NTFS: Fix a mount time deadlock.

Big thanks go to Mathias Kolehmainen for reporting the bug, providing
debug output and testing the patches I sent him to get it working.

The fix was to stop calling ntfs_attr_set() at mount time as that causes
balance_dirty_pages_ratelimited() to be called which on systems with
little memory actually tries to go and balance the dirty pages which tries
to take the s_umount semaphore but because we are still in fill_super()
across which the VFS holds s_umount for writing this results in a
deadlock.

We now do the dirty work by hand by submitting individual buffers.  This
has the annoying "feature" that mounting can take a few seconds if the
journal is large as we have clear it all.  One day someone should improve
on this by deferring the journal clearing to a helper kernel thread so it
can be done in the background but I don't have time for this at the moment
and the current solution works fine so I am leaving it like this for now.

Signed-off-by: Anton Altaparmakov <aia21@cantab.net>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

1 files changed, 129 insertions, 14 deletions

diff --git a/fs/ntfs/logfile.c b/fs/ntfs/logfile.c
index acfed325f4ec..d7932e95b1fd 100644
--- a/fs/ntfs/logfile.c
+++ b/fs/ntfs/logfile.c
@@ -1,7 +1,7 @@
 /*
  * logfile.c - NTFS kernel journal handling. Part of the Linux-NTFS project.
  *
- * Copyright (c) 2002-2005 Anton Altaparmakov
+ * Copyright (c) 2002-2007 Anton Altaparmakov
  *
  * This program/include file is free software; you can redistribute it and/or
  * modify it under the terms of the GNU General Public License as published
@@ -724,24 +724,139 @@ bool ntfs_is_logfile_clean(struct inode *log_vi, const RESTART_PAGE_HEADER *rp)
  */
 bool ntfs_empty_logfile(struct inode *log_vi)
 {
-        ntfs_volume *vol = NTFS_SB(log_vi->i_sb);
+        VCN vcn, end_vcn;
+        ntfs_inode *log_ni = NTFS_I(log_vi);
+        ntfs_volume *vol = log_ni->vol;
+        struct super_block *sb = vol->sb;
+        runlist_element *rl;
+        unsigned long flags;
+        unsigned block_size, block_size_bits;
+        int err;
+        bool should_wait = true;
 
