]> err.no Git - linux-2.6/blobdiff - arch/ppc64/mm/hugetlbpage.c
[PATCH] Avoiding mmap fragmentation
[linux-2.6] / arch / ppc64 / mm / hugetlbpage.c
index d3bf86a5c1ad40a745ded13bdf9d9b6aa80da0c8..fdcfe97c75c1dbc388cec0fe261ce446e82ab3e3 100644 (file)
@@ -121,7 +121,7 @@ static pte_t *hugepte_alloc(struct mm_struct *mm, pud_t *dir, unsigned long addr
        return hugepte_offset(dir, addr);
 }
 
-static pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
+pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
 {
        pud_t *pud;
 
@@ -134,7 +134,7 @@ static pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
        return hugepte_offset(pud, addr);
 }
 
-static pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr)
+pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr)
 {
        pud_t *pud;
 
@@ -147,25 +147,6 @@ static pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr)
        return hugepte_alloc(mm, pud, addr);
 }
 
-static void set_huge_pte(struct mm_struct *mm, struct vm_area_struct *vma,
-                        unsigned long addr, struct page *page,
-                        pte_t *ptep, int write_access)
-{
-       pte_t entry;
-
-       add_mm_counter(mm, rss, HPAGE_SIZE / PAGE_SIZE);
-       if (write_access) {
-               entry =
-                   pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
-       } else {
-               entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
-       }
-       entry = pte_mkyoung(entry);
-       entry = pte_mkhuge(entry);
-
-       set_pte_at(mm, addr, ptep, entry);
-}
-
 /*
  * This function checks for proper alignment of input addr and len parameters.
  */
@@ -259,80 +240,6 @@ int prepare_hugepage_range(unsigned long addr, unsigned long len)
        return -EINVAL;
 }
 
-int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
-                       struct vm_area_struct *vma)
-{
-       pte_t *src_pte, *dst_pte, entry;
-       struct page *ptepage;
-       unsigned long addr = vma->vm_start;
-       unsigned long end = vma->vm_end;
-       int err = -ENOMEM;
-
-       while (addr < end) {
-               dst_pte = huge_pte_alloc(dst, addr);
-               if (!dst_pte)
-                       goto out;
-
-               src_pte = huge_pte_offset(src, addr);
-               entry = *src_pte;
-               
-               ptepage = pte_page(entry);
-               get_page(ptepage);
-               add_mm_counter(dst, rss, HPAGE_SIZE / PAGE_SIZE);
-               set_pte_at(dst, addr, dst_pte, entry);
-
-               addr += HPAGE_SIZE;
-       }
-
-       err = 0;
- out:
-       return err;
-}
-
-int
-follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
-                   struct page **pages, struct vm_area_struct **vmas,
-                   unsigned long *position, int *length, int i)
-{
-       unsigned long vpfn, vaddr = *position;
-       int remainder = *length;
-
-       WARN_ON(!is_vm_hugetlb_page(vma));
-
-       vpfn = vaddr/PAGE_SIZE;
-       while (vaddr < vma->vm_end && remainder) {
-               if (pages) {
-                       pte_t *pte;
-                       struct page *page;
-
-                       pte = huge_pte_offset(mm, vaddr);
-
-                       /* hugetlb should be locked, and hence, prefaulted */
-                       WARN_ON(!pte || pte_none(*pte));
-
-                       page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
-
-                       WARN_ON(!PageCompound(page));
-
-                       get_page(page);
-                       pages[i] = page;
-               }
-
-               if (vmas)
-                       vmas[i] = vma;
-
-               vaddr += PAGE_SIZE;
-               ++vpfn;
-               --remainder;
-               ++i;
-       }
-
-       *length = remainder;
-       *position = vaddr;
-
-       return i;
-}
-
 struct page *
 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
 {
@@ -363,89 +270,6 @@ follow_huge_pmd(struct mm_struct *mm, unsigned long address,
        return NULL;
 }
 
