[GFS2] Fix lack of buffers in writepage bug
[linux-2.6-block.git] / fs / gfs2 / ops_address.c
index d611b2ad2e97f087c0bb91863698b738009075fd..45afd0508689a4ef39174fae366f9db2b621db74 100644 (file)
@@ -1,6 +1,6 @@
 /*
  * Copyright (C) Sistina Software, Inc.  1997-2003 All rights reserved.
- * Copyright (C) 2004-2005 Red Hat, Inc.  All rights reserved.
+ * Copyright (C) 2004-2006 Red Hat, Inc.  All rights reserved.
  *
  * This copyrighted material is made available to anyone wishing to use,
  * modify, copy, or redistribute it subject to the terms and conditions
 #include <linux/completion.h>
 #include <linux/buffer_head.h>
 #include <linux/pagemap.h>
+#include <linux/pagevec.h>
 #include <linux/mpage.h>
-#include <asm/semaphore.h>
+#include <linux/fs.h>
+#include <linux/gfs2_ondisk.h>
 
 #include "gfs2.h"
+#include "lm_interface.h"
+#include "incore.h"
 #include "bmap.h"
 #include "glock.h"
 #include "inode.h"
-#include "jdata.h"
 #include "log.h"
 #include "meta_io.h"
 #include "ops_address.h"
-#include "page.h"
 #include "quota.h"
 #include "trans.h"
+#include "rgrp.h"
+#include "ops_file.h"
+#include "util.h"
+#include "glops.h"
+
+
+static void gfs2_page_add_databufs(struct gfs2_inode *ip, struct page *page,
+                                  unsigned int from, unsigned int to)
+{
+       struct buffer_head *head = page_buffers(page);
+       unsigned int bsize = head->b_size;
+       struct buffer_head *bh;
+       unsigned int start, end;
+
+       for (bh = head, start = 0; bh != head || !start;
+            bh = bh->b_this_page, start = end) {
+               end = start + bsize;
+               if (end <= from || start >= to)
+                       continue;
+               gfs2_trans_add_bh(ip->i_gl, bh, 0);
+       }
+}
 
 /**
  * gfs2_get_block - Fills in a buffer head with details about a block
 int gfs2_get_block(struct inode *inode, sector_t lblock,
                   struct buffer_head *bh_result, int create)
 {
-       struct gfs2_inode *ip = get_v2ip(inode);
        int new = create;
        uint64_t dblock;
        int error;
+       int boundary;
 
-       error = gfs2_block_map(ip, lblock, &new, &dblock, NULL);
+       error = gfs2_block_map(inode, lblock, &new, &dblock, &boundary);
        if (error)
                return error;
 
@@ -56,6 +80,8 @@ int gfs2_get_block(struct inode *inode, sector_t lblock,
        map_bh(bh_result, inode->i_sb, dblock);
        if (new)
                set_buffer_new(bh_result);
+       if (boundary)
+               set_buffer_boundary(bh_result);
 
        return 0;
 }
@@ -73,72 +99,21 @@ int gfs2_get_block(struct inode *inode, sector_t lblock,
 static int get_block_noalloc(struct inode *inode, sector_t lblock,
                             struct buffer_head *bh_result, int create)
 {
-       struct gfs2_inode *ip = get_v2ip(inode);
        int new = 0;
        uint64_t dblock;
        int error;
+       int boundary;
 
-       error = gfs2_block_map(ip, lblock, &new, &dblock, NULL);
+       error = gfs2_block_map(inode, lblock, &new, &dblock, &boundary);
        if (error)
                return error;
 
