xref: /openbmc/linux/drivers/gpu/drm/gma500/mmu.c (revision 5b394b2d)
1 /**************************************************************************
2  * Copyright (c) 2007, Intel Corporation.
3  *
4  * This program is free software; you can redistribute it and/or modify it
5  * under the terms and conditions of the GNU General Public License,
6  * version 2, as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope it will be useful, but WITHOUT
9  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
10  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
11  * more details.
12  *
13  * You should have received a copy of the GNU General Public License along with
14  * this program; if not, write to the Free Software Foundation, Inc.,
15  * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
16  *
17  **************************************************************************/
18 #include <drm/drmP.h>
19 #include "psb_drv.h"
20 #include "psb_reg.h"
21 #include "mmu.h"
22 
23 /*
24  * Code for the SGX MMU:
25  */
26 
27 /*
28  * clflush on one processor only:
29  * clflush should apparently flush the cache line on all processors in an
30  * SMP system.
31  */
32 
33 /*
34  * kmap atomic:
35  * The usage of the slots must be completely encapsulated within a spinlock, and
36  * no other functions that may be using the locks for other purposed may be
37  * called from within the locked region.
38  * Since the slots are per processor, this will guarantee that we are the only
39  * user.
40  */
41 
42 /*
43  * TODO: Inserting ptes from an interrupt handler:
44  * This may be desirable for some SGX functionality where the GPU can fault in
45  * needed pages. For that, we need to make an atomic insert_pages function, that
46  * may fail.
47  * If it fails, the caller need to insert the page using a workqueue function,
48  * but on average it should be fast.
49  */
50 
51 static inline uint32_t psb_mmu_pt_index(uint32_t offset)
52 {
53 	return (offset >> PSB_PTE_SHIFT) & 0x3FF;
54 }
55 
56 static inline uint32_t psb_mmu_pd_index(uint32_t offset)
57 {
58 	return offset >> PSB_PDE_SHIFT;
59 }
60 
61 #if defined(CONFIG_X86)
62 static inline void psb_clflush(void *addr)
63 {
64 	__asm__ __volatile__("clflush (%0)\n" : : "r"(addr) : "memory");
65 }
66 
67 static inline void psb_mmu_clflush(struct psb_mmu_driver *driver, void *addr)
68 {
69 	if (!driver->has_clflush)
70 		return;
71 
72 	mb();
73 	psb_clflush(addr);
74 	mb();
75 }
76 #else
77 
78 static inline void psb_mmu_clflush(struct psb_mmu_driver *driver, void *addr)
79 {;
80 }
81 
82 #endif
83 
84 static void psb_mmu_flush_pd_locked(struct psb_mmu_driver *driver, int force)
85 {
86 	struct drm_device *dev = driver->dev;
87 	struct drm_psb_private *dev_priv = dev->dev_private;
88 
89 	if (atomic_read(&driver->needs_tlbflush) || force) {
90 		uint32_t val = PSB_RSGX32(PSB_CR_BIF_CTRL);
91 		PSB_WSGX32(val | _PSB_CB_CTRL_INVALDC, PSB_CR_BIF_CTRL);
92 
93 		/* Make sure data cache is turned off before enabling it */
94 		wmb();
95 		PSB_WSGX32(val & ~_PSB_CB_CTRL_INVALDC, PSB_CR_BIF_CTRL);
96 		(void)PSB_RSGX32(PSB_CR_BIF_CTRL);
