[PATCH] mm: page fault handlers tidyup
authorHugh Dickins <hugh@veritas.com>
Sun, 30 Oct 2005 01:15:59 +0000 (18:15 -0700)
committerLinus Torvalds <torvalds@g5.osdl.org>
Sun, 30 Oct 2005 04:40:37 +0000 (21:40 -0700)
Impose a little more consistency on the page fault handlers do_wp_page,
do_swap_page, do_anonymous_page, do_no_page, do_file_page: why not pass their
arguments in the same order, called the same names?

break_cow is all very well, but what it did was inlined elsewhere: easier to
compare if it's brought back into do_wp_page.

do_file_page's fallback to do_no_page dates from a time when we were testing
pte_file by using it wherever possible: currently it's peculiar to nonlinear
vmas, so just check that.  BUG_ON if not?  Better not, it's probably page
table corruption, so just show the pte: hmm, there's a pte_ERROR macro, let's
use that for do_wp_page's invalid pfn too.

Hah!  Someone in the ppc64 world noticed pte_ERROR was unused so removed it:
restored (and say "pud" not "pmd" in its pud_ERROR).

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/asm-ppc64/pgtable.h
mm/filemap.c
mm/memory.c
mm/shmem.c

index c83679c..2eb1778 100644 (file)
@@ -478,10 +478,12 @@ extern pgprot_t phys_mem_access_prot(struct file *file, unsigned long addr,
 #define __HAVE_ARCH_PTE_SAME
 #define pte_same(A,B)  (((pte_val(A) ^ pte_val(B)) & ~_PAGE_HPTEFLAGS) == 0)
 
+#define pte_ERROR(e) \
+       printk("%s:%d: bad pte %08lx.\n", __FILE__, __LINE__, pte_val(e))
 #define pmd_ERROR(e) \
        printk("%s:%d: bad pmd %08lx.\n", __FILE__, __LINE__, pmd_val(e))
 #define pud_ERROR(e) \
-       printk("%s:%d: bad pmd %08lx.\n", __FILE__, __LINE__, pud_val(e))
+       printk("%s:%d: bad pud %08lx.\n", __FILE__, __LINE__, pud_val(e))
 #define pgd_ERROR(e) \
        printk("%s:%d: bad pgd %08lx.\n", __FILE__, __LINE__, pgd_val(e))
 
index 1c31b2f..8aa344e 100644 (file)
@@ -1520,7 +1520,7 @@ repeat:
                        page_cache_release(page);
                        return err;
                }
-       } else {
+       } else if (vma->vm_flags & VM_NONLINEAR) {
                /* No page was found just because we can't read it in now (being
                 * here implies nonblock != 0), but the page may exist, so set
                 * the PTE to fault it in later. */
index 1366768..eaf7903 100644 (file)
@@ -1213,28 +1213,10 @@ static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
 }
 
 /*
- * We hold the mm semaphore for reading and vma->vm_mm->page_table_lock
- */
-static inline void break_cow(struct vm_area_struct * vma, struct page * new_page, unsigned long address, 
-               pte_t *page_table)
-{
-       pte_t entry;
-
-       entry = maybe_mkwrite(pte_mkdirty(mk_pte(new_page, vma->vm_page_prot)),
-                             vma);
-       ptep_establish(vma, address, page_table, entry);
-       update_mmu_cache(vma, address, entry);
-       lazy_mmu_prot_update(entry);
-}
-
-/*
  * This routine handles present pages, when users try to write
  * to a shared page. It is done by copying the page to a new address
  * and decrementing the shared-page counter for the old page.
  *
- * Goto-purists beware: the only reason for goto's here is that it results
- * in better assembly code.. The "default" path will see no jumps at all.
- *
  * Note that this routine assumes that the protection checks have been
  * done by the caller (the low-level page fault routine in most cases).
  * Thus we can safely just mark it writable once we've done any necessary
@@ -1247,25 +1229,22 @@ static inline void break_cow(struct vm_area_struct * vma, struct page * new_page
  * We hold the mm semaphore and the page_table_lock on entry and exit
  * with the page_table_lock released.
  */
-static int do_wp_page(struct mm_struct *mm, struct vm_area_struct * vma,
-       unsigned long address, pte_t *page_table, pmd_t *pmd, pte_t pte)
+static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
+               unsigned long address, pte_t *page_table, pmd_t *pmd,
+               pte_t orig_pte)
 {
        struct page *old_page, *new_page;
-       unsigned long pfn = pte_pfn(pte);
+       unsigned long pfn = pte_pfn(orig_pte);
        pte_t entry;
-       int ret;
+       int ret = VM_FAULT_MINOR;
 