        if (dblock)
                map_bh(bh_result, inode->i_sb, dblock);
-       else if (gfs2_assert_withdraw(ip->i_sbd, !create))
-               error = -EIO;
-
-       return error;
-}
-
-static int get_blocks(struct inode *inode, sector_t lblock,
-                     unsigned long max_blocks, struct buffer_head *bh_result,
-                     int create)
-{
-       struct gfs2_inode *ip = get_v2ip(inode);
-       int new = create;
-       uint64_t dblock;
-       uint32_t extlen;
-       int error;
-
-       error = gfs2_block_map(ip, lblock, &new, &dblock, &extlen);
-       if (error)
-               return error;
-
-       if (!dblock)
-               return 0;
-
-       map_bh(bh_result, inode->i_sb, dblock);
-       if (new)
-               set_buffer_new(bh_result);
-
-       if (extlen > max_blocks)
-               extlen = max_blocks;
-       bh_result->b_size = extlen << inode->i_blkbits;
-
-       return 0;
-}
-
-static int get_blocks_noalloc(struct inode *inode, sector_t lblock,
-                             unsigned long max_blocks,
-                             struct buffer_head *bh_result, int create)
-{
-       struct gfs2_inode *ip = get_v2ip(inode);
-       int new = 0;
-       uint64_t dblock;
-       uint32_t extlen;
-       int error;
-
-       error = gfs2_block_map(ip, lblock, &new, &dblock, &extlen);
-       if (error)
-               return error;
-
-       if (dblock) {
-               map_bh(bh_result, inode->i_sb, dblock);
-               if (extlen > max_blocks)
-                       extlen = max_blocks;
-               bh_result->b_size = extlen << inode->i_blkbits;
-       } else if (gfs2_assert_withdraw(ip->i_sbd, !create))
+       else if (gfs2_assert_withdraw(GFS2_SB(inode), !create))
                error = -EIO;
+       if (boundary)
+               set_buffer_boundary(bh_result);
 
        return error;
 }
@@ -149,33 +124,70 @@ static int get_blocks_noalloc(struct inode *inode, sector_t lblock,
  *
  * Returns: errno
  *
- * Use Linux VFS block_write_full_page() to write one page,
- *   using GFS2's get_block_noalloc to find which blocks to write.
+ * Some of this is copied from block_write_full_page() although we still
+ * call it to do most of the work.
  */
 
 static int gfs2_writepage(struct page *page, struct writeback_control *wbc)
 {
-       struct gfs2_inode *ip = get_v2ip(page->mapping->host);
-       struct gfs2_sbd *sdp = ip->i_sbd;
+       struct inode *inode = page->mapping->host;
+       struct gfs2_inode *ip = GFS2_I(inode);
+       struct gfs2_sbd *sdp = GFS2_SB(inode);
+       loff_t i_size = i_size_read(inode);
+       pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
+       unsigned offset;
        int error;
-
-       atomic_inc(&sdp->sd_ops_address);
+       int done_trans = 0;
 
        if (gfs2_assert_withdraw(sdp, gfs2_glock_is_held_excl(ip->i_gl))) {
                unlock_page(page);
                return -EIO;
        }
-       if (get_transaction) {
-               redirty_page_for_writepage(wbc, page);
+       if (current->journal_info)
+               goto out_ignore;
+
+       /* Is the page fully outside i_size? (truncate in progress) */
+        offset = i_size & (PAGE_CACHE_SIZE-1);
+       if (page->index > end_index || (page->index == end_index && !offset)) {
+               page->mapping->a_ops->invalidatepage(page, 0);
                unlock_page(page);
-               return 0;
+               return 0; /* don't care */
        }
 
+       if (sdp->sd_args.ar_data == GFS2_DATA_ORDERED || gfs2_is_jdata(ip)) {
+               error = gfs2_trans_begin(sdp, RES_DINODE + 1, 0);
+               if (error)
+                       goto out_ignore;
+               if (!page_has_buffers(page)) {
+                       create_empty_buffers(page, inode->i_sb->s_blocksize,
+                                            (1 << BH_Dirty)|(1 << BH_Uptodate));
+               }
+               gfs2_page_add_databufs(ip, page, 0, sdp->sd_vfs->s_blocksize-1);
+               done_trans = 1;
+       }
        error = block_write_full_page(page, get_block_noalloc, wbc);
-
+       if (done_trans)
+               gfs2_trans_end(sdp);
        gfs2_meta_cache_flush(ip);
-
        return error;
+
+out_ignore:
+       redirty_page_for_writepage(wbc, page);
+       unlock_page(page);
+       return 0;
+}
+
+static int zero_readpage(struct page *page)
+{
+       void *kaddr;
+
+       kaddr = kmap_atomic(page, KM_USER0);
+       memset(kaddr, 0, PAGE_CACHE_SIZE);
+       kunmap_atomic(page, KM_USER0);
+
+       SetPageUptodate(page);
+
+       return 0;
 }
 
