ocfs2: Fix ocfs2_page_mkwrite()
authorWengang Wang <wen.gang.wang@oracle.com>
Sun, 24 Jul 2011 17:36:54 +0000 (10:36 -0700)
committerSunil Mushran <sunil.mushran@oracle.com>
Sun, 24 Jul 2011 17:36:54 +0000 (10:36 -0700)
This patch address two shortcomings in ocfs2_page_mkwrite():
1. Makes the function return better VM_FAULT_* errors.
2. It handles a error that is triggered when a page is dropped from the mapping
due to memory pressure. This patch locks the page to prevent that.

[Patch was cleaned up by Sunil Mushran.]

Signed-off-by: Wengang Wang <wen.gang.wang@oracle.com>
Signed-off-by: Sunil Mushran <sunil.mushran@oracle.com>
fs/ocfs2/aops.c
fs/ocfs2/mmap.c

index c1efe939c774e2c9b909892f6c95434d6da760ee..ff98c169b631efb1a0e2d2467f9a31eec9cc5e85 100644 (file)
@@ -862,6 +862,12 @@ struct ocfs2_write_ctxt {
        struct page                     *w_pages[OCFS2_MAX_CTXT_PAGES];
        struct page                     *w_target_page;
 
+       /*
+        * w_target_locked is used for page_mkwrite path indicating no unlocking
+        * against w_target_page in ocfs2_write_end_nolock.
+        */
+       unsigned int                    w_target_locked:1;
+
        /*
         * ocfs2_write_end() uses this to know what the real range to
         * write in the target should be.
@@ -895,6 +901,24 @@ void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
 
 static void ocfs2_free_write_ctxt(struct ocfs2_write_ctxt *wc)
 {
+       int i;
+
+       /*
+        * w_target_locked is only set to true in the page_mkwrite() case.
+        * The intent is to allow us to lock the target page from write_begin()
+        * to write_end(). The caller must hold a ref on w_target_page.
+        */
+       if (wc->w_target_locked) {
+               BUG_ON(!wc->w_target_page);
+               for (i = 0; i < wc->w_num_pages; i++) {
+                       if (wc->w_target_page == wc->w_pages[i]) {
+                               wc->w_pages[i] = NULL;
+                               break;
+                       }
+               }
+               mark_page_accessed(wc->w_target_page);
+               page_cache_release(wc->w_target_page);
+       }
        ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
 
        brelse(wc->w_di_bh);
@@ -1132,20 +1156,17 @@ static int ocfs2_grab_pages_for_write(struct address_space *mapping,
                         */
                        lock_page(mmap_page);
 
+                       /* Exit and let the caller retry */
                        if (mmap_page->mapping != mapping) {
+                               WARN_ON(mmap_page->mapping);
                                unlock_page(mmap_page);
-                               /*
-                                * Sanity check - the locking in
-                                * ocfs2_pagemkwrite() should ensure
-                                * that this code doesn't trigger.
-                                */
-                               ret = -EINVAL;
-                               mlog_errno(ret);
+                               ret = -EAGAIN;
                                goto out;
                        }
 
                        page_cache_get(mmap_page);
                        wc->w_pages[i] = mmap_page;
+                       wc->w_target_locked = true;
                } else {
                        wc->w_pages[i] = find_or_create_page(mapping, index,
                                                             GFP_NOFS);
@@ -1160,6 +1181,8 @@ static int ocfs2_grab_pages_for_write(struct address_space *mapping,
                        wc->w_target_page = wc->w_pages[i];
        }
 out:
+       if (ret)
+               wc->w_target_locked = false;
        return ret;
 }
 
@@ -1817,11 +1840,23 @@ try_again:
         */
        ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos, len,
                                         cluster_of_pages, mmap_page);
-       if (ret) {
+       if (ret && ret != -EAGAIN) {
                mlog_errno(ret);
                goto out_quota;
        }
 
