]>
git.proxmox.com Git - mirror_ubuntu-zesty-kernel.git/blob - arch/powerpc/mm/hugetlbpage.c
2 * PPC64 (POWER4) Huge TLB Page Support for Kernel.
4 * Copyright (C) 2003 David Gibson, IBM Corporation.
6 * Based on the IA-32 version:
7 * Copyright (C) 2002, Rohit Seth <rohit.seth@intel.com>
10 #include <linux/init.h>
13 #include <linux/hugetlb.h>
14 #include <linux/pagemap.h>
15 #include <linux/smp_lock.h>
16 #include <linux/slab.h>
17 #include <linux/err.h>
18 #include <linux/sysctl.h>
20 #include <asm/pgalloc.h>
22 #include <asm/tlbflush.h>
23 #include <asm/mmu_context.h>
24 #include <asm/machdep.h>
25 #include <asm/cputable.h>
29 #include <linux/sysctl.h>
31 #define NUM_LOW_AREAS (0x100000000UL >> SID_SHIFT)
32 #define NUM_HIGH_AREAS (PGTABLE_RANGE >> HTLB_AREA_SHIFT)
34 #ifdef CONFIG_PPC_64K_PAGES
35 #define HUGEPTE_INDEX_SIZE (PMD_SHIFT-HPAGE_SHIFT)
37 #define HUGEPTE_INDEX_SIZE (PUD_SHIFT-HPAGE_SHIFT)
39 #define PTRS_PER_HUGEPTE (1 << HUGEPTE_INDEX_SIZE)
40 #define HUGEPTE_TABLE_SIZE (sizeof(pte_t) << HUGEPTE_INDEX_SIZE)
42 #define HUGEPD_SHIFT (HPAGE_SHIFT + HUGEPTE_INDEX_SIZE)
43 #define HUGEPD_SIZE (1UL << HUGEPD_SHIFT)
44 #define HUGEPD_MASK (~(HUGEPD_SIZE-1))
46 #define huge_pgtable_cache (pgtable_cache[HUGEPTE_CACHE_NUM])
48 /* Flag to mark huge PD pointers. This means pmd_bad() and pud_bad()
49 * will choke on pointers to hugepte tables, which is handy for
50 * catching screwups early. */
53 typedef struct { unsigned long pd
; } hugepd_t
;
55 #define hugepd_none(hpd) ((hpd).pd == 0)
57 static inline pte_t
*hugepd_page(hugepd_t hpd
)
59 BUG_ON(!(hpd
.pd
& HUGEPD_OK
));
60 return (pte_t
*)(hpd
.pd
& ~HUGEPD_OK
);
63 static inline pte_t
*hugepte_offset(hugepd_t
*hpdp
, unsigned long addr
)
65 unsigned long idx
= ((addr
>> HPAGE_SHIFT
) & (PTRS_PER_HUGEPTE
-1));
66 pte_t
*dir
= hugepd_page(*hpdp
);
71 static int __hugepte_alloc(struct mm_struct
*mm
, hugepd_t
*hpdp
,
72 unsigned long address
)
74 pte_t
*new = kmem_cache_alloc(huge_pgtable_cache
,
75 GFP_KERNEL
|__GFP_REPEAT
);
80 spin_lock(&mm
->page_table_lock
);
81 if (!hugepd_none(*hpdp
))
82 kmem_cache_free(huge_pgtable_cache
, new);
84 hpdp
->pd
= (unsigned long)new | HUGEPD_OK
;
85 spin_unlock(&mm
->page_table_lock
);
89 /* Modelled after find_linux_pte() */
90 pte_t
*huge_pte_offset(struct mm_struct
*mm
, unsigned long addr
)
95 BUG_ON(! in_hugepage_area(mm
->context
, addr
));
99 pg
= pgd_offset(mm
, addr
);
100 if (!pgd_none(*pg
)) {
101 pu
= pud_offset(pg
, addr
);
102 if (!pud_none(*pu
)) {
103 #ifdef CONFIG_PPC_64K_PAGES
105 pm
= pmd_offset(pu
, addr
);
107 return hugepte_offset((hugepd_t
*)pm
, addr
);
109 return hugepte_offset((hugepd_t
*)pu
, addr
);
117 pte_t
*huge_pte_alloc(struct mm_struct
*mm
, unsigned long addr
)
121 hugepd_t
*hpdp
= NULL
;
123 BUG_ON(! in_hugepage_area(mm
->context
, addr
));
127 pg
= pgd_offset(mm
, addr
);
128 pu
= pud_alloc(mm
, pg
, addr
