xref: /linux/drivers/gpu/drm/gma500/mmu.c (revision b43ab901d671e3e3cad425ea5e9a3c74e266dcdd)
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 
22 /*
23  * Code for the SGX MMU:
24  */
25 
26 /*
27  * clflush on one processor only:
28  * clflush should apparently flush the cache line on all processors in an
29  * SMP system.
30  */
31 
32 /*
33  * kmap atomic:
34  * The usage of the slots must be completely encapsulated within a spinlock, and
35  * no other functions that may be using the locks for other purposed may be
36  * called from within the locked region.
37  * Since the slots are per processor, this will guarantee that we are the only
38  * user.
39  */
40 
41 /*
42  * TODO: Inserting ptes from an interrupt handler:
43  * This may be desirable for some SGX functionality where the GPU can fault in
44  * needed pages. For that, we need to make an atomic insert_pages function, that
45  * may fail.
46  * If it fails, the caller need to insert the page using a workqueue function,
47  * but on average it should be fast.
48  */
49 
50 struct psb_mmu_driver {
51 	/* protects driver- and pd structures. Always take in read mode
52 	 * before taking the page table spinlock.
53 	 */
54 	struct rw_semaphore sem;
55 
56 	/* protects page tables, directory tables and pt tables.
57 	 * and pt structures.
58 	 */
59 	spinlock_t lock;
60 
61 	atomic_t needs_tlbflush;
62 
63 	uint8_t __iomem *register_map;
64 	struct psb_mmu_pd *default_pd;
65 	/*uint32_t bif_ctrl;*/
66 	int has_clflush;
67 	int clflush_add;
68 	unsigned long clflush_mask;
69 
70 	struct drm_psb_private *dev_priv;
71 };
72 
73 struct psb_mmu_pd;
74 
75 struct psb_mmu_pt {
76 	struct psb_mmu_pd *pd;
77 	uint32_t index;
78 	uint32_t count;
79 	struct page *p;
80 	uint32_t *v;
81 };
82 
83 struct psb_mmu_pd {
84 	struct psb_mmu_driver *driver;
85 	int hw_context;
86 	struct psb_mmu_pt **tables;
87 	struct page *p;
88 	struct page *dummy_pt;
89 	struct page *dummy_page;
90 	uint32_t pd_mask;
91 	uint32_t invalid_pde;
92 	uint32_t invalid_pte;
93 };
94 
95 static inline uint32_t psb_mmu_pt_index(uint32_t offset)
96 {
97 	return (offset >> PSB_PTE_SHIFT) & 0x3FF;
98 }
99 
100 static inline uint32_t psb_mmu_pd_index(uint32_t offset)
101 {
102 	return offset >> PSB_PDE_SHIFT;
103 }
104 
105 static inline void psb_clflush(void *addr)
106 {
107 	__asm__ __volatile__("clflush (%0)\n" : : "r"(addr) : "memory");
108 }
109 
110 static inline void psb_mmu_clflush(struct psb_mmu_driver *driver,
111 				   void *addr)
112 {
113 	if (!driver->has_clflush)
114 		return;
115 
116 	mb();
117 	psb_clflush(addr);
118 	mb();
119 }
120 
121 static void psb_page_clflush(struct psb_mmu_driver *driver, struct page* page)
122 {
123 	uint32_t clflush_add = driver->clflush_add >> PAGE_SHIFT;
124 	uint32_t clflush_count = PAGE_SIZE / clflush_add;
125 	int i;
126 	uint8_t *clf;
127 
128 	clf = kmap_atomic(page, KM_USER0);
129 	mb();
130 	for (i = 0; i < clflush_count; ++i) {
131 		psb_clflush(clf);
132 		clf += clflush_add;
133 	}
134 	mb();
135 	kunmap_atomic(clf, KM_USER0);
