ocfs2: shared writeable mmap

Implement cluster consistent shared writeable mappings using the
->page_mkwrite() callback.

Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>

fs/ocfs2/aops.c

@@ -1034,7 +1034,8 @@ out:
  */
 static int ocfs2_grab_pages_for_write(struct address_space *mapping,
 				      struct ocfs2_write_ctxt *wc,
-				      u32 cpos, loff_t user_pos, int new)
+				      u32 cpos, loff_t user_pos, int new,
+				      struct page *mmap_page)
 {
 	int ret = 0, i;
 	unsigned long start, target_index, index;
@@ -1058,11 +1059,36 @@ static int ocfs2_grab_pages_for_write(struct address_space *mapping,
 	for(i = 0; i < wc->w_num_pages; i++) {
 		index = start + i;
 
-		wc->w_pages[i] = find_or_create_page(mapping, index, GFP_NOFS);
-		if (!wc->w_pages[i]) {
-			ret = -ENOMEM;
-			mlog_errno(ret);
-			goto out;
+		if (index == target_index && mmap_page) {
+			/*
+			 * ocfs2_pagemkwrite() is a little different
+			 * and wants us to directly use the page
+			 * passed in.
+			 */
+			lock_page(mmap_page);
+
+			if (mmap_page->mapping != 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);
+				goto out;
+			}
+
+			page_cache_get(mmap_page);
+			wc->w_pages[i] = mmap_page;
+		} else {
+			wc->w_pages[i] = find_or_create_page(mapping, index,
+							     GFP_NOFS);
+			if (!wc->w_pages[i]) {
+				ret = -ENOMEM;
+				mlog_errno(ret);
+				goto out;
+			}
 		}
 
 		if (index == target_index)
@@ -1213,10 +1239,10 @@ static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
 	}
 }
 
-static int ocfs2_write_begin_nolock(struct address_space *mapping,
-				    loff_t pos, unsigned len, unsigned flags,
-				    struct page **pagep, void **fsdata,
-				    struct buffer_head *di_bh)
+int ocfs2_write_begin_nolock(struct address_space *mapping,
+			     loff_t pos, unsigned len, unsigned flags,
+			     struct page **pagep, void **fsdata,
+			     struct buffer_head *di_bh, struct page *mmap_page)
 {
 	int ret, i, credits = OCFS2_INODE_UPDATE_CREDITS;
 	unsigned int num_clusters = 0, clusters_to_alloc = 0;
@@ -1318,7 +1344,7 @@ static int ocfs2_write_begin_nolock(struct address_space *mapping,
 	 * extent.
 	 */
 	ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos,
-					 clusters_to_alloc);
+					 clusters_to_alloc, mmap_page);
 	if (ret) {
 		mlog_errno(ret);
 		goto out_commit;
@@ -1386,7 +1412,7 @@ int ocfs2_write_begin(struct file *file, struct address_space *mapping,
 	}
 
 	ret = ocfs2_write_begin_nolock(mapping, pos, len, flags, pagep,
-				       fsdata, di_bh);
+				       fsdata, di_bh, NULL);
 	if (ret) {
 		mlog_errno(ret);
 		goto out_fail_data;
@@ -1407,9 +1433,9 @@ out_fail:
 	return ret;
 }
 
-static int ocfs2_write_end_nolock(struct address_space *mapping,
-				  loff_t pos, unsigned len, unsigned copied,
-				  struct page *page, void *fsdata)
+int ocfs2_write_end_nolock(struct address_space *mapping,
+			   loff_t pos, unsigned len, unsigned copied,
+			   struct page *page, void *fsdata)
 {
 	int i;
 	unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1);

fs/ocfs2/aops.h

@@ -50,6 +50,15 @@ int ocfs2_write_end(struct file *file, struct address_space *mapping,
 		    loff_t pos, unsigned len, unsigned copied,
 		    struct page *page, void *fsdata);
 
