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/module.h>
37 static DEFINE_SPINLOCK(slice_convert_lock);
43 static void slice_print_mask(const char *label, struct slice_mask mask)
45 char *p, buf[16 + 3 + 16 + 1];
51 for (i = 0; i < SLICE_NUM_LOW; i++)
52 *(p++) = (mask.low_slices & (1 << i)) ? '1' : '0';
56 for (i = 0; i < SLICE_NUM_HIGH; i++)
57 *(p++) = (mask.high_slices & (1 << i)) ? '1' : '0';
60 printk(KERN_DEBUG "%s:%s\n", label, buf);
63 #define slice_dbg(fmt...) do { if (_slice_debug) pr_debug(fmt); } while(0)
67 static void slice_print_mask(const char *label, struct slice_mask mask) {}
68 #define slice_dbg(fmt...)
72 static struct slice_mask slice_range_to_mask(unsigned long start,
75 unsigned long end = start + len - 1;
76 struct slice_mask ret = { 0, 0 };
78 if (start < SLICE_LOW_TOP) {
79 unsigned long mend = min(end, SLICE_LOW_TOP);
80 unsigned long mstart = min(start, SLICE_LOW_TOP);
82 ret.low_slices = (1u << (GET_LOW_SLICE_INDEX(mend) + 1))
83 - (1u << GET_LOW_SLICE_INDEX(mstart));
86 if ((start + len) > SLICE_LOW_TOP)
87 ret.high_slices = (1u << (GET_HIGH_SLICE_INDEX(end) + 1))
88 - (1u << GET_HIGH_SLICE_INDEX(start));
93 static int slice_area_is_free(struct mm_struct *mm, unsigned long addr,
96 struct vm_area_struct *vma;
98 if ((mm->task_size - len) < addr)
100 vma = find_vma(mm, addr);
101 return (!vma || (addr + len) <= vma->vm_start);
104 static int slice_low_has_vma(struct mm_struct *mm, unsigned long slice)
106 return !slice_area_is_free(mm, slice << SLICE_LOW_SHIFT,
107 1ul << SLICE_LOW_SHIFT);
110 static int slice_high_has_vma(struct mm_struct *mm, unsigned long slice)
112 unsigned long start = slice << SLICE_HIGH_SHIFT;
113 unsigned long end = start + (1ul << SLICE_HIGH_SHIFT);
115 /* Hack, so that each addresses is controlled by exactly one
116 * of the high or low area bitmaps, the first high area starts
119 start = SLICE_LOW_TOP;
121 return !slice_area_is_free(mm, start, end - start);
124 static struct slice_mask slice_mask_for_free(struct mm_struct *mm)
126 struct slice_mask ret = { 0, 0 };
129 for (i = 0; i < SLICE_NUM_LOW; i++)
130 if (!slice_low_has_vma(mm, i))
131 ret.low_slices |= 1u << i;
133 if (mm->task_size <= SLICE_LOW_TOP)
136 for (i = 0; i < SLICE_NUM_HIGH; i++)
137 if (!slice_high_has_vma(mm, i))
138 ret.high_slices |= 1u << i;
143 static struct slice_mask slice_mask_for_size(struct mm_struct *mm, int psize)
145 struct slice_mask ret = { 0, 0 };
149 psizes = mm->context.low_slices_psize;
150 for (i = 0; i < SLICE_NUM_LOW; i++)
151 if (((psizes >> (i * 4)) & 0xf) == psize)
152 ret.low_slices |= 1u << i;
154 psizes = mm->context.high_slices_psize;
155 for (i = 0; i < SLICE_NUM_HIGH; i++)
156 if (((psizes >> (i * 4)) & 0xf) == psize)
157 ret.high_slices |= 1u << i;
162 static int slice_check_fit(struct slice_mask mask, struct slice_mask available)
164 return (mask.low_slices & available.low_slices) == mask.low_slices &&
165 (mask.high_slices & available.high_slices) == mask.high_slices;
168 static void slice_flush_segments(void *parm)
170 struct mm_struct *mm = parm;
173 if (mm != current->active_mm)
176 /* update the paca copy of the context struct */
177 get_paca()->context = current->active_mm->context;
179 local_irq_save(flags);
180 slb_flush_and_rebolt();
181 local_irq_restore(flags);
184 static void slice_convert(struct mm_struct *mm, struct slice_mask mask, int psize)
186 /* Write the new slice psize bits */
187 u64 lpsizes, hpsizes;
188 unsigned long i, flags;
190 slice_dbg("slice_convert(mm=%p, psize=%d)\n", mm, psize);
191 slice_print_mask(" mask", mask);
193 /* We need to use a spinlock here to protect against
194 * concurrent 64k -> 4k demotion ...