136 }
137 
138 static void psb_pages_clflush(struct psb_mmu_driver *driver,
139 				struct page *page[], unsigned long num_pages)
140 {
141 	int i;
142 
143 	if (!driver->has_clflush)
144 		return ;
145 
146 	for (i = 0; i < num_pages; i++)
147 		psb_page_clflush(driver, *page++);
148 }
149 
150 static void psb_mmu_flush_pd_locked(struct psb_mmu_driver *driver,
151 				    int force)
152 {
153 	atomic_set(&driver->needs_tlbflush, 0);
154 }
155 
156 static void psb_mmu_flush_pd(struct psb_mmu_driver *driver, int force)
157 {
158 	down_write(&driver->sem);
159 	psb_mmu_flush_pd_locked(driver, force);
160 	up_write(&driver->sem);
161 }
162 
163 void psb_mmu_flush(struct psb_mmu_driver *driver, int rc_prot)
164 {
165 	if (rc_prot)
166 		down_write(&driver->sem);
167 	if (rc_prot)
168 		up_write(&driver->sem);
169 }
170 
171 void psb_mmu_set_pd_context(struct psb_mmu_pd *pd, int hw_context)
172 {
173 	/*ttm_tt_cache_flush(&pd->p, 1);*/
174 	psb_pages_clflush(pd->driver, &pd->p, 1);
175 	down_write(&pd->driver->sem);
176 	wmb();
177 	psb_mmu_flush_pd_locked(pd->driver, 1);
178 	pd->hw_context = hw_context;
179 	up_write(&pd->driver->sem);
180 
181 }
182 
183 static inline unsigned long psb_pd_addr_end(unsigned long addr,
184 					    unsigned long end)
185 {
186 
187 	addr = (addr + PSB_PDE_MASK + 1) & ~PSB_PDE_MASK;
188 	return (addr < end) ? addr : end;
189 }
190 
191 static inline uint32_t psb_mmu_mask_pte(uint32_t pfn, int type)
192 {
193 	uint32_t mask = PSB_PTE_VALID;
194 
195 	if (type & PSB_MMU_CACHED_MEMORY)
196 		mask |= PSB_PTE_CACHED;
197 	if (type & PSB_MMU_RO_MEMORY)
198 		mask |= PSB_PTE_RO;
199 	if (type & PSB_MMU_WO_MEMORY)
200 		mask |= PSB_PTE_WO;
201 
202 	return (pfn << PAGE_SHIFT) | mask;
203 }
204 
205 struct psb_mmu_pd *psb_mmu_alloc_pd(struct psb_mmu_driver *driver,
206 				    int trap_pagefaults, int invalid_type)
207 {
208 	struct psb_mmu_pd *pd = kmalloc(sizeof(*pd), GFP_KERNEL);
209 	uint32_t *v;
210 	int i;
211 
212 	if (!pd)
213 		return NULL;
214 
215 	pd->p = alloc_page(GFP_DMA32);
216 	if (!pd->p)
217 		goto out_err1;
218 	pd->dummy_pt = alloc_page(GFP_DMA32);
219 	if (!pd->dummy_pt)
220 		goto out_err2;
221 	pd->dummy_page = alloc_page(GFP_DMA32);
222 	if (!pd->dummy_page)
223 		goto out_err3;
224 
225 	if (!trap_pagefaults) {
226 		pd->invalid_pde =
227 		    psb_mmu_mask_pte(page_to_pfn(pd->dummy_pt),
228 				     invalid_type);
229 		pd->invalid_pte =
230 		    psb_mmu_mask_pte(page_to_pfn(pd->dummy_page),
231 				     invalid_type);
232 	} else {
233 		pd->invalid_pde = 0;
234 		pd->invalid_pte = 0;
235 	}
236 
237 	v = kmap(pd->dummy_pt);
238 	for (i = 0; i < (PAGE_SIZE / sizeof(uint32_t)); ++i)
239 		v[i] = pd->invalid_pte;
240 
241 	kunmap(pd->dummy_pt);
242 
243 	v = kmap(pd->p);
244 	for (i = 0; i < (PAGE_SIZE / sizeof(uint32_t)); ++i)
245 		v[i] = pd->invalid_pde;
246 
247 	kunmap(pd->p);
248 
249 	clear_page(kmap(pd->dummy_page));
250 	kunmap(pd->dummy_page);
251 
252 	pd->tables = vmalloc_user(sizeof(struct psb_mmu_pt *) * 1024);
253 	if (!pd->tables)
