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>
28 #include <linux/sysctl.h>
30 #define HUGEPGDIR_SHIFT (HPAGE_SHIFT + PAGE_SHIFT - 3)
31 #define HUGEPGDIR_SIZE (1UL << HUGEPGDIR_SHIFT)
32 #define HUGEPGDIR_MASK (~(HUGEPGDIR_SIZE-1))
34 #define HUGEPTE_INDEX_SIZE 9
35 #define HUGEPGD_INDEX_SIZE 10
37 #define PTRS_PER_HUGEPTE (1 << HUGEPTE_INDEX_SIZE)
38 #define PTRS_PER_HUGEPGD (1 << HUGEPGD_INDEX_SIZE)
40 static inline int hugepgd_index(unsigned long addr)
42 return (addr & ~REGION_MASK) >> HUGEPGDIR_SHIFT;
45 static pgd_t *hugepgd_offset(struct mm_struct *mm, unsigned long addr)
49 if (! mm->context.huge_pgdir)
53 index = hugepgd_index(addr);
54 BUG_ON(index >= PTRS_PER_HUGEPGD);
55 return mm->context.huge_pgdir + index;
58 static inline pte_t *hugepte_offset(pgd_t *dir, unsigned long addr)
65 index = (addr >> HPAGE_SHIFT) % PTRS_PER_HUGEPTE;
66 return (pte_t *)pgd_page(*dir) + index;
69 static pgd_t *hugepgd_alloc(struct mm_struct *mm, unsigned long addr)
71 BUG_ON(! in_hugepage_area(mm->context, addr));
73 if (! mm->context.huge_pgdir) {
75 spin_unlock(&mm->page_table_lock);
76 /* Don't use pgd_alloc(), because we want __GFP_REPEAT */
77 new = kmem_cache_alloc(zero_cache, GFP_KERNEL | __GFP_REPEAT);
78 BUG_ON(memcmp(new, empty_zero_page, PAGE_SIZE));
79 spin_lock(&mm->page_table_lock);
82 * Because we dropped the lock, we should re-check the
83 * entry, as somebody else could have populated it..
85 if (mm->context.huge_pgdir)
88 mm->context.huge_pgdir = new;
90 return hugepgd_offset(mm, addr);
93 static pte_t *hugepte_alloc(struct mm_struct *mm, pgd_t *dir,
96 if (! pgd_present(*dir)) {
99 spin_unlock(&mm->page_table_lock);
100 new = kmem_cache_alloc(zero_cache, GFP_KERNEL | __GFP_REPEAT);
101 BUG_ON(memcmp(new, empty_zero_page, PAGE_SIZE));
102 spin_lock(&mm->page_table_lock);
104 * Because we dropped the lock, we should re-check the
105 * entry, as somebody else could have populated it..
107 if (pgd_present(*dir)) {
109 kmem_cache_free(zero_cache, new);
111 struct page *ptepage;
115 ptepage = virt_to_page(new);
116 ptepage->mapping = (void *) mm;
117 ptepage->index = addr & HUGEPGDIR_MASK;
118 pgd_populate(mm, dir, new);
122 return hugepte_offset(dir, addr);
125 static pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
129 BUG_ON(! in_hugepage_area(mm->context, addr));
131 pgd = hugepgd_offset(mm, addr);
135 return hugepte_offset(pgd, addr);
138 static pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr)
142 BUG_ON(! in_hugepage_area(mm->context, addr));
144 pgd = hugepgd_alloc(mm, addr);
148 return hugepte_alloc(mm, pgd, addr);
151 static void set_huge_pte(struct mm_struct *mm, struct vm_area_struct *vma,
152 unsigned long addr, struct page *page,
153 pte_t *ptep, int write_access)
157 add_mm_counter(mm, rss, HPAGE_SIZE / PAGE_SIZE);
160 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
162 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
164 entry = pte_mkyoung(entry);
165 entry = pte_mkhuge(entry);
167 set_pte_at(mm, addr, ptep, entry);
171 * This function checks for proper alignment of input addr and len parameters.
