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 spinlock_t slice_convert_lock = SPIN_LOCK_UNLOCKED;
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);
220 /* XXX this is sub-optimal but will do for now */
221 on_each_cpu(slice_flush_segments, mm, 0, 1);
222 #ifdef CONFIG_SPU_BASE
223 spu_flush_all_slbs(mm);
227 static unsigned long slice_find_area_bottomup(struct mm_struct *mm,
229 struct slice_mask available,
230 int psize, int use_cache)
232 struct vm_area_struct *vma;
233 unsigned long start_addr, addr;
234 struct slice_mask mask;
235 int pshift = max_t(int, mmu_psize_defs[psize].shift, PAGE_SHIFT);
238 if (len <= mm->cached_hole_size) {
239 start_addr = addr = TASK_UNMAPPED_BASE;
240 mm->cached_hole_size = 0;
242 start_addr = addr = mm->free_area_cache;
244 start_addr = addr = TASK_UNMAPPED_BASE;
248 addr = _ALIGN_UP(addr, 1ul << pshift);
249 if ((TASK_SIZE - len) < addr)
251 vma = find_vma(mm, addr);
252 BUG_ON(vma && (addr >= vma->vm_end));
254 mask = slice_range_to_mask(addr, len);
255 if (!slice_check_fit(mask, available)) {
256 if (addr < SLICE_LOW_TOP)
257 addr = _ALIGN_UP(addr + 1, 1ul << SLICE_LOW_SHIFT);
259 addr = _ALIGN_UP(addr + 1, 1ul << SLICE_HIGH_SHIFT);
262 if (!vma || addr + len <= vma->vm_start) {
264 * Remember the place where we stopped the search:
267 mm->free_area_cache = addr + len;
270 if (use_cache && (addr + mm->cached_hole_size) < vma->vm_start)
271 mm->cached_hole_size = vma->vm_start - addr;
275 /* Make sure we didn't miss any holes */
276 if (use_cache && start_addr != TASK_UNMAPPED_BASE) {
277 start_addr = addr = TASK_UNMAPPED_BASE;
278 mm->cached_hole_size = 0;
284 static unsigned long slice_find_area_topdown(struct mm_struct *mm,
286 struct slice_mask available,
287 int psize, int use_cache)
289 struct vm_area_struct *vma;
291 struct slice_mask mask;
292 int pshift = max_t(int, mmu_psize_defs[psize].shift, PAGE_SHIFT);
294 /* check if free_area_cache is useful for us */
296 if (len <= mm->cached_hole_size) {
297 mm->cached_hole_size = 0;
298 mm->free_area_cache = mm->mmap_base;
301 /* either no address requested or can't fit in requested
304 addr = mm->free_area_cache;
306 /* make sure it can fit in the remaining address space */
308 addr = _ALIGN_DOWN(addr - len, 1ul << pshift);
309 mask = slice_range_to_mask(addr, len);
310 if (slice_check_fit(mask, available) &&
311 slice_area_is_free(mm, addr, len))
312 /* remember the address as a hint for
315 return (mm->free_area_cache = addr);
319 addr = mm->mmap_base;
321 /* Go down by chunk size */
322 addr = _ALIGN_DOWN(addr - len, 1ul << pshift);
324 /* Check for hit with different page size */
325 mask = slice_range_to_mask(addr, len);
326 if (!slice_check_fit(mask, available)) {
327 if (addr < SLICE_LOW_TOP)
328 addr = _ALIGN_DOWN(addr, 1ul << SLICE_LOW_SHIFT);
329 else if (addr < (1ul << SLICE_HIGH_SHIFT))
330 addr = SLICE_LOW_TOP;
332 addr = _ALIGN_DOWN(addr, 1ul << SLICE_HIGH_SHIFT);
337 * Lookup failure means no vma is above this address,
338 * else if new region fits below vma->vm_start,
339 * return with success:
341 vma = find_vma(mm, addr);
342 if (!vma || (addr + len) <= vma->vm_start) {
343 /* remember the address as a hint for next time */
345 mm->free_area_cache = addr;
349 /* remember the largest hole we saw so far */
350 if (use_cache && (addr + mm->cached_hole_size) < vma->vm_start)
351 mm->cached_hole_size = vma->vm_start - addr;
353 /* try just below the current vma->vm_start */
354 addr = vma->vm_start;
358 * A failed mmap() very likely causes application failure,
359 * so fall back to the bottom-up function here. This scenario
360 * can happen with large stack limits and large mmap()
363 addr = slice_find_area_bottomup(mm, len, available, psize, 0);
366 * Restore the topdown base:
369 mm->free_area_cache = mm->mmap_base;
370 mm->cached_hole_size = ~0UL;
377 static unsigned long slice_find_area(struct mm_struct *mm, unsigned long len,
378 struct slice_mask mask, int psize,
379 int topdown, int use_cache)
382 return slice_find_area_topdown(mm, len, mask, psize, use_cache);
384 return slice_find_area_bottomup(mm, len, mask, psize, use_cache);