 /**
@@ -192,18 +204,19 @@ static int stuffed_readpage(struct gfs2_inode *ip, struct page *page)
        void *kaddr;
        int error;
 
+       /* Only the first page of a stuffed file might contain data */
+       if (unlikely(page->index))
+               return zero_readpage(page);
+
        error = gfs2_meta_inode_buffer(ip, &dibh);
        if (error)
                return error;
 
-       kaddr = kmap(page);
-       memcpy((char *)kaddr,
-              dibh->b_data + sizeof(struct gfs2_dinode),
+       kaddr = kmap_atomic(page, KM_USER0);
+       memcpy(kaddr, dibh->b_data + sizeof(struct gfs2_dinode),
               ip->i_di.di_size);
-       memset((char *)kaddr + ip->i_di.di_size,
-              0,
-              PAGE_CACHE_SIZE - ip->i_di.di_size);
-       kunmap(page);
+       memset(kaddr + ip->i_di.di_size, 0, PAGE_CACHE_SIZE - ip->i_di.di_size);
+       kunmap_atomic(page, KM_USER0);
 
        brelse(dibh);
 
@@ -212,89 +225,147 @@ static int stuffed_readpage(struct gfs2_inode *ip, struct page *page)
        return 0;
 }
 
-static int zero_readpage(struct page *page)
-{
-       void *kaddr;
-
-       kaddr = kmap(page);
-       memset(kaddr, 0, PAGE_CACHE_SIZE);
-       kunmap(page);
-
-       SetPageUptodate(page);
-       unlock_page(page);
-
-       return 0;
-}
 
 /**
- * jdata_readpage - readpage that goes through gfs2_jdata_read_mem()
- * @ip:
+ * gfs2_readpage - readpage with locking
+ * @file: The file to read a page for. N.B. This may be NULL if we are
+ * reading an internal file.
  * @page: The page to read
  *
  * Returns: errno
  */
 
-static int jdata_readpage(struct gfs2_inode *ip, struct page *page)
+static int gfs2_readpage(struct file *file, struct page *page)
 {
-       void *kaddr;
-       int ret;
-
-       kaddr = kmap(page);
-
-       ret = gfs2_jdata_read_mem(ip, kaddr,
-                                 (uint64_t)page->index << PAGE_CACHE_SHIFT,
-                                 PAGE_CACHE_SIZE);
-       if (ret >= 0) {
-               if (ret < PAGE_CACHE_SIZE)
-                       memset(kaddr + ret, 0, PAGE_CACHE_SIZE - ret);
-               SetPageUptodate(page);
-               ret = 0;
+       struct gfs2_inode *ip = GFS2_I(page->mapping->host);
+       struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
+       struct gfs2_holder gh;
+       int error;
+       int do_unlock = 0;
+
+       if (likely(file != &gfs2_internal_file_sentinal)) {
+               if (file) {
+                       struct gfs2_file *gf = file->private_data;
+                       if (test_bit(GFF_EXLOCK, &gf->f_flags))
+                               goto skip_lock;
+               }
+               gfs2_holder_init(ip->i_gl, LM_ST_SHARED, GL_ATIME|GL_AOP, &gh);
+               do_unlock = 1;
+               error = gfs2_glock_nq_m_atime(1, &gh);
+               if (unlikely(error))
+                       goto out_unlock;
        }
 
-       kunmap(page);
+skip_lock:
+       if (gfs2_is_stuffed(ip)) {
+               error = stuffed_readpage(ip, page);
+               unlock_page(page);
+       } else
+               error = mpage_readpage(page, gfs2_get_block);
 
-       unlock_page(page);
+       if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
+               error = -EIO;
 
-       return ret;
+       if (file != &gfs2_internal_file_sentinal) {
+               gfs2_glock_dq_m(1, &gh);
+               gfs2_holder_uninit(&gh);
+       }
+out:
+       return error;
+out_unlock:
+       unlock_page(page);
+       if (do_unlock)
+               gfs2_holder_uninit(&gh);
+       goto out;
 }
 