);
131 #ifdef CONFIG_PPC_64K_PAGES
133 pm
= pmd_alloc(mm
, pu
, addr
);
135 hpdp
= (hugepd_t
*)pm
;
137 hpdp
= (hugepd_t
*)pu
;
144 if (hugepd_none(*hpdp
) && __hugepte_alloc(mm
, hpdp
, addr
))
147 return hugepte_offset(hpdp
, addr
);
150 int huge_pmd_unshare(struct mm_struct
*mm
, unsigned long *addr
, pte_t
*ptep
)
155 static void free_hugepte_range(struct mmu_gather
*tlb
, hugepd_t
*hpdp
)
157 pte_t
*hugepte
= hugepd_page(*hpdp
);
161 pgtable_free_tlb(tlb
, pgtable_free_cache(hugepte
, HUGEPTE_CACHE_NUM
,
165 #ifdef CONFIG_PPC_64K_PAGES
166 static void hugetlb_free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
167 unsigned long addr
, unsigned long end
,
168 unsigned long floor
, unsigned long ceiling
)
175 pmd
= pmd_offset(pud
, addr
);
177 next
= pmd_addr_end(addr
, end
);
180 free_hugepte_range(tlb
, (hugepd_t
*)pmd
);
181 } while (pmd
++, addr
= next
, addr
!= end
);
191 if (end
- 1 > ceiling
- 1)
194 pmd
= pmd_offset(pud
, start
);
196 pmd_free_tlb(tlb
, pmd
);
200 static void hugetlb_free_pud_range(struct mmu_gather
*tlb
, pgd_t
*pgd
,
201 unsigned long addr
, unsigned long end
,
202 unsigned long floor
, unsigned long ceiling
)
209 pud
= pud_offset(pgd
, addr
);
211 next
= pud_addr_end(addr
, end
);
212 #ifdef CONFIG_PPC_64K_PAGES
213 if (pud_none_or_clear_bad(pud
))
215 hugetlb_free_pmd_range(tlb
, pud
, addr
, next
, floor
, ceiling
);
219 free_hugepte_range(tlb
, (hugepd_t
*)pud
);
221 } while (pud
++, addr
= next
, addr
!= end
);
227 ceiling
&= PGDIR_MASK
;
231 if (end
- 1 > ceiling
- 1)
234 pud
= pud_offset(pgd
, start
);
236 pud_free_tlb(tlb
, pud
);
240 * This function frees user-level page tables of a process.
242 * Must be called with pagetable lock held.
244 void hugetlb_free_pgd_range(struct mmu_gather
**tlb
,
245 unsigned long addr
, unsigned long end
,
246 unsigned long floor
, unsigned long ceiling
)
253 * Comments below take from the normal free_pgd_range(). They
254 * apply here too. The tests against HUGEPD_MASK below are
255 * essential, because we *don't* test for this at the bottom
256 * level. Without them we'll attempt to free a hugepte table
257 * when we unmap just part of it, even if there are other
258 * active mappings using it.
260 * The next few lines have given us lots of grief...
262 * Why are we testing HUGEPD* at this top level? Because
263 * often there will be no work to do at all, and we'd prefer
264 * not to go all the way down to the bottom just to discover
267 * Why all these "- 1"s? Because 0 represents both the bottom
268 * of the address space and the top of it (using -1 for the
269 * top wouldn't help much: the masks would do the wrong thing).
270 * The rule is that addr 0 and floor 0 refer to the bottom of
271 * the address space, but end 0 and ceiling 0 refer to the top
272 * Comparisons need to use "end - 1" and "ceiling - 1" (though
273 * that end 0 case should be mythical).