254 		goto out_err4;
255 
256 	pd->hw_context = -1;
257 	pd->pd_mask = PSB_PTE_VALID;
258 	pd->driver = driver;
259 
260 	return pd;
261 
262 out_err4:
263 	__free_page(pd->dummy_page);
264 out_err3:
265 	__free_page(pd->dummy_pt);
266 out_err2:
267 	__free_page(pd->p);
268 out_err1:
269 	kfree(pd);
270 	return NULL;
271 }
272 
273 void psb_mmu_free_pt(struct psb_mmu_pt *pt)
274 {
275 	__free_page(pt->p);
276 	kfree(pt);
277 }
278 
279 void psb_mmu_free_pagedir(struct psb_mmu_pd *pd)
280 {
281 	struct psb_mmu_driver *driver = pd->driver;
282 	struct psb_mmu_pt *pt;
283 	int i;
284 
285 	down_write(&driver->sem);
286 	if (pd->hw_context != -1)
287 		psb_mmu_flush_pd_locked(driver, 1);
288 
289 	/* Should take the spinlock here, but we don't need to do that
290 	   since we have the semaphore in write mode. */
291 
292 	for (i = 0; i < 1024; ++i) {
293 		pt = pd->tables[i];
294 		if (pt)
295 			psb_mmu_free_pt(pt);
296 	}
297 
298 	vfree(pd->tables);
299 	__free_page(pd->dummy_page);
300 	__free_page(pd->dummy_pt);
301 	__free_page(pd->p);
302 	kfree(pd);
303 	up_write(&driver->sem);
304 }
305 
306 static struct psb_mmu_pt *psb_mmu_alloc_pt(struct psb_mmu_pd *pd)
307 {
308 	struct psb_mmu_pt *pt = kmalloc(sizeof(*pt), GFP_KERNEL);
309 	void *v;
310 	uint32_t clflush_add = pd->driver->clflush_add >> PAGE_SHIFT;
311 	uint32_t clflush_count = PAGE_SIZE / clflush_add;
312 	spinlock_t *lock = &pd->driver->lock;
313 	uint8_t *clf;
314 	uint32_t *ptes;
315 	int i;
316 
317 	if (!pt)
318 		return NULL;
319 
320 	pt->p = alloc_page(GFP_DMA32);
321 	if (!pt->p) {
322 		kfree(pt);
323 		return NULL;
324 	}
325 
326 	spin_lock(lock);
327 
328 	v = kmap_atomic(pt->p, KM_USER0);
329 	clf = (uint8_t *) v;
330 	ptes = (uint32_t *) v;
331 	for (i = 0; i < (PAGE_SIZE / sizeof(uint32_t)); ++i)
332 		*ptes++ = pd->invalid_pte;
333 
334 
335 	if (pd->driver->has_clflush && pd->hw_context != -1) {
336 		mb();
337 		for (i = 0; i < clflush_count; ++i) {
338 			psb_clflush(clf);
339 			clf += clflush_add;
340 		}
341 		mb();
342 	}
343 
344 	kunmap_atomic(v, KM_USER0);
345 	spin_unlock(lock);
346 
347 	pt->count = 0;
348 	pt->pd = pd;
349 	pt->index = 0;
350 
351 	return pt;
352 }
353 
354 struct psb_mmu_pt *psb_mmu_pt_alloc_map_lock(struct psb_mmu_pd *pd,
355 					     unsigned long addr)
356 {
357 	uint32_t index = psb_mmu_pd_index(addr);
358 	struct psb_mmu_pt *pt;
359 	uint32_t *v;
360 	spinlock_t *lock = &pd->driver->lock;
361 
362 	spin_lock(lock);
363 	pt = pd->tables[index];
364 	while (!pt) {
365 		spin_unlock(lock);
366 		pt = psb_mmu_alloc_pt(pd);
367 		if (!pt)
368 			return NULL;
369 		spin_lock(lock);
370 
371 		if (pd->tables[index]) {
372 			spin_unlock(lock);
373 			psb_mmu_free_pt(pt);
374 			spin_lock(lock);
375 			pt = pd->tables[index];
376 			continue;
377 		}
378 
379 		v = kmap_atomic(pd->p, KM_USER0);
380 		pd->tables[index] = pt;
381 		v[index] = (page_to_pfn(pt->p) << 12) | pd->pd_mask;
382 		pt->index = index;
383 		kunmap_atomic((void *) v, KM_USER0);
384 
385 		if (pd->hw_context != -1) {