173 int is_aligned_hugepage_range(unsigned long addr, unsigned long len)
175 if (len & ~HPAGE_MASK)
177 if (addr & ~HPAGE_MASK)
179 if (! (within_hugepage_low_range(addr, len)
180 || within_hugepage_high_range(addr, len)) )
185 static void flush_segments(void *parm)
187 u16 segs = (unsigned long) parm;
190 asm volatile("isync" : : : "memory");
192 for (i = 0; i < 16; i++) {
193 if (! (segs & (1U << i)))
195 asm volatile("slbie %0" : : "r" (i << SID_SHIFT));
198 asm volatile("isync" : : : "memory");
201 static int prepare_low_seg_for_htlb(struct mm_struct *mm, unsigned long seg)
203 unsigned long start = seg << SID_SHIFT;
204 unsigned long end = (seg+1) << SID_SHIFT;
205 struct vm_area_struct *vma;
207 struct mmu_gather *tlb;
211 /* Check no VMAs are in the region */
212 vma = find_vma(mm, start);
213 if (vma && (vma->vm_start < end))
216 /* Clean up any leftover PTE pages in the region */
217 spin_lock(&mm->page_table_lock);
218 tlb = tlb_gather_mmu(mm, 0);
219 for (addr = start; addr < end; addr += PMD_SIZE) {
220 pgd_t *pgd = pgd_offset(mm, addr);
228 pmd = pmd_offset(pgd, addr);
229 if (!pmd || pmd_none(*pmd))
236 pte = (pte_t *)pmd_page_kernel(*pmd);
237 /* No VMAs, so there should be no PTEs, check just in case. */
238 for (i = 0; i < PTRS_PER_PTE; i++) {
239 BUG_ON(!pte_none(*pte));
242 page = pmd_page(*pmd);
245 dec_page_state(nr_page_table_pages);
246 pte_free_tlb(tlb, page);
248 tlb_finish_mmu(tlb, start, end);
249 spin_unlock(&mm->page_table_lock);
254 static int open_low_hpage_segs(struct mm_struct *mm, u16 newsegs)
258 newsegs &= ~(mm->context.htlb_segs);
260 return 0; /* The segments we want are already open */
262 for (i = 0; i < 16; i++)
263 if ((1 << i) & newsegs)
264 if (prepare_low_seg_for_htlb(mm, i) != 0)
267 mm->context.htlb_segs |= newsegs;
269 /* update the paca copy of the context struct */
270 get_paca()->context = mm->context;
272 /* the context change must make it to memory before the flush,
273 * so that further SLB misses do the right thing. */
275 on_each_cpu(flush_segments, (void *)(unsigned long)newsegs, 0, 1);
280 int prepare_hugepage_range(unsigned long addr, unsigned long len)
282 if (within_hugepage_high_range(addr, len))
284 else if ((addr < 0x100000000UL) && ((addr+len) < 0x100000000UL)) {
286 /* Yes, we need both tests, in case addr+len overflows
287 * 64-bit arithmetic */
288 err = open_low_hpage_segs(current->mm,
289 LOW_ESID_MASK(addr, len));
291 printk(KERN_DEBUG "prepare_hugepage_range(%lx, %lx)"
292 " failed (segs: 0x%04hx)\n", addr, len,
293 LOW_ESID_MASK(addr, len));
300 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
301 struct vm_area_struct *vma)
303 pte_t *src_pte, *dst_pte, entry;
304 struct page *ptepage;
305 unsigned long addr = vma->vm_start;
306 unsigned long end = vma->vm_end;
310 dst_pte = huge_pte_alloc(dst, addr);
314 src_pte = huge_pte_offset(src, addr);
317 ptepage = pte_page(entry);
319 add_mm_counter(dst, rss, HPAGE_SIZE / PAGE_SIZE);
320 set_pte_at(dst, addr, dst_pte, entry);
331 follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
332 struct page **pages, struct vm_area_struct **vmas,
333 unsigned long *position, int *length, int i)
335 unsigned long vpfn, vaddr = *position;
336 int remainder = *length;
338 WARN_ON(!is_vm_hugetlb_page(vma));
340 vpfn = vaddr/PAGE_SIZE;
341 while (vaddr < vma->vm_end && remainder) {
346 pte = huge_pte_offset(mm, vaddr);
348 /* hugetlb should be locked, and hence, prefaulted */
349 WARN_ON(!pte || pte_none(*pte));