275 * Wherever addr is brought up or ceiling brought down, we
276 * must be careful to reject "the opposite 0" before it
277 * confuses the subsequent tests. But what about where end is
278 * brought down by HUGEPD_SIZE below? no, end can't go down to
281 * Whereas we round start (addr) and ceiling down, by different
282 * masks at different levels, in order to test whether a table
283 * now has no other vmas using it, so can be freed, we don't
284 * bother to round floor or end up - the tests don't need that.
294 ceiling
&= HUGEPD_MASK
;
298 if (end
- 1 > ceiling
- 1)
304 pgd
= pgd_offset((*tlb
)->mm
, addr
);
306 BUG_ON(! in_hugepage_area((*tlb
)->mm
->context
, addr
));
307 next
= pgd_addr_end(addr
, end
);
308 if (pgd_none_or_clear_bad(pgd
))
310 hugetlb_free_pud_range(*tlb
, pgd
, addr
, next
, floor
, ceiling
);
311 } while (pgd
++, addr
= next
, addr
!= end
);
314 void set_huge_pte_at(struct mm_struct
*mm
, unsigned long addr
,
315 pte_t
*ptep
, pte_t pte
)
317 if (pte_present(*ptep
)) {
318 /* We open-code pte_clear because we need to pass the right
319 * argument to hpte_need_flush (huge / !huge). Might not be
320 * necessary anymore if we make hpte_need_flush() get the
321 * page size from the slices
323 pte_update(mm
, addr
& HPAGE_MASK
, ptep
, ~0UL, 1);
325 *ptep
= __pte(pte_val(pte
) & ~_PAGE_HPTEFLAGS
);
328 pte_t
huge_ptep_get_and_clear(struct mm_struct
*mm
, unsigned long addr
,
331 unsigned long old
= pte_update(mm
, addr
, ptep
, ~0UL, 1);
335 struct slb_flush_info
{
336 struct mm_struct
*mm
;
340 static void flush_low_segments(void *parm
)
342 struct slb_flush_info
*fi
= parm
;
345 BUILD_BUG_ON((sizeof(fi
->newareas
)*8) != NUM_LOW_AREAS
);
347 if (current
->active_mm
!= fi
->mm
)
350 /* Only need to do anything if this CPU is working in the same
351 * mm as the one which has changed */
353 /* update the paca copy of the context struct */
354 get_paca()->context
= current
->active_mm
->context
;
356 asm volatile("isync" : : : "memory");
357 for (i
= 0; i
< NUM_LOW_AREAS
; i
++) {
358 if (! (fi
->newareas
& (1U << i
)))
360 asm volatile("slbie %0"
361 : : "r" ((i
<< SID_SHIFT
) | SLBIE_C
));
363 asm volatile("isync" : : : "memory");
366 static void flush_high_segments(void *parm
)
368 struct slb_flush_info
*fi
= parm
;
372 BUILD_BUG_ON((sizeof(fi
->newareas
)*8) != NUM_HIGH_AREAS
);
374 if (current
->active_mm
!= fi
->mm
)
377 /* Only need to do anything if this CPU is working in the same
378 * mm as the one which has changed */
380 /* update the paca copy of the context struct */
381 get_paca()->context
= current
->active_mm
->context
;
383 asm volatile("isync" : : : "memory");
384 for (i
= 0; i
< NUM_HIGH_AREAS
; i
++) {
385 if (! (fi
->newareas
& (1U << i
)))
387 for (j
= 0; j
< (1UL << (HTLB_AREA_SHIFT
-SID_SHIFT
)); j
++)
388 asm volatile("slbie %0"
389 :: "r" (((i
<< HTLB_AREA_SHIFT
)
390 + (j
<< SID_SHIFT
)) | SLBIE_C
));
392 asm volatile("isync" : : : "memory");
395 static int prepare_low_area_for_htlb(struct mm_struct
*mm
, unsigned long area
)
397 unsigned long start
= area
<< SID_SHIFT
;
398 unsigned long end
= (area
+1) << SID_SHIFT
;
399 struct vm_area_struct
*vma
;
401 BUG_ON(area
>= NUM_LOW_AREAS
);
403 /* Check no VMAs are in the region */
