2 * address space "slices" (meta-segments) support
4 * Copyright (C) 2007 Benjamin Herrenschmidt, IBM Corporation.
6 * Based on hugetlb implementation
8 * Copyright (C) 2003 David Gibson, IBM Corporation.
10 * This program is free software; you can redistribute it and/or modify
11 * it under the terms of the GNU General Public License as published by
12 * the Free Software Foundation; either version 2 of the License, or
13 * (at your option) any later version.
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 * GNU General Public License for more details.
20 * You should have received a copy of the GNU General Public License
21 * along with this program; if not, write to the Free Software
22 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
27 #include <linux/kernel.h>
29 #include <linux/pagemap.h>
30 #include <linux/err.h>
31 #include <linux/spinlock.h>
32 #include <linux/export.h>
33 #include <linux/hugetlb.h>
36 #include <asm/copro.h>
37 #include <asm/hugetlb.h>
39 static DEFINE_SPINLOCK(slice_convert_lock);
41 * One bit per slice. We have lower slices which cover 256MB segments
42 * upto 4G range. That gets us 16 low slices. For the rest we track slices
47 DECLARE_BITMAP(high_slices, SLICE_NUM_HIGH);
53 static void slice_print_mask(const char *label, struct slice_mask mask)
57 pr_devel("%s low_slice: %*pbl\n", label, (int)SLICE_NUM_LOW, &mask.low_slices);
58 pr_devel("%s high_slice: %*pbl\n", label, (int)SLICE_NUM_HIGH, mask.high_slices);
61 #define slice_dbg(fmt...) do { if (_slice_debug) pr_devel(fmt); } while (0)
65 static void slice_print_mask(const char *label, struct slice_mask mask) {}
66 #define slice_dbg(fmt...)
70 static void slice_range_to_mask(unsigned long start, unsigned long len,
71 struct slice_mask *ret)
73 unsigned long end = start + len - 1;
76 bitmap_zero(ret->high_slices, SLICE_NUM_HIGH);
78 if (start < SLICE_LOW_TOP) {
79 unsigned long mend = min(end, (SLICE_LOW_TOP - 1));
81 ret->low_slices = (1u << (GET_LOW_SLICE_INDEX(mend) + 1))
82 - (1u << GET_LOW_SLICE_INDEX(start));
85 if ((start + len) > SLICE_LOW_TOP) {
86 unsigned long start_index = GET_HIGH_SLICE_INDEX(start);
87 unsigned long align_end = ALIGN(end, (1UL << SLICE_HIGH_SHIFT));
88 unsigned long count = GET_HIGH_SLICE_INDEX(align_end) - start_index;
90 bitmap_set(ret->high_slices, start_index, count);
94 static int slice_area_is_free(struct mm_struct *mm, unsigned long addr,
97 struct vm_area_struct *vma;
99 if ((mm->task_size - len) < addr)
101 vma = find_vma(mm, addr);
102 return (!vma || (addr + len) <= vma->vm_start);
105 static int slice_low_has_vma(struct mm_struct *mm, unsigned long slice)
107 return !slice_area_is_free(mm, slice << SLICE_LOW_SHIFT,
108 1ul << SLICE_LOW_SHIFT);
111 static int slice_high_has_vma(struct mm_struct *mm, unsigned long slice)
113 unsigned long start = slice << SLICE_HIGH_SHIFT;
114 unsigned long end = start + (1ul << SLICE_HIGH_SHIFT);
116 /* Hack, so that each addresses is controlled by exactly one
117 * of the high or low area bitmaps, the first high area starts
120 start = SLICE_LOW_TOP;
122 return !slice_area_is_free(mm, start, end - start);
125 static void slice_mask_for_free(struct mm_struct *mm, struct slice_mask *ret)
130 bitmap_zero(ret->high_slices, SLICE_NUM_HIGH);
132 for (i = 0; i < SLICE_NUM_LOW; i++)
133 if (!slice_low_has_vma(mm, i))
134 ret->low_slices |= 1u << i;
