[PATCH] mm: follow_page with inner ptlock
authorHugh Dickins <hugh@veritas.com>
Sun, 30 Oct 2005 01:16:33 +0000 (18:16 -0700)
committerLinus Torvalds <torvalds@g5.osdl.org>
Sun, 30 Oct 2005 04:40:41 +0000 (21:40 -0700)
Final step in pushing down common core's page_table_lock.  follow_page no
longer wants caller to hold page_table_lock, uses pte_offset_map_lock itself;
and so no page_table_lock is taken in get_user_pages itself.

But get_user_pages (and get_futex_key) do then need follow_page to pin the
page for them: take Daniel's suggestion of bitflags to follow_page.

Need one for WRITE, another for TOUCH (it was the accessed flag before:
vanished along with check_user_page_readable, but surely get_numa_maps is
wrong to mark every page it finds as accessed), another for GET.

And another, ANON to dispose of untouched_anonymous_page: it seems silly for
that to descend a second time, let follow_page observe if there was no page
table and return ZERO_PAGE if so.  Fix minor bug in that: check VM_LOCKED -
make_pages_present ought to make readonly anonymous present.

Give get_numa_maps a cond_resched while we're there.

Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
fs/proc/task_mmu.c
include/linux/mm.h
kernel/futex.c
mm/memory.c
mm/nommu.c

index 7e5e7ec..d2fa420 100644 (file)
@@ -419,7 +419,6 @@ static struct numa_maps *get_numa_maps(const struct vm_area_struct *vma)
        for_each_node(i)
                md->node[i] =0;
 
-       spin_lock(&mm->page_table_lock);
        for (vaddr = vma->vm_start; vaddr < vma->vm_end; vaddr += PAGE_SIZE) {
                page = follow_page(mm, vaddr, 0);
                if (page) {
@@ -434,8 +433,8 @@ static struct numa_maps *get_numa_maps(const struct vm_area_struct *vma)
                                md->anon++;
                        md->node[page_to_nid(page)]++;
                }
+               cond_resched();
        }
-       spin_unlock(&mm->page_table_lock);
        return md;
 }
 
index aa8de20..e8d1424 100644 (file)
@@ -938,14 +938,18 @@ static inline unsigned long vma_pages(struct vm_area_struct *vma)
        return (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
 }
 
-extern struct vm_area_struct *find_extend_vma(struct mm_struct *mm, unsigned long addr);
-
-extern struct page * vmalloc_to_page(void *addr);
-extern unsigned long vmalloc_to_pfn(void *addr);
-extern struct page * follow_page(struct mm_struct *mm, unsigned long address,
-               int write);
-int remap_pfn_range(struct vm_area_struct *, unsigned long,
-               unsigned long, unsigned long, pgprot_t);
+struct vm_area_struct *find_extend_vma(struct mm_struct *, unsigned long addr);
+struct page *vmalloc_to_page(void *addr);
+unsigned long vmalloc_to_pfn(void *addr);
+int remap_pfn_range(struct vm_area_struct *, unsigned long addr,
+                       unsigned long pfn, unsigned long size, pgprot_t);
+
+struct page *follow_page(struct mm_struct *, unsigned long address,
+                       unsigned int foll_flags);
+#define FOLL_WRITE     0x01    /* check pte is writable */
+#define FOLL_TOUCH     0x02    /* mark page accessed */
+#define FOLL_GET       0x04    /* do get_page on page */
+#define FOLL_ANON      0x08    /* give ZERO_PAGE if no pgtable */
 
 #ifdef CONFIG_PROC_FS
 void vm_stat_account(struct mm_struct *, unsigned long, struct file *, long);
index ca05fe6..3b4d5ad 100644 (file)
@@ -205,15 +205,13 @@ static int get_futex_key(unsigned long uaddr, union futex_key *key)
        /*
         * Do a quick atomic lookup first - this is the fastpath.
         */
-       spin_lock(&current->mm->page_table_lock);
-       page = follow_page(mm, uaddr, 0);
+       page = follow_page(mm, uaddr, FOLL_TOUCH|FOLL_GET);
        if (likely(page != NULL)) {
                key->shared.pgoff =
                        page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
-               spin_unlock(&current->mm->page_table_lock);
+               put_page(page);
                return 0;
        }
-       spin_unlock(&current->mm->page_table_lock);
 