404 vma
= find_vma(mm
, start
);
405 if (vma
&& (vma
->vm_start
< end
))
411 static int prepare_high_area_for_htlb(struct mm_struct
*mm
, unsigned long area
)
413 unsigned long start
= area
<< HTLB_AREA_SHIFT
;
414 unsigned long end
= (area
+1) << HTLB_AREA_SHIFT
;
415 struct vm_area_struct
*vma
;
417 BUG_ON(area
>= NUM_HIGH_AREAS
);
419 /* Hack, so that each addresses is controlled by exactly one
420 * of the high or low area bitmaps, the first high area starts
423 start
= 0x100000000UL
;
425 /* Check no VMAs are in the region */
426 vma
= find_vma(mm
, start
);
427 if (vma
&& (vma
->vm_start
< end
))
433 static int open_low_hpage_areas(struct mm_struct
*mm
, u16 newareas
)
436 struct slb_flush_info fi
;
438 BUILD_BUG_ON((sizeof(newareas
)*8) != NUM_LOW_AREAS
);
439 BUILD_BUG_ON((sizeof(mm
->context
.low_htlb_areas
)*8) != NUM_LOW_AREAS
);
441 newareas
&= ~(mm
->context
.low_htlb_areas
);
443 return 0; /* The segments we want are already open */
445 for (i
= 0; i
< NUM_LOW_AREAS
; i
++)
446 if ((1 << i
) & newareas
)
447 if (prepare_low_area_for_htlb(mm
, i
) != 0)
450 mm
->context
.low_htlb_areas
|= newareas
;
452 /* the context change must make it to memory before the flush,
453 * so that further SLB misses do the right thing. */
457 fi
.newareas
= newareas
;
458 on_each_cpu(flush_low_segments
, &fi
, 0, 1);
463 static int open_high_hpage_areas(struct mm_struct
*mm
, u16 newareas
)
465 struct slb_flush_info fi
;
468 BUILD_BUG_ON((sizeof(newareas
)*8) != NUM_HIGH_AREAS
);
469 BUILD_BUG_ON((sizeof(mm
->context
.high_htlb_areas
)*8)
472 newareas
&= ~(mm
->context
.high_htlb_areas
);
474 return 0; /* The areas we want are already open */
476 for (i
= 0; i
< NUM_HIGH_AREAS
; i
++)
477 if ((1 << i
) & newareas
)
478 if (prepare_high_area_for_htlb(mm
, i
) != 0)
481 mm
->context
.high_htlb_areas
|= newareas
;
483 /* the context change must make it to memory before the flush,
484 * so that further SLB misses do the right thing. */
488 fi
.newareas
= newareas
;
489 on_each_cpu(flush_high_segments
, &fi
, 0, 1);
494 int prepare_hugepage_range(unsigned long addr
, unsigned long len
, pgoff_t pgoff
)
498 if (pgoff
& (~HPAGE_MASK
>> PAGE_SHIFT
))
500 if (len
& ~HPAGE_MASK
)
502 if (addr
& ~HPAGE_MASK
)
505 if (addr
< 0x100000000UL
)
506 err
= open_low_hpage_areas(current
->mm
,
507 LOW_ESID_MASK(addr
, len
));
508 if ((addr
+ len
) > 0x100000000UL
)
509 err
= open_high_hpage_areas(current
->mm
,
510 HTLB_AREA_MASK(addr
, len
));
511 #ifdef CONFIG_SPE_BASE
512 spu_flush_all_slbs(current
->mm
);
515 printk(KERN_DEBUG
"prepare_hugepage_range(%lx, %lx)"
516 " failed (lowmask: 0x%04hx, highmask: 0x%04hx)\n",
518 LOW_ESID_MASK(addr
, len
), HTLB_AREA_MASK(addr
, len
));
526 follow_huge_addr(struct mm_struct
*mm
, unsigned long address
, int write
)
531 if (! in_hugepage_area(mm
->context
, address
))
532 return ERR_PTR(-EINVAL
);
534 ptep
= huge_pte_offset(mm
, address
);
535 page
= pte_page(*ptep
);
537 page
+= (address
% HPAGE_SIZE
) / PAGE_SIZE
;
542 int pmd_huge(pmd_t pmd
)
548 follow_huge_pmd(struct mm_struct
*mm
, unsigned long address
,
549 pmd_t
*pmd
, int write
)
555 /* Because we have an exclusive hugepage region which lies within the
556 * normal user address space, we have to take special measures to make