387 unsigned long slice_get_unmapped_area(unsigned long addr, unsigned long len,
388 unsigned long flags, unsigned int psize,
389 int topdown, int use_cache)
391 struct slice_mask mask;
392 struct slice_mask good_mask;
393 struct slice_mask potential_mask = {0,0} /* silence stupid warning */;
395 int fixed = (flags & MAP_FIXED);
396 int pshift = max_t(int, mmu_psize_defs[psize].shift, PAGE_SHIFT);
397 struct mm_struct *mm = current->mm;
400 BUG_ON(mm->task_size == 0);
402 slice_dbg("slice_get_unmapped_area(mm=%p, psize=%d...\n", mm, psize);
403 slice_dbg(" addr=%lx, len=%lx, flags=%lx, topdown=%d, use_cache=%d\n",
404 addr, len, flags, topdown, use_cache);
406 if (len > mm->task_size)
408 if (len & ((1ul << pshift) - 1))
410 if (fixed && (addr & ((1ul << pshift) - 1)))
412 if (fixed && addr > (mm->task_size - len))
415 /* If hint, make sure it matches our alignment restrictions */
416 if (!fixed && addr) {
417 addr = _ALIGN_UP(addr, 1ul << pshift);
418 slice_dbg(" aligned addr=%lx\n", addr);
421 /* First makeup a "good" mask of slices that have the right size
424 good_mask = slice_mask_for_size(mm, psize);
425 slice_print_mask(" good_mask", good_mask);
427 /* First check hint if it's valid or if we have MAP_FIXED */
428 if ((addr != 0 || fixed) && (mm->task_size - len) >= addr) {
430 /* Don't bother with hint if it overlaps a VMA */
431 if (!fixed && !slice_area_is_free(mm, addr, len))
434 /* Build a mask for the requested range */
435 mask = slice_range_to_mask(addr, len);
436 slice_print_mask(" mask", mask);
438 /* Check if we fit in the good mask. If we do, we just return,
441 if (slice_check_fit(mask, good_mask)) {
442 slice_dbg(" fits good !\n");
446 /* We don't fit in the good mask, check what other slices are
447 * empty and thus can be converted
449 potential_mask = slice_mask_for_free(mm);
450 potential_mask.low_slices |= good_mask.low_slices;
451 potential_mask.high_slices |= good_mask.high_slices;
453 slice_print_mask(" potential", potential_mask);
454 if (slice_check_fit(mask, potential_mask)) {
455 slice_dbg(" fits potential !\n");
460 /* If we have MAP_FIXED and failed the above step, then error out */
465 slice_dbg(" search...\n");
467 /* Now let's see if we can find something in the existing slices
470 addr = slice_find_area(mm, len, good_mask, psize, topdown, use_cache);
471 if (addr != -ENOMEM) {
472 /* Found within the good mask, we don't have to setup,
473 * we thus return directly
475 slice_dbg(" found area at 0x%lx\n", addr);
479 /* Won't fit, check what can be converted */
481 potential_mask = slice_mask_for_free(mm);
482 potential_mask.low_slices |= good_mask.low_slices;
483 potential_mask.high_slices |= good_mask.high_slices;
485 slice_print_mask(" potential", potential_mask);
488 /* Now let's see if we can find something in the existing slices
491 addr = slice_find_area(mm, len, potential_mask, psize, topdown,
496 mask = slice_range_to_mask(addr, len);
497 slice_dbg(" found potential area at 0x%lx\n", addr);
498 slice_print_mask(" mask", mask);
501 slice_convert(mm, mask, psize);
505 EXPORT_SYMBOL_GPL(slice_get_unmapped_area);
507 unsigned long arch_get_unmapped_area(struct file *filp,
513 return slice_get_unmapped_area(addr, len, flags,
514 current->mm->context.user_psize,
518 unsigned long arch_get_unmapped_area_topdown(struct file *filp,
519 const unsigned long addr0,
520 const unsigned long len,
521 const unsigned long pgoff,
522 const unsigned long flags)
524 return slice_get_unmapped_area(addr0, len, flags,
525 current->mm->context.user_psize,
529 unsigned int get_slice_psize(struct mm_struct *mm, unsigned long addr)
534 if (addr < SLICE_LOW_TOP) {
535 psizes = mm->context.low_slices_psize;
536 index = GET_LOW_SLICE_INDEX(addr);
538 psizes = mm->context.high_slices_psize;
539 index = GET_HIGH_SLICE_INDEX(addr);
542 return (psizes >> (index * 4)) & 0xf;
544 EXPORT_SYMBOL_GPL(get_slice_psize);
547 * This is called by hash_page when it needs to do a lazy conversion of
548 * an address space from real 64K pages to combo 4K pages (typically
549 * when hitting a non cacheable mapping on a processor or hypervisor
550 * that won't allow them for 64K pages).