97 		if (driver->msvdx_mmu_invaldc)
98 			atomic_set(driver->msvdx_mmu_invaldc, 1);
99 	}
100 	atomic_set(&driver->needs_tlbflush, 0);
101 }
102 
103 #if 0
104 static void psb_mmu_flush_pd(struct psb_mmu_driver *driver, int force)
105 {
106 	down_write(&driver->sem);
107 	psb_mmu_flush_pd_locked(driver, force);
108 	up_write(&driver->sem);
109 }
110 #endif
111 
112 void psb_mmu_flush(struct psb_mmu_driver *driver)
113 {
114 	struct drm_device *dev = driver->dev;
115 	struct drm_psb_private *dev_priv = dev->dev_private;
116 	uint32_t val;
117 
118 	down_write(&driver->sem);
119 	val = PSB_RSGX32(PSB_CR_BIF_CTRL);
120 	if (atomic_read(&driver->needs_tlbflush))
121 		PSB_WSGX32(val | _PSB_CB_CTRL_INVALDC, PSB_CR_BIF_CTRL);
122 	else
123 		PSB_WSGX32(val | _PSB_CB_CTRL_FLUSH, PSB_CR_BIF_CTRL);
124 
125 	/* Make sure data cache is turned off and MMU is flushed before
126 	   restoring bank interface control register */
127 	wmb();
128 	PSB_WSGX32(val & ~(_PSB_CB_CTRL_FLUSH | _PSB_CB_CTRL_INVALDC),
129 		   PSB_CR_BIF_CTRL);
130 	(void)PSB_RSGX32(PSB_CR_BIF_CTRL);
131 
132 	atomic_set(&driver->needs_tlbflush, 0);
133 	if (driver->msvdx_mmu_invaldc)
134 		atomic_set(driver->msvdx_mmu_invaldc, 1);
135 	up_write(&driver->sem);
136 }
137 
138 void psb_mmu_set_pd_context(struct psb_mmu_pd *pd, int hw_context)
139 {
140 	struct drm_device *dev = pd->driver->dev;
141 	struct drm_psb_private *dev_priv = dev->dev_private;
142 	uint32_t offset = (hw_context == 0) ? PSB_CR_BIF_DIR_LIST_BASE0 :
143 			  PSB_CR_BIF_DIR_LIST_BASE1 + hw_context * 4;
144 
145 	down_write(&pd->driver->sem);
146 	PSB_WSGX32(page_to_pfn(pd->p) << PAGE_SHIFT, offset);
147 	wmb();
148 	psb_mmu_flush_pd_locked(pd->driver, 1);
149 	pd->hw_context = hw_context;
150 	up_write(&pd->driver->sem);
151 
152 }
153 
154 static inline unsigned long psb_pd_addr_end(unsigned long addr,
155 					    unsigned long end)
156 {
157 	addr = (addr + PSB_PDE_MASK + 1) & ~PSB_PDE_MASK;
158 	return (addr < end) ? addr : end;
159 }
160 
161 static inline uint32_t psb_mmu_mask_pte(uint32_t pfn, int type)
162 {
163 	uint32_t mask = PSB_PTE_VALID;
164 
165 	if (type & PSB_MMU_CACHED_MEMORY)
166 		mask |= PSB_PTE_CACHED;
167 	if (type & PSB_MMU_RO_MEMORY)
168 		mask |= PSB_PTE_RO;
169 	if (type & PSB_MMU_WO_MEMORY)
170 		mask |= PSB_PTE_WO;
171 
172 	return (pfn << PAGE_SHIFT) | mask;
173 }
174 
175 struct psb_mmu_pd *psb_mmu_alloc_pd(struct psb_mmu_driver *driver,
176 				    int trap_pagefaults, int invalid_type)
177 {
178 	struct psb_mmu_pd *pd = kmalloc(sizeof(*pd), GFP_KERNEL);
179 	uint32_t *v;
180 	int i;
181 
182 	if (!pd)
183 		return NULL;
184 
185 	pd->p = alloc_page(GFP_DMA32);
186 	if (!pd->p)
187 		goto out_err1;
188 	pd->dummy_pt = alloc_page(GFP_DMA32);
189 	if (!pd->dummy_pt)
190 		goto out_err2;
191 	pd->dummy_page = alloc_page(GFP_DMA32);
192 	if (!pd->dummy_page)
193 		goto out_err3;
194 
195 	if (!trap_pagefaults) {
196 		pd->invalid_pde = psb_mmu_mask_pte(page_to_pfn(pd->dummy_pt),
197 						   invalid_type);
198 		pd->invalid_pte = psb_mmu_mask_pte(page_to_pfn(pd->dummy_page),
199 						   invalid_type);
200 	} else {
