xref: /openbmc/linux/drivers/gpu/drm/gma500/mmu.c (revision 9cdb81c7)
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);
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);
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 static 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);
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);
345 	spin_unlock(lock);
346 
347 	pt->count = 0;
348 	pt->pd = pd;
349 	pt->index = 0;
350 
351 	return pt;
352 }
353 
354 static 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);
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);
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);
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);
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);
417 	if (pt->count == 0) {
418 		v = kmap_atomic(pd->p);
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);
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);
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);
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 void psb_mmu_driver_takedown(struct psb_mmu_driver *driver)
492 {
493 	psb_mmu_free_pagedir(driver->default_pd);
494 	kfree(driver);
495 }
496 
497 struct psb_mmu_driver *psb_mmu_driver_init(uint8_t __iomem * registers,
498 					int trap_pagefaults,
499 					int invalid_type,
500 					struct drm_psb_private *dev_priv)
501 {
502 	struct psb_mmu_driver *driver;
503 
504 	driver = kmalloc(sizeof(*driver), GFP_KERNEL);
505 
506 	if (!driver)
507 		return NULL;
508 	driver->dev_priv = dev_priv;
509 
510 	driver->default_pd = psb_mmu_alloc_pd(driver, trap_pagefaults,
511 					      invalid_type);
512 	if (!driver->default_pd)
513 		goto out_err1;
514 
515 	spin_lock_init(&driver->lock);
516 	init_rwsem(&driver->sem);
517 	down_write(&driver->sem);
518 	driver->register_map = registers;
519 	atomic_set(&driver->needs_tlbflush, 1);
520 
521 	driver->has_clflush = 0;
522 
523 	if (boot_cpu_has(X86_FEATURE_CLFLSH)) {
524 		uint32_t tfms, misc, cap0, cap4, clflush_size;
525 
526 		/*
527 		 * clflush size is determined at kernel setup for x86_64
528 		 *  but not for i386. We have to do it here.
529 		 */
530 
531 		cpuid(0x00000001, &tfms, &misc, &cap0, &cap4);
532 		clflush_size = ((misc >> 8) & 0xff) * 8;
533 		driver->has_clflush = 1;
534 		driver->clflush_add =
535 		    PAGE_SIZE * clflush_size / sizeof(uint32_t);
536 		driver->clflush_mask = driver->clflush_add - 1;
537 		driver->clflush_mask = ~driver->clflush_mask;
538 	}
539 
540 	up_write(&driver->sem);
541 	return driver;
542 
543 out_err1:
544 	kfree(driver);
545 	return NULL;
546 }
547 
548 static void psb_mmu_flush_ptes(struct psb_mmu_pd *pd,
549 			       unsigned long address, uint32_t num_pages,
550 			       uint32_t desired_tile_stride,
551 			       uint32_t hw_tile_stride)
552 {
553 	struct psb_mmu_pt *pt;
554 	uint32_t rows = 1;
555 	uint32_t i;
556 	unsigned long addr;
557 	unsigned long end;
558 	unsigned long next;
559 	unsigned long add;
560 	unsigned long row_add;
561 	unsigned long clflush_add = pd->driver->clflush_add;
562 	unsigned long clflush_mask = pd->driver->clflush_mask;
563 
564 	if (!pd->driver->has_clflush) {
565 		/*ttm_tt_cache_flush(&pd->p, num_pages);*/
566 		psb_pages_clflush(pd->driver, &pd->p, num_pages);
567 		return;
568 	}
569 
570 	if (hw_tile_stride)
571 		rows = num_pages / desired_tile_stride;
572 	else
573 		desired_tile_stride = num_pages;
574 
