The block validity framework does a more comprehensive set of checks,
and it saves object code space to use the ext4_data_block_valid() than
the limited open-coded version that had been in ext4_free_blocks().
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Add the facility for ext4_forget() to be called from
ext4_free_blocks(). This simplifies the code in a large number of
places, and centralizes most of the work of calling ext4_forget() into
a single place.
Also fix a bug in the extents migration code; it wasn't calling
ext4_forget() when releasing the indirect blocks during the
conversion. As a result, if the system cashed during or shortly after
the extents migration, and the released indirect blocks get reused as
data blocks, the journal replay would corrupt the data blocks. With
this new patch, fixing this bug was as simple as adding the
EXT4_FREE_BLOCKS_FORGET flags to the call to ext4_free_blocks().
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
ext4_mb_free_blocks() is only called by ext4_free_blocks(), and the
latter function doesn't really do much. So merge the two functions
together, such that ext4_free_blocks() is now found in
fs/ext4/mballoc.c. This saves about 200 bytes of compiled text space.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Convert the last two callers of ext4_journal_forget() to use
ext4_forget() instead, and then fold ext4_journal_forget() into
ext4_forget(). This reduces are code complexity and shortens our call
stack.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
The only caller of ext4_journal_revoke() is ext4_forget(), so we can
fold ext4_journal_revoke() into ext4_forget() to simplify the code and
shorten the call stack.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
The ext4_forget() function better belongs in ext4_jbd2.c. This will
allow us to do some cleanup of the ext4_journal_revoke() and
ext4_journal_forget() functions, as well as giving us better error
reporting since we can report the caller of ext4_forget() when things
go wrong.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Users on the linux-ext4 list recently complained about differences
across filesystems w.r.t. how to mount without a journal replay.
In the discussion it was noted that xfs's "norecovery" option is
perhaps more descriptively accurate than "noload," so let's make
that an alias for ext4.
Also show this status in /proc/mounts
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
It is anticipated that when sb_issue_discard starts doing
real work on trim-capable devices, we may see issues. Make
this mount-time optional, and default it to off until we know
that things are working out OK.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
When an error happened in ext4_splice_branch we failed to notice that
in ext4_ind_get_blocks and mapped the buffer anyway. Fix the problem
by checking for error properly.
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Cc: stable@kernel.org
We don't to issue an I/O barrier on an error or if we force commit
because we are doing data journaling.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Cc: Jan Kara <jack@suse.cz>
Cc: stable@kernel.org
The block validity checks used by ext4_data_block_valid() wasn't
correctly written to check file systems with the meta_bg feature. Fix
this.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Cc: stable@kernel.org
The number of old-style block group descriptor blocks is
s_meta_first_bg when the meta_bg feature flag is set.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Cc: stable@kernel.org
commit a71ce8c6c9 updated ext4_statfs()
to update the on-disk superblock counters, but modified this buffer
directly without any journaling of the change. This is one of the
accesses that was causing the crc errors in journal replay as seen in
kernel.org bugzilla #14354.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Cc: stable@kernel.org
ext4_xattr_set_handle() was zeroing out an inode outside
of journaling constraints; this is one of the accesses that
was causing the crc errors in journal replay as seen in
kernel.org bugzilla #14354.
Reviewed-by: Andreas Dilger <adilger@sun.com>
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Cc: stable@kernel.org
We need to be testing the i_flags field in the ext4 specific portion
of the inode, instead of the (confusingly aliased) i_flags field in
the generic struct inode.
Signed-off-by: Julia Lawall <julia@diku.dk>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Cc: stable@kernel.org
When an inode gets unlinked, the functions ext4_clear_blocks() and
ext4_remove_blocks() call ext4_forget() for all the buffer heads
corresponding to the deleted inode's data blocks. If the inode is a
directory or a symlink, the is_metadata parameter must be non-zero so
ext4_forget() will revoke them via jbd2_journal_revoke(). Otherwise,
if these blocks are reused for a data file, and the system crashes
before a journal checkpoint, the journal replay could end up
corrupting these data blocks.