136 if (mm->task_size <= SLICE_LOW_TOP)
139 for (i = 0; i < GET_HIGH_SLICE_INDEX(mm->context.addr_limit); i++)
140 if (!slice_high_has_vma(mm, i))
141 __set_bit(i, ret->high_slices);
144 static void slice_mask_for_size(struct mm_struct *mm, int psize, struct slice_mask *ret)
146 unsigned char *hpsizes;
147 int index, mask_index;
152 bitmap_zero(ret->high_slices, SLICE_NUM_HIGH);
154 lpsizes = mm->context.low_slices_psize;
155 for (i = 0; i < SLICE_NUM_LOW; i++)
156 if (((lpsizes >> (i * 4)) & 0xf) == psize)
157 ret->low_slices |= 1u << i;
159 hpsizes = mm->context.high_slices_psize;
160 for (i = 0; i < GET_HIGH_SLICE_INDEX(mm->context.addr_limit); i++) {
161 mask_index = i & 0x1;
163 if (((hpsizes[index] >> (mask_index * 4)) & 0xf) == psize)
164 __set_bit(i, ret->high_slices);
168 static int slice_check_fit(struct mm_struct *mm,
169 struct slice_mask mask, struct slice_mask available)
171 DECLARE_BITMAP(result, SLICE_NUM_HIGH);
172 unsigned long slice_count = GET_HIGH_SLICE_INDEX(mm->context.addr_limit);
174 bitmap_and(result, mask.high_slices,
175 available.high_slices, slice_count);
177 return (mask.low_slices & available.low_slices) == mask.low_slices &&
178 bitmap_equal(result, mask.high_slices, slice_count);
181 static void slice_flush_segments(void *parm)
183 struct mm_struct *mm = parm;
186 if (mm != current->active_mm)
189 copy_mm_to_paca(current->active_mm);
191 local_irq_save(flags);
192 slb_flush_and_rebolt();
193 local_irq_restore(flags);
196 static void slice_convert(struct mm_struct *mm, struct slice_mask mask, int psize)
198 int index, mask_index;
199 /* Write the new slice psize bits */
200 unsigned char *hpsizes;
202 unsigned long i, flags;
204 slice_dbg("slice_convert(mm=%p, psize=%d)\n", mm, psize);
205 slice_print_mask(" mask", mask);
207 /* We need to use a spinlock here to protect against
208 * concurrent 64k -> 4k demotion ...
210 spin_lock_irqsave(&slice_convert_lock, flags);
212 lpsizes = mm->context.low_slices_psize;
213 for (i = 0; i < SLICE_NUM_LOW; i++)
214 if (mask.low_slices & (1u << i))
215 lpsizes = (lpsizes & ~(0xful << (i * 4))) |
216 (((unsigned long)psize) << (i * 4));
218 /* Assign the value back */
219 mm->context.low_slices_psize = lpsizes;
221 hpsizes = mm->context.high_slices_psize;
222 for (i = 0; i < GET_HIGH_SLICE_INDEX(mm->context.addr_limit); i++) {
223 mask_index = i & 0x1;
225 if (test_bit(i, mask.high_slices))
226 hpsizes[index] = (hpsizes[index] &
227 ~(0xf << (mask_index * 4))) |
228 (((unsigned long)psize) << (mask_index * 4));
231 slice_dbg(" lsps=%lx, hsps=%lx\n",
232 (unsigned long)mm->context.low_slices_psize,
233 (unsigned long)mm->context.high_slices_psize);
235 spin_unlock_irqrestore(&slice_convert_lock, flags);
237 copro_flush_all_slbs(mm);
241 * Compute which slice addr is part of;
242 * set *boundary_addr to the start or end boundary of that slice
243 * (depending on 'end' parameter);
244 * return boolean indicating if the slice is marked as available in the
245 * 'available' slice_mark.
247 static bool slice_scan_available(unsigned long addr,
248 struct slice_mask available,
250 unsigned long *boundary_addr)
253 if (addr < SLICE_LOW_TOP) {
254 slice = GET_LOW_SLICE_INDEX(addr);
255 *boundary_addr = (slice + end) << SLICE_LOW_SHIFT;
256 return !!(available.low_slices & (1u << slice));
258 slice = GET_HIGH_SLICE_INDEX(addr);
259 *boundary_addr = (slice + end) ?