        /*
         * Do it the general way.
index 51f7c0a..8461e2d 100644 (file)
@@ -807,86 +807,82 @@ unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
 
 /*
  * Do a quick page-table lookup for a single page.
- * mm->page_table_lock must be held.
  */
-struct page *follow_page(struct mm_struct *mm, unsigned long address, int write)
+struct page *follow_page(struct mm_struct *mm, unsigned long address,
+                       unsigned int flags)
 {
        pgd_t *pgd;
        pud_t *pud;
        pmd_t *pmd;
        pte_t *ptep, pte;
+       spinlock_t *ptl;
        unsigned long pfn;
        struct page *page;
 
-       page = follow_huge_addr(mm, address, write);
-       if (! IS_ERR(page))
-               return page;
+       page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
+       if (!IS_ERR(page)) {
+               BUG_ON(flags & FOLL_GET);
+               goto out;
+       }
 
+       page = NULL;
        pgd = pgd_offset(mm, address);
        if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
-               goto out;
+               goto no_page_table;
 
        pud = pud_offset(pgd, address);
        if (pud_none(*pud) || unlikely(pud_bad(*pud)))
-               goto out;
+               goto no_page_table;
        
        pmd = pmd_offset(pud, address);
        if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
+               goto no_page_table;
+
+       if (pmd_huge(*pmd)) {
+               BUG_ON(flags & FOLL_GET);
+               page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
                goto out;
-       if (pmd_huge(*pmd))
-               return follow_huge_pmd(mm, address, pmd, write);
+       }
 
-       ptep = pte_offset_map(pmd, address);
+       ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
        if (!ptep)
                goto out;
 
        pte = *ptep;
-       pte_unmap(ptep);
-       if (pte_present(pte)) {
-               if (write && !pte_write(pte))
-                       goto out;
-               pfn = pte_pfn(pte);
-               if (pfn_valid(pfn)) {
-                       page = pfn_to_page(pfn);
-                       if (write && !pte_dirty(pte) &&!PageDirty(page))
-                               set_page_dirty(page);
-                       mark_page_accessed(page);
-                       return page;
-               }
-       }
+       if (!pte_present(pte))
+               goto unlock;
+       if ((flags & FOLL_WRITE) && !pte_write(pte))
+               goto unlock;
+       pfn = pte_pfn(pte);
+       if (!pfn_valid(pfn))
+               goto unlock;
 
+       page = pfn_to_page(pfn);
+       if (flags & FOLL_GET)
+               get_page(page);
+       if (flags & FOLL_TOUCH) {
+               if ((flags & FOLL_WRITE) &&
+                   !pte_dirty(pte) && !PageDirty(page))
+                       set_page_dirty(page);
+               mark_page_accessed(page);
+       }
+unlock:
+       pte_unmap_unlock(ptep, ptl);
 out:
-       return NULL;
-}
-
-static inline int
-untouched_anonymous_page(struct mm_struct* mm, struct vm_area_struct *vma,
-                        unsigned long address)
-{
-       pgd_t *pgd;
-       pud_t *pud;
-       pmd_t *pmd;
-
-       /* Check if the vma is for an anonymous mapping. */
-       if (vma->vm_ops && vma->vm_ops->nopage)
-               return 0;
-
-       /* Check if page directory entry exists. */
-       pgd = pgd_offset(mm, address);
-       if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
-               return 1;
-
-       pud = pud_offset(pgd, address);
-       if (pud_none(*pud) || unlikely(pud_bad(*pud)))
-               return 1;
-
-       /* Check if page middle directory entry exists. */
-       pmd = pmd_offset(pud, address);
-       if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
-               return 1;
+       return page;
 
-       /* There is a pte slot for 'address' in 'mm'. */
-       return 0;
+no_page_table:
+       /*
+        * When core dumping an enormous anonymous area that nobody
+        * has touched so far, we don't want to allocate page tables.
+        */
+       if (flags & FOLL_ANON) {
+               page = ZERO_PAGE(address);
+               if (flags & FOLL_GET)
+                       get_page(page);
+               BUG_ON(flags & FOLL_WRITE);
+       }
+       return page;
 }
 