Thanks to Curt Wohlgemuth for pointing out potential problems in this
area.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Cc: stable@kernel.org
Now that we are checking for failed journal checksums in the jbd2
layer, we don't need to check in the ext4 mount path --- since a
checksum fail will result in ext4_load_journal() returning an error,
causing the file system to refuse to be mounted until e2fsck can deal
with the problem.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
If there is a failed journal checksum, don't reset the journal. This
allows for userspace programs to decide how to recover from this
situation. It may be that ignoring the journal checksum failure might
be a better way of recovering the file system. Once we add per-block
checksums, we can definitely do better. Until then, a system
administrator can try backing up the file system image (or taking a
snapshot) and and trying to determine experimentally whether ignoring
the checksum failure or aborting the journal replay results in less
data loss.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Cc: stable@kernel.org
One of the invalid error paths in ext4_iget() forgot to brelse() the
inode buffer head. Fix it by adding a brelse() in the common error
return path, which also simplifies function.
Thanks to Andi Kleen <ak@linux.intel.com> reporting the problem.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
If CONFIG_PROVE_LOCKING is enabled, the double_down_write_data_sem()
will trigger a false-positive warning of a recursive lock. Since we
take i_data_sem for the two inodes ordered by their inode numbers,
this isn't a problem. Use of down_write_nested() will notify the lock
dependency checker machinery that there is no problem here.
This problem was reported by Brian Rogers:
http://marc.info/?l=linux-ext4&m=125115356928011&w=1
Reported-by: Brian Rogers <brian@xyzw.org>
Signed-off-by: Akira Fujita <a-fujita@rs.jp.nec.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
ext4_move_extents() checks the logical block contiguousness
of original file with ext4_find_extent() and mext_next_extent().
Therefore the extent which ext4_ext_path structure indicates
must not be changed between above functions.
But in current implementation, there is no i_data_sem protection
between ext4_ext_find_extent() and mext_next_extent(). So the extent
which ext4_ext_path structure indicates may be overwritten by
delalloc. As a result, ext4_move_extents() will exchange wrong blocks
between original and donor files. I change the place where
acquire/release i_data_sem to solve this problem.
Moreover, I changed move_extent_per_page() to start transaction first,
and then acquire i_data_sem. Without this change, there is a
possibility of the deadlock between mmap() and ext4_move_extents():
* NOTE: "A", "B" and "C" mean different processes
A-1: ext4_ext_move_extents() acquires i_data_sem of two inodes.
B: do_page_fault() starts the transaction (T),
and then tries to acquire i_data_sem.
But process "A" is already holding it, so it is kept waiting.
C: While "A" and "B" running, kjournald2 tries to commit transaction (T)
but it is under updating, so kjournald2 waits for it.
A-2: Call ext4_journal_start with holding i_data_sem,
but transaction (T) is locked.
Signed-off-by: Akira Fujita <a-fujita@rs.jp.nec.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
If the EXT4_IOC_MOVE_EXT ioctl fails, the number of blocks that were
exchanged before the failure should be returned to the userspace
caller. Unfortunately, currently if the block size is not the same as
the page size, the returned block count that is returned is the
page-aligned block count instead of the actual block count. This
commit addresses this bug.
Signed-off-by: Akira Fujita <a-fujita@rs.jp.nec.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
If s_log_groups_per_flex is greater than 31, then groups_per_flex will
will overflow and cause a divide by zero error. This can cause kernel
BUG if such a file system is mounted.
Thanks to Nageswara R Sastry for analyzing the failure and providing
an initial patch.
http://bugzilla.kernel.org/show_bug.cgi?id=14287
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Cc: stable@kernel.org
Previously add_dirent_to_buf() did not free its passed-in buffer head
in the case of ENOSPC, since in some cases the caller still needed it.
However, this led to potential buffer head leaks since not all callers
dealt with this correctly. Fix this by making simplifying the freeing
convention; now add_dirent_to_buf() *never* frees the passed-in buffer
head, and leaves that to the responsibility of its caller. This makes
things cleaner and easier to prove that the code is neither leaking
buffer heads or calling brelse() one time too many.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Cc: Curt Wohlgemuth <curtw@google.com>
Cc: stable@kernel.org
Because of an integer overflow on start_blk, various kind of wrong results
would be returned by the generic_block_fiemap() handler, such as no
extents when there is a 4GB+ hole at the beginning of the file, or wrong
fe_logical when an extent starts after the first 4GB.
Signed-off-by: Mike Hommey <mh@glandium.org>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Cc: Theodore Ts'o <tytso@mit.edu>
Cc: Eric Sandeen <sandeen@sgi.com>
Cc: Josef Bacik <jbacik@redhat.com>
Cc: Mark Fasheh <mfasheh@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
In setup_arg_pages we work hard to assign a value to ret, but on exit we
always return 0.