 /**
- * gfs2_readpage - readpage with locking
- * @file: The file to read a page for
- * @page: The page to read
+ * gfs2_readpages - Read a bunch of pages at once
  *
- * Returns: errno
+ * Some notes:
+ * 1. This is only for readahead, so we can simply ignore any things
+ *    which are slightly inconvenient (such as locking conflicts between
+ *    the page lock and the glock) and return having done no I/O. Its
+ *    obviously not something we'd want to do on too regular a basis.
+ *    Any I/O we ignore at this time will be done via readpage later.
+ * 2. We have to handle stuffed files here too.
+ * 3. mpage_readpages() does most of the heavy lifting in the common case.
+ * 4. gfs2_get_block() is relied upon to set BH_Boundary in the right places.
+ * 5. We use LM_FLAG_TRY_1CB here, effectively we then have lock-ahead as
+ *    well as read-ahead.
  */
-
-static int gfs2_readpage(struct file *file, struct page *page)
+static int gfs2_readpages(struct file *file, struct address_space *mapping,
+                         struct list_head *pages, unsigned nr_pages)
 {
-       struct gfs2_inode *ip = get_v2ip(page->mapping->host);
-       struct gfs2_sbd *sdp = ip->i_sbd;
-       int error;
-
-       atomic_inc(&sdp->sd_ops_address);
-
-       if (gfs2_assert_warn(sdp, gfs2_glock_is_locked_by_me(ip->i_gl))) {
-               unlock_page(page);
-               return -EOPNOTSUPP;
+       struct inode *inode = mapping->host;
+       struct gfs2_inode *ip = GFS2_I(inode);
+       struct gfs2_sbd *sdp = GFS2_SB(inode);
+       struct gfs2_holder gh;
+       unsigned page_idx;
+       int ret;
+       int do_unlock = 0;
+
+       if (likely(file != &gfs2_internal_file_sentinal)) {
+               if (file) {
+                       struct gfs2_file *gf = file->private_data;
+                       if (test_bit(GFF_EXLOCK, &gf->f_flags))
+                               goto skip_lock;
+               }
+               gfs2_holder_init(ip->i_gl, LM_ST_SHARED,
+                                LM_FLAG_TRY_1CB|GL_ATIME|GL_AOP, &gh);
+               do_unlock = 1;
+               ret = gfs2_glock_nq_m_atime(1, &gh);
+               if (ret == GLR_TRYFAILED) 
+                       goto out_noerror;
+               if (unlikely(ret))
+                       goto out_unlock;
        }
-
-       if (!gfs2_is_jdata(ip)) {
-               if (gfs2_is_stuffed(ip)) {
-                       if (!page->index) {
-                               error = stuffed_readpage(ip, page);
+skip_lock:
+       if (gfs2_is_stuffed(ip)) {
+               struct pagevec lru_pvec;
+               pagevec_init(&lru_pvec, 0);
+               for (page_idx = 0; page_idx < nr_pages; page_idx++) {
+                       struct page *page = list_entry(pages->prev, struct page, lru);
+                       prefetchw(&page->flags);
+                       list_del(&page->lru);
+                       if (!add_to_page_cache(page, mapping,
+                                              page->index, GFP_KERNEL)) {
+                               ret = stuffed_readpage(ip, page);
                                unlock_page(page);
-                       } else
-                               error = zero_readpage(page);
-               } else
-                       error = mpage_readpage(page, gfs2_get_block);
-       } else
-               error = jdata_readpage(ip, page);
+                               if (!pagevec_add(&lru_pvec, page))
+                                        __pagevec_lru_add(&lru_pvec);
+                       } else {
+                               page_cache_release(page);
+                       }
+               }
+               pagevec_lru_add(&lru_pvec);
+               ret = 0;
+       } else {
+               /* What we really want to do .... */
+               ret = mpage_readpages(mapping, pages, nr_pages, gfs2_get_block);
+       }
 
+       if (do_unlock) {
+               gfs2_glock_dq_m(1, &gh);
+               gfs2_holder_uninit(&gh);
+       }
+out:
        if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
-               error = -EIO;
-
-       return error;
+               ret = -EIO;
+       return ret;
+out_noerror:
+       ret = 0;
+out_unlock:
+       /* unlock all pages, we can't do any I/O right now */
+       for (page_idx = 0; page_idx < nr_pages; page_idx++) {
+               struct page *page = list_entry(pages->prev, struct page, lru);
+               list_del(&page->lru);
+               unlock_page(page);
+               page_cache_release(page);
+       }
+       if (do_unlock)
+               gfs2_holder_uninit(&gh);
+       goto out;
 }
 