260 ((slice + end) << SLICE_HIGH_SHIFT) : SLICE_LOW_TOP;
261 return !!test_bit(slice, available.high_slices);
265 static unsigned long slice_find_area_bottomup(struct mm_struct *mm,
267 struct slice_mask available,
268 int psize, unsigned long high_limit)
270 int pshift = max_t(int, mmu_psize_defs[psize].shift, PAGE_SHIFT);
271 unsigned long addr, found, next_end;
272 struct vm_unmapped_area_info info;
276 info.align_mask = PAGE_MASK & ((1ul << pshift) - 1);
277 info.align_offset = 0;
279 addr = TASK_UNMAPPED_BASE;
281 * Check till the allow max value for this mmap request
283 while (addr < high_limit) {
284 info.low_limit = addr;
285 if (!slice_scan_available(addr, available, 1, &addr))
290 * At this point [info.low_limit; addr) covers
291 * available slices only and ends at a slice boundary.
292 * Check if we need to reduce the range, or if we can
293 * extend it to cover the next available slice.
295 if (addr >= high_limit)
297 else if (slice_scan_available(addr, available, 1, &next_end)) {
301 info.high_limit = addr;
303 found = vm_unmapped_area(&info);
304 if (!(found & ~PAGE_MASK))
311 static unsigned long slice_find_area_topdown(struct mm_struct *mm,
313 struct slice_mask available,
314 int psize, unsigned long high_limit)
316 int pshift = max_t(int, mmu_psize_defs[psize].shift, PAGE_SHIFT);
317 unsigned long addr, found, prev;
318 struct vm_unmapped_area_info info;
320 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
322 info.align_mask = PAGE_MASK & ((1ul << pshift) - 1);
323 info.align_offset = 0;
325 addr = mm->mmap_base;
327 * If we are trying to allocate above DEFAULT_MAP_WINDOW
328 * Add the different to the mmap_base.
329 * Only for that request for which high_limit is above
330 * DEFAULT_MAP_WINDOW we should apply this.
332 if (high_limit > DEFAULT_MAP_WINDOW)
333 addr += mm->context.addr_limit - DEFAULT_MAP_WINDOW;
335 while (addr > PAGE_SIZE) {
336 info.high_limit = addr;
337 if (!slice_scan_available(addr - 1, available, 0, &addr))
342 * At this point [addr; info.high_limit) covers
343 * available slices only and starts at a slice boundary.
344 * Check if we need to reduce the range, or if we can
345 * extend it to cover the previous available slice.
347 if (addr < PAGE_SIZE)
349 else if (slice_scan_available(addr - 1, available, 0, &prev)) {
353 info.low_limit = addr;
355 found = vm_unmapped_area(&info);
356 if (!(found & ~PAGE_MASK))
361 * A failed mmap() very likely causes application failure,
362 * so fall back to the bottom-up function here. This scenario
363 * can happen with large stack limits and large mmap()
366 return slice_find_area_bottomup(mm, len, available, psize, high_limit);
370 static unsigned long slice_find_area(struct mm_struct *mm, unsigned long len,
371 struct slice_mask mask, int psize,
372 int topdown, unsigned long high_limit)
375 return slice_find_area_topdown(mm, len, mask, psize, high_limit);
377 return slice_find_area_bottomup(mm, len, mask, psize, high_limit);
380 static inline void slice_or_mask(struct slice_mask *dst, struct slice_mask *src)
382 DECLARE_BITMAP(result, SLICE_NUM_HIGH);
384 dst->low_slices |= src->low_slices;
385 bitmap_or(result, dst->high_slices, src->high_slices, SLICE_NUM_HIGH);
386 bitmap_copy(dst->high_slices, result, SLICE_NUM_HIGH);
389 static inline void slice_andnot_mask(struct slice_mask *dst, struct slice_mask *src)
391 DECLARE_BITMAP(result, SLICE_NUM_HIGH);
393 dst->low_slices &= ~src->low_slices;
395 bitmap_andnot(result, dst->high_slices, src->high_slices, SLICE_NUM_HIGH);
396 bitmap_copy(dst->high_slices, result, SLICE_NUM_HIGH);
399 #ifdef CONFIG_PPC_64K_PAGES
400 #define MMU_PAGE_BASE MMU_PAGE_64K
402 #define MMU_PAGE_BASE MMU_PAGE_4K
405 unsigned long slice_get_unmapped_area(unsigned long addr, unsigned long len,
406 unsigned long flags, unsigned int psize,
409 struct slice_mask mask;
410 struct slice_mask good_mask;
411 struct slice_mask potential_mask;
412 struct slice_mask compat_mask;
413 int fixed = (flags & MAP_FIXED);
414 int pshift = max_t(int, mmu_psize_defs[psize].shift, PAGE_SHIFT);
415 struct mm_struct *mm = current->mm;
416 unsigned long newaddr;
417 unsigned long high_limit;
420 * Check if we need to expland slice area.
