGE_SIZE, addr < next);
649 		psb_mmu_pt_unmap_unlock(pt);
650 
651 	} while (addr = next, next != end);
652 	ret = 0;
653 
654 out:
655 	if (pd->hw_context != -1)
656 		psb_mmu_flush_ptes(pd, f_address, num_pages, 1, 1);
657 
658 	up_read(&pd->driver->sem);
659 
660 	if (pd->hw_context != -1)
661 		psb_mmu_flush(pd->driver);
662 
663 	return ret;
664 }
665 
666 int psb_mmu_insert_pages(struct psb_mmu_pd *pd, struct page **pages,
667 			 unsigned long address, uint32_t num_pages,
668 			 uint32_t desired_tile_stride, uint32_t hw_tile_stride,
669 			 int type)
670 {
671 	struct psb_mmu_pt *pt;
672 	uint32_t rows = 1;
673 	uint32_t i;
674 	uint32_t pte;
675 	unsigned long addr;
676 	unsigned long end;
677 	unsigned long next;
678 	unsigned long add;
679 	unsigned long row_add;
680 	unsigned long f_address = address;
681 	int ret = -ENOMEM;
682 
683 	if (hw_tile_stride) {
684 		if (num_pages % desired_tile_stride != 0)
685 			return -EINVAL;
686 		rows = num_pages / desired_tile_stride;
687 	} else {
688 		desired_tile_stride = num_pages;
689 	}
690 
691 	add = desired_tile_stride << PAGE_SHIFT;
692 	row_add = hw_tile_stride << PAGE_SHIFT;
693 
694 	down_read(&pd->driver->sem);
695 
696 	for (i = 0; i < rows; ++i) {
697 
698 		addr = address;
699 		end = addr + add;
700 
701 		do {
702 			next = psb_pd_addr_end(addr, end);
703 			pt = psb_mmu_pt_alloc_map_lock(pd, addr);
704 			if (!pt)
705 				goto out;
706 			do {
707 				pte = psb_mmu_mask_pte(page_to_pfn(*pages++),
708 						       type);
709 				psb_mmu_set_pte(pt, addr, pte);
710 				pt->count++;
711 			} while (addr += PAGE_SIZE, addr < next);
712 			psb_mmu_pt_unmap_unlock(pt);
713 
714 		} while (addr = next, next != end);
715 
716 		address += row_add;
717 	}
718 
719 	ret = 0;
720 out:
721 	if (pd->hw_context != -1)
722 		psb_mmu_flush_ptes(pd, f_address, num_pages,
723 				   desired_tile_stride, hw_tile_stride);
724 
725 	up_read(&pd->driver->sem);
726 
727 	if (pd->hw_context != -1)
728 		psb_mmu_flush(pd->driver);
729 
730 	return ret;
731 }
732 
733 int psb_mmu_virtual_to_pfn(struct psb_mmu_pd *pd, uint32_t virtual,
734 			   unsigned long *pfn)
735 {
736 	int ret;
737 	struct psb_mmu_pt *pt;
738 	uint32_t tmp;
739 	spinlock_t *lock = &pd->driver->lock;
740 
741 	down_read(&pd->driver->sem);
742 	pt = psb_mmu_pt_map_lock(pd, virtual);
743 	if (!pt) {
744 		uint32_t *v;
745 
746 		spin_lock(lock);
747 		v = kmap_atomic(pd->p);
748 		tmp = v[psb_mmu_pd_index(virtual)];
749 		kunmap_atomic(v);
750 		spin_unlock(lock);
751 
752 		if (tmp != pd->invalid_pde || !(tmp & PSB_PTE_VALID) ||
753 		    !(pd->invalid_pte & PSB_PTE_VALID)) {
754 			ret = -EINVAL;
755 			goto out;
756 		}
757 		ret = 0;
758 		*pfn = pd->invalid_pte >> PAGE_SHIFT;
759 		goto out;
760 	}
761 	tmp = pt->v[psb_mmu_pt_index(virtual)];
762 	if (!(tmp & PSB_PTE_VALID)) {
763 		ret = -EINVAL;
764 	} else {
765 		ret = 0;
766 		*pfn = tmp >> PAGE_SHIFT;
767 	}
768 	psb_mmu_pt_unmap_unlock(pt);
769 out:
770 	up_read(&pd->driver->sem);
771 	return ret;
772 }
773