552 * This is also called in init_new_context() to change back the user
553 * psize from whatever the parent context had it set to
554 * N.B. This may be called before mm->context.id has been set.
556 * This function will only change the content of the {low,high)_slice_psize
557 * masks, it will not flush SLBs as this shall be handled lazily by the
560 void slice_set_user_psize(struct mm_struct *mm, unsigned int psize)
562 unsigned long flags, lpsizes, hpsizes;
563 unsigned int old_psize;
566 slice_dbg("slice_set_user_psize(mm=%p, psize=%d)\n", mm, psize);
568 spin_lock_irqsave(&slice_convert_lock, flags);
570 old_psize = mm->context.user_psize;
571 slice_dbg(" old_psize=%d\n", old_psize);
572 if (old_psize == psize)
575 mm->context.user_psize = psize;
578 lpsizes = mm->context.low_slices_psize;
579 for (i = 0; i < SLICE_NUM_LOW; i++)
580 if (((lpsizes >> (i * 4)) & 0xf) == old_psize)
581 lpsizes = (lpsizes & ~(0xful << (i * 4))) |
582 (((unsigned long)psize) << (i * 4));
584 hpsizes = mm->context.high_slices_psize;
585 for (i = 0; i < SLICE_NUM_HIGH; i++)
586 if (((hpsizes >> (i * 4)) & 0xf) == old_psize)
587 hpsizes = (hpsizes & ~(0xful << (i * 4))) |
588 (((unsigned long)psize) << (i * 4));
590 mm->context.low_slices_psize = lpsizes;
591 mm->context.high_slices_psize = hpsizes;
593 slice_dbg(" lsps=%lx, hsps=%lx\n",
594 mm->context.low_slices_psize,
595 mm->context.high_slices_psize);
598 spin_unlock_irqrestore(&slice_convert_lock, flags);
602 * is_hugepage_only_range() is used by generic code to verify wether
603 * a normal mmap mapping (non hugetlbfs) is valid on a given area.
605 * until the generic code provides a more generic hook and/or starts
606 * calling arch get_unmapped_area for MAP_FIXED (which our implementation
607 * here knows how to deal with), we hijack it to keep standard mappings
610 * because of that generic code limitation, MAP_FIXED mapping cannot
611 * "convert" back a slice with no VMAs to the standard page size, only
612 * get_unmapped_area() can. It would be possible to fix it here but I
613 * prefer working on fixing the generic code instead.
615 * WARNING: This will not work if hugetlbfs isn't enabled since the
616 * generic code will redefine that function as 0 in that. This is ok
617 * for now as we only use slices with hugetlbfs enabled. This should
618 * be fixed as the generic code gets fixed.
620 int is_hugepage_only_range(struct mm_struct *mm, unsigned long addr,
623 struct slice_mask mask, available;
625 mask = slice_range_to_mask(addr, len);
626 available = slice_mask_for_size(mm, mm->context.user_psize);
628 #if 0 /* too verbose */
629 slice_dbg("is_hugepage_only_range(mm=%p, addr=%lx, len=%lx)\n",
631 slice_print_mask(" mask", mask);
632 slice_print_mask(" available", available);
634 return !slice_check_fit(mask, available);