386 			psb_mmu_clflush(pd->driver, (void *) &v[index]);
387 			atomic_set(&pd->driver->needs_tlbflush, 1);
388 		}
389 	}
390 	pt->v = kmap_atomic(pt->p, KM_USER0);
391 	return pt;
392 }
393 
394 static struct psb_mmu_pt *psb_mmu_pt_map_lock(struct psb_mmu_pd *pd,
395 					      unsigned long addr)
396 {
397 	uint32_t index = psb_mmu_pd_index(addr);
398 	struct psb_mmu_pt *pt;
399 	spinlock_t *lock = &pd->driver->lock;
400 
401 	spin_lock(lock);
402 	pt = pd->tables[index];
403 	if (!pt) {
404 		spin_unlock(lock);
405 		return NULL;
406 	}
407 	pt->v = kmap_atomic(pt->p, KM_USER0);
408 	return pt;
409 }
410 
411 static void psb_mmu_pt_unmap_unlock(struct psb_mmu_pt *pt)
412 {
413 	struct psb_mmu_pd *pd = pt->pd;
414 	uint32_t *v;
415 
416 	kunmap_atomic(pt->v, KM_USER0);
417 	if (pt->count == 0) {
418 		v = kmap_atomic(pd->p, KM_USER0);
419 		v[pt->index] = pd->invalid_pde;
420 		pd->tables[pt->index] = NULL;
421 
422 		if (pd->hw_context != -1) {
423 			psb_mmu_clflush(pd->driver,
424 					(void *) &v[pt->index]);
425 			atomic_set(&pd->driver->needs_tlbflush, 1);
426 		}
427 		kunmap_atomic(pt->v, KM_USER0);
428 		spin_unlock(&pd->driver->lock);
429 		psb_mmu_free_pt(pt);
430 		return;
431 	}
432 	spin_unlock(&pd->driver->lock);
433 }
434 
435 static inline void psb_mmu_set_pte(struct psb_mmu_pt *pt,
436 				   unsigned long addr, uint32_t pte)
437 {
438 	pt->v[psb_mmu_pt_index(addr)] = pte;
439 }
440 
441 static inline void psb_mmu_invalidate_pte(struct psb_mmu_pt *pt,
442 					  unsigned long addr)
443 {
444 	pt->v[psb_mmu_pt_index(addr)] = pt->pd->invalid_pte;
445 }
446 
447 
448 void psb_mmu_mirror_gtt(struct psb_mmu_pd *pd,
449 			uint32_t mmu_offset, uint32_t gtt_start,
450 			uint32_t gtt_pages)
451 {
452 	uint32_t *v;
453 	uint32_t start = psb_mmu_pd_index(mmu_offset);
454 	struct psb_mmu_driver *driver = pd->driver;
455 	int num_pages = gtt_pages;
456 
457 	down_read(&driver->sem);
458 	spin_lock(&driver->lock);
459 
460 	v = kmap_atomic(pd->p, KM_USER0);
461 	v += start;
462 
463 	while (gtt_pages--) {
464 		*v++ = gtt_start | pd->pd_mask;
465 		gtt_start += PAGE_SIZE;
466 	}
467 
468 	/*ttm_tt_cache_flush(&pd->p, num_pages);*/
469 	psb_pages_clflush(pd->driver, &pd->p, num_pages);
470 	kunmap_atomic(v, KM_USER0);
471 	spin_unlock(&driver->lock);
472 
473 	if (pd->hw_context != -1)
474 		atomic_set(&pd->driver->needs_tlbflush, 1);
475 
476 	up_read(&pd->driver->sem);
477 	psb_mmu_flush_pd(pd->driver, 0);
478 }
479 
480 struct psb_mmu_pd *psb_mmu_get_default_pd(struct psb_mmu_driver *driver)
481 {
482 	struct psb_mmu_pd *pd;
483 
484 	/* down_read(&driver->sem); */
485 	pd = driver->default_pd;
486 	/* up_read(&driver->sem); */
487 
488 	return pd;
489 }
490 
491 /* Returns the physical address of the PD shared by sgx/msvdx */
492 uint32_t psb_get_default_pd_addr(struct psb_mmu_driver *driver)
493 {
494 	struct psb_mmu_pd *pd;
495 
496 	pd = psb_mmu_get_default_pd(driver);
497 	return page_to_pfn(pd->p) << PAGE_SHIFT;
498 }
499 
500 void psb_mmu_driver_takedown(struct psb_mmu_driver *driver)
501 {