557 * non-huge mmap()s evade the hugepage reserved regions. */
558 unsigned long arch_get_unmapped_area(struct file
*filp
, unsigned long addr
,
559 unsigned long len
, unsigned long pgoff
,
562 struct mm_struct
*mm
= current
->mm
;
563 struct vm_area_struct
*vma
;
564 unsigned long start_addr
;
570 addr
= PAGE_ALIGN(addr
);
571 vma
= find_vma(mm
, addr
);
572 if (((TASK_SIZE
- len
) >= addr
)
573 && (!vma
|| (addr
+len
) <= vma
->vm_start
)
574 && !is_hugepage_only_range(mm
, addr
,len
))
577 if (len
> mm
->cached_hole_size
) {
578 start_addr
= addr
= mm
->free_area_cache
;
580 start_addr
= addr
= TASK_UNMAPPED_BASE
;
581 mm
->cached_hole_size
= 0;
585 vma
= find_vma(mm
, addr
);
586 while (TASK_SIZE
- len
>= addr
) {
587 BUG_ON(vma
&& (addr
>= vma
->vm_end
));
589 if (touches_hugepage_low_range(mm
, addr
, len
)) {
590 addr
= ALIGN(addr
+1, 1<<SID_SHIFT
);
591 vma
= find_vma(mm
, addr
);
594 if (touches_hugepage_high_range(mm
, addr
, len
)) {
595 addr
= ALIGN(addr
+1, 1UL<<HTLB_AREA_SHIFT
);
596 vma
= find_vma(mm
, addr
);
599 if (!vma
|| addr
+ len
<= vma
->vm_start
) {
601 * Remember the place where we stopped the search:
603 mm
->free_area_cache
= addr
+ len
;
606 if (addr
+ mm
->cached_hole_size
< vma
->vm_start
)
607 mm
->cached_hole_size
= vma
->vm_start
- addr
;
612 /* Make sure we didn't miss any holes */
613 if (start_addr
!= TASK_UNMAPPED_BASE
) {
614 start_addr
= addr
= TASK_UNMAPPED_BASE
;
615 mm
->cached_hole_size
= 0;
622 * This mmap-allocator allocates new areas top-down from below the
623 * stack's low limit (the base):
625 * Because we have an exclusive hugepage region which lies within the
626 * normal user address space, we have to take special measures to make
627 * non-huge mmap()s evade the hugepage reserved regions.
630 arch_get_unmapped_area_topdown(struct file
*filp
, const unsigned long addr0
,
631 const unsigned long len
, const unsigned long pgoff
,
632 const unsigned long flags
)
634 struct vm_area_struct
*vma
, *prev_vma
;
635 struct mm_struct
*mm
= current
->mm
;
636 unsigned long base
= mm
->mmap_base
, addr
= addr0
;
637 unsigned long largest_hole
= mm
->cached_hole_size
;
640 /* requested length too big for entire address space */
644 /* dont allow allocations above current base */
645 if (mm
->free_area_cache
> base
)
646 mm
->free_area_cache
= base
;
648 /* requesting a specific address */
650 addr
= PAGE_ALIGN(addr
);
651 vma
= find_vma(mm
, addr
);
652 if (TASK_SIZE
- len
>= addr
&&
653 (!vma
|| addr
+ len
<= vma
->vm_start
)
654 && !is_hugepage_only_range(mm
, addr
,len
))
658 if (len
<= largest_hole
) {
660 mm
->free_area_cache
= base
;
663 /* make sure it can fit in the remaining address space */
664 if (mm
->free_area_cache
< len
)
667 /* either no address requested or cant fit in requested address hole */
668 addr
= (mm
->free_area_cache
- len
) & PAGE_MASK
;
671 if (touches_hugepage_low_range(mm
, addr
, len
)) {
672 addr
= (addr
& ((~0) << SID_SHIFT
)) - len
;
673 goto hugepage_recheck
;
674 } else if (touches_hugepage_high_range(mm
, addr
, len
)) {
675 addr
= (addr
& ((~0UL) << HTLB_AREA_SHIFT
)) - len
;
676 goto hugepage_recheck
;
680 * Lookup failure means no vma is above this address,