+int ocfs2_write_end_nolock(struct address_space *mapping,
+			   loff_t pos, unsigned len, unsigned copied,
+			   struct page *page, void *fsdata);
+
+int ocfs2_write_begin_nolock(struct address_space *mapping,
+			     loff_t pos, unsigned len, unsigned flags,
+			     struct page **pagep, void **fsdata,
+			     struct buffer_head *di_bh, struct page *mmap_page);
+
 /* all ocfs2_dio_end_io()'s fault */
 #define ocfs2_iocb_is_rw_locked(iocb) \
 	test_bit(0, (unsigned long *)&iocb->private)

fs/ocfs2/file.c

@@ -1001,6 +1001,13 @@ int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
 		goto bail_unlock;
 	}
 
+	/*
+	 * This will intentionally not wind up calling vmtruncate(),
+	 * since all the work for a size change has been done above.
+	 * Otherwise, we could get into problems with truncate as
+	 * ip_alloc_sem is used there to protect against i_size
+	 * changes.
+	 */
 	status = inode_setattr(inode, attr);
 	if (status < 0) {
 		mlog_errno(status);

fs/ocfs2/mmap.c

@@ -37,11 +37,29 @@
 
 #include "ocfs2.h"
 
+#include "aops.h"
 #include "dlmglue.h"
 #include "file.h"
 #include "inode.h"
 #include "mmap.h"
 
+static inline int ocfs2_vm_op_block_sigs(sigset_t *blocked, sigset_t *oldset)
+{
+	/* The best way to deal with signals in the vm path is
+	 * to block them upfront, rather than allowing the
+	 * locking paths to return -ERESTARTSYS. */
+	sigfillset(blocked);
+
+	/* We should technically never get a bad return value
+	 * from sigprocmask */
+	return sigprocmask(SIG_BLOCK, blocked, oldset);
+}
+
+static inline int ocfs2_vm_op_unblock_sigs(sigset_t *oldset)
+{
+	return sigprocmask(SIG_SETMASK, oldset, NULL);
+}
+
 static struct page *ocfs2_nopage(struct vm_area_struct * area,
 				 unsigned long address,
 				 int *type)
@@ -53,14 +71,7 @@ static struct page *ocfs2_nopage(struct vm_area_struct * area,
 	mlog_entry("(area=%p, address=%lu, type=%p)\n", area, address,
 		   type);
 
-	/* The best way to deal with signals in this path is
-	 * to block them upfront, rather than allowing the
-	 * locking paths to return -ERESTARTSYS. */
-	sigfillset(&blocked);
-
-	/* We should technically never get a bad ret return
-	 * from sigprocmask */
-	ret = sigprocmask(SIG_BLOCK, &blocked, &oldset);
+	ret = ocfs2_vm_op_block_sigs(&blocked, &oldset);
 	if (ret < 0) {
 		mlog_errno(ret);
 		goto out;
@@ -68,7 +79,7 @@ static struct page *ocfs2_nopage(struct vm_area_struct * area,
 
 	page = filemap_nopage(area, address, type);
 
-	ret = sigprocmask(SIG_SETMASK, &oldset, NULL);
+	ret = ocfs2_vm_op_unblock_sigs(&oldset);
 	if (ret < 0)
 		mlog_errno(ret);
 out:
@@ -76,28 +87,136 @@ out:
 	return page;
 }
 
-static struct vm_operations_struct ocfs2_file_vm_ops = {
-	.nopage = ocfs2_nopage,
-};
+static int __ocfs2_page_mkwrite(struct inode *inode, struct buffer_head *di_bh,
+				struct page *page)
+{
+	int ret;
+	struct address_space *mapping = inode->i_mapping;
+	loff_t pos = page->index << PAGE_CACHE_SHIFT;
+	unsigned int len = PAGE_CACHE_SIZE;
+	pgoff_t last_index;
+	struct page *locked_page = NULL;
+	void *fsdata;
+	loff_t size = i_size_read(inode);
 