575 	add = desired_tile_stride << PAGE_SHIFT;
576 	row_add = hw_tile_stride << PAGE_SHIFT;
577 	mb();
578 	for (i = 0; i < rows; ++i) {
579 
580 		addr = address;
581 		end = addr + add;
582 
583 		do {
584 			next = psb_pd_addr_end(addr, end);
585 			pt = psb_mmu_pt_map_lock(pd, addr);
586 			if (!pt)
587 				continue;
588 			do {
589 				psb_clflush(&pt->v
590 					    [psb_mmu_pt_index(addr)]);
591 			} while (addr +=
592 				 clflush_add,
593 				 (addr & clflush_mask) < next);
594 
595 			psb_mmu_pt_unmap_unlock(pt);
596 		} while (addr = next, next != end);
597 		address += row_add;
598 	}
599 	mb();
600 }
601 
602 void psb_mmu_remove_pfn_sequence(struct psb_mmu_pd *pd,
603 				 unsigned long address, uint32_t num_pages)
604 {
605 	struct psb_mmu_pt *pt;
606 	unsigned long addr;
607 	unsigned long end;
608 	unsigned long next;
609 	unsigned long f_address = address;
610 
611 	down_read(&pd->driver->sem);
612 
613 	addr = address;
614 	end = addr + (num_pages << PAGE_SHIFT);
615 
616 	do {
617 		next = psb_pd_addr_end(addr, end);
618 		pt = psb_mmu_pt_alloc_map_lock(pd, addr);
619 		if (!pt)
620 			goto out;
621 		do {
622 			psb_mmu_invalidate_pte(pt, addr);
623 			--pt->count;
624 		} while (addr += PAGE_SIZE, addr < next);
625 		psb_mmu_pt_unmap_unlock(pt);
626 
627 	} while (addr = next, next != end);
628 
629 out:
630 	if (pd->hw_context != -1)
631 		psb_mmu_flush_ptes(pd, f_address, num_pages, 1, 1);
632 
633 	up_read(&pd->driver->sem);
634 
635 	if (pd->hw_context != -1)
636 		psb_mmu_flush(pd->driver, 0);
637 
638 	return;
639 }
640 
641 void psb_mmu_remove_pages(struct psb_mmu_pd *pd, unsigned long address,
642 			  uint32_t num_pages, uint32_t desired_tile_stride,
643 			  uint32_t hw_tile_stride)
644 {
645 	struct psb_mmu_pt *pt;
646 	uint32_t rows = 1;
647 	uint32_t i;
648 	unsigned long addr;
649 	unsigned long end;
650 	unsigned long next;
651 	unsigned long add;
652 	unsigned long row_add;
653 	unsigned long f_address = address;
654 
655 	if (hw_tile_stride)
656 		rows = num_pages / desired_tile_stride;
657 	else
658 		desired_tile_stride = num_pages;
659 
660 	add = desired_tile_stride << PAGE_SHIFT;
661 	row_add = hw_tile_stride << PAGE_SHIFT;
662 
663 	/* down_read(&pd->driver->sem); */
664 
665 	/* Make sure we only need to flush this processor's cache */
666 
667 	for (i = 0; i < rows; ++i) {
668 
669 		addr = address;
670 		end = addr + add;
671 
672 		do {
673 			next = psb_pd_addr_end(addr, end);
674 			pt = psb_mmu_pt_map_lock(pd, addr);
675 			if (!pt)
676 				continue;
677 			do {
678 				psb_mmu_invalidate_pte(pt, addr);
679 				--pt->count;
680 
681 			} while (addr += PAGE_SIZE, addr < next);
682 			psb_mmu_pt_unmap_unlock(pt);
683 
684 		} while (addr = next, next != end);
685 		address += row_add;
686 	}
687 	if (pd->hw_context != -1)
688 		psb_mmu_flush_ptes(pd, f_address, num_pages,
689 				   desired_tile_stride, hw_tile_stride);
690 
691 	/* up_read(&pd->driver->sem); */
692 
693 	if (pd->hw_context != -1)
694 		psb_mmu_flush(pd->driver, 0);
695 }
696 
697 int psb_mmu_insert_pfn_sequence(struct psb_mmu_pd *pd, uint32_t start_pfn,
698 				unsigned long address, uint32_t num_pages,
699 				int type)
700 {
701 	struct psb_mmu_pt *pt;
702 	uint32_t pte;
703 	unsigned long addr;
704 	unsigned long end;
705 	unsigned long next;
706 	unsigned long f_address = address;
707 	int ret = 0;
708 
709 	down_read(&pd->driver->sem);
710 
