Mostly supporting