201 		pd->invalid_pde = 0;
202 		pd->invalid_pte = 0;
203 	}
204 
205 	v = kmap(pd->dummy_pt);
206 	for (i = 0; i < (PAGE_SIZE / sizeof(uint32_t)); ++i)
207 		v[i] = pd->invalid_pte;
208 
209 	kunmap(pd->dummy_pt);
210 
211 	v = kmap(pd->p);
212 	for (i = 0; i < (PAGE_SIZE / sizeof(uint32_t)); ++i)
213 		v[i] = pd->invalid_pde;
214 
215 	kunmap(pd->p);
216 
217 	clear_page(kmap(pd->dummy_page));
218 	kunmap(pd->dummy_page);
219 
220 	pd->tables = vmalloc_user(sizeof(struct psb_mmu_pt *) * 1024);
221 	if (!pd->tables)
222 		goto out_err4;
223 
224 	pd->hw_context = -1;
225 	pd->pd_mask = PSB_PTE_VALID;
226 	pd->driver = driver;
227 
228 	return pd;
229 
230 out_err4:
231 	__free_page(pd->dummy_page);
232 out_err3:
233 	__free_page(pd->dummy_pt);
234 out_err2:
235 	__free_page(pd->p);
236 out_err1:
237 	kfree(pd);
238 	return NULL;
239 }
240 
241 static void psb_mmu_free_pt(struct psb_mmu_pt *pt)
242 {
243 	__free_page(pt->p);
244 	kfree(pt);
245 }
246 
247 void psb_mmu_free_pagedir(struct psb_mmu_pd *pd)
248 {
249 	struct psb_mmu_driver *driver = pd->driver;
250 	struct drm_device *dev = driver->dev;
251 	struct drm_psb_private *dev_priv = dev->dev_private;
252 	struct psb_mmu_pt *pt;
253 	int i;
254 
255 	down_write(&driver->sem);
256 	if (pd->hw_context != -1) {
257 		PSB_WSGX32(0, PSB_CR_BIF_DIR_LIST_BASE0 + pd->hw_context * 4);
258 		psb_mmu_flush_pd_locked(driver, 1);
259 	}
260 
261 	/* Should take the spinlock here, but we don't need to do that
262 	   since we have the semaphore in write mode. */
263 
264 	for (i = 0; i < 1024; ++i) {
265 		pt = pd->tables[i];
266 		if (pt)
267 			psb_mmu_free_pt(pt);
268 	}
269 
270 	vfree(pd->tables);
271 	__free_page(pd->dummy_page);
272 	__free_page(pd->dummy_pt);
273 	__free_page(pd->p);
274 	kfree(pd);
275 	up_write(&driver->sem);
276 }
277 
278 static struct psb_mmu_pt *psb_mmu_alloc_pt(struct psb_mmu_pd *pd)
279 {
280 	struct psb_mmu_pt *pt = kmalloc(sizeof(*pt), GFP_KERNEL);
281 	void *v;
282 	uint32_t clflush_add = pd->driver->clflush_add >> PAGE_SHIFT;
283 	uint32_t clflush_count = PAGE_SIZE / clflush_add;
284 	spinlock_t *lock = &pd->driver->lock;
285 	uint8_t *clf;
286 	uint32_t *ptes;
287 	int i;
288 
289 	if (!pt)
290 		return NULL;
291 
292 	pt->p = alloc_page(GFP_DMA32);
293 	if (!pt->p) {
294 		kfree(pt);
295 		return NULL;
296 	}
297 
298 	spin_lock(lock);
299 
300 	v = kmap_atomic(pt->p);
301 	clf = (uint8_t *) v;
302 	ptes = (uint32_t *) v;
303 	for (i = 0; i < (PAGE_SIZE / sizeof(uint32_t)); ++i)
304 		*ptes++ = pd->invalid_pte;
305 
306 #if defined(CONFIG_X86)
307 	if (pd->driver->has_clflush && pd->hw_context != -1) {
308 		mb();
309 		for (i = 0; i < clflush_count; ++i) {
310 			psb_clflush(clf);
311 			clf += clflush_add;
312 		}
313 		mb();
314 	}
315 #endif
316 	kunmap_atomic(v);
317 	spin_unlock(lock);
318 
319 	pt->count = 0;
320 	pt->pd = pd;
321 	pt->index = 0;
322 
323 	return pt;
324 }
325 
326 struct psb_mmu_pt *psb_mmu_pt_alloc_map_lock(struct psb_mmu_pd *pd,
327 					     unsigned long addr)
328 {
329 	uint32_t index = psb_mmu_pd_index(addr);
330 	struct psb_mmu_pt *pt;
331 	uint32_t *v;