Also remove a now duplicated exit path and branch to out_unlock instead.
Signed-off-by: Anton Blanchard <anton@samba.org>
Acked-by: Serge Hallyn <serue@us.ibm.com>
Reviewed-by: WANG Cong <xiyou.wangcong@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
For FS_IOC_RESVSP and FS_IOC_RESVSP64 compat_sys_ioctl() uses its
arg argument as a pointer to userspace. However it is missing a
a call to compat_ptr() which will do a proper pointer conversion.
This was introduced with 3e63cbb1 "fs: Add new pre-allocation ioctls
to vfs for compatibility with legacy xfs ioctls".
Signed-off-by: Heiko Carstens <heiko.carstens@de.ibm.com>
Cc: Ankit Jain <me@ankitjain.org>
Acked-by: Christoph Hellwig <hch@lst.de>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Acked-by: Arnd Bergmann <arndbergmann@googlemail.com>
Acked-by: David S. Miller <davem@davemloft.net>
Cc: <stable@kernel.org> [2.6.31.x]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Daniel Lezcano reported a leak in 'struct pid' and 'struct pid_namespace'
that is discussed in:
http://lkml.org/lkml/2009/10/2/159.
To summarize the thread, when container-init is terminated, it sets the
PF_EXITING flag, zaps other processes in the container and waits to reap
them. As a part of reaping, the container-init should flush any /proc
dentries associated with the processes. But because the container-init is
itself exiting and the following PF_EXITING check, the dentries are not
flushed, resulting in leak in /proc inodes and dentries.
This fix reverts the commit 7766755a2f ("Fix /proc dcache deadlock
in do_exit") which introduced the check for PF_EXITING. At the time of
the commit, shrink_dcache_parent() flushed dentries from other filesystems
also and could have caused a deadlock which the commit fixed. But as
pointed out by Eric Biederman, after commit 0feae5c47a,
shrink_dcache_parent() no longer affects other filesystems. So reverting
the commit is now safe.
As pointed out by Jan Kara, the leak is not as critical since the
unclaimed space will be reclaimed under memory pressure or by:
echo 3 > /proc/sys/vm/drop_caches
But since this check is no longer required, its best to remove it.
Signed-off-by: Sukadev Bhattiprolu <sukadev@us.ibm.com>
Reported-by: Daniel Lezcano <dlezcano@fr.ibm.com>
Acked-by: Eric W. Biederman <ebiederm@xmission.com>
Acked-by: Jan Kara <jack@ucw.cz>
Cc: Andrea Arcangeli <andrea@cpushare.com>
Cc: Serge Hallyn <serue@us.ibm.com>
Cc: <stable@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
* git://git.kernel.org/pub/scm/linux/kernel/git/mason/btrfs-unstable:
Btrfs: fix panic when trying to destroy a newly allocated
Btrfs: allow more metadata chunk preallocation
Btrfs: fallback on uncompressed io if compressed io fails
Btrfs: find ideal block group for caching
Btrfs: avoid null deref in unpin_extent_cache()
Btrfs: skip btrfs_release_path in btrfs_update_root and btrfs_del_root
Btrfs: fix some metadata enospc issues
Btrfs: fix how we set max_size for free space clusters
Btrfs: cleanup transaction starting and fix journal_info usage
Btrfs: fix data allocation hint start
There is a problem where iget5_locked will look for an inode, not find it, and
then subsequently try to allocate it. Another CPU will have raced in and
allocated the inode instead, so when iget5_locked gets the inode spin lock again
and does a search, it finds the new inode. So it goes ahead and calls
destroy_inode on the inode it just allocated. The problem is we don't set
BTRFS_I(inode)->root until the new inode is completely initialized. This patch
makes us set root to NULL when alloc'ing a new inode, so when we get to
btrfs_destroy_inode and we see that root is NULL we can just free up the memory
and continue on. This fixes the panic
http://www.kerneloops.org/submitresult.php?number=812690
Thanks,
Signed-off-by: Josef Bacik <josef@redhat.com>
Signed-off-by: Chris Mason <chris.mason@oracle.com>
* 'for_linus' of git://git.kernel.org/pub/scm/linux/kernel/git/jack/linux-fs-2.6:
JBD/JBD2: free j_wbuf if journal init fails.
ext3: Wait for proper transaction commit on fsync
ext3: retry failed direct IO allocations
* 'for_linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tytso/ext4:
ext4: partial revert to fix double brelse WARNING()
ext4: Fix return value of ext4_split_unwritten_extents() to fix direct I/O
ext4: code clean up for dio fallocate handling
ext4: skip conversion of uninit extents after direct IO if there isn't any
ext4: fix ext4_ext_direct_IO()'s return value after converting uninit extents
ext4: discard preallocation when restarting a transaction during truncate
On an FS where all of the space has not been allocated into chunks yet,
the enospc can return enospc just because the existing metadata chunks
are full.