681 * i.e. return with success:
683 if (!(vma
= find_vma_prev(mm
, addr
, &prev_vma
)))
687 * new region fits between prev_vma->vm_end and
688 * vma->vm_start, use it:
690 if (addr
+len
<= vma
->vm_start
&&
691 (!prev_vma
|| (addr
>= prev_vma
->vm_end
))) {
692 /* remember the address as a hint for next time */
693 mm
->cached_hole_size
= largest_hole
;
694 return (mm
->free_area_cache
= addr
);
696 /* pull free_area_cache down to the first hole */
697 if (mm
->free_area_cache
== vma
->vm_end
) {
698 mm
->free_area_cache
= vma
->vm_start
;
699 mm
->cached_hole_size
= largest_hole
;
703 /* remember the largest hole we saw so far */
704 if (addr
+ largest_hole
< vma
->vm_start
)
705 largest_hole
= vma
->vm_start
- addr
;
707 /* try just below the current vma->vm_start */
708 addr
= vma
->vm_start
-len
;
709 } while (len
<= vma
->vm_start
);
713 * if hint left us with no space for the requested
714 * mapping then try again:
717 mm
->free_area_cache
= base
;
723 * A failed mmap() very likely causes application failure,
724 * so fall back to the bottom-up function here. This scenario
725 * can happen with large stack limits and large mmap()
728 mm
->free_area_cache
= TASK_UNMAPPED_BASE
;
729 mm
->cached_hole_size
= ~0UL;
730 addr
= arch_get_unmapped_area(filp
, addr0
, len
, pgoff
, flags
);
732 * Restore the topdown base:
734 mm
->free_area_cache
= base
;
735 mm
->cached_hole_size
= ~0UL;
740 static int htlb_check_hinted_area(unsigned long addr
, unsigned long len
)
742 struct vm_area_struct
*vma
;
744 vma
= find_vma(current
->mm
, addr
);
745 if (TASK_SIZE
- len
>= addr
&&
746 (!vma
|| ((addr
+ len
) <= vma
->vm_start
)))
752 static unsigned long htlb_get_low_area(unsigned long len
, u16 segmask
)
754 unsigned long addr
= 0;
755 struct vm_area_struct
*vma
;
757 vma
= find_vma(current
->mm
, addr
);
758 while (addr
+ len
<= 0x100000000UL
) {
759 BUG_ON(vma
&& (addr
>= vma
->vm_end
)); /* invariant */
761 if (! __within_hugepage_low_range(addr
, len
, segmask
)) {
762 addr
= ALIGN(addr
+1, 1<<SID_SHIFT
);
763 vma
= find_vma(current
->mm
, addr
);
767 if (!vma
|| (addr
+ len
) <= vma
->vm_start
)
769 addr
= ALIGN(vma
->vm_end
, HPAGE_SIZE
);
770 /* Depending on segmask this might not be a confirmed
771 * hugepage region, so the ALIGN could have skipped
773 vma
= find_vma(current
->mm
, addr
);
779 static unsigned long htlb_get_high_area(unsigned long len
, u16 areamask
)
781 unsigned long addr
= 0x100000000UL
;
782 struct vm_area_struct
*vma
;
784 vma
= find_vma(current
->mm
, addr
);
785 while (addr
+ len
<= TASK_SIZE_USER64
) {
786 BUG_ON(vma
&& (addr
>= vma
->vm_end
)); /* invariant */
788 if (! __within_hugepage_high_range(addr
, len
, areamask
)) {
789 addr
= ALIGN(addr
+1, 1UL<<HTLB_AREA_SHIFT
);
790 vma
= find_vma(current
->mm
, addr
);
794 if (!vma
|| (addr
+ len
) <= vma
->vm_start
)
796 addr
= ALIGN(vma
->vm_end
, HPAGE_SIZE
);
797 /* Depending on segmask this might not be a confirmed
798 * hugepage region, so the ALIGN could have skipped
800 vma
= find_vma(current
->mm
, addr
);
806 unsigned long hugetlb_get_unmapped_area(struct file
*file
, unsigned long addr
,
807 unsigned long len
, unsigned long pgoff
,
811 u16 areamask
, curareas
;
813 if (HPAGE_SHIFT
== 0)