332 	spinlock_t *lock = &pd->driver->lock;
333 
334 	spin_lock(lock);
335 	pt = pd->tables[index];
336 	while (!pt) {
337 		spin_unlock(lock);
338 		pt = psb_mmu_alloc_pt(pd);
339 		if (!pt)
340 			return NULL;
341 		spin_lock(lock);
342 
343 		if (pd->tables[index]) {
344 			spin_unlock(lock);
345 			psb_mmu_free_pt(pt);
346 			spin_lock(lock);
347 			pt = pd->tables[index];
348 			continue;
349 		}
350 
351 		v = kmap_atomic(pd->p);
352 		pd->tables[index] = pt;
353 		v[index] = (page_to_pfn(pt->p) << 12) | pd->pd_mask;
354 		pt->index = index;
355 		kunmap_atomic((void *) v);
356 
357 		if (pd->hw_context != -1) {
358 			psb_mmu_clflush(pd->driver, (void *)&v[index]);
359 			atomic_set(&pd->driver->needs_tlbflush, 1);
360 		}
361 	}
362 	pt->v = kmap_atomic(pt->p);
363 	return pt;
364 }
365 
366 static struct psb_mmu_pt *psb_mmu_pt_map_lock(struct psb_mmu_pd *pd,
367 					      unsigned long addr)
368 {
369 	uint32_t index = psb_mmu_pd_index(addr);
370 	struct psb_mmu_pt *pt;
371 	spinlock_t *lock = &pd->driver->lock;
372 
373 	spin_lock(lock);
374 	pt = pd->tables[index];
375 	if (!pt) {
376 		spin_unlock(lock);
377 		return NULL;
378 	}
379 	pt->v = kmap_atomic(pt->p);
380 	return pt;
381 }
382 
383 static void psb_mmu_pt_unmap_unlock(struct psb_mmu_pt *pt)
384 {
385 	struct psb_mmu_pd *pd = pt->pd;
386 	uint32_t *v;
387 
388 	kunmap_atomic(pt->v);
389 	if (pt->count == 0) {
390 		v = kmap_atomic(pd->p);
391 		v[pt->index] = pd->invalid_pde;
392 		pd->tables[pt->index] = NULL;
393 
394 		if (pd->hw_context != -1) {
395 			psb_mmu_clflush(pd->driver, (void *)&v[pt->index]);
396 			atomic_set(&pd->driver->needs_tlbflush, 1);
397 		}
398 		kunmap_atomic(v);
399 		spin_unlock(&pd->driver->lock);
400 		psb_mmu_free_pt(pt);
401 		return;
402 	}
403 	spin_unlock(&pd->driver->lock);
404 }
405 
406 static inline void psb_mmu_set_pte(struct psb_mmu_pt *pt, unsigned long addr,
407 				   uint32_t pte)
408 {
409 	pt->v[psb_mmu_pt_index(addr)] = pte;
410 }
411 
412 static inline void psb_mmu_invalidate_pte(struct psb_mmu_pt *pt,
413 					  unsigned long addr)
414 {
415 	pt->v[psb_mmu_pt_index(addr)] = pt->pd->invalid_pte;
416 }
417 
418 struct psb_mmu_pd *psb_mmu_get_default_pd(struct psb_mmu_driver *driver)
419 {
420 	struct psb_mmu_pd *pd;
421 
422 	down_read(&driver->sem);
423 	pd = driver->default_pd;
424 	up_read(&driver->sem);
425 
426 	return pd;
427 }
428 
429 /* Returns the physical address of the PD shared by sgx/msvdx */
430 uint32_t psb_get_default_pd_addr(struct psb_mmu_driver *driver)
431 {
432 	struct psb_mmu_pd *pd;
433 
434 	pd = psb_mmu_get_default_pd(driver);
435 	return page_to_pfn(pd->p) << PAGE_SHIFT;
436 }
437 
438 void psb_mmu_driver_takedown(struct psb_mmu_driver *driver)
439 {
440 	struct drm_device *dev = driver->dev;
441 	struct drm_psb_private *dev_priv = dev->dev_private;
442 
443 	PSB_WSGX32(driver->bif_ctrl, PSB_CR_BIF_CTRL);
444 	psb_mmu_free_pagedir(driver->default_pd);
445 	kfree(driver);
446 }
447 
448 struct psb_mmu_driver *psb_mmu_driver_init(struct drm_device *dev,
449 					   int trap_pagefaults,
450 					   int invalid_type,