422 if (unlikely(addr > mm->context.addr_limit && addr < TASK_SIZE)) {
423 mm->context.addr_limit = TASK_SIZE;
424 on_each_cpu(slice_flush_segments, mm, 1);
427 * This mmap request can allocate upt to 512TB
429 if (addr > DEFAULT_MAP_WINDOW)
430 high_limit = mm->context.addr_limit;
432 high_limit = DEFAULT_MAP_WINDOW;
434 * init different masks
437 bitmap_zero(mask.high_slices, SLICE_NUM_HIGH);
439 /* silence stupid warning */;
440 potential_mask.low_slices = 0;
441 bitmap_zero(potential_mask.high_slices, SLICE_NUM_HIGH);
443 compat_mask.low_slices = 0;
444 bitmap_zero(compat_mask.high_slices, SLICE_NUM_HIGH);
447 BUG_ON(mm->task_size == 0);
448 VM_BUG_ON(radix_enabled());
450 slice_dbg("slice_get_unmapped_area(mm=%p, psize=%d...\n", mm, psize);
451 slice_dbg(" addr=%lx, len=%lx, flags=%lx, topdown=%d\n",
452 addr, len, flags, topdown);
454 if (len > mm->task_size)
456 if (len & ((1ul << pshift) - 1))
458 if (fixed && (addr & ((1ul << pshift) - 1)))
460 if (fixed && addr > (mm->task_size - len))
463 /* If hint, make sure it matches our alignment restrictions */
464 if (!fixed && addr) {
465 addr = _ALIGN_UP(addr, 1ul << pshift);
466 slice_dbg(" aligned addr=%lx\n", addr);
467 /* Ignore hint if it's too large or overlaps a VMA */
468 if (addr > mm->task_size - len ||
469 !slice_area_is_free(mm, addr, len))
473 /* First make up a "good" mask of slices that have the right size
476 slice_mask_for_size(mm, psize, &good_mask);
477 slice_print_mask(" good_mask", good_mask);
480 * Here "good" means slices that are already the right page size,
481 * "compat" means slices that have a compatible page size (i.e.
482 * 4k in a 64k pagesize kernel), and "free" means slices without
486 * check if fits in good | compat => OK
487 * check if fits in good | compat | free => convert free
490 * check if hint fits in good => OK
491 * check if hint fits in good | free => convert free
493 * search in good, found => OK
494 * search in good | free, found => convert free
495 * search in good | compat | free, found => convert free.