        if (unlikely(!pfn_valid(pfn))) {
                /*
-                * This should really halt the system so it can be debugged or
-                * at least the kernel stops what it's doing before it corrupts
-                * data, but for the moment just pretend this is OOM.
+                * Page table corrupted: show pte and kill process.
                 */
-               pte_unmap(page_table);
-               printk(KERN_ERR "do_wp_page: bogus page at address %08lx\n",
-                               address);
-               spin_unlock(&mm->page_table_lock);
-               return VM_FAULT_OOM;
+               pte_ERROR(orig_pte);
+               ret = VM_FAULT_OOM;
+               goto unlock;
        }
        old_page = pfn_to_page(pfn);
 
@@ -1274,52 +1253,57 @@ static int do_wp_page(struct mm_struct *mm, struct vm_area_struct * vma,
                unlock_page(old_page);
                if (reuse) {
                        flush_cache_page(vma, address, pfn);
-                       entry = maybe_mkwrite(pte_mkyoung(pte_mkdirty(pte)),
-                                             vma);
+                       entry = pte_mkyoung(orig_pte);
+                       entry = maybe_mkwrite(pte_mkdirty(entry), vma);
                        ptep_set_access_flags(vma, address, page_table, entry, 1);
                        update_mmu_cache(vma, address, entry);
                        lazy_mmu_prot_update(entry);
-                       pte_unmap(page_table);
-                       spin_unlock(&mm->page_table_lock);
-                       return VM_FAULT_MINOR|VM_FAULT_WRITE;
+                       ret |= VM_FAULT_WRITE;
+                       goto unlock;
                }
        }
-       pte_unmap(page_table);
 
        /*
         * Ok, we need to copy. Oh, well..
         */
        if (!PageReserved(old_page))
                page_cache_get(old_page);
+       pte_unmap(page_table);
        spin_unlock(&mm->page_table_lock);
 
        if (unlikely(anon_vma_prepare(vma)))
-               goto no_new_page;
+               goto oom;
        if (old_page == ZERO_PAGE(address)) {
                new_page = alloc_zeroed_user_highpage(vma, address);
                if (!new_page)
-                       goto no_new_page;
+                       goto oom;
        } else {
                new_page = alloc_page_vma(GFP_HIGHUSER, vma, address);
                if (!new_page)
-                       goto no_new_page;
+                       goto oom;
                copy_user_highpage(new_page, old_page, address);
        }
+
        /*
         * Re-check the pte - we dropped the lock
         */
-       ret = VM_FAULT_MINOR;
        spin_lock(&mm->page_table_lock);
        page_table = pte_offset_map(pmd, address);
-       if (likely(pte_same(*page_table, pte))) {
+       if (likely(pte_same(*page_table, orig_pte))) {
                if (PageAnon(old_page))
                        dec_mm_counter(mm, anon_rss);
                if (PageReserved(old_page))
                        inc_mm_counter(mm, rss);
                else
                        page_remove_rmap(old_page);
+
                flush_cache_page(vma, address, pfn);
-               break_cow(vma, new_page, address, page_table);
+               entry = mk_pte(new_page, vma->vm_page_prot);
+               entry = maybe_mkwrite(pte_mkdirty(entry), vma);
+               ptep_establish(vma, address, page_table, entry);
+               update_mmu_cache(vma, address, entry);
+               lazy_mmu_prot_update(entry);
+
                lru_cache_add_active(new_page);
                page_add_anon_rmap(new_page, vma, address);
 