196 spin_lock_irqsave(&slice_convert_lock, flags);
198 lpsizes = mm->context.low_slices_psize;
199 for (i = 0; i < SLICE_NUM_LOW; i++)
200 if (mask.low_slices & (1u << i))
201 lpsizes = (lpsizes & ~(0xful << (i * 4))) |
202 (((unsigned long)psize) << (i * 4));
204 hpsizes = mm->context.high_slices_psize;
205 for (i = 0; i < SLICE_NUM_HIGH; i++)
206 if (mask.high_slices & (1u << i))
207 hpsizes = (hpsizes & ~(0xful << (i * 4))) |
208 (((unsigned long)psize) << (i * 4));
210 mm->context.low_slices_psize = lpsizes;
211 mm->context.high_slices_psize = hpsizes;
213 slice_dbg(" lsps=%lx, hsps=%lx\n",
214 mm->context.low_slices_psize,
215 mm->context.high_slices_psize);
217 spin_unlock_irqrestore(&slice_convert_lock, flags);
219 #ifdef CONFIG_SPU_BASE
220 spu_flush_all_slbs(mm);
224 static unsigned long slice_find_area_bottomup(struct mm_struct *mm,
226 struct slice_mask available,
227 int psize, int use_cache)
229 struct vm_area_struct *vma;
230 unsigned long start_addr, addr;
231 struct slice_mask mask;
232 int pshift = max_t(int, mmu_psize_defs[psize].shift, PAGE_SHIFT);
235 if (len <= mm->cached_hole_size) {
236 start_addr = addr = TASK_UNMAPPED_BASE;
237 mm->cached_hole_size = 0;
239 start_addr = addr = mm->free_area_cache;
241 start_addr = addr = TASK_UNMAPPED_BASE;
245 addr = _ALIGN_UP(addr, 1ul << pshift);
246 if ((TASK_SIZE - len) < addr)
248 vma = find_vma(mm, addr);
249 BUG_ON(vma && (addr >= vma->vm_end));
251 mask = slice_range_to_mask(addr, len);
252 if (!slice_check_fit(mask, available)) {
253 if (addr < SLICE_LOW_TOP)
254 addr = _ALIGN_UP(addr + 1, 1ul << SLICE_LOW_SHIFT);
256 addr = _ALIGN_UP(addr + 1, 1ul << SLICE_HIGH_SHIFT);
259 if (!vma || addr + len <= vma->vm_start) {
261 * Remember the place where we stopped the search:
264 mm->free_area_cache = addr + len;
267 if (use_cache && (addr + mm->cached_hole_size) < vma->vm_start)
268 mm->cached_hole_size = vma->vm_start - addr;
272 /* Make sure we didn't miss any holes */
273 if (use_cache && start_addr != TASK_UNMAPPED_BASE) {
274 start_addr = addr = TASK_UNMAPPED_BASE;
275 mm->cached_hole_size = 0;
281 static unsigned long slice_find_area_topdown(struct mm_struct *mm,
283 struct slice_mask available,
284 int psize, int use_cache)
286 struct vm_area_struct *vma;
288 struct slice_mask mask;
289 int pshift = max_t(int, mmu_psize_defs[psize].shift, PAGE_SHIFT);
291 /* check if free_area_cache is useful for us */
293 if (len <= mm->cached_hole_size) {
294 mm->cached_hole_size = 0;
295 mm->free_area_cache = mm->mmap_base;
298 /* either no address requested or can't fit in requested
301 addr = mm->free_area_cache;
303 /* make sure it can fit in the remaining address space */
305 addr = _ALIGN_DOWN(addr - len, 1ul << pshift);
306 mask = slice_range_to_mask(addr, len);
307 if (slice_check_fit(mask, available) &&
308 slice_area_is_free(mm, addr, len))
309 /* remember the address as a hint for
312 return (mm->free_area_cache = addr);
316 addr = mm->mmap_base;
318 /* Go down by chunk size */
319 addr = _ALIGN_DOWN(addr - len, 1ul << pshift);
321 /* Check for hit with different page size */
322 mask = slice_range_to_mask(addr, len);
323 if (!slice_check_fit(mask, available)) {
324 if (addr < SLICE_LOW_TOP)
325 addr = _ALIGN_DOWN(addr, 1ul << SLICE_LOW_SHIFT);
326 else if (addr < (1ul << SLICE_HIGH_SHIFT))
327 addr = SLICE_LOW_TOP;
329 addr = _ALIGN_DOWN(addr, 1ul << SLICE_HIGH_SHIFT);
334 * Lookup failure means no vma is above this address,