351 page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
353 WARN_ON(!PageCompound(page));
375 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
380 if (! in_hugepage_area(mm->context, address))
381 return ERR_PTR(-EINVAL);
383 ptep = huge_pte_offset(mm, address);
384 page = pte_page(*ptep);
386 page += (address % HPAGE_SIZE) / PAGE_SIZE;
391 int pmd_huge(pmd_t pmd)
397 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
398 pmd_t *pmd, int write)
404 void unmap_hugepage_range(struct vm_area_struct *vma,
405 unsigned long start, unsigned long end)
407 struct mm_struct *mm = vma->vm_mm;
412 WARN_ON(!is_vm_hugetlb_page(vma));
413 BUG_ON((start % HPAGE_SIZE) != 0);
414 BUG_ON((end % HPAGE_SIZE) != 0);
416 for (addr = start; addr < end; addr += HPAGE_SIZE) {
419 ptep = huge_pte_offset(mm, addr);
420 if (!ptep || pte_none(*ptep))
424 page = pte_page(pte);
425 pte_clear(mm, addr, ptep);
429 add_mm_counter(mm, rss, -((end - start) >> PAGE_SHIFT));
433 int hugetlb_prefault(struct address_space *mapping, struct vm_area_struct *vma)
435 struct mm_struct *mm = current->mm;
439 WARN_ON(!is_vm_hugetlb_page(vma));
440 BUG_ON((vma->vm_start % HPAGE_SIZE) != 0);
441 BUG_ON((vma->vm_end % HPAGE_SIZE) != 0);
443 spin_lock(&mm->page_table_lock);
444 for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
446 pte_t *pte = huge_pte_alloc(mm, addr);
453 if (! pte_none(*pte))
456 idx = ((addr - vma->vm_start) >> HPAGE_SHIFT)
457 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
458 page = find_get_page(mapping, idx);
460 /* charge the fs quota first */
461 if (hugetlb_get_quota(mapping)) {
465 page = alloc_huge_page();
467 hugetlb_put_quota(mapping);
471 ret = add_to_page_cache(page, mapping, idx, GFP_ATOMIC);
475 hugetlb_put_quota(mapping);
476 free_huge_page(page);
480 set_huge_pte(mm, vma, addr, page, pte, vma->vm_flags & VM_WRITE);
483 spin_unlock(&mm->page_table_lock);
487 /* Because we have an exclusive hugepage region which lies within the
488 * normal user address space, we have to take special measures to make
489 * non-huge mmap()s evade the hugepage reserved regions. */
490 unsigned long arch_get_unmapped_area(struct file *filp, unsigned long addr,
491 unsigned long len, unsigned long pgoff,
494 struct mm_struct *mm = current->mm;
495 struct vm_area_struct *vma;
496 unsigned long start_addr;
502 addr = PAGE_ALIGN(addr);
503 vma = find_vma(mm, addr);
504 if (((TASK_SIZE - len) >= addr)
505 && (!vma || (addr+len) <= vma->vm_start)
506 && !is_hugepage_only_range(mm, addr,len))
509 start_addr = addr = mm->free_area_cache;
512 vma = find_vma(mm, addr);
513 while (TASK_SIZE - len >= addr) {
514 BUG_ON(vma && (addr >= vma->vm_end));
516 if (touches_hugepage_low_range(mm, addr, len)) {
517 addr = ALIGN(addr+1, 1<<SID_SHIFT);
518 vma = find_vma(mm, addr);
521 if (touches_hugepage_high_range(addr, len)) {
522 addr = TASK_HPAGE_END;
523 vma = find_vma(mm, addr);
526 if (!vma || addr + len <= vma->vm_start) {
528 * Remember the place where we stopped the search:
530 mm->free_area_cache = addr + len;
537 /* Make sure we didn't miss any holes */
538 if (start_addr != TASK_UNMAPPED_BASE) {
539 start_addr = addr = TASK_UNMAPPED_BASE;
546 * This mmap-allocator allocates new areas top-down from below the
547 * stack's low limit (the base):
549 * Because we have an exclusive hugepage region which lies within the
550 * normal user address space, we have to take special measures to make
551 * non-huge mmap()s evade the hugepage reserved regions.