 /**
@@ -310,30 +381,82 @@ static int gfs2_readpage(struct file *file, struct page *page)
 static int gfs2_prepare_write(struct file *file, struct page *page,
                              unsigned from, unsigned to)
 {
-       struct gfs2_inode *ip = get_v2ip(page->mapping->host);
-       struct gfs2_sbd *sdp = ip->i_sbd;
+       struct gfs2_inode *ip = GFS2_I(page->mapping->host);
+       struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
+       unsigned int data_blocks, ind_blocks, rblocks;
+       int alloc_required;
        int error = 0;
+       loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + from;
+       loff_t end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
+       struct gfs2_alloc *al;
+
+       gfs2_holder_init(ip->i_gl, LM_ST_EXCLUSIVE, GL_ATIME|GL_AOP, &ip->i_gh);
+       error = gfs2_glock_nq_m_atime(1, &ip->i_gh);
+       if (error)
+               goto out_uninit;
 
-       atomic_inc(&sdp->sd_ops_address);
+       gfs2_write_calc_reserv(ip, to - from, &data_blocks, &ind_blocks);
 
-       if (gfs2_assert_warn(sdp, gfs2_glock_is_locked_by_me(ip->i_gl)))
-               return -EOPNOTSUPP;
+       error = gfs2_write_alloc_required(ip, pos, from - to, &alloc_required);
+       if (error)
+               goto out_unlock;
 
-       if (gfs2_is_stuffed(ip)) {
-               uint64_t file_size;
-               file_size = ((uint64_t)page->index << PAGE_CACHE_SHIFT) + to;
 
-               if (file_size > sdp->sd_sb.sb_bsize -
-                               sizeof(struct gfs2_dinode)) {
-                       error = gfs2_unstuff_dinode(ip, gfs2_unstuffer_page,
-                                                   page);
-                       if (!error)
-                               error = block_prepare_write(page, from, to,
-                                                           gfs2_get_block);
+       if (alloc_required) {
+               al = gfs2_alloc_get(ip);
+
+               error = gfs2_quota_lock(ip, NO_QUOTA_CHANGE, NO_QUOTA_CHANGE);
+               if (error)
+                       goto out_alloc_put;
+
+               error = gfs2_quota_check(ip, ip->i_di.di_uid, ip->i_di.di_gid);
+               if (error)
+                       goto out_qunlock;
+
+               al->al_requested = data_blocks + ind_blocks;
+               error = gfs2_inplace_reserve(ip);
+               if (error)
+                       goto out_qunlock;
+       }
+
+       rblocks = RES_DINODE + ind_blocks;
+       if (gfs2_is_jdata(ip))
+               rblocks += data_blocks ? data_blocks : 1;
+       if (ind_blocks || data_blocks)
+               rblocks += RES_STATFS + RES_QUOTA;
+
+       error = gfs2_trans_begin(sdp, rblocks, 0);
+       if (error)
+               goto out;
+
+       if (gfs2_is_stuffed(ip)) {
+               if (end > sdp->sd_sb.sb_bsize - sizeof(struct gfs2_dinode)) {
+                       error = gfs2_unstuff_dinode(ip, page);
+                       if (error == 0)
+                               goto prepare_write;
                } else if (!PageUptodate(page))
                        error = stuffed_readpage(ip, page);
-       } else
-               error = block_prepare_write(page, from, to, gfs2_get_block);
+               goto out;
+       }
+
+prepare_write:
+       error = block_prepare_write(page, from, to, gfs2_get_block);
+
+out:
+       if (error) {
+               gfs2_trans_end(sdp);
+               if (alloc_required) {
+                       gfs2_inplace_release(ip);
+out_qunlock:
+                       gfs2_quota_unlock(ip);
+out_alloc_put:
+                       gfs2_alloc_put(ip);
+               }
+out_unlock:
+               gfs2_glock_dq_m(1, &ip->i_gh);
+out_uninit:
+               gfs2_holder_uninit(&ip->i_gh);
+       }
 
        return error;
 }
@@ -352,50 +475,73 @@ static int gfs2_commit_write(struct file *file, struct page *page,
                             unsigned from, unsigned to)
 {
        struct inode *inode = page->mapping->host;
-       struct gfs2_inode *ip = get_v2ip(inode);
-       struct gfs2_sbd *sdp = ip->i_sbd;
-       int error;
+       struct gfs2_inode *ip = GFS2_I(inode);
+       struct gfs2_sbd *sdp = GFS2_SB(inode);
+       int error = -EOPNOTSUPP;
+       struct buffer_head *dibh;
+       struct gfs2_alloc *al = &ip->i_alloc;;
 