+       /*
+        * ocfs2_grab_pages_for_write() returns -EAGAIN if it could not lock
+        * the target page. In this case, we exit with no error and no target
+        * page. This will trigger the caller, page_mkwrite(), to re-try
+        * the operation.
+        */
+       if (ret == -EAGAIN) {
+               BUG_ON(wc->w_target_page);
+               ret = 0;
+               goto out_quota;
+       }
+
        ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
                                          len);
        if (ret) {
index 3e9393ca39ebd823772ae910cda1b4516d46fa62..9cd41083e99123eca1c48085fb39809e6b906b40 100644 (file)
@@ -61,7 +61,7 @@ static int ocfs2_fault(struct vm_area_struct *area, struct vm_fault *vmf)
 static int __ocfs2_page_mkwrite(struct file *file, struct buffer_head *di_bh,
                                struct page *page)
 {
-       int ret;
+       int ret = VM_FAULT_NOPAGE;
        struct inode *inode = file->f_path.dentry->d_inode;
        struct address_space *mapping = inode->i_mapping;
        loff_t pos = page_offset(page);
@@ -71,32 +71,25 @@ static int __ocfs2_page_mkwrite(struct file *file, struct buffer_head *di_bh,
        void *fsdata;
        loff_t size = i_size_read(inode);
 
-       /*
-        * Another node might have truncated while we were waiting on
-        * cluster locks.
-        * We don't check size == 0 before the shift. This is borrowed
-        * from do_generic_file_read.
-        */
        last_index = (size - 1) >> PAGE_CACHE_SHIFT;
-       if (unlikely(!size || page->index > last_index)) {
-               ret = -EINVAL;
-               goto out;
-       }
 
        /*
-        * The i_size check above doesn't catch the case where nodes
-        * truncated and then re-extended the file. We'll re-check the
-        * page mapping after taking the page lock inside of
-        * ocfs2_write_begin_nolock().
+        * There are cases that lead to the page no longer bebongs to the
+        * mapping.
+        * 1) pagecache truncates locally due to memory pressure.
+        * 2) pagecache truncates when another is taking EX lock against 
+        * inode lock. see ocfs2_data_convert_worker.
+        * 
+        * The i_size check doesn't catch the case where nodes truncated and
+        * then re-extended the file. We'll re-check the page mapping after
+        * taking the page lock inside of ocfs2_write_begin_nolock().
+        *
+        * Let VM retry with these cases.
         */
-       if (!PageUptodate(page) || page->mapping != inode->i_mapping) {
-               /*
-                * the page has been umapped in ocfs2_data_downconvert_worker.
-                * So return 0 here and let VFS retry.
-                */
-               ret = 0;
+       if ((page->mapping != inode->i_mapping) ||
+           (!PageUptodate(page)) ||
+           (page_offset(page) >= size))
                goto out;
-       }
 
        /*
         * Call ocfs2_write_begin() and ocfs2_write_end() to take
@@ -116,17 +109,21 @@ static int __ocfs2_page_mkwrite(struct file *file, struct buffer_head *di_bh,
        if (ret) {
                if (ret != -ENOSPC)
                        mlog_errno(ret);
+               if (ret == -ENOMEM)
+                       ret = VM_FAULT_OOM;
+               else
+                       ret = VM_FAULT_SIGBUS;
                goto out;
        }
 
-       ret = ocfs2_write_end_nolock(mapping, pos, len, len, locked_page,
-                                    fsdata);
-       if (ret < 0) {
-               mlog_errno(ret);
+       if (!locked_page) {
+               ret = VM_FAULT_NOPAGE;
                goto out;
        }
+       ret = ocfs2_write_end_nolock(mapping, pos, len, len, locked_page,
+                                    fsdata);
        BUG_ON(ret != len);
-       ret = 0;
+       ret = VM_FAULT_LOCKED;
 out:
        return ret;
 }
@@ -168,8 +165,6 @@ static int ocfs2_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
 
 out:
        ocfs2_unblock_signals(&oldset);
-       if (ret)
-               ret = VM_FAULT_SIGBUS;
        return ret;
 }