         ntfs_debug("Entering.");
-        if (!NVolLogFileEmpty(vol)) {
-                int err;
-                
-                err = ntfs_attr_set(NTFS_I(log_vi), 0, i_size_read(log_vi),
-                                0xff);
-                if (unlikely(err)) {
-                        ntfs_error(vol->sb, "Failed to fill $LogFile with "
-                                        "0xff bytes (error code %i).", err);
-                        return false;
-                }
-                /* Set the flag so we do not have to do it again on remount. */
-                NVolSetLogFileEmpty(vol);
+        if (NVolLogFileEmpty(vol)) {
+                ntfs_debug("Done.");
+                return true;
         }
+        /*
+         * We cannot use ntfs_attr_set() because we may be still in the middle
+         * of a mount operation.  Thus we do the emptying by hand by first
+         * zapping the page cache pages for the $LogFile/$DATA attribute and
+         * then emptying each of the buffers in each of the clusters specified
+         * by the runlist by hand.
+         */
+        block_size = sb->s_blocksize;
+        block_size_bits = sb->s_blocksize_bits;
+        vcn = 0;
+        read_lock_irqsave(&log_ni->size_lock, flags);
+        end_vcn = (log_ni->initialized_size + vol->cluster_size_mask) >>
+                        vol->cluster_size_bits;
+        read_unlock_irqrestore(&log_ni->size_lock, flags);
+        truncate_inode_pages(log_vi->i_mapping, 0);
+        down_write(&log_ni->runlist.lock);
+        rl = log_ni->runlist.rl;
+        if (unlikely(!rl || vcn < rl->vcn || !rl->length)) {
+map_vcn:
+                err = ntfs_map_runlist_nolock(log_ni, vcn, NULL);
+                if (err) {
+                        ntfs_error(sb, "Failed to map runlist fragment (error "
+                                        "%d).", -err);
+                        goto err;
+                }
+                rl = log_ni->runlist.rl;
+                BUG_ON(!rl || vcn < rl->vcn || !rl->length);
+        }
+        /* Seek to the runlist element containing @vcn. */
+        while (rl->length && vcn >= rl[1].vcn)
+                rl++;
+        do {
+                LCN lcn;
+                sector_t block, end_block;
+                s64 len;
+
+                /*
+                 * If this run is not mapped map it now and start again as the
+                 * runlist will have been updated.
+                 */
+                lcn = rl->lcn;
+                if (unlikely(lcn == LCN_RL_NOT_MAPPED)) {
+                        vcn = rl->vcn;
+                        goto map_vcn;
+                }
+                /* If this run is not valid abort with an error. */
+                if (unlikely(!rl->length || lcn < LCN_HOLE))
+                        goto rl_err;
+                /* Skip holes. */
+                if (lcn == LCN_HOLE)
+                        continue;
+                block = lcn << vol->cluster_size_bits >> block_size_bits;
+                len = rl->length;
+                if (rl[1].vcn > end_vcn)
+                        len = end_vcn - rl->vcn;
+                end_block = (lcn + len) << vol->cluster_size_bits >>
+                                block_size_bits;
+                /* Iterate over the blocks in the run and empty them. */
+                do {
+                        struct buffer_head *bh;
+
+                        /* Obtain the buffer, possibly not uptodate. */
+                        bh = sb_getblk(sb, block);
+                        BUG_ON(!bh);
+                        /* Setup buffer i/o submission. */
+                        lock_buffer(bh);
+                        bh->b_end_io = end_buffer_write_sync;
+                        get_bh(bh);
+                        /* Set the entire contents of the buffer to 0xff. */
+                        memset(bh->b_data, -1, block_size);
+                        if (!buffer_uptodate(bh))
+                                set_buffer_uptodate(bh);
+                        if (buffer_dirty(bh))
+                                clear_buffer_dirty(bh);
+                        /*
+                         * Submit the buffer and wait for i/o to complete but
+                         * only for the first buffer so we do not miss really
+                         * serious i/o errors.  Once the first buffer has
+                         * completed ignore errors afterwards as we can assume
+                         * that if one buffer worked all of them will work.
+                         */
+                        submit_bh(WRITE, bh);
+                        if (should_wait) {
+                                should_wait = false;
+                                wait_on_buffer(bh);
+                                if (unlikely(!buffer_uptodate(bh)))
+                                        goto io_err;
+                        }
+                        brelse(bh);
+                } while (++block < end_block);
+        } while ((++rl)->vcn < end_vcn);
+        up_write(&log_ni->runlist.lock);
+        /*
+         * Zap the pages again just in case any got instantiated whilst we were
+         * emptying the blocks by hand.  FIXME: We may not have completed
+         * writing to all the buffer heads yet so this may happen too early.
+         * We really should use a kernel thread to do the emptying
+         * asynchronously and then we can also set the volume dirty and output
+         * an error message if emptying should fail.
+         */
+        truncate_inode_pages(log_vi->i_mapping, 0);
+        /* Set the flag so we do not have to do it again on remount. */
+        NVolSetLogFileEmpty(vol);
         ntfs_debug("Done.");
         return true;
+io_err:
+        ntfs_error(sb, "Failed to write buffer.  Unmount and run chkdsk.");
+        goto dirty_err;
+rl_err:
+        ntfs_error(sb, "Runlist is corrupt.  Unmount and run chkdsk.");
+dirty_err:
+        NVolSetErrors(vol);
+        err = -EIO;
+err:
+        up_write(&log_ni->runlist.lock);
+        ntfs_error(sb, "Failed to fill $LogFile with 0xff bytes (error %d).",
+                        -err);
+        return false;
 }
 
 #endif /* NTFS_RW */