-       atomic_inc(&sdp->sd_ops_address);
+       if (gfs2_assert_withdraw(sdp, gfs2_glock_is_locked_by_me(ip->i_gl)))
+                goto fail_nounlock;
+
+       error = gfs2_meta_inode_buffer(ip, &dibh);
+       if (error)
+               goto fail_endtrans;
+
+       gfs2_trans_add_bh(ip->i_gl, dibh, 1);
 
        if (gfs2_is_stuffed(ip)) {
-               struct buffer_head *dibh;
                uint64_t file_size;
                void *kaddr;
 
                file_size = ((uint64_t)page->index << PAGE_CACHE_SHIFT) + to;
 
-               error = gfs2_meta_inode_buffer(ip, &dibh);
-               if (error)
-                       goto fail;
-
-               gfs2_trans_add_bh(ip->i_gl, dibh, 1);
-
-               kaddr = kmap(page);
+               kaddr = kmap_atomic(page, KM_USER0);
                memcpy(dibh->b_data + sizeof(struct gfs2_dinode) + from,
-                      (char *)kaddr + from,
-                      to - from);
-               kunmap(page);
-
-               brelse(dibh);
+                      (char *)kaddr + from, to - from);
+               kunmap_atomic(page, KM_USER0);
 
                SetPageUptodate(page);
 
                if (inode->i_size < file_size)
                        i_size_write(inode, file_size);
        } else {
-               if (sdp->sd_args.ar_data == GFS2_DATA_ORDERED)
+               if (sdp->sd_args.ar_data == GFS2_DATA_ORDERED ||
+                   gfs2_is_jdata(ip))
                        gfs2_page_add_databufs(ip, page, from, to);
                error = generic_commit_write(file, page, from, to);
                if (error)
                        goto fail;
        }
 
+       if (ip->i_di.di_size < inode->i_size)
+               ip->i_di.di_size = inode->i_size;
+
+       gfs2_dinode_out(&ip->i_di, dibh->b_data);
+       brelse(dibh);
+       gfs2_trans_end(sdp);
+       if (al->al_requested) {
+               gfs2_inplace_release(ip);
+               gfs2_quota_unlock(ip);
+               gfs2_alloc_put(ip);
+       }
+       gfs2_glock_dq_m(1, &ip->i_gh);
+       gfs2_holder_uninit(&ip->i_gh);
        return 0;
 
- fail:
+fail:
+       brelse(dibh);
+fail_endtrans:
+       gfs2_trans_end(sdp);
+       if (al->al_requested) {
+               gfs2_inplace_release(ip);
+               gfs2_quota_unlock(ip);
+               gfs2_alloc_put(ip);
+       }
+       gfs2_glock_dq_m(1, &ip->i_gh);
+       gfs2_holder_uninit(&ip->i_gh);
+fail_nounlock:
        ClearPageUptodate(page);
-
        return error;
 }
 
@@ -409,13 +555,11 @@ static int gfs2_commit_write(struct file *file, struct page *page,
 
 static sector_t gfs2_bmap(struct address_space *mapping, sector_t lblock)
 {
-       struct gfs2_inode *ip = get_v2ip(mapping->host);
+       struct gfs2_inode *ip = GFS2_I(mapping->host);
        struct gfs2_holder i_gh;
        sector_t dblock = 0;
        int error;
 
-       atomic_inc(&ip->i_sbd->sd_ops_address);
-
        error = gfs2_glock_nq_init(ip->i_gl, LM_ST_SHARED, LM_FLAG_ANY, &i_gh);
        if (error)
                return 0;
@@ -433,10 +577,10 @@ static void discard_buffer(struct gfs2_sbd *sdp, struct buffer_head *bh)
        struct gfs2_bufdata *bd;
 
        gfs2_log_lock(sdp);
-       bd = get_v2bd(bh);
+       bd = bh->b_private;
        if (bd) {
                bd->bd_bh = NULL;
-               set_v2bd(bh, NULL);
+               bh->b_private = NULL;
                gfs2_log_unlock(sdp);
                brelse(bh);
        } else
@@ -452,16 +596,15 @@ static void discard_buffer(struct gfs2_sbd *sdp, struct buffer_head *bh)
        unlock_buffer(bh);
 }
 