451 					   atomic_t *msvdx_mmu_invaldc)
452 {
453 	struct psb_mmu_driver *driver;
454 	struct drm_psb_private *dev_priv = dev->dev_private;
455 
456 	driver = kmalloc(sizeof(*driver), GFP_KERNEL);
457 
458 	if (!driver)
459 		return NULL;
460 
461 	driver->dev = dev;
462 	driver->default_pd = psb_mmu_alloc_pd(driver, trap_pagefaults,
463 					      invalid_type);
464 	if (!driver->default_pd)
465 		goto out_err1;
466 
467 	spin_lock_init(&driver->lock);
468 	init_rwsem(&driver->sem);
469 	down_write(&driver->sem);
470 	atomic_set(&driver->needs_tlbflush, 1);
471 	driver->msvdx_mmu_invaldc = msvdx_mmu_invaldc;
472 
473 	driver->bif_ctrl = PSB_RSGX32(PSB_CR_BIF_CTRL);
474 	PSB_WSGX32(driver->bif_ctrl | _PSB_CB_CTRL_CLEAR_FAULT,
475 		   PSB_CR_BIF_CTRL);
476 	PSB_WSGX32(driver->bif_ctrl & ~_PSB_CB_CTRL_CLEAR_FAULT,
477 		   PSB_CR_BIF_CTRL);
478 
479 	driver->has_clflush = 0;
480 
481 #if defined(CONFIG_X86)
482 	if (boot_cpu_has(X86_FEATURE_CLFLUSH)) {
483 		uint32_t tfms, misc, cap0, cap4, clflush_size;
484 
485 		/*
486 		 * clflush size is determined at kernel setup for x86_64 but not
487 		 * for i386. We have to do it here.
488 		 */
489 
490 		cpuid(0x00000001, &tfms, &misc, &cap0, &cap4);
491 		clflush_size = ((misc >> 8) & 0xff) * 8;
492 		driver->has_clflush = 1;
493 		driver->clflush_add =
494 		    PAGE_SIZE * clflush_size / sizeof(uint32_t);
495 		driver->clflush_mask = driver->clflush_add - 1;
496 		driver->clflush_mask = ~driver->clflush_mask;
497 	}
498 #endif
499 
500 	up_write(&driver->sem);
501 	return driver;
502 
503 out_err1:
504 	kfree(driver);
505 	return NULL;
506 }
507 
508 #if defined(CONFIG_X86)
509 static void psb_mmu_flush_ptes(struct psb_mmu_pd *pd, unsigned long address,
510 			       uint32_t num_pages, uint32_t desired_tile_stride,
511 			       uint32_t hw_tile_stride)
512 {
513 	struct psb_mmu_pt *pt;
514 	uint32_t rows = 1;
515 	uint32_t i;
516 	unsigned long addr;
517 	unsigned long end;
518 	unsigned long next;
519 	unsigned long add;
520 	unsigned long row_add;
521 	unsigned long clflush_add = pd->driver->clflush_add;
522 	unsigned long clflush_mask = pd->driver->clflush_mask;
523 
524 	if (!pd->driver->has_clflush)
525 		return;
526 
527 	if (hw_tile_stride)
528 		rows = num_pages / desired_tile_stride;
529 	else
530 		desired_tile_stride = num_pages;
531 
532 	add = desired_tile_stride << PAGE_SHIFT;
533 	row_add = hw_tile_stride << PAGE_SHIFT;
534 	mb();
535 	for (i = 0; i < rows; ++i) {
536 
537 		addr = address;
538 		end = addr + add;
539 
540 		do {
541 			next = psb_pd_addr_end(addr, end);
542 			pt = psb_mmu_pt_map_lock(pd, addr);
543 			if (!pt)
544 				continue;
545 			do {
546 				psb_clflush(&pt->v[psb_mmu_pt_index(addr)]);
547 			} while (addr += clflush_add,
548 				 (addr & clflush_mask) < next);
549 
550 			psb_mmu_pt_unmap_unlock(pt);
551 		} while (addr = next, next != end);
552 		address += row_add;
553 	}
554 	mb();
555 }
556 #else
557 static void psb_mmu_flush_ptes(struct psb_mmu_pd *pd, unsigned long address,
558 			       uint32_t num_pages, uint32_t desired_tile_stride,
559 			       uint32_t hw_tile_stride)
560 {