 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
@@ -894,18 +890,19 @@ int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
                struct page **pages, struct vm_area_struct **vmas)
 {
        int i;
-       unsigned int flags;
+       unsigned int vm_flags;
 
        /* 
         * Require read or write permissions.
         * If 'force' is set, we only require the "MAY" flags.
         */
-       flags = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
-       flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
+       vm_flags  = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
+       vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
        i = 0;
 
        do {
-               struct vm_area_struct * vma;
+               struct vm_area_struct *vma;
+               unsigned int foll_flags;
 
                vma = find_extend_vma(mm, start);
                if (!vma && in_gate_area(tsk, start)) {
@@ -946,7 +943,7 @@ int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
                }
 
                if (!vma || (vma->vm_flags & (VM_IO | VM_RESERVED))
-                               || !(flags & vma->vm_flags))
+                               || !(vm_flags & vma->vm_flags))
                        return i ? : -EFAULT;
 
                if (is_vm_hugetlb_page(vma)) {
@@ -954,29 +951,25 @@ int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
                                                &start, &len, i);
                        continue;
                }
-               spin_lock(&mm->page_table_lock);
+
+               foll_flags = FOLL_TOUCH;
+               if (pages)
+                       foll_flags |= FOLL_GET;
+               if (!write && !(vma->vm_flags & VM_LOCKED) &&
+                   (!vma->vm_ops || !vma->vm_ops->nopage))
+                       foll_flags |= FOLL_ANON;
+
                do {
-                       int write_access = write;
                        struct page *page;
 
-                       cond_resched_lock(&mm->page_table_lock);
-                       while (!(page = follow_page(mm, start, write_access))) {
-                               int ret;
-
-                               /*
-                                * Shortcut for anonymous pages. We don't want
-                                * to force the creation of pages tables for
-                                * insanely big anonymously mapped areas that
-                                * nobody touched so far. This is important
-                                * for doing a core dump for these mappings.
-                                */
-                               if (!write && untouched_anonymous_page(mm,vma,start)) {
-                                       page = ZERO_PAGE(start);
-                                       break;
-                               }
-                               spin_unlock(&mm->page_table_lock);
-                               ret = __handle_mm_fault(mm, vma, start, write_access);
+                       if (write)
+                               foll_flags |= FOLL_WRITE;
 
+                       cond_resched();
+                       while (!(page = follow_page(mm, start, foll_flags))) {
+                               int ret;
+                               ret = __handle_mm_fault(mm, vma, start,
+                                               foll_flags & FOLL_WRITE);
                                /*
                                 * The VM_FAULT_WRITE bit tells us that do_wp_page has
                                 * broken COW when necessary, even if maybe_mkwrite
@@ -984,7 +977,7 @@ int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
                                 * subsequent page lookups as if they were reads.
                                 */
                                if (ret & VM_FAULT_WRITE)
-                                       write_access = 0;
+                                       foll_flags &= ~FOLL_WRITE;
                                
                                switch (ret & ~VM_FAULT_WRITE) {
                                case VM_FAULT_MINOR:
@@ -1000,12 +993,10 @@ int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
                                default:
                                        BUG();
                                }
-                               spin_lock(&mm->page_table_lock);
                        }
                        if (pages) {
                                pages[i] = page;
                                flush_dcache_page(page);
-                               page_cache_get(page);
                        }
                        if (vmas)
                                vmas[i] = vma;
@@ -1013,7 +1004,6 @@ int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
                        start += PAGE_SIZE;
                        len--;
                } while (len && start < vma->vm_end);
-               spin_unlock(&mm->page_table_lock);
        } while (len);
        return i;
 }
index dfb124f..d1e076a 100644 (file)
@@ -1049,7 +1049,8 @@ struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
 
 EXPORT_SYMBOL(find_vma);
 
-struct page * follow_page(struct mm_struct *mm, unsigned long addr, int write)
+struct page *follow_page(struct mm_struct *mm, unsigned long address,
+                       unsigned int foll_flags)
 {
        return NULL;
 }