502 	psb_mmu_free_pagedir(driver->default_pd);
503 	kfree(driver);
504 }
505 
506 struct psb_mmu_driver *psb_mmu_driver_init(uint8_t __iomem * registers,
507 					int trap_pagefaults,
508 					int invalid_type,
509 					struct drm_psb_private *dev_priv)
510 {
511 	struct psb_mmu_driver *driver;
512 
513 	driver = kmalloc(sizeof(*driver), GFP_KERNEL);
514 
515 	if (!driver)
516 		return NULL;
517 	driver->dev_priv = dev_priv;
518 
519 	driver->default_pd = psb_mmu_alloc_pd(driver, trap_pagefaults,
520 					      invalid_type);
521 	if (!driver->default_pd)
522 		goto out_err1;
523 
524 	spin_lock_init(&driver->lock);
525 	init_rwsem(&driver->sem);
526 	down_write(&driver->sem);
527 	driver->register_map = registers;
528 	atomic_set(&driver->needs_tlbflush, 1);
529 
530 	driver->has_clflush = 0;
531 
532 	if (boot_cpu_has(X86_FEATURE_CLFLSH)) {
533 		uint32_t tfms, misc, cap0, cap4, clflush_size;
534 
535 		/*
536 		 * clflush size is determined at kernel setup for x86_64
537 		 *  but not for i386. We have to do it here.
538 		 */
539 
540 		cpuid(0x00000001, &tfms, &misc, &cap0, &cap4);
541 		clflush_size = ((misc >> 8) & 0xff) * 8;
542 		driver->has_clflush = 1;
543 		driver->clflush_add =
544 		    PAGE_SIZE * clflush_size / sizeof(uint32_t);
545 		driver->clflush_mask = driver->clflush_add - 1;
546 		driver->clflush_mask = ~driver->clflush_mask;
547 	}
548 
549 	up_write(&driver->sem);
550 	return driver;
551 
552 out_err1:
553 	kfree(driver);
554 	return NULL;
555 }
556 
557 static void psb_mmu_flush_ptes(struct psb_mmu_pd *pd,
558 			       unsigned long address, uint32_t num_pages,
559 			       uint32_t desired_tile_stride,
560 			       uint32_t hw_tile_stride)
561 {
562 	struct psb_mmu_pt *pt;
563 	uint32_t rows = 1;
564 	uint32_t i;
565 	unsigned long addr;
566 	unsigned long end;
567 	unsigned long next;
568 	unsigned long add;
569 	unsigned long row_add;
570 	unsigned long clflush_add = pd->driver->clflush_add;
571 	unsigned long clflush_mask = pd->driver->clflush_mask;
572 
573 	if (!pd->driver->has_clflush) {
574 		/*ttm_tt_cache_flush(&pd->p, num_pages);*/
575 		psb_pages_clflush(pd->driver, &pd->p, num_pages);
576 		return;
577 	}
578 
579 	if (hw_tile_stride)
580 		rows = num_pages / desired_tile_stride;
581 	else
582 		desired_tile_stride = num_pages;
583 
584 	add = desired_tile_stride << PAGE_SHIFT;
585 	row_add = hw_tile_stride << PAGE_SHIFT;
586 	mb();
587 	for (i = 0; i < rows; ++i) {
588 
589 		addr = address;
590 		end = addr + add;
591 
592 		do {
593 			next = psb_pd_addr_end(addr, end);
594 			pt = psb_mmu_pt_map_lock(pd, addr);
595 			if (!pt)
596 				continue;
597 			do {
598 				psb_clflush(&pt->v
599 					    [psb_mmu_pt_index(addr)]);
600 			} while (addr +=
601 				 clflush_add,
602 				 (addr & clflush_mask) < next);
603 
604 			psb_mmu_pt_unmap_unlock(pt);
605 		} while (addr = next, next != end);
606 		address += row_add;
607 	}
608 	mb();
609 }
610 
611 void psb_mmu_remove_pfn_sequence(struct psb_mmu_pd *pd,
612 				 unsigned long address, uint32_t num_pages)
613 {