815 if (len
& ~HPAGE_MASK
)
820 if (!cpu_has_feature(CPU_FTR_16M_PAGE
))
823 /* Paranoia, caller should have dealt with this */
824 BUG_ON((addr
+ len
) < addr
);
826 if (test_thread_flag(TIF_32BIT
)) {
827 curareas
= current
->mm
->context
.low_htlb_areas
;
829 /* First see if we can use the hint address */
830 if (addr
&& (htlb_check_hinted_area(addr
, len
) == 0)) {
831 areamask
= LOW_ESID_MASK(addr
, len
);
832 if (open_low_hpage_areas(current
->mm
, areamask
) == 0)
836 /* Next see if we can map in the existing low areas */
837 addr
= htlb_get_low_area(len
, curareas
);
841 /* Finally go looking for areas to open */
843 for (areamask
= LOW_ESID_MASK(0x100000000UL
-len
, len
);
844 ! lastshift
; areamask
>>=1) {
848 addr
= htlb_get_low_area(len
, curareas
| areamask
);
849 if ((addr
!= -ENOMEM
)
850 && open_low_hpage_areas(current
->mm
, areamask
) == 0)
854 curareas
= current
->mm
->context
.high_htlb_areas
;
856 /* First see if we can use the hint address */
857 /* We discourage 64-bit processes from doing hugepage
858 * mappings below 4GB (must use MAP_FIXED) */
859 if ((addr
>= 0x100000000UL
)
860 && (htlb_check_hinted_area(addr
, len
) == 0)) {
861 areamask
= HTLB_AREA_MASK(addr
, len
);
862 if (open_high_hpage_areas(current
->mm
, areamask
) == 0)
866 /* Next see if we can map in the existing high areas */
867 addr
= htlb_get_high_area(len
, curareas
);
871 /* Finally go looking for areas to open */
873 for (areamask
= HTLB_AREA_MASK(TASK_SIZE_USER64
-len
, len
);
874 ! lastshift
; areamask
>>=1) {
878 addr
= htlb_get_high_area(len
, curareas
| areamask
);
879 if ((addr
!= -ENOMEM
)
880 && open_high_hpage_areas(current
->mm
, areamask
) == 0)
884 printk(KERN_DEBUG
"hugetlb_get_unmapped_area() unable to open"
890 * Called by asm hashtable.S for doing lazy icache flush
892 static unsigned int hash_huge_page_do_lazy_icache(unsigned long rflags
,
898 if (!pfn_valid(pte_pfn(pte
)))
901 page
= pte_page(pte
);
904 if (!test_bit(PG_arch_1
, &page
->flags
) && !PageReserved(page
)) {
906 for (i
= 0; i
< (HPAGE_SIZE
/ PAGE_SIZE
); i
++)
907 __flush_dcache_icache(page_address(page
+i
));
908 set_bit(PG_arch_1
, &page
->flags
);
916 int hash_huge_page(struct mm_struct
*mm
, unsigned long access
,
917 unsigned long ea
, unsigned long vsid
, int local
,
921 unsigned long old_pte
, new_pte
;
922 unsigned long va
, rflags
, pa
;
926 ptep
= huge_pte_offset(mm
, ea
);
928 /* Search the Linux page table for a match with va */
929 va
= (vsid
<< 28) | (ea
& 0x0fffffff);
932 * If no pte found or not present, send the problem up to
935 if (unlikely(!ptep
|| pte_none(*ptep
)))
939 * Check the user's access rights to the page. If access should be
940 * prevented then send the problem up to do_page_fault.
942 if (unlikely(access
& ~pte_val(*ptep
)))
945 * At this point, we have a pte (old_pte) which can be used to build
946 * or update an HPTE. There are 2 cases:
948 * 1. There is a valid (present) pte with no associated HPTE (this is
949 * the most common case)
950 * 2. There is a valid (present) pte with an associated HPTE. The
951 * current values of the pp bits in the HPTE prevent access
952 * because we are doing software DIRTY bit management and the
953 * page is currently not DIRTY.