We get around this by allowing more metadata chunks to be allocated up
to a certain limit, and finding the right limit is a little fuzzy. The
problem is the reservations for delalloc would preallocate way too much
of the FS as metadata. We need to start saying no and just force some
IO to happen.
But we also need to let a reasonable amount of the FS become metadata.
This bumps the hard limit up, later releases will have a better system.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
Currently compressed IO does not deal with not having its entire extent able to
be allocated. So if we have enough free space to allocate for the extent, but
its not contiguous, it will fail spectacularly. This patch fixes this by
falling back on uncompressed IO which lets us spread the delalloc extent across
multiple extents. I tested this by making us randomly think the reservation had
failed to make it fallback on the uncompressed io way and it seemed to work
fine. Thanks,
Signed-off-by: Josef Bacik <josef@redhat.com>
Signed-off-by: Chris Mason <chris.mason@oracle.com>
This patch changes a few things. Hopefully the comments are helpfull, but
I'll try and be as verbose here.
Problem:
My fedora box was taking 1 minute and 21 seconds to boot with btrfs as root.
Part of this problem was we pick the first block group we can find and start
caching it, even if it may not have enough free space. The other problem is
we only search for cached block groups the first time around, which we won't
find any cached block groups because this is a newly mounted fs, so we end up
caching several block groups during bootup, which with alot of fragmentation
takes around 30-45 seconds to complete, which bogs down the system. So
Solution:
1) Don't cache block groups willy-nilly at first. Instead try and figure out
which block group has the most free, and therefore will take the least amount
of time to cache.
2) Don't be so picky about cached block groups. The other problem is once
we've filled up a cluster, if the block group isn't finished caching the next
time we try and do the allocation we'll completely ignore the cluster and
start searching from the beginning of the space, which makes us cache more
block groups, which slows us down even more. So instead of skipping block
groups that are not finished caching when we have a hint, only skip the block
group if it hasn't started caching yet.
There is one other tweak in here. Before if we allocated a chunk and still
couldn't find new space, we'd end up switching the space info to force another
chunk allocation. This could make us end up with way too many chunks, so keep
track of this particular case.
With this patch and my previous cluster fixes my fedora box now boots in 43
seconds, and according to the bootchart is not held up by our block group
caching at all.
Signed-off-by: Josef Bacik <josef@redhat.com>
Signed-off-by: Chris Mason <chris.mason@oracle.com>
I re-orderred the checks to avoid dereferencing "em" if it was null.
Found by smatch static checker.
Signed-off-by: Dan Carpenter <error27@gmail.com>
Signed-off-by: Chris Mason <chris.mason@oracle.com>
We don't need to call btrfs_release_path because btrfs_free_path will do
that for us.
Signed-off-by: Li Dongyang <Jerry87905@gmail.com>
Signed-off-by: Chris Mason <chris.mason@oracle.com>
We weren't reserving metadata space for rename, rmdir and unlink, which could
cause problems.
Signed-off-by: Josef Bacik <josef@redhat.com>
Signed-off-by: Chris Mason <chris.mason@oracle.com>
This patch fixes a problem where max_size can be set to 0 even though we
filled the cluster properly. We set max_size to 0 if we restart the cluster
window, but if the new start entry is big enough to be our new cluster then we
could return with a max_size set to 0, which will mean the next time we try to
allocate from this cluster it will fail. So set max_extent to the entry's
size. Tested this on my box and now we actually allocate from the cluster
after we fill it. Thanks,
Signed-off-by: Josef Bacik <josef@redhat.com>
Signed-off-by: Chris Mason <chris.mason@oracle.com>
We use journal_info to tell if we're in a nested transaction to make sure we
don't commit the transaction within a nested transaction. We use another
method to see if there are any outstanding ioctl trans handles, so if we're
starting one do not set current->journal_info, since it will screw with other
filesystems. This patch also cleans up the starting stuff so there aren't any
magic numbers.