-int ocfs2_mmap(struct file *file, struct vm_area_struct *vma)
+	/*
+	 * Another node might have truncated while we were waiting on
+	 * cluster locks.
+	 */
+	last_index = size >> PAGE_CACHE_SHIFT;
+	if (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().
+	 */
+	if (!PageUptodate(page) || page->mapping != inode->i_mapping) {
+		ret = -EINVAL;
+		goto out;
+	}
+
+	/*
+	 * Call ocfs2_write_begin() and ocfs2_write_end() to take
+	 * advantage of the allocation code there. We pass a write
+	 * length of the whole page (chopped to i_size) to make sure
+	 * the whole thing is allocated.
+	 *
+	 * Since we know the page is up to date, we don't have to
+	 * worry about ocfs2_write_begin() skipping some buffer reads
+	 * because the "write" would invalidate their data.
+	 */
+	if (page->index == last_index)
+		len = size & ~PAGE_CACHE_MASK;
+
+	ret = ocfs2_write_begin_nolock(mapping, pos, len, 0, &locked_page,
+				       &fsdata, di_bh, page);
+	if (ret) {
+		if (ret != -ENOSPC)
+			mlog_errno(ret);
+		goto out;
+	}
+
+	ret = ocfs2_write_end_nolock(mapping, pos, len, len, locked_page,
+				     fsdata);
+	if (ret < 0) {
+		mlog_errno(ret);
+		goto out;
+	}
+	BUG_ON(ret != len);
+	ret = 0;
+out:
+	return ret;
+}
+
+static int ocfs2_page_mkwrite(struct vm_area_struct *vma, struct page *page)
 {
-	int ret = 0, lock_level = 0;
-	struct ocfs2_super *osb = OCFS2_SB(file->f_dentry->d_inode->i_sb);
+	struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
+	struct buffer_head *di_bh = NULL;
+	sigset_t blocked, oldset;
+	int ret, ret2;
+
+	ret = ocfs2_vm_op_block_sigs(&blocked, &oldset);
+	if (ret < 0) {
+		mlog_errno(ret);
+		return ret;
+	}
+
+	/*
+	 * The cluster locks taken will block a truncate from another
+	 * node. Taking the data lock will also ensure that we don't
+	 * attempt page truncation as part of a downconvert.
+	 */
+	ret = ocfs2_meta_lock(inode, &di_bh, 1);
+	if (ret < 0) {
+		mlog_errno(ret);
+		goto out;
+	}
 
 	/*
-	 * Only support shared writeable mmap for local mounts which
-	 * don't know about holes.
+	 * The alloc sem should be enough to serialize with
+	 * ocfs2_truncate_file() changing i_size as well as any thread
+	 * modifying the inode btree.
 	 */
-	if ((!ocfs2_mount_local(osb) || ocfs2_sparse_alloc(osb)) &&
-	    ((vma->vm_flags & VM_SHARED) || (vma->vm_flags & VM_MAYSHARE)) &&
-	    ((vma->vm_flags & VM_WRITE) || (vma->vm_flags & VM_MAYWRITE))) {
-		mlog(0, "disallow shared writable mmaps %lx\n", vma->vm_flags);
-		/* This is -EINVAL because generic_file_readonly_mmap
-		 * returns it in a similar situation. */
-		return -EINVAL;
+	down_write(&OCFS2_I(inode)->ip_alloc_sem);
+
+	ret = ocfs2_data_lock(inode, 1);
+	if (ret < 0) {
+		mlog_errno(ret);
+		goto out_meta_unlock;
 	}
 
+	ret = __ocfs2_page_mkwrite(inode, di_bh, page);
+
+	ocfs2_data_unlock(inode, 1);
+
+out_meta_unlock:
+	up_write(&OCFS2_I(inode)->ip_alloc_sem);
+
+	brelse(di_bh);
+	ocfs2_meta_unlock(inode, 1);
+
+out:
+	ret2 = ocfs2_vm_op_unblock_sigs(&oldset);
+	if (ret2 < 0)
+		mlog_errno(ret2);
+
+	return ret;
+}
+
+static struct vm_operations_struct ocfs2_file_vm_ops = {
+	.nopage		= ocfs2_nopage,
+	.page_mkwrite	= ocfs2_page_mkwrite,
+};
+
+int ocfs2_mmap(struct file *file, struct vm_area_struct *vma)
+{
+	int ret = 0, lock_level = 0;
+
 	ret = ocfs2_meta_lock_atime(file->f_dentry->d_inode,
 				    file->f_vfsmnt, &lock_level);
 	if (ret < 0) {