498 #ifdef CONFIG_PPC_64K_PAGES
499 /* If we support combo pages, we can allow 64k pages in 4k slices */
500 if (psize == MMU_PAGE_64K) {
501 slice_mask_for_size(mm, MMU_PAGE_4K, &compat_mask);
503 slice_or_mask(&good_mask, &compat_mask);
507 /* First check hint if it's valid or if we have MAP_FIXED */
508 if (addr != 0 || fixed) {
509 /* Build a mask for the requested range */
510 slice_range_to_mask(addr, len, &mask);
511 slice_print_mask(" mask", mask);
513 /* Check if we fit in the good mask. If we do, we just return,
516 if (slice_check_fit(mm, mask, good_mask)) {
517 slice_dbg(" fits good !\n");
521 /* Now let's see if we can find something in the existing
522 * slices for that size
524 newaddr = slice_find_area(mm, len, good_mask,
525 psize, topdown, high_limit);
526 if (newaddr != -ENOMEM) {
527 /* Found within the good mask, we don't have to setup,
528 * we thus return directly
530 slice_dbg(" found area at 0x%lx\n", newaddr);
535 /* We don't fit in the good mask, check what other slices are
536 * empty and thus can be converted
538 slice_mask_for_free(mm, &potential_mask);
539 slice_or_mask(&potential_mask, &good_mask);
540 slice_print_mask(" potential", potential_mask);
542 if ((addr != 0 || fixed) && slice_check_fit(mm, mask, potential_mask)) {
543 slice_dbg(" fits potential !\n");
547 /* If we have MAP_FIXED and failed the above steps, then error out */
551 slice_dbg(" search...\n");
553 /* If we had a hint that didn't work out, see if we can fit
554 * anywhere in the good area.
557 addr = slice_find_area(mm, len, good_mask,
558 psize, topdown, high_limit);
559 if (addr != -ENOMEM) {
560 slice_dbg(" found area at 0x%lx\n", addr);
565 /* Now let's see if we can find something in the existing slices
566 * for that size plus free slices
568 addr = slice_find_area(mm, len, potential_mask,
569 psize, topdown, high_limit);
571 #ifdef CONFIG_PPC_64K_PAGES
572 if (addr == -ENOMEM && psize == MMU_PAGE_64K) {
573 /* retry the search with 4k-page slices included */
574 slice_or_mask(&potential_mask, &compat_mask);
575 addr = slice_find_area(mm, len, potential_mask,
576 psize, topdown, high_limit);
583 slice_range_to_mask(addr, len, &mask);
584 slice_dbg(" found potential area at 0x%lx\n", addr);
585 slice_print_mask(" mask", mask);
588 slice_andnot_mask(&mask, &good_mask);
589 slice_andnot_mask(&mask, &compat_mask);
590 if (mask.low_slices || !bitmap_empty(mask.high_slices, SLICE_NUM_HIGH)) {
591 slice_convert(mm, mask, psize);
592 if (psize > MMU_PAGE_BASE)
593 on_each_cpu(slice_flush_segments, mm, 1);
598 EXPORT_SYMBOL_GPL(slice_get_unmapped_area);
600 unsigned long arch_get_unmapped_area(struct file *filp,
606 return slice_get_unmapped_area(addr, len, flags,
607 current->mm->context.user_psize, 0);
610 unsigned long arch_get_unmapped_area_topdown(struct file *filp,
611 const unsigned long addr0,
612 const unsigned long len,
613 const unsigned long pgoff,
614 const unsigned long flags)
616 return slice_get_unmapped_area(addr0, len, flags,
617 current->mm->context.user_psize, 1);
620 unsigned int get_slice_psize(struct mm_struct *mm, unsigned long addr)
622 unsigned char *hpsizes;
623 int index, mask_index;
626 * Radix doesn't use slice, but can get enabled along with MMU_SLICE
628 if (radix_enabled()) {
629 #ifdef CONFIG_PPC_64K_PAGES
635 if (addr < SLICE_LOW_TOP) {
637 lpsizes = mm->context.low_slices_psize;
638 index = GET_LOW_SLICE_INDEX(addr);
639 return (lpsizes >> (index * 4)) & 0xf;
641 hpsizes = mm->context.high_slices_psize;
642 index = GET_HIGH_SLICE_INDEX(addr);
643 mask_index = index & 0x1;
644 return (hpsizes[index >> 1] >> (mask_index * 4)) & 0xf;
646 EXPORT_SYMBOL_GPL(get_slice_psize);
649 * This is called by hash_page when it needs to do a lazy conversion of
650 * an address space from real 64K pages to combo 4K pages (typically
651 * when hitting a non cacheable mapping on a processor or hypervisor
652 * that won't allow them for 64K pages).