-static int gfs2_invalidatepage(struct page *page, unsigned long offset)
+static void gfs2_invalidatepage(struct page *page, unsigned long offset)
 {
-       struct gfs2_sbd *sdp = get_v2sdp(page->mapping->host->i_sb);
+       struct gfs2_sbd *sdp = page->mapping->host->i_sb->s_fs_info;
        struct buffer_head *head, *bh, *next;
        unsigned int curr_off = 0;
-       int ret = 1;
 
        BUG_ON(!PageLocked(page));
        if (!page_has_buffers(page))
-               return 1;
+               return;
 
        bh = head = page_buffers(page);
        do {
@@ -476,41 +619,179 @@ static int gfs2_invalidatepage(struct page *page, unsigned long offset)
        } while (bh != head);
 
        if (!offset)
-               ret = try_to_release_page(page, 0);
+               try_to_release_page(page, 0);
 
-       return ret;
+       return;
 }
 
-static ssize_t gfs2_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
-                         loff_t offset, unsigned long nr_segs)
+static ssize_t gfs2_direct_IO(int rw, struct kiocb *iocb,
+                             const struct iovec *iov, loff_t offset,
+                             unsigned long nr_segs)
 {
        struct file *file = iocb->ki_filp;
        struct inode *inode = file->f_mapping->host;
-       struct gfs2_inode *ip = get_v2ip(inode);
-       struct gfs2_sbd *sdp = ip->i_sbd;
-       get_blocks_t *gb = get_blocks;
+       struct gfs2_inode *ip = GFS2_I(inode);
+       struct gfs2_holder gh;
+       int rv;
+
+       if (rw == READ)
+               mutex_lock(&inode->i_mutex);
+       /*
+        * Shared lock, even if its a write, since we do no allocation
+        * on this path. All we need change is atime.
+        */
+       gfs2_holder_init(ip->i_gl, LM_ST_SHARED, GL_ATIME, &gh);
+       rv = gfs2_glock_nq_m_atime(1, &gh);
+       if (rv)
+               goto out;
+
+       if (offset > i_size_read(inode))
+               goto out;
+
+       /*
+        * Should we return an error here? I can't see that O_DIRECT for
+        * a journaled file makes any sense. For now we'll silently fall
+        * back to buffered I/O, likewise we do the same for stuffed
+        * files since they are (a) small and (b) unaligned.
+        */
+       if (gfs2_is_jdata(ip))
+               goto out;
+
+       if (gfs2_is_stuffed(ip))
+               goto out;
+
+       rv = blockdev_direct_IO_own_locking(rw, iocb, inode,
+                                           inode->i_sb->s_bdev,
+                                           iov, offset, nr_segs,
+                                           gfs2_get_block, NULL);
+out:
+       gfs2_glock_dq_m(1, &gh);
+       gfs2_holder_uninit(&gh);
+       if (rw == READ)
+               mutex_unlock(&inode->i_mutex);
+
+       return rv;
+}
 
-       atomic_inc(&sdp->sd_ops_address);
+/**
+ * stuck_releasepage - We're stuck in gfs2_releasepage().  Print stuff out.
+ * @bh: the buffer we're stuck on
+ *
+ */
+
+static void stuck_releasepage(struct buffer_head *bh)
+{
+       struct inode *inode = bh->b_page->mapping->host;
+       struct gfs2_sbd *sdp = inode->i_sb->s_fs_info;
+       struct gfs2_bufdata *bd = bh->b_private;
+       struct gfs2_glock *gl;
 