614 	struct psb_mmu_pt *pt;
615 	unsigned long addr;
616 	unsigned long end;
617 	unsigned long next;
618 	unsigned long f_address = address;
619 
620 	down_read(&pd->driver->sem);
621 
622 	addr = address;
623 	end = addr + (num_pages << PAGE_SHIFT);
624 
625 	do {
626 		next = psb_pd_addr_end(addr, end);
627 		pt = psb_mmu_pt_alloc_map_lock(pd, addr);
628 		if (!pt)
629 			goto out;
630 		do {
631 			psb_mmu_invalidate_pte(pt, addr);
632 			--pt->count;
633 		} while (addr += PAGE_SIZE, addr < next);
634 		psb_mmu_pt_unmap_unlock(pt);
635 
636 	} while (addr = next, next != end);
637 
638 out:
639 	if (pd->hw_context != -1)
640 		psb_mmu_flush_ptes(pd, f_address, num_pages, 1, 1);
641 
642 	up_read(&pd->driver->sem);
643 
644 	if (pd->hw_context != -1)
645 		psb_mmu_flush(pd->driver, 0);
646 
647 	return;
648 }
649 
650 void psb_mmu_remove_pages(struct psb_mmu_pd *pd, unsigned long address,
651 			  uint32_t num_pages, uint32_t desired_tile_stride,
652 			  uint32_t hw_tile_stride)
653 {
654 	struct psb_mmu_pt *pt;
655 	uint32_t rows = 1;
656 	uint32_t i;
657 	unsigned long addr;
658 	unsigned long end;
659 	unsigned long next;
660 	unsigned long add;
661 	unsigned long row_add;
662 	unsigned long f_address = address;
663 
664 	if (hw_tile_stride)
665 		rows = num_pages / desired_tile_stride;
666 	else
667 		desired_tile_stride = num_pages;
668 
669 	add = desired_tile_stride << PAGE_SHIFT;
670 	row_add = hw_tile_stride << PAGE_SHIFT;
671 
672 	/* down_read(&pd->driver->sem); */
673 
674 	/* Make sure we only need to flush this processor's cache */
675 
676 	for (i = 0; i < rows; ++i) {
677 
678 		addr = address;
679 		end = addr + add;
680 
681 		do {
682 			next = psb_pd_addr_end(addr, end);
683 			pt = psb_mmu_pt_map_lock(pd, addr);
684 			if (!pt)
685 				continue;
686 			do {
687 				psb_mmu_invalidate_pte(pt, addr);
688 				--pt->count;
689 
690 			} while (addr += PAGE_SIZE, addr < next);
691 			psb_mmu_pt_unmap_unlock(pt);
692 
693 		} while (addr = next, next != end);
694 		address += row_add;
695 	}
696 	if (pd->hw_context != -1)
697 		psb_mmu_flush_ptes(pd, f_address, num_pages,
698 				   desired_tile_stride, hw_tile_stride);
699 
700 	/* up_read(&pd->driver->sem); */
701 
702 	if (pd->hw_context != -1)
703 		psb_mmu_flush(pd->driver, 0);
704 }
705 
706 int psb_mmu_insert_pfn_sequence(struct psb_mmu_pd *pd, uint32_t start_pfn,
707 				unsigned long address, uint32_t num_pages,
708 				int type)
709 {
710 	struct psb_mmu_pt *pt;
711 	uint32_t pte;
712 	unsigned long addr;
713 	unsigned long end;
714 	unsigned long next;
715 	unsigned long f_address = address;
716 	int ret = 0;
717 
718 	down_read(&pd->driver->sem);
719 
720 	addr = address;
721 	end = addr + (num_pages << PAGE_SHIFT);
722 
723 	do {
724 		next = psb_pd_addr_end(addr, end);
725 		pt = psb_mmu_pt_alloc_map_lock(pd, addr);
726 		if (!pt) {
727 			ret = -ENOMEM;
728 			goto out;
729 		}
730 		do {
731 			pte = psb_mmu_mask_pte(start_pfn++, type);
732 			psb_mmu_set_pte(pt, addr, pte);
733 			pt->count++;
734 		} while (addr += PAGE_SIZE, addr < next);