654 * This is also called in init_new_context() to change back the user
655 * psize from whatever the parent context had it set to
656 * N.B. This may be called before mm->context.id has been set.
658 * This function will only change the content of the {low,high)_slice_psize
659 * masks, it will not flush SLBs as this shall be handled lazily by the
662 void slice_set_user_psize(struct mm_struct *mm, unsigned int psize)
664 int index, mask_index;
665 unsigned char *hpsizes;
666 unsigned long flags, lpsizes;
667 unsigned int old_psize;
670 slice_dbg("slice_set_user_psize(mm=%p, psize=%d)\n", mm, psize);
672 VM_BUG_ON(radix_enabled());
673 spin_lock_irqsave(&slice_convert_lock, flags);
675 old_psize = mm->context.user_psize;
676 slice_dbg(" old_psize=%d\n", old_psize);
677 if (old_psize == psize)
680 mm->context.user_psize = psize;
683 lpsizes = mm->context.low_slices_psize;
684 for (i = 0; i < SLICE_NUM_LOW; i++)
685 if (((lpsizes >> (i * 4)) & 0xf) == old_psize)
686 lpsizes = (lpsizes & ~(0xful << (i * 4))) |
687 (((unsigned long)psize) << (i * 4));
688 /* Assign the value back */
689 mm->context.low_slices_psize = lpsizes;
691 hpsizes = mm->context.high_slices_psize;
692 for (i = 0; i < SLICE_NUM_HIGH; i++) {
693 mask_index = i & 0x1;
695 if (((hpsizes[index] >> (mask_index * 4)) & 0xf) == old_psize)
696 hpsizes[index] = (hpsizes[index] &
697 ~(0xf << (mask_index * 4))) |
698 (((unsigned long)psize) << (mask_index * 4));
704 slice_dbg(" lsps=%lx, hsps=%lx\n",
705 (unsigned long)mm->context.low_slices_psize,
706 (unsigned long)mm->context.high_slices_psize);
709 spin_unlock_irqrestore(&slice_convert_lock, flags);
712 void slice_set_range_psize(struct mm_struct *mm, unsigned long start,
713 unsigned long len, unsigned int psize)
715 struct slice_mask mask;
717 VM_BUG_ON(radix_enabled());
719 slice_range_to_mask(start, len, &mask);
720 slice_convert(mm, mask, psize);
723 #ifdef CONFIG_HUGETLB_PAGE
725 * is_hugepage_only_range() is used by generic code to verify whether
726 * a normal mmap mapping (non hugetlbfs) is valid on a given area.
728 * until the generic code provides a more generic hook and/or starts
729 * calling arch get_unmapped_area for MAP_FIXED (which our implementation
730 * here knows how to deal with), we hijack it to keep standard mappings
733 * because of that generic code limitation, MAP_FIXED mapping cannot
734 * "convert" back a slice with no VMAs to the standard page size, only
735 * get_unmapped_area() can. It would be possible to fix it here but I
736 * prefer working on fixing the generic code instead.
738 * WARNING: This will not work if hugetlbfs isn't enabled since the
739 * generic code will redefine that function as 0 in that. This is ok
740 * for now as we only use slices with hugetlbfs enabled. This should
741 * be fixed as the generic code gets fixed.
743 int is_hugepage_only_range(struct mm_struct *mm, unsigned long addr,
746 struct slice_mask mask, available;
747 unsigned int psize = mm->context.user_psize;
752 slice_range_to_mask(addr, len, &mask);
753 slice_mask_for_size(mm, psize, &available);
754 #ifdef CONFIG_PPC_64K_PAGES
755 /* We need to account for 4k slices too */
756 if (psize == MMU_PAGE_64K) {
757 struct slice_mask compat_mask;
758 slice_mask_for_size(mm, MMU_PAGE_4K, &compat_mask);
759 slice_or_mask(&available, &compat_mask);
763 #if 0 /* too verbose */
764 slice_dbg("is_hugepage_only_range(mm=%p, addr=%lx, len=%lx)\n",
766 slice_print_mask(" mask", mask);
767 slice_print_mask(" available", available);
769 return !slice_check_fit(mm, mask, available);