561 	drm_ttm_cache_flush();
562 }
563 #endif
564 
565 void psb_mmu_remove_pfn_sequence(struct psb_mmu_pd *pd,
566 				 unsigned long address, uint32_t num_pages)
567 {
568 	struct psb_mmu_pt *pt;
569 	unsigned long addr;
570 	unsigned long end;
571 	unsigned long next;
572 	unsigned long f_address = address;
573 
574 	down_read(&pd->driver->sem);
575 
576 	addr = address;
577 	end = addr + (num_pages << PAGE_SHIFT);
578 
579 	do {
580 		next = psb_pd_addr_end(addr, end);
581 		pt = psb_mmu_pt_alloc_map_lock(pd, addr);
582 		if (!pt)
583 			goto out;
584 		do {
585 			psb_mmu_invalidate_pte(pt, addr);
586 			--pt->count;
587 		} while (addr += PAGE_SIZE, addr < next);
588 		psb_mmu_pt_unmap_unlock(pt);
589 
590 	} while (addr = next, next != end);
591 
592 out:
593 	if (pd->hw_context != -1)
594 		psb_mmu_flush_ptes(pd, f_address, num_pages, 1, 1);
595 
596 	up_read(&pd->driver->sem);
597 
598 	if (pd->hw_context != -1)
599 		psb_mmu_flush(pd->driver);
600 
601 	return;
602 }
603 
604 void psb_mmu_remove_pages(struct psb_mmu_pd *pd, unsigned long address,
605 			  uint32_t num_pages, uint32_t desired_tile_stride,
606 			  uint32_t hw_tile_stride)
607 {
608 	struct psb_mmu_pt *pt;
609 	uint32_t rows = 1;
610 	uint32_t i;
611 	unsigned long addr;
612 	unsigned long end;
613 	unsigned long next;
614 	unsigned long add;
615 	unsigned long row_add;
616 	unsigned long f_address = address;
617 
618 	if (hw_tile_stride)
619 		rows = num_pages / desired_tile_stride;
620 	else
621 		desired_tile_stride = num_pages;
622 
623 	add = desired_tile_stride << PAGE_SHIFT;
624 	row_add = hw_tile_stride << PAGE_SHIFT;
625 
626 	down_read(&pd->driver->sem);
627 
628 	/* Make sure we only need to flush this processor's cache */
629 
630 	for (i = 0; i < rows; ++i) {
631 
632 		addr = address;
633 		end = addr + add;
634 
635 		do {
636 			next = psb_pd_addr_end(addr, end);
637 			pt = psb_mmu_pt_map_lock(pd, addr);
638 			if (!pt)
639 				continue;
640 			do {
641 				psb_mmu_invalidate_pte(pt, addr);
642 				--pt->count;
643 
644 			} while (addr += PAGE_SIZE, addr < next);
645 			psb_mmu_pt_unmap_unlock(pt);
646 
647 		} while (addr = next, next != end);
648 		address += row_add;
649 	}
650 	if (pd->hw_context != -1)
651 		psb_mmu_flush_ptes(pd, f_address, num_pages,
652 				   desired_tile_stride, hw_tile_stride);
653 
654 	up_read(&pd->driver->sem);
655 
656 	if (pd->hw_context != -1)
657 		psb_mmu_flush(pd->driver);
658 }
659 
660 int psb_mmu_insert_pfn_sequence(struct psb_mmu_pd *pd, uint32_t start_pfn,
661 				unsigned long address, uint32_t num_pages,
662 				int type)
663 {
664 	struct psb_mmu_pt *pt;
665 	uint32_t pte;
666 	unsigned long addr;
667 	unsigned long end;
668 	unsigned long next;
669 	unsigned long f_address = address;
670 	int ret = -ENOMEM;
671 
672 	down_read(&pd->driver->sem);
673 
674 	addr = address;
675 	end = addr + (num_pages << PAGE_SHIFT);
676 
677 	do {
678 		next = psb_pd_addr_end(addr, end);
679 		pt = psb_mmu_pt_alloc_map_lock(pd, addr);
680 		if (!pt) {
681 			ret = -ENOMEM;
682 			goto out;
683 		}
684 		do {
685 			pte = psb_mmu_mask_pte(start_pfn++, type);