-void unmap_hugepage_range(struct vm_area_struct *vma,
-                         unsigned long start, unsigned long end)
-{
-       struct mm_struct *mm = vma->vm_mm;
-       unsigned long addr;
-       pte_t *ptep;
-       struct page *page;
-
-       WARN_ON(!is_vm_hugetlb_page(vma));
-       BUG_ON((start % HPAGE_SIZE) != 0);
-       BUG_ON((end % HPAGE_SIZE) != 0);
-
-       for (addr = start; addr < end; addr += HPAGE_SIZE) {
-               pte_t pte;
-
-               ptep = huge_pte_offset(mm, addr);
-               if (!ptep || pte_none(*ptep))
-                       continue;
-
-               pte = *ptep;
-               page = pte_page(pte);
-               pte_clear(mm, addr, ptep);
-
-               put_page(page);
-       }
-       add_mm_counter(mm, rss, -((end - start) >> PAGE_SHIFT));
-       flush_tlb_pending();
-}
-
-int hugetlb_prefault(struct address_space *mapping, struct vm_area_struct *vma)
-{
-       struct mm_struct *mm = current->mm;
-       unsigned long addr;
-       int ret = 0;
-
-       WARN_ON(!is_vm_hugetlb_page(vma));
-       BUG_ON((vma->vm_start % HPAGE_SIZE) != 0);
-       BUG_ON((vma->vm_end % HPAGE_SIZE) != 0);
-
-       spin_lock(&mm->page_table_lock);
-       for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
-               unsigned long idx;
-               pte_t *pte = huge_pte_alloc(mm, addr);
-               struct page *page;
-
-               if (!pte) {
-                       ret = -ENOMEM;
-                       goto out;
-               }
-               if (! pte_none(*pte))
-                       continue;
-
-               idx = ((addr - vma->vm_start) >> HPAGE_SHIFT)
-                       + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
-               page = find_get_page(mapping, idx);
-               if (!page) {
-                       /* charge the fs quota first */
-                       if (hugetlb_get_quota(mapping)) {
-                               ret = -ENOMEM;
-                               goto out;
-                       }
-                       page = alloc_huge_page();
-                       if (!page) {
-                               hugetlb_put_quota(mapping);
-                               ret = -ENOMEM;
-                               goto out;
-                       }
-                       ret = add_to_page_cache(page, mapping, idx, GFP_ATOMIC);
-                       if (! ret) {
-                               unlock_page(page);
-                       } else {
-                               hugetlb_put_quota(mapping);
-                               free_huge_page(page);
-                               goto out;
-                       }
-               }
-               set_huge_pte(mm, vma, addr, page, pte, vma->vm_flags & VM_WRITE);
-       }
-out:
-       spin_unlock(&mm->page_table_lock);
-       return ret;
-}
-
 /* Because we have an exclusive hugepage region which lies within the
  * normal user address space, we have to take special measures to make
  * non-huge mmap()s evade the hugepage reserved regions. */
@@ -468,7 +292,12 @@ unsigned long arch_get_unmapped_area(struct file *filp, unsigned long addr,
                    && !is_hugepage_only_range(mm, addr,len))
                        return addr;
        }
-       start_addr = addr = mm->free_area_cache;
+       if (len > mm->cached_hole_size) {
+               start_addr = addr = mm->free_area_cache;
+       } else {
+               start_addr = addr = TASK_UNMAPPED_BASE;
+               mm->cached_hole_size = 0;
+       }
 
 full_search:
        vma = find_vma(mm, addr);
@@ -492,6 +321,8 @@ full_search:
                        mm->free_area_cache = addr + len;
                        return addr;
                }
+               if (addr + mm->cached_hole_size < vma->vm_start)
+                       mm->cached_hole_size = vma->vm_start - addr;
                addr = vma->vm_end;
                vma = vma->vm_next;
        }
@@ -499,6 +330,7 @@ full_search:
        /* Make sure we didn't miss any holes */
        if (start_addr != TASK_UNMAPPED_BASE) {
                start_addr = addr = TASK_UNMAPPED_BASE;
+               mm->cached_hole_size = 0;
                goto full_search;
        }
        return -ENOMEM;
@@ -520,6 +352,7 @@ arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
        struct vm_area_struct *vma, *prev_vma;
        struct mm_struct *mm = current->mm;
        unsigned long base = mm->mmap_base, addr = addr0;
+       unsigned long largest_hole = mm->cached_hole_size;
        int first_time = 1;
 
        /* requested length too big for entire address space */
@@ -540,6 +373,10 @@ arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
                        return addr;
        }
 
+       if (len <= largest_hole) {
+               largest_hole = 0;
+               mm->free_area_cache = base;
+       }
 try_again:
        /* make sure it can fit in the remaining address space */
        if (mm->free_area_cache < len)
@@ -568,13 +405,21 @@ hugepage_recheck:
                 * vma->vm_start, use it:
                 */
                if (addr+len <= vma->vm_start &&
-                               (!prev_vma || (addr >= prev_vma->vm_end)))
+                         (!prev_vma || (addr >= prev_vma->vm_end))) {
                        /* remember the address as a hint for next time */
-                       return (mm->free_area_cache = addr);
-               else
+                       mm->cached_hole_size = largest_hole;
+                       return (mm->free_area_cache = addr);
+               } else {
                        /* pull free_area_cache down to the first hole */
-                       if (mm->free_area_cache == vma->vm_end)
+                       if (mm->free_area_cache == vma->vm_end) {
                                mm->free_area_cache = vma->vm_start;
+                               mm->cached_hole_size = largest_hole;
+                       }
+               }
+
+               /* remember the largest hole we saw so far */
+               if (addr + largest_hole < vma->vm_start)
+                       largest_hole = vma->vm_start - addr;
 
                /* try just below the current vma->vm_start */
                addr = vma->vm_start-len;
@@ -587,6 +432,7 @@ fail:
         */
        if (first_time) {
                mm->free_area_cache = base;
+               largest_hole = 0;
                first_time = 0;
                goto try_again;
        }
@@ -597,11 +443,13 @@ fail:
         * allocations.
         */
        mm->free_area_cache = TASK_UNMAPPED_BASE;
+       mm->cached_hole_size = ~0UL;
        addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
        /*
         * Restore the topdown base:
         */
        mm->free_area_cache = base;
+       mm->cached_hole_size = ~0UL;
 
        return addr;
 }