[PATCH] mm: rmap with inner ptlock
Hugh Dickins [Sun, 30 Oct 2005 01:16:31 +0000 (18:16 -0700)]
rmap's page_check_address descend without page_table_lock.  First just
pte_offset_map in case there's no pte present worth locking for, then take
page_table_lock for the full check, and pass ptl back to caller in the same
style as pte_offset_map_lock.  __xip_unmap, page_referenced_one and
try_to_unmap_one use pte_unmap_unlock.  try_to_unmap_cluster also.

page_check_address reformatted to avoid progressive indentation.  No use is
made of its one error code, return NULL when it fails.

Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

include/linux/rmap.h
mm/filemap_xip.c
mm/rmap.c

index e80fb7e..35b30e6 100644 (file)
@@ -95,8 +95,8 @@ int try_to_unmap(struct page *);
 /*
  * Called from mm/filemap_xip.c to unmap empty zero page
  */
-pte_t *page_check_address(struct page *, struct mm_struct *, unsigned long);
-
+pte_t *page_check_address(struct page *, struct mm_struct *,
+                               unsigned long, spinlock_t **);
 
 /*
  * Used by swapoff to help locate where page is expected in vma.
index 4e74ad6..9cf687e 100644 (file)
@@ -174,6 +174,7 @@ __xip_unmap (struct address_space * mapping,
        unsigned long address;
        pte_t *pte;
        pte_t pteval;
+       spinlock_t *ptl;
        struct page *page;
 
        spin_lock(&mapping->i_mmap_lock);
@@ -183,20 +184,15 @@ __xip_unmap (struct address_space * mapping,
                        ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
                BUG_ON(address < vma->vm_start || address >= vma->vm_end);
                page = ZERO_PAGE(address);
-               /*
-                * We need the page_table_lock to protect us from page faults,
-                * munmap, fork, etc...
-                */
-               pte = page_check_address(page, mm, address);
-               if (!IS_ERR(pte)) {
+               pte = page_check_address(page, mm, address, &ptl);
+               if (pte) {
                        /* Nuke the page table entry. */
                        flush_cache_page(vma, address, pte_pfn(*pte));
                        pteval = ptep_clear_flush(vma, address, pte);
                        page_remove_rmap(page);
                        dec_mm_counter(mm, file_rss);
                        BUG_ON(pte_dirty(pteval));
-                       pte_unmap(pte);
-                       spin_unlock(&mm->page_table_lock);
+                       pte_unmap_unlock(pte, ptl);
                        page_cache_release(page);
                }
        }
index 4c52c56..a84bdfe 100644 (file)
--- a/mm/rmap.c
+++ b/mm/rmap.c
@@ -247,34 +247,41 @@ unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
  * On success returns with mapped pte and locked mm->page_table_lock.
  */
 pte_t *page_check_address(struct page *page, struct mm_struct *mm,
-                         unsigned long address)
+                         unsigned long address, spinlock_t **ptlp)
 {
        pgd_t *pgd;
        pud_t *pud;
        pmd_t *pmd;
        pte_t *pte;
+       spinlock_t *ptl;
 
-       /*
-        * We need the page_table_lock to protect us from page faults,
-        * munmap, fork, etc...
-        */
-       spin_lock(&mm->page_table_lock);
        pgd = pgd_offset(mm, address);
-       if (likely(pgd_present(*pgd))) {
-               pud = pud_offset(pgd, address);
-               if (likely(pud_present(*pud))) {
-                       pmd = pmd_offset(pud, address);
-                       if (likely(pmd_present(*pmd))) {
-                               pte = pte_offset_map(pmd, address);
-                               if (likely(pte_present(*pte) &&
-                                          page_to_pfn(page) == pte_pfn(*pte)))
-                                       return pte;
-                               pte_unmap(pte);
-                       }
-               }
+       if (!pgd_present(*pgd))
+               return NULL;
+
+       pud = pud_offset(pgd, address);
+       if (!pud_present(*pud))
+               return NULL;
+
+       pmd = pmd_offset(pud, address);
+       if (!pmd_present(*pmd))
+               return NULL;
+
+       pte = pte_offset_map(pmd, address);
+       /* Make a quick check before getting the lock */
+       if (!pte_present(*pte)) {
+               pte_unmap(pte);
+               return NULL;
        }
-       spin_unlock(&mm->page_table_lock);
-       return ERR_PTR(-ENOENT);
+
+       ptl = &mm->page_table_lock;
+       spin_lock(ptl);
+       if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
+               *ptlp = ptl;
+               return pte;
+       }
+       pte_unmap_unlock(pte, ptl);
+       return NULL;
 }
 