@@ -1327,13 +1311,13 @@ static int do_wp_page(struct mm_struct *mm, struct vm_area_struct * vma,
                new_page = old_page;
                ret |= VM_FAULT_WRITE;
        }
-       pte_unmap(page_table);
        page_cache_release(new_page);
        page_cache_release(old_page);
+unlock:
+       pte_unmap(page_table);
        spin_unlock(&mm->page_table_lock);
        return ret;
-
-no_new_page:
+oom:
        page_cache_release(old_page);
        return VM_FAULT_OOM;
 }
@@ -1661,17 +1645,19 @@ void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struc
  * We hold the mm semaphore and the page_table_lock on entry and
  * should release the pagetable lock on exit..
  */
-static int do_swap_page(struct mm_struct * mm,
-       struct vm_area_struct * vma, unsigned long address,
-       pte_t *page_table, pmd_t *pmd, pte_t orig_pte, int write_access)
+static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
+               unsigned long address, pte_t *page_table, pmd_t *pmd,
+               int write_access, pte_t orig_pte)
 {
        struct page *page;
-       swp_entry_t entry = pte_to_swp_entry(orig_pte);
+       swp_entry_t entry;
        pte_t pte;
        int ret = VM_FAULT_MINOR;
 
        pte_unmap(page_table);
        spin_unlock(&mm->page_table_lock);
+
+       entry = pte_to_swp_entry(orig_pte);
        page = lookup_swap_cache(entry);
        if (!page) {
                swapin_readahead(entry, address, vma);
@@ -1685,11 +1671,7 @@ static int do_swap_page(struct mm_struct * mm,
                        page_table = pte_offset_map(pmd, address);
                        if (likely(pte_same(*page_table, orig_pte)))
                                ret = VM_FAULT_OOM;
-                       else
-                               ret = VM_FAULT_MINOR;
-                       pte_unmap(page_table);
-                       spin_unlock(&mm->page_table_lock);
-                       goto out;
+                       goto unlock;
                }
 
                /* Had to read the page from swap area: Major fault */
@@ -1745,6 +1727,7 @@ static int do_swap_page(struct mm_struct * mm,
        /* No need to invalidate - it was non-present before */
        update_mmu_cache(vma, address, pte);
        lazy_mmu_prot_update(pte);
+unlock:
        pte_unmap(page_table);
        spin_unlock(&mm->page_table_lock);
 out:
@@ -1754,7 +1737,7 @@ out_nomap:
        spin_unlock(&mm->page_table_lock);
        unlock_page(page);
        page_cache_release(page);
-       goto out;
+       return ret;
 }
 
 /*
@@ -1762,17 +1745,15 @@ out_nomap:
  * spinlock held to protect against concurrent faults in
  * multithreaded programs. 
  */
-static int
-do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
-               pte_t *page_table, pmd_t *pmd, int write_access,
-               unsigned long addr)
+static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
+               unsigned long address, pte_t *page_table, pmd_t *pmd,
+               int write_access)
 {
        pte_t entry;
 
        /* Mapping of ZERO_PAGE - vm_page_prot is readonly */
        entry = mk_pte(ZERO_PAGE(addr), vma->vm_page_prot);
 
-       /* ..except if it's a write access */
        if (write_access) {
                struct page *page;
 
@@ -1781,39 +1762,36 @@ do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
                spin_unlock(&mm->page_table_lock);
 
                if (unlikely(anon_vma_prepare(vma)))
-                       goto no_mem;
-               page = alloc_zeroed_user_highpage(vma, addr);
+                       goto oom;
+               page = alloc_zeroed_user_highpage(vma, address);
                if (!page)
-                       goto no_mem;
+                       goto oom;
 
                spin_lock(&mm->page_table_lock);
-               page_table = pte_offset_map(pmd, addr);
+               page_table = pte_offset_map(pmd, address);
 