554 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
555 const unsigned long len, const unsigned long pgoff,
556 const unsigned long flags)
558 struct vm_area_struct *vma, *prev_vma;
559 struct mm_struct *mm = current->mm;
560 unsigned long base = mm->mmap_base, addr = addr0;
563 /* requested length too big for entire address space */
567 /* dont allow allocations above current base */
568 if (mm->free_area_cache > base)
569 mm->free_area_cache = base;
571 /* requesting a specific address */
573 addr = PAGE_ALIGN(addr);
574 vma = find_vma(mm, addr);
575 if (TASK_SIZE - len >= addr &&
576 (!vma || addr + len <= vma->vm_start)
577 && !is_hugepage_only_range(mm, addr,len))
582 /* make sure it can fit in the remaining address space */
583 if (mm->free_area_cache < len)
586 /* either no address requested or cant fit in requested address hole */
587 addr = (mm->free_area_cache - len) & PAGE_MASK;
590 if (touches_hugepage_low_range(mm, addr, len)) {
591 addr = (addr & ((~0) << SID_SHIFT)) - len;
592 goto hugepage_recheck;
593 } else if (touches_hugepage_high_range(addr, len)) {
594 addr = TASK_HPAGE_BASE - len;
598 * Lookup failure means no vma is above this address,
599 * i.e. return with success:
601 if (!(vma = find_vma_prev(mm, addr, &prev_vma)))
605 * new region fits between prev_vma->vm_end and
606 * vma->vm_start, use it:
608 if (addr+len <= vma->vm_start &&
609 (!prev_vma || (addr >= prev_vma->vm_end)))
610 /* remember the address as a hint for next time */
611 return (mm->free_area_cache = addr);
613 /* pull free_area_cache down to the first hole */
614 if (mm->free_area_cache == vma->vm_end)
615 mm->free_area_cache = vma->vm_start;
617 /* try just below the current vma->vm_start */
618 addr = vma->vm_start-len;
619 } while (len <= vma->vm_start);
623 * if hint left us with no space for the requested
624 * mapping then try again:
627 mm->free_area_cache = base;
632 * A failed mmap() very likely causes application failure,
633 * so fall back to the bottom-up function here. This scenario
634 * can happen with large stack limits and large mmap()
637 mm->free_area_cache = TASK_UNMAPPED_BASE;
638 addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
640 * Restore the topdown base:
642 mm->free_area_cache = base;
647 static unsigned long htlb_get_low_area(unsigned long len, u16 segmask)
649 unsigned long addr = 0;
650 struct vm_area_struct *vma;
652 vma = find_vma(current->mm, addr);
653 while (addr + len <= 0x100000000UL) {
654 BUG_ON(vma && (addr >= vma->vm_end)); /* invariant */
656 if (! __within_hugepage_low_range(addr, len, segmask)) {
657 addr = ALIGN(addr+1, 1<<SID_SHIFT);
658 vma = find_vma(current->mm, addr);
662 if (!vma || (addr + len) <= vma->vm_start)
664 addr = ALIGN(vma->vm_end, HPAGE_SIZE);
665 /* Depending on segmask this might not be a confirmed
666 * hugepage region, so the ALIGN could have skipped
668 vma = find_vma(current->mm, addr);
674 static unsigned long htlb_get_high_area(unsigned long len)
676 unsigned long addr = TASK_HPAGE_BASE;
677 struct vm_area_struct *vma;
679 vma = find_vma(current->mm, addr);
680 for (vma = find_vma(current->mm, addr);
681 addr + len <= TASK_HPAGE_END;
682 vma = vma->vm_next) {
683 BUG_ON(vma && (addr >= vma->vm_end)); /* invariant */
684 BUG_ON(! within_hugepage_high_range(addr, len));
686 if (!vma || (addr + len) <= vma->vm_start)
688 addr = ALIGN(vma->vm_end, HPAGE_SIZE);
689 /* Because we're in a hugepage region, this alignment
690 * should not skip us over any VMAs */
696 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
697 unsigned long len, unsigned long pgoff,
700 if (len & ~HPAGE_MASK)
703 if (!cpu_has_feature(CPU_FTR_16M_PAGE))
706 if (test_thread_flag(TIF_32BIT)) {
708 u16 segmask, cursegs = current->mm->context.htlb_segs;
710 /* First see if we can do the mapping in the existing
711 * low hpage segments */
712 addr = htlb_get_low_area(len, cursegs);
716 for (segmask = LOW_ESID_MASK(0x100000000UL-len, len);