735 		psb_mmu_pt_unmap_unlock(pt);
736 
737 	} while (addr = next, next != end);
738 
739 out:
740 	if (pd->hw_context != -1)
741 		psb_mmu_flush_ptes(pd, f_address, num_pages, 1, 1);
742 
743 	up_read(&pd->driver->sem);
744 
745 	if (pd->hw_context != -1)
746 		psb_mmu_flush(pd->driver, 1);
747 
748 	return ret;
749 }
750 
751 int psb_mmu_insert_pages(struct psb_mmu_pd *pd, struct page **pages,
752 			 unsigned long address, uint32_t num_pages,
753 			 uint32_t desired_tile_stride,
754 			 uint32_t hw_tile_stride, int type)
755 {
756 	struct psb_mmu_pt *pt;
757 	uint32_t rows = 1;
758 	uint32_t i;
759 	uint32_t pte;
760 	unsigned long addr;
761 	unsigned long end;
762 	unsigned long next;
763 	unsigned long add;
764 	unsigned long row_add;
765 	unsigned long f_address = address;
766 	int ret = 0;
767 
768 	if (hw_tile_stride) {
769 		if (num_pages % desired_tile_stride != 0)
770 			return -EINVAL;
771 		rows = num_pages / desired_tile_stride;
772 	} else {
773 		desired_tile_stride = num_pages;
774 	}
775 
776 	add = desired_tile_stride << PAGE_SHIFT;
777 	row_add = hw_tile_stride << PAGE_SHIFT;
778 
779 	down_read(&pd->driver->sem);
780 
781 	for (i = 0; i < rows; ++i) {
782 
783 		addr = address;
784 		end = addr + add;
785 
786 		do {
787 			next = psb_pd_addr_end(addr, end);
788 			pt = psb_mmu_pt_alloc_map_lock(pd, addr);
789 			if (!pt) {
790 				ret = -ENOMEM;
791 				goto out;
792 			}
793 			do {
794 				pte =
795 				    psb_mmu_mask_pte(page_to_pfn(*pages++),
796 						     type);
797 				psb_mmu_set_pte(pt, addr, pte);
798 				pt->count++;
799 			} while (addr += PAGE_SIZE, addr < next);
800 			psb_mmu_pt_unmap_unlock(pt);
801 
802 		} while (addr = next, next != end);
803 
804 		address += row_add;
805 	}
806 out:
807 	if (pd->hw_context != -1)
808 		psb_mmu_flush_ptes(pd, f_address, num_pages,
809 				   desired_tile_stride, hw_tile_stride);
810 
811 	up_read(&pd->driver->sem);
812 
813 	if (pd->hw_context != -1)
814 		psb_mmu_flush(pd->driver, 1);
815 
816 	return ret;
817 }
818 
819 int psb_mmu_virtual_to_pfn(struct psb_mmu_pd *pd, uint32_t virtual,
820 			   unsigned long *pfn)
821 {
822 	int ret;
823 	struct psb_mmu_pt *pt;
824 	uint32_t tmp;
825 	spinlock_t *lock = &pd->driver->lock;
826 
827 	down_read(&pd->driver->sem);
828 	pt = psb_mmu_pt_map_lock(pd, virtual);
829 	if (!pt) {
830 		uint32_t *v;
831 
832 		spin_lock(lock);
833 		v = kmap_atomic(pd->p, KM_USER0);
834 		tmp = v[psb_mmu_pd_index(virtual)];
835 		kunmap_atomic(v, KM_USER0);
836 		spin_unlock(lock);
837 
838 		if (tmp != pd->invalid_pde || !(tmp & PSB_PTE_VALID) ||
839 		    !(pd->invalid_pte & PSB_PTE_VALID)) {
840 			ret = -EINVAL;
841 			goto out;
842 		}
843 		ret = 0;
844 		*pfn = pd->invalid_pte >> PAGE_SHIFT;
845 		goto out;
846 	}
847 	tmp = pt->v[psb_mmu_pt_index(virtual)];
848 	if (!(tmp & PSB_PTE_VALID)) {
849 		ret = -EINVAL;
850 	} else {
851 		ret = 0;
852 		*pfn = tmp >> PAGE_SHIFT;
853 	}
854 	psb_mmu_pt_unmap_unlock(pt);
855 out:
856 	up_read(&pd->driver->sem);
857 	return ret;
858 }
859