-       if (gfs2_assert_warn(sdp, gfs2_glock_is_locked_by_me(ip->i_gl)) ||
-           gfs2_assert_warn(sdp, !gfs2_is_stuffed(ip)))
-               return -EINVAL;
+       fs_warn(sdp, "stuck in gfs2_releasepage() %p\n", inode);
+       fs_warn(sdp, "blkno = %llu, bh->b_count = %d\n",
+               (unsigned long long)bh->b_blocknr, atomic_read(&bh->b_count));
+       fs_warn(sdp, "pinned = %u\n", buffer_pinned(bh));
+       fs_warn(sdp, "bh->b_private = %s\n", (bd) ? "!NULL" : "NULL");
 
-       if (rw == WRITE && !get_transaction)
-               gb = get_blocks_noalloc;
+       if (!bd)
+               return;
+
+       gl = bd->bd_gl;
+
+       fs_warn(sdp, "gl = (%u, %llu)\n", 
+               gl->gl_name.ln_type, (unsigned long long)gl->gl_name.ln_number);
+
+       fs_warn(sdp, "bd_list_tr = %s, bd_le.le_list = %s\n",
+               (list_empty(&bd->bd_list_tr)) ? "no" : "yes",
+               (list_empty(&bd->bd_le.le_list)) ? "no" : "yes");
+
+       if (gl->gl_ops == &gfs2_inode_glops) {
+               struct gfs2_inode *ip = gl->gl_object;
+               unsigned int x;
+
+               if (!ip)
+                       return;
+
+               fs_warn(sdp, "ip = %llu %llu\n",
+                       (unsigned long long)ip->i_num.no_formal_ino,
+                       (unsigned long long)ip->i_num.no_addr);
+
+               for (x = 0; x < GFS2_MAX_META_HEIGHT; x++)
+                       fs_warn(sdp, "ip->i_cache[%u] = %s\n",
+                               x, (ip->i_cache[x]) ? "!NULL" : "NULL");
+       }
+}
+
+/**
+ * gfs2_aspace_releasepage - free the metadata associated with a page
+ * @page: the page that's being released
+ * @gfp_mask: passed from Linux VFS, ignored by us
+ *
+ * Call try_to_free_buffers() if the buffers in this page can be
+ * released.
+ *
+ * Returns: 0
+ */
+
+int gfs2_releasepage(struct page *page, gfp_t gfp_mask)
+{
+       struct inode *aspace = page->mapping->host;
+       struct gfs2_sbd *sdp = aspace->i_sb->s_fs_info;
+       struct buffer_head *bh, *head;
+       struct gfs2_bufdata *bd;
+       unsigned long t;
+
+       if (!page_has_buffers(page))
+               goto out;
+
+       head = bh = page_buffers(page);
+       do {
+               t = jiffies;
+
+               while (atomic_read(&bh->b_count)) {
+                       if (atomic_read(&aspace->i_writecount)) {
+                               if (time_after_eq(jiffies, t +
+                                   gfs2_tune_get(sdp, gt_stall_secs) * HZ)) {
+                                       stuck_releasepage(bh);
+                                       t = jiffies;
+                               }
+
+                               yield();
+                               continue;
+                       }
+
+                       return 0;
+               }
+
+               gfs2_assert_warn(sdp, !buffer_pinned(bh));
+
+               bd = bh->b_private;
+               if (bd) {
+                       gfs2_assert_warn(sdp, bd->bd_bh == bh);
+                       gfs2_assert_warn(sdp, list_empty(&bd->bd_list_tr));
+                       gfs2_assert_warn(sdp, list_empty(&bd->bd_le.le_list));
+                       gfs2_assert_warn(sdp, !bd->bd_ail);
+                       kmem_cache_free(gfs2_bufdata_cachep, bd);
+                       bh->b_private = NULL;
+               }
+
+               bh = bh->b_this_page;
+       }
+       while (bh != head);
 
-       return blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov,
-                                 offset, nr_segs, gb, NULL);
+out:
+       return try_to_free_buffers(page);
 }
 
-struct address_space_operations gfs2_file_aops = {
+const struct address_space_operations gfs2_file_aops = {
        .writepage = gfs2_writepage,
        .readpage = gfs2_readpage,
+       .readpages = gfs2_readpages,
        .sync_page = block_sync_page,
        .prepare_write = gfs2_prepare_write,
        .commit_write = gfs2_commit_write,
        .bmap = gfs2_bmap,
        .invalidatepage = gfs2_invalidatepage,
+       .releasepage = gfs2_releasepage,
        .direct_IO = gfs2_direct_IO,
 };