                if (!pte_none(*page_table)) {
-                       pte_unmap(page_table);
                        page_cache_release(page);
-                       spin_unlock(&mm->page_table_lock);
-                       goto out;
+                       goto unlock;
                }
                inc_mm_counter(mm, rss);
-               entry = maybe_mkwrite(pte_mkdirty(mk_pte(page,
-                                                        vma->vm_page_prot)),
-                                     vma);
+               entry = mk_pte(page, vma->vm_page_prot);
+               entry = maybe_mkwrite(pte_mkdirty(entry), vma);
                lru_cache_add_active(page);
                SetPageReferenced(page);
-               page_add_anon_rmap(page, vma, addr);
+               page_add_anon_rmap(page, vma, address);
        }
 
-       set_pte_at(mm, addr, page_table, entry);
-       pte_unmap(page_table);
+       set_pte_at(mm, address, page_table, entry);
 
        /* No need to invalidate - it was non-present before */
-       update_mmu_cache(vma, addr, entry);
+       update_mmu_cache(vma, address, entry);
        lazy_mmu_prot_update(entry);
+unlock:
+       pte_unmap(page_table);
        spin_unlock(&mm->page_table_lock);
-out:
        return VM_FAULT_MINOR;
-no_mem:
+oom:
        return VM_FAULT_OOM;
 }
 
@@ -1829,20 +1807,17 @@ no_mem:
  * This is called with the MM semaphore held and the page table
  * spinlock held. Exit with the spinlock released.
  */
-static int
-do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
-       unsigned long address, int write_access, pte_t *page_table, pmd_t *pmd)
+static int do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
+               unsigned long address, pte_t *page_table, pmd_t *pmd,
+               int write_access)
 {
-       struct page * new_page;
+       struct page *new_page;
        struct address_space *mapping = NULL;
        pte_t entry;
        unsigned int sequence = 0;
        int ret = VM_FAULT_MINOR;
        int anon = 0;
 
-       if (!vma->vm_ops || !vma->vm_ops->nopage)
-               return do_anonymous_page(mm, vma, page_table,
-                                       pmd, write_access, address);
        pte_unmap(page_table);
        spin_unlock(&mm->page_table_lock);
 
@@ -1852,7 +1827,6 @@ do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
                smp_rmb(); /* serializes i_size against truncate_count */
        }
 retry:
-       cond_resched();
        new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
        /*
         * No smp_rmb is needed here as long as there's a full
@@ -1892,9 +1866,11 @@ retry:
         * retry getting the page.
         */
        if (mapping && unlikely(sequence != mapping->truncate_count)) {
-               sequence = mapping->truncate_count;
                spin_unlock(&mm->page_table_lock);
                page_cache_release(new_page);
+               cond_resched();
+               sequence = mapping->truncate_count;
+               smp_rmb();
                goto retry;
        }
        page_table = pte_offset_map(pmd, address);
@@ -1924,25 +1900,22 @@ retry:
                        page_add_anon_rmap(new_page, vma, address);
                } else
                        page_add_file_rmap(new_page);
-               pte_unmap(page_table);
        } else {
                /* One of our sibling threads was faster, back out. */
-               pte_unmap(page_table);
                page_cache_release(new_page);
-               spin_unlock(&mm->page_table_lock);
-               goto out;
+               goto unlock;
        }
 
        /* no need to invalidate: a not-present page shouldn't be cached */
        update_mmu_cache(vma, address, entry);
        lazy_mmu_prot_update(entry);
+unlock:
+       pte_unmap(page_table);
        spin_unlock(&mm->page_table_lock);
-out:
        return ret;
 oom:
        page_cache_release(new_page);
-       ret = VM_FAULT_OOM;
-       goto out;
+       return VM_FAULT_OOM;
 }
 
 /*
@@ -1950,29 +1923,28 @@ oom:
  * from the encoded file_pte if possible. This enables swappable
  * nonlinear vmas.
  */
-static int do_file_page(struct mm_struct * mm, struct vm_area_struct * vma,
-       unsigned long address, int write_access, pte_t *pte, pmd_t *pmd)
+static int do_file_page(struct mm_struct *mm, struct vm_area_struct *vma,
+               unsigned long address, pte_t *page_table, pmd_t *pmd,
+               int write_access, pte_t orig_pte)
 {
-       unsigned long pgoff;
+       pgoff_t pgoff;
        int err;
 