686 			psb_mmu_set_pte(pt, addr, pte);
687 			pt->count++;
688 		} while (addr += PAGE_SIZE, addr < next);
689 		psb_mmu_pt_unmap_unlock(pt);
690 
691 	} while (addr = next, next != end);
692 	ret = 0;
693 
694 out:
695 	if (pd->hw_context != -1)
696 		psb_mmu_flush_ptes(pd, f_address, num_pages, 1, 1);
697 
698 	up_read(&pd->driver->sem);
699 
700 	if (pd->hw_context != -1)
701 		psb_mmu_flush(pd->driver);
702 
703 	return 0;
704 }
705 
706 int psb_mmu_insert_pages(struct psb_mmu_pd *pd, struct page **pages,
707 			 unsigned long address, uint32_t num_pages,
708 			 uint32_t desired_tile_stride, uint32_t hw_tile_stride,
709 			 int type)
710 {
711 	struct psb_mmu_pt *pt;
712 	uint32_t rows = 1;
713 	uint32_t i;
714 	uint32_t pte;
715 	unsigned long addr;
716 	unsigned long end;
717 	unsigned long next;
718 	unsigned long add;
719 	unsigned long row_add;
720 	unsigned long f_address = address;
721 	int ret = -ENOMEM;
722 
723 	if (hw_tile_stride) {
724 		if (num_pages % desired_tile_stride != 0)
725 			return -EINVAL;
726 		rows = num_pages / desired_tile_stride;
727 	} else {
728 		desired_tile_stride = num_pages;
729 	}
730 
731 	add = desired_tile_stride << PAGE_SHIFT;
732 	row_add = hw_tile_stride << PAGE_SHIFT;
733 
734 	down_read(&pd->driver->sem);
735 
736 	for (i = 0; i < rows; ++i) {
737 
738 		addr = address;
739 		end = addr + add;
740 
741 		do {
742 			next = psb_pd_addr_end(addr, end);
743 			pt = psb_mmu_pt_alloc_map_lock(pd, addr);
744 			if (!pt)
745 				goto out;
746 			do {
747 				pte = psb_mmu_mask_pte(page_to_pfn(*pages++),
748 						       type);
749 				psb_mmu_set_pte(pt, addr, pte);
750 				pt->count++;
751 			} while (addr += PAGE_SIZE, addr < next);
752 			psb_mmu_pt_unmap_unlock(pt);
753 
754 		} while (addr = next, next != end);
755 
756 		address += row_add;
757 	}
758 
759 	ret = 0;
760 out:
761 	if (pd->hw_context != -1)
762 		psb_mmu_flush_ptes(pd, f_address, num_pages,
763 				   desired_tile_stride, hw_tile_stride);
764 
765 	up_read(&pd->driver->sem);
766 
767 	if (pd->hw_context != -1)
768 		psb_mmu_flush(pd->driver);
769 
770 	return ret;
771 }
772 
773 int psb_mmu_virtual_to_pfn(struct psb_mmu_pd *pd, uint32_t virtual,
774 			   unsigned long *pfn)
775 {
776 	int ret;
777 	struct psb_mmu_pt *pt;
778 	uint32_t tmp;
779 	spinlock_t *lock = &pd->driver->lock;
780 
781 	down_read(&pd->driver->sem);
782 	pt = psb_mmu_pt_map_lock(pd, virtual);
783 	if (!pt) {
784 		uint32_t *v;
785 
786 		spin_lock(lock);
787 		v = kmap_atomic(pd->p);
788 		tmp = v[psb_mmu_pd_index(virtual)];
789 		kunmap_atomic(v);
790 		spin_unlock(lock);
791 
792 		if (tmp != pd->invalid_pde || !(tmp & PSB_PTE_VALID) ||
793 		    !(pd->invalid_pte & PSB_PTE_VALID)) {
794 			ret = -EINVAL;
795 			goto out;
796 		}
797 		ret = 0;
798 		*pfn = pd->invalid_pte >> PAGE_SHIFT;
799 		goto out;
800 	}
801 	tmp = pt->v[psb_mmu_pt_index(virtual)];
802 	if (!(tmp & PSB_PTE_VALID)) {
803 		ret = -EINVAL;
804 	} else {
805 		ret = 0;
806 		*pfn = tmp >> PAGE_SHIFT;
807 	}
808 	psb_mmu_pt_unmap_unlock(pt);
809 out:
810 	up_read(&pd->driver->sem);
811 	return ret;
812 }
813