 /*
@@ -287,28 +294,28 @@ static int page_referenced_one(struct page *page,
        struct mm_struct *mm = vma->vm_mm;
        unsigned long address;
        pte_t *pte;
+       spinlock_t *ptl;
        int referenced = 0;
 
        address = vma_address(page, vma);
        if (address == -EFAULT)
                goto out;
 
-       pte = page_check_address(page, mm, address);
-       if (!IS_ERR(pte)) {
-               if (ptep_clear_flush_young(vma, address, pte))
-                       referenced++;
+       pte = page_check_address(page, mm, address, &ptl);
+       if (!pte)
+               goto out;
 
-               /* Pretend the page is referenced if the task has the
-                  swap token and is in the middle of a page fault. */
-               if (mm != current->mm && !ignore_token &&
-                               has_swap_token(mm) &&
-                               rwsem_is_locked(&mm->mmap_sem))
-                       referenced++;
+       if (ptep_clear_flush_young(vma, address, pte))
+               referenced++;
 
-               (*mapcount)--;
-               pte_unmap(pte);
-               spin_unlock(&mm->page_table_lock);
-       }
+       /* Pretend the page is referenced if the task has the
+          swap token and is in the middle of a page fault. */
+       if (mm != current->mm && !ignore_token && has_swap_token(mm) &&
+                       rwsem_is_locked(&mm->mmap_sem))
+               referenced++;
+
+       (*mapcount)--;
+       pte_unmap_unlock(pte, ptl);
 out:
        return referenced;
 }
@@ -507,14 +514,15 @@ static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma)
        unsigned long address;
        pte_t *pte;
        pte_t pteval;
+       spinlock_t *ptl;
        int ret = SWAP_AGAIN;
 
        address = vma_address(page, vma);
        if (address == -EFAULT)
                goto out;
 
-       pte = page_check_address(page, mm, address);
-       if (IS_ERR(pte))
+       pte = page_check_address(page, mm, address, &ptl);
+       if (!pte)
                goto out;
 
        /*
@@ -564,8 +572,7 @@ static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma)
        page_cache_release(page);
 
 out_unmap:
-       pte_unmap(pte);
-       spin_unlock(&mm->page_table_lock);
+       pte_unmap_unlock(pte, ptl);
 out:
        return ret;
 }
@@ -599,19 +606,14 @@ static void try_to_unmap_cluster(unsigned long cursor,
        pgd_t *pgd;
        pud_t *pud;
        pmd_t *pmd;
-       pte_t *pte, *original_pte;
+       pte_t *pte;
        pte_t pteval;
+       spinlock_t *ptl;
        struct page *page;
        unsigned long address;
        unsigned long end;
        unsigned long pfn;
 
-       /*
-        * We need the page_table_lock to protect us from page faults,
-        * munmap, fork, etc...
-        */
-       spin_lock(&mm->page_table_lock);
-
        address = (vma->vm_start + cursor) & CLUSTER_MASK;
        end = address + CLUSTER_SIZE;
        if (address < vma->vm_start)
@@ -621,22 +623,22 @@ static void try_to_unmap_cluster(unsigned long cursor,
 
        pgd = pgd_offset(mm, address);
        if (!pgd_present(*pgd))
-               goto out_unlock;
+               return;
 
        pud = pud_offset(pgd, address);
        if (!pud_present(*pud))
-               goto out_unlock;
+               return;
 
        pmd = pmd_offset(pud, address);
        if (!pmd_present(*pmd))
-               goto out_unlock;
+               return;
+
+       pte = pte_offset_map_lock(mm, pmd, address, &ptl);
 
        /* Update high watermark before we lower rss */
        update_hiwater_rss(mm);
 
-       for (original_pte = pte = pte_offset_map(pmd, address);
-                       address < end; pte++, address += PAGE_SIZE) {
-
+       for (; address < end; pte++, address += PAGE_SIZE) {
                if (!pte_present(*pte))
                        continue;
 
@@ -669,10 +671,7 @@ static void try_to_unmap_cluster(unsigned long cursor,
                dec_mm_counter(mm, file_rss);
                (*mapcount)--;
        }
-
-       pte_unmap(original_pte);
-out_unlock:
-       spin_unlock(&mm->page_table_lock);
+       pte_unmap_unlock(pte - 1, ptl);
 }
 
 static int try_to_unmap_anon(struct page *page)