-       BUG_ON(!vma->vm_ops || !vma->vm_ops->nopage);
-       /*
-        * Fall back to the linear mapping if the fs does not support
-        * ->populate:
-        */
-       if (!vma->vm_ops->populate ||
-                       (write_access && !(vma->vm_flags & VM_SHARED))) {
-               pte_clear(mm, address, pte);
-               return do_no_page(mm, vma, address, write_access, pte, pmd);
-       }
-
-       pgoff = pte_to_pgoff(*pte);
-
-       pte_unmap(pte);
+       pte_unmap(page_table);
        spin_unlock(&mm->page_table_lock);
 
-       err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE, vma->vm_page_prot, pgoff, 0);
+       if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
+               /*
+                * Page table corrupted: show pte and kill process.
+                */
+               pte_ERROR(orig_pte);
+               return VM_FAULT_OOM;
+       }
+       /* We can then assume vm->vm_ops && vma->vm_ops->populate */
+
+       pgoff = pte_to_pgoff(orig_pte);
+       err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE,
+                                       vma->vm_page_prot, pgoff, 0);
        if (err == -ENOMEM)
                return VM_FAULT_OOM;
        if (err)
@@ -2002,23 +1974,25 @@ static int do_file_page(struct mm_struct * mm, struct vm_area_struct * vma,
  * release it when done.
  */
 static inline int handle_pte_fault(struct mm_struct *mm,
-       struct vm_area_struct * vma, unsigned long address,
-       int write_access, pte_t *pte, pmd_t *pmd)
+               struct vm_area_struct *vma, unsigned long address,
+               pte_t *pte, pmd_t *pmd, int write_access)
 {
        pte_t entry;
 
        entry = *pte;
        if (!pte_present(entry)) {
-               /*
-                * If it truly wasn't present, we know that kswapd
-                * and the PTE updates will not touch it later. So
-                * drop the lock.
-                */
-               if (pte_none(entry))
-                       return do_no_page(mm, vma, address, write_access, pte, pmd);
+               if (pte_none(entry)) {
+                       if (!vma->vm_ops || !vma->vm_ops->nopage)
+                               return do_anonymous_page(mm, vma, address,
+                                       pte, pmd, write_access);
+                       return do_no_page(mm, vma, address,
+                                       pte, pmd, write_access);
+               }
                if (pte_file(entry))
-                       return do_file_page(mm, vma, address, write_access, pte, pmd);
-               return do_swap_page(mm, vma, address, pte, pmd, entry, write_access);
+                       return do_file_page(mm, vma, address,
+                                       pte, pmd, write_access, entry);
+               return do_swap_page(mm, vma, address,
+                                       pte, pmd, write_access, entry);
        }
 
        if (write_access) {
@@ -2038,7 +2012,7 @@ static inline int handle_pte_fault(struct mm_struct *mm,
 /*
  * By the time we get here, we already hold the mm semaphore
  */
-int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct * vma,
+int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
                unsigned long address, int write_access)
 {
        pgd_t *pgd;
@@ -2072,7 +2046,7 @@ int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct * vma,
        if (!pte)
                goto oom;
        
-       return handle_pte_fault(mm, vma, address, write_access, pte, pmd);
+       return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
 
  oom:
        spin_unlock(&mm->page_table_lock);
index 55e04a0..6796311 100644 (file)
@@ -1201,7 +1201,7 @@ static int shmem_populate(struct vm_area_struct *vma,
                                page_cache_release(page);
                                return err;
                        }
-               } else {
+               } else if (vma->vm_flags & VM_NONLINEAR) {
                        /* No page was found just because we can't read it in
                         * now (being here implies nonblock != 0), but the page
                         * may exist, so set the PTE to fault it in later. */