711 	addr = address;
712 	end = addr + (num_pages << PAGE_SHIFT);
713 
714 	do {
715 		next = psb_pd_addr_end(addr, end);
716 		pt = psb_mmu_pt_alloc_map_lock(pd, addr);
717 		if (!pt) {
718 			ret = -ENOMEM;
719 			goto out;
720 		}
721 		do {
722 			pte = psb_mmu_mask_pte(start_pfn++, type);
723 			psb_mmu_set_pte(pt, addr, pte);
724 			pt->count++;
725 		} while (addr += PAGE_SIZE, addr < next);
726 		psb_mmu_pt_unmap_unlock(pt);
727 
728 	} while (addr = next, next != end);
729 
730 out:
731 	if (pd->hw_context != -1)
732 		psb_mmu_flush_ptes(pd, f_address, num_pages, 1, 1);
733 
734 	up_read(&pd->driver->sem);
735 
736 	if (pd->hw_context != -1)
737 		psb_mmu_flush(pd->driver, 1);
738 
739 	return ret;
740 }
741 
742 int psb_mmu_insert_pages(struct psb_mmu_pd *pd, struct page **pages,
743 			 unsigned long address, uint32_t num_pages,
744 			 uint32_t desired_tile_stride,
745 			 uint32_t hw_tile_stride, int type)
746 {
747 	struct psb_mmu_pt *pt;
748 	uint32_t rows = 1;
749 	uint32_t i;
750 	uint32_t pte;
751 	unsigned long addr;
752 	unsigned long end;
753 	unsigned long next;
754 	unsigned long add;
755 	unsigned long row_add;
756 	unsigned long f_address = address;
757 	int ret = 0;
758 
759 	if (hw_tile_stride) {
760 		if (num_pages % desired_tile_stride != 0)
761 			return -EINVAL;
762 		rows = num_pages / desired_tile_stride;
763 	} else {
764 		desired_tile_stride = num_pages;
765 	}
766 
767 	add = desired_tile_stride << PAGE_SHIFT;
768 	row_add = hw_tile_stride << PAGE_SHIFT;
769 
770 	down_read(&pd->driver->sem);
771 
772 	for (i = 0; i < rows; ++i) {
773 
774 		addr = address;
775 		end = addr + add;
776 
777 		do {
778 			next = psb_pd_addr_end(addr, end);
779 			pt = psb_mmu_pt_alloc_map_lock(pd, addr);
780 			if (!pt) {
781 				ret = -ENOMEM;
782 				goto out;
783 			}
784 			do {
785 				pte =
786 				    psb_mmu_mask_pte(page_to_pfn(*pages++),
787 						     type);
788 				psb_mmu_set_pte(pt, addr, pte);
789 				pt->count++;
790 			} while (addr += PAGE_SIZE, addr < next);
791 			psb_mmu_pt_unmap_unlock(pt);
792 
793 		} while (addr = next, next != end);
794 
795 		address += row_add;
796 	}
797 out:
798 	if (pd->hw_context != -1)
799 		psb_mmu_flush_ptes(pd, f_address, num_pages,
800 				   desired_tile_stride, hw_tile_stride);
801 
802 	up_read(&pd->driver->sem);
803 
804 	if (pd->hw_context != -1)
805 		psb_mmu_flush(pd->driver, 1);
806 
807 	return ret;
808 }
809 
810 int psb_mmu_virtual_to_pfn(struct psb_mmu_pd *pd, uint32_t virtual,
811 			   unsigned long *pfn)
812 {
813 	int ret;
814 	struct psb_mmu_pt *pt;
815 	uint32_t tmp;
816 	spinlock_t *lock = &pd->driver->lock;
817 
818 	down_read(&pd->driver->sem);
819 	pt = psb_mmu_pt_map_lock(pd, virtual);
820 	if (!pt) {
821 		uint32_t *v;
822 
823 		spin_lock(lock);
824 		v = kmap_atomic(pd->p);
825 		tmp = v[psb_mmu_pd_index(virtual)];
826 		kunmap_atomic(v);
827 		spin_unlock(lock);
828 
829 		if (tmp != pd->invalid_pde || !(tmp & PSB_PTE_VALID) ||
830 		    !(pd->invalid_pte & PSB_PTE_VALID)) {
831 			ret = -EINVAL;
832 			goto out;
833 		}
834 		ret = 0;
835 		*pfn = pd->invalid_pte >> PAGE_SHIFT;
836 		goto out;
837 	}
838 	tmp = pt->v[psb_mmu_pt_index(virtual)];
839 	if (!(tmp & PSB_PTE_VALID)) {
840 		ret = -EINVAL;
841 	} else {
842 		ret = 0;
843 		*pfn = tmp >> PAGE_SHIFT;
844 	}
845 	psb_mmu_pt_unmap_unlock(pt);
846 out:
847 	up_read(&pd->driver->sem);
848 	return ret;
849 }
850