Signed-off-by: Josef Bacik <josef@redhat.com>
Signed-off-by: Chris Mason <chris.mason@oracle.com>
Sometimes our start allocation hint when we cow a file can be either
EXTENT_HOLE or some other such place holder, which is not optimal. So if we
find that our em->block_start is one of these special values, check to see
where the first block of the inode is stored, and use that as a hint. If that
block is also a special value, just fallback on a hint of 0 and let the
allocator figure out a good place to put the data.
Signed-off-by: Josef Bacik <josef@redhat.com>
Signed-off-by: Chris Mason <chris.mason@oracle.com>
If journal init fails, we need to free j_wbuf.
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Jan Kara <jack@suse.cz>
Signed-off-by: Tao Ma <tao.ma@oracle.com>
Signed-off-by: Jan Kara <jack@suse.cz>
We cannot rely on buffer dirty bits during fsync because pdflush can come
before fsync is called and clear dirty bits without forcing a transaction
commit. What we do is that we track which transaction has last changed
the inode and which transaction last changed allocation and force it to
disk on fsync.
Signed-off-by: Jan Kara <jack@suse.cz>
Reviewed-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
On a 256M 4k block filesystem, doing this in a loop:
dd if=/dev/zero of=test oflag=direct bs=1M count=64
rm -f test
eventually leads to spurious ENOSPC:
dd: writing `test': No space left on device
As with other block allocation callers, it looks like we need to
potentially retry the allocations on the initial ENOSPC.
A similar patch went into ext4 (commit
fbbf694566)
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: Jan Kara <jack@suse.cz>
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/ryusuke/nilfs2:
nilfs2: fix missing cleanup of gc cache on error cases
nilfs2: fix kernel oops in error case of nilfs_ioctl_move_blocks
This is a partial revert of commit 6487a9d (only the changes made to
fs/ext4/namei.c), since it is causing the following brelse()
double-free warning when running fsstress on a file system with 1k
blocksize and we run into a block allocation failure while converting
a single-block directory to a multi-block hash-tree indexed directory.
WARNING: at fs/buffer.c:1197 __brelse+0x2e/0x33()
Hardware name:
VFS: brelse: Trying to free free buffer
Modules linked in:
Pid: 2226, comm: jbd2/sdd-8 Not tainted 2.6.32-rc6-00577-g0003f55 #101
Call Trace:
[<c01587fb>] warn_slowpath_common+0x65/0x95
[<c0158869>] warn_slowpath_fmt+0x29/0x2c
[<c021168e>] __brelse+0x2e/0x33
[<c0288a9f>] jbd2_journal_refile_buffer+0x67/0x6c
[<c028a9ed>] jbd2_journal_commit_transaction+0x319/0x14d8
[<c0164d73>] ? try_to_del_timer_sync+0x58/0x60
[<c0175bcc>] ? sched_clock_cpu+0x12a/0x13e
[<c017f6b4>] ? trace_hardirqs_off+0xb/0xd
[<c0175c1f>] ? cpu_clock+0x3f/0x5b
[<c017f6ec>] ? lock_release_holdtime+0x36/0x137
[<c0664ad0>] ? _spin_unlock_irqrestore+0x44/0x51
[<c0180af3>] ? trace_hardirqs_on_caller+0x103/0x124
[<c0180b1f>] ? trace_hardirqs_on+0xb/0xd
[<c0164d73>] ? try_to_del_timer_sync+0x58/0x60
[<c0290d1c>] kjournald2+0x11a/0x310
[<c017118e>] ? autoremove_wake_function+0x0/0x38
[<c0290c02>] ? kjournald2+0x0/0x310
[<c0170ee6>] kthread+0x66/0x6b
[<c0170e80>] ? kthread+0x0/0x6b
[<c01251b3>] kernel_thread_helper+0x7/0x10
---[ end trace 5579351b86af61e3 ]---
Commit 6487a9d was an attempt some buffer head leaks in an ENOSPC
error path, but in some cases it actually results in an excess ENOSPC,
as shown above. Fixing this means cleaning up who is responsible for
releasing the buffer heads from the callee to the caller of
add_dirent_to_buf().
Since that's a relatively complex change, and we're late in the rcX
development cycle, I'm reverting this now, and holding back a more
complete fix until after 2.6.32 ships. We've lived with this
buffer_head leak on ENOSPC in ext3 and ext4 for a very long time; a
few more months won't kill us.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Cc: Curt Wohlgemuth <curtw@google.com>