335 * else if new region fits below vma->vm_start,
336 * return with success:
338 vma = find_vma(mm, addr);
339 if (!vma || (addr + len) <= vma->vm_start) {
340 /* remember the address as a hint for next time */
342 mm->free_area_cache = addr;
346 /* remember the largest hole we saw so far */
347 if (use_cache && (addr + mm->cached_hole_size) < vma->vm_start)
348 mm->cached_hole_size = vma->vm_start - addr;
350 /* try just below the current vma->vm_start */
351 addr = vma->vm_start;
355 * A failed mmap() very likely causes application failure,
356 * so fall back to the bottom-up function here. This scenario
357 * can happen with large stack limits and large mmap()
360 addr = slice_find_area_bottomup(mm, len, available, psize, 0);
363 * Restore the topdown base:
366 mm->free_area_cache = mm->mmap_base;
367 mm->cached_hole_size = ~0UL;
374 static unsigned long slice_find_area(struct mm_struct *mm, unsigned long len,
375 struct slice_mask mask, int psize,
376 int topdown, int use_cache)
379 return slice_find_area_topdown(mm, len, mask, psize, use_cache);
381 return slice_find_area_bottomup(mm, len, mask, psize, use_cache);
384 #define or_mask(dst, src) do { \
385 (dst).low_slices |= (src).low_slices; \
386 (dst).high_slices |= (src).high_slices; \
389 #define andnot_mask(dst, src) do { \
390 (dst).low_slices &= ~(src).low_slices; \
391 (dst).high_slices &= ~(src).high_slices; \
394 #ifdef CONFIG_PPC_64K_PAGES
395 #define MMU_PAGE_BASE MMU_PAGE_64K
397 #define MMU_PAGE_BASE MMU_PAGE_4K
400 unsigned long slice_get_unmapped_area(unsigned long addr, unsigned long len,
401 unsigned long flags, unsigned int psize,
402 int topdown, int use_cache)
404 struct slice_mask mask = {0, 0};
405 struct slice_mask good_mask;
406 struct slice_mask potential_mask = {0,0} /* silence stupid warning */;
407 struct slice_mask compat_mask = {0, 0};
408 int fixed = (flags & MAP_FIXED);
409 int pshift = max_t(int, mmu_psize_defs[psize].shift, PAGE_SHIFT);
410 struct mm_struct *mm = current->mm;
411 unsigned long newaddr;
414 BUG_ON(mm->task_size == 0);
416 slice_dbg("slice_get_unmapped_area(mm=%p, psize=%d...\n", mm, psize);
417 slice_dbg(" addr=%lx, len=%lx, flags=%lx, topdown=%d, use_cache=%d\n",
418 addr, len, flags, topdown, use_cache);
420 if (len > mm->task_size)
422 if (len & ((1ul << pshift) - 1))
424 if (fixed && (addr & ((1ul << pshift) - 1)))
426 if (fixed && addr > (mm->task_size - len))
429 /* If hint, make sure it matches our alignment restrictions */
430 if (!fixed && addr) {
431 addr = _ALIGN_UP(addr, 1ul << pshift);
432 slice_dbg(" aligned addr=%lx\n", addr);
433 /* Ignore hint if it's too large or overlaps a VMA */
434 if (addr > mm->task_size - len ||
435 !slice_area_is_free(mm, addr, len))
439 /* First make up a "good" mask of slices that have the right size
442 good_mask = slice_mask_for_size(mm, psize);
443 slice_print_mask(" good_mask", good_mask);
446 * Here "good" means slices that are already the right page size,
447 * "compat" means slices that have a compatible page size (i.e.
448 * 4k in a 64k pagesize kernel), and "free" means slices without
452 * check if fits in good | compat => OK
453 * check if fits in good | compat | free => convert free
456 * check if hint fits in good => OK
457 * check if hint fits in good | free => convert free
459 * search in good, found => OK
460 * search in good | free, found => convert free
461 * search in good | compat | free, found => convert free.