958 old_pte
= pte_val(*ptep
);
959 if (old_pte
& _PAGE_BUSY
)
961 new_pte
= old_pte
| _PAGE_BUSY
|
962 _PAGE_ACCESSED
| _PAGE_HASHPTE
;
963 } while(old_pte
!= __cmpxchg_u64((unsigned long *)ptep
,
966 rflags
= 0x2 | (!(new_pte
& _PAGE_RW
));
967 /* _PAGE_EXEC -> HW_NO_EXEC since it's inverted */
968 rflags
|= ((new_pte
& _PAGE_EXEC
) ? 0 : HPTE_R_N
);
969 if (!cpu_has_feature(CPU_FTR_COHERENT_ICACHE
))
970 /* No CPU has hugepages but lacks no execute, so we
971 * don't need to worry about that case */
972 rflags
= hash_huge_page_do_lazy_icache(rflags
, __pte(old_pte
),
975 /* Check if pte already has an hpte (case 2) */
976 if (unlikely(old_pte
& _PAGE_HASHPTE
)) {
977 /* There MIGHT be an HPTE for this pte */
978 unsigned long hash
, slot
;
980 hash
= hpt_hash(va
, HPAGE_SHIFT
);
981 if (old_pte
& _PAGE_F_SECOND
)
983 slot
= (hash
& htab_hash_mask
) * HPTES_PER_GROUP
;
984 slot
+= (old_pte
& _PAGE_F_GIX
) >> 12;
986 if (ppc_md
.hpte_updatepp(slot
, rflags
, va
, mmu_huge_psize
,
988 old_pte
&= ~_PAGE_HPTEFLAGS
;
991 if (likely(!(old_pte
& _PAGE_HASHPTE
))) {
992 unsigned long hash
= hpt_hash(va
, HPAGE_SHIFT
);
993 unsigned long hpte_group
;
995 pa
= pte_pfn(__pte(old_pte
)) << PAGE_SHIFT
;
998 hpte_group
= ((hash
& htab_hash_mask
) *
999 HPTES_PER_GROUP
) & ~0x7UL
;
1001 /* clear HPTE slot informations in new PTE */
1002 new_pte
= (new_pte
& ~_PAGE_HPTEFLAGS
) | _PAGE_HASHPTE
;
1004 /* Add in WIMG bits */
1005 /* XXX We should store these in the pte */
1006 /* --BenH: I think they are ... */
1007 rflags
|= _PAGE_COHERENT
;
1009 /* Insert into the hash table, primary slot */
1010 slot
= ppc_md
.hpte_insert(hpte_group
, va
, pa
, rflags
, 0,
1013 /* Primary is full, try the secondary */
1014 if (unlikely(slot
== -1)) {
1015 hpte_group
= ((~hash
& htab_hash_mask
) *
1016 HPTES_PER_GROUP
) & ~0x7UL
;
1017 slot
= ppc_md
.hpte_insert(hpte_group
, va
, pa
, rflags
,
1022 hpte_group
= ((hash
& htab_hash_mask
) *
1023 HPTES_PER_GROUP
)&~0x7UL
;
1025 ppc_md
.hpte_remove(hpte_group
);
1030 if (unlikely(slot
== -2))
1031 panic("hash_huge_page: pte_insert failed\n");
1033 new_pte
|= (slot
<< 12) & (_PAGE_F_SECOND
| _PAGE_F_GIX
);
1037 * No need to use ldarx/stdcx here
1039 *ptep
= __pte(new_pte
& ~_PAGE_BUSY
);
1047 static void zero_ctor(void *addr
, struct kmem_cache
*cache
, unsigned long flags
)
1049 memset(addr
, 0, kmem_cache_size(cache
));
1052 static int __init
hugetlbpage_init(void)
1054 if (!cpu_has_feature(CPU_FTR_16M_PAGE
))
1057 huge_pgtable_cache
= kmem_cache_create("hugepte_cache",
1060 SLAB_HWCACHE_ALIGN
|
1061 SLAB_MUST_HWCACHE_ALIGN
,
1063 if (! huge_pgtable_cache
)
1064 panic("hugetlbpage_init(): could not create hugepte cache\n");
1069 module_init(hugetlbpage_init
);