717 ! lastshift; segmask >>=1) {
721 addr = htlb_get_low_area(len, cursegs | segmask);
722 if ((addr != -ENOMEM)
723 && open_low_hpage_segs(current->mm, segmask) == 0)
726 printk(KERN_DEBUG "hugetlb_get_unmapped_area() unable to open"
727 " enough segments\n");
730 return htlb_get_high_area(len);
734 void hugetlb_mm_free_pgd(struct mm_struct *mm)
739 spin_lock(&mm->page_table_lock);
741 pgdir = mm->context.huge_pgdir;
745 mm->context.huge_pgdir = NULL;
747 /* cleanup any hugepte pages leftover */
748 for (i = 0; i < PTRS_PER_HUGEPGD; i++) {
749 pgd_t *pgd = pgdir + i;
751 if (! pgd_none(*pgd)) {
752 pte_t *pte = (pte_t *)pgd_page(*pgd);
753 struct page *ptepage = virt_to_page(pte);
755 ptepage->mapping = NULL;
757 BUG_ON(memcmp(pte, empty_zero_page, PAGE_SIZE));
758 kmem_cache_free(zero_cache, pte);
763 BUG_ON(memcmp(pgdir, empty_zero_page, PAGE_SIZE));
764 kmem_cache_free(zero_cache, pgdir);
767 spin_unlock(&mm->page_table_lock);
770 int hash_huge_page(struct mm_struct *mm, unsigned long access,
771 unsigned long ea, unsigned long vsid, int local)
774 unsigned long va, vpn;
775 pte_t old_pte, new_pte;
776 unsigned long hpteflags, prpn;
780 spin_lock(&mm->page_table_lock);
782 ptep = huge_pte_offset(mm, ea);
784 /* Search the Linux page table for a match with va */
785 va = (vsid << 28) | (ea & 0x0fffffff);
786 vpn = va >> HPAGE_SHIFT;
789 * If no pte found or not present, send the problem up to
792 if (unlikely(!ptep || pte_none(*ptep)))
795 /* BUG_ON(pte_bad(*ptep)); */
798 * Check the user's access rights to the page. If access should be
799 * prevented then send the problem up to do_page_fault.
801 if (unlikely(access & ~pte_val(*ptep)))
804 * At this point, we have a pte (old_pte) which can be used to build
805 * or update an HPTE. There are 2 cases:
807 * 1. There is a valid (present) pte with no associated HPTE (this is
808 * the most common case)
809 * 2. There is a valid (present) pte with an associated HPTE. The
810 * current values of the pp bits in the HPTE prevent access
811 * because we are doing software DIRTY bit management and the
812 * page is currently not DIRTY.
819 hpteflags = 0x2 | (! (pte_val(new_pte) & _PAGE_RW));
820 /* _PAGE_EXEC -> HW_NO_EXEC since it's inverted */
821 hpteflags |= ((pte_val(new_pte) & _PAGE_EXEC) ? 0 : HW_NO_EXEC);
823 /* Check if pte already has an hpte (case 2) */
824 if (unlikely(pte_val(old_pte) & _PAGE_HASHPTE)) {
825 /* There MIGHT be an HPTE for this pte */
826 unsigned long hash, slot;
828 hash = hpt_hash(vpn, 1);
829 if (pte_val(old_pte) & _PAGE_SECONDARY)
831 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
832 slot += (pte_val(old_pte) & _PAGE_GROUP_IX) >> 12;
834 if (ppc_md.hpte_updatepp(slot, hpteflags, va, 1, local) == -1)
835 pte_val(old_pte) &= ~_PAGE_HPTEFLAGS;
838 if (likely(!(pte_val(old_pte) & _PAGE_HASHPTE))) {
839 unsigned long hash = hpt_hash(vpn, 1);
840 unsigned long hpte_group;
842 prpn = pte_pfn(old_pte);
845 hpte_group = ((hash & htab_hash_mask) *
846 HPTES_PER_GROUP) & ~0x7UL;
848 /* Update the linux pte with the HPTE slot */
849 pte_val(new_pte) &= ~_PAGE_HPTEFLAGS;
850 pte_val(new_pte) |= _PAGE_HASHPTE;
852 /* Add in WIMG bits */
853 /* XXX We should store these in the pte */
854 hpteflags |= _PAGE_COHERENT;
856 slot = ppc_md.hpte_insert(hpte_group, va, prpn, 0,
859 /* Primary is full, try the secondary */
860 if (unlikely(slot == -1)) {
861 pte_val(new_pte) |= _PAGE_SECONDARY;
862 hpte_group = ((~hash & htab_hash_mask) *
863 HPTES_PER_GROUP) & ~0x7UL;
864 slot = ppc_md.hpte_insert(hpte_group, va, prpn,
868 hpte_group = ((hash & htab_hash_mask) * HPTES_PER_GROUP) & ~0x7UL;
870 ppc_md.hpte_remove(hpte_group);
875 if (unlikely(slot == -2))
876 panic("hash_huge_page: pte_insert failed\n");
878 pte_val(new_pte) |= (slot<<12) & _PAGE_GROUP_IX;
881 * No need to use ldarx/stdcx here because all who
882 * might be updating the pte will hold the
891 spin_unlock(&mm->page_table_lock);