464 #ifdef CONFIG_PPC_64K_PAGES
465 /* If we support combo pages, we can allow 64k pages in 4k slices */
466 if (psize == MMU_PAGE_64K) {
467 compat_mask = slice_mask_for_size(mm, MMU_PAGE_4K);
469 or_mask(good_mask, compat_mask);
473 /* First check hint if it's valid or if we have MAP_FIXED */
474 if (addr != 0 || fixed) {
475 /* Build a mask for the requested range */
476 mask = slice_range_to_mask(addr, len);
477 slice_print_mask(" mask", mask);
479 /* Check if we fit in the good mask. If we do, we just return,
482 if (slice_check_fit(mask, good_mask)) {
483 slice_dbg(" fits good !\n");
487 /* Now let's see if we can find something in the existing
488 * slices for that size
490 newaddr = slice_find_area(mm, len, good_mask, psize, topdown,
492 if (newaddr != -ENOMEM) {
493 /* Found within the good mask, we don't have to setup,
494 * we thus return directly
496 slice_dbg(" found area at 0x%lx\n", newaddr);
501 /* We don't fit in the good mask, check what other slices are
502 * empty and thus can be converted
504 potential_mask = slice_mask_for_free(mm);
505 or_mask(potential_mask, good_mask);
506 slice_print_mask(" potential", potential_mask);
508 if ((addr != 0 || fixed) && slice_check_fit(mask, potential_mask)) {
509 slice_dbg(" fits potential !\n");
513 /* If we have MAP_FIXED and failed the above steps, then error out */
517 slice_dbg(" search...\n");
519 /* If we had a hint that didn't work out, see if we can fit
520 * anywhere in the good area.
523 addr = slice_find_area(mm, len, good_mask, psize, topdown,
525 if (addr != -ENOMEM) {
526 slice_dbg(" found area at 0x%lx\n", addr);
531 /* Now let's see if we can find something in the existing slices
532 * for that size plus free slices
534 addr = slice_find_area(mm, len, potential_mask, psize, topdown,
537 #ifdef CONFIG_PPC_64K_PAGES
538 if (addr == -ENOMEM && psize == MMU_PAGE_64K) {
539 /* retry the search with 4k-page slices included */
540 or_mask(potential_mask, compat_mask);
541 addr = slice_find_area(mm, len, potential_mask, psize,
549 mask = slice_range_to_mask(addr, len);
550 slice_dbg(" found potential area at 0x%lx\n", addr);
551 slice_print_mask(" mask", mask);
554 andnot_mask(mask, good_mask);
555 andnot_mask(mask, compat_mask);
556 if (mask.low_slices || mask.high_slices) {
557 slice_convert(mm, mask, psize);
558 if (psize > MMU_PAGE_BASE)
559 on_each_cpu(slice_flush_segments, mm, 1);
564 EXPORT_SYMBOL_GPL(slice_get_unmapped_area);
566 unsigned long arch_get_unmapped_area(struct file *filp,
572 return slice_get_unmapped_area(addr, len, flags,
573 current->mm->context.user_psize,
577 unsigned long arch_get_unmapped_area_topdown(struct file *filp,
578 const unsigned long addr0,
579 const unsigned long len,
580 const unsigned long pgoff,
581 const unsigned long flags)
583 return slice_get_unmapped_area(addr0, len, flags,
584 current->mm->context.user_psize,
588 unsigned int get_slice_psize(struct mm_struct *mm, unsigned long addr)
593 if (addr < SLICE_LOW_TOP) {
594 psizes = mm->context.low_slices_psize;
595 index = GET_LOW_SLICE_INDEX(addr);
597 psizes = mm->context.high_slices_psize;
598 index = GET_HIGH_SLICE_INDEX(addr);
601 return (psizes >> (index * 4)) & 0xf;
603 EXPORT_SYMBOL_GPL(get_slice_psize);
606 * This is called by hash_page when it needs to do a lazy conversion of
607 * an address space from real 64K pages to combo 4K pages (typically
608 * when hitting a non cacheable mapping on a processor or hypervisor
609 * that won't allow them for 64K pages).
611 * This is also called in init_new_context() to change back the user
612 * psize from whatever the parent context had it set to
613 * N.B. This may be called before mm->context.id has been set.
615 * This function will only change the content of the {low,high)_slice_psize
616 * masks, it will not flush SLBs as this shall be handled lazily by the
619 void slice_set_user_psize(struct mm_struct *mm, unsigned int psize)
621 unsigned long flags, lpsizes, hpsizes;
622 unsigned int old_psize;
625 slice_dbg("slice_set_user_psize(mm=%p, psize=%d)\n", mm, psize);
627 spin_lock_irqsave(&slice_convert_lock, flags);
629 old_psize = mm->context.user_psize;
630 slice_dbg(" old_psize=%d\n", old_psize);
631 if (old_psize == psize)
634 mm->context.user_psize = psize;
637 lpsizes = mm->context.low_slices_psize;
638 for (i = 0; i < SLICE_NUM_LOW; i++)
639 if (((lpsizes >> (i * 4)) & 0xf) == old_psize)
640 lpsizes = (lpsizes & ~(0xful << (i * 4))) |
641 (((unsigned long)psize) << (i * 4));
643 hpsizes = mm->context.high_slices_psize;
644 for (i = 0; i < SLICE_NUM_HIGH; i++)
645 if (((hpsizes >> (i * 4)) & 0xf) == old_psize)
646 hpsizes = (hpsizes & ~(0xful << (i * 4))) |
647 (((unsigned long)psize) << (i * 4));
649 mm->context.low_slices_psize = lpsizes;
650 mm->context.high_slices_psize = hpsizes;
652 slice_dbg(" lsps=%lx, hsps=%lx\n",
653 mm->context.low_slices_psize,
654 mm->context.high_slices_psize);
657 spin_unlock_irqrestore(&slice_convert_lock, flags);
660 void slice_set_psize(struct mm_struct *mm, unsigned long address,
663 unsigned long i, flags;
666 spin_lock_irqsave(&slice_convert_lock, flags);
667 if (address < SLICE_LOW_TOP) {
668 i = GET_LOW_SLICE_INDEX(address);
669 p = &mm->context.low_slices_psize;
671 i = GET_HIGH_SLICE_INDEX(address);
672 p = &mm->context.high_slices_psize;
674 *p = (*p & ~(0xful << (i * 4))) | ((unsigned long) psize << (i * 4));
675 spin_unlock_irqrestore(&slice_convert_lock, flags);
677 #ifdef CONFIG_SPU_BASE
678 spu_flush_all_slbs(mm);
682 void slice_set_range_psize(struct mm_struct *mm, unsigned long start,
683 unsigned long len, unsigned int psize)
685 struct slice_mask mask = slice_range_to_mask(start, len);
687 slice_convert(mm, mask, psize);
691 * is_hugepage_only_range() is used by generic code to verify wether
692 * a normal mmap mapping (non hugetlbfs) is valid on a given area.
694 * until the generic code provides a more generic hook and/or starts
695 * calling arch get_unmapped_area for MAP_FIXED (which our implementation
696 * here knows how to deal with), we hijack it to keep standard mappings
699 * because of that generic code limitation, MAP_FIXED mapping cannot
700 * "convert" back a slice with no VMAs to the standard page size, only
701 * get_unmapped_area() can. It would be possible to fix it here but I
702 * prefer working on fixing the generic code instead.
704 * WARNING: This will not work if hugetlbfs isn't enabled since the
705 * generic code will redefine that function as 0 in that. This is ok
706 * for now as we only use slices with hugetlbfs enabled. This should
707 * be fixed as the generic code gets fixed.
709 int is_hugepage_only_range(struct mm_struct *mm, unsigned long addr,
712 struct slice_mask mask, available;
713 unsigned int psize = mm->context.user_psize;
715 mask = slice_range_to_mask(addr, len);
716 available = slice_mask_for_size(mm, psize);
717 #ifdef CONFIG_PPC_64K_PAGES
718 /* We need to account for 4k slices too */
719 if (psize == MMU_PAGE_64K) {
720 struct slice_mask compat_mask;
721 compat_mask = slice_mask_for_size(mm, MMU_PAGE_4K);
722 or_mask(available, compat_mask);
726 #if 0 /* too verbose */
727 slice_dbg("is_hugepage_only_range(mm=%p, addr=%lx, len=%lx)\n",
729 slice_print_mask(" mask", mask);
730 slice_print_mask(" available", available);
732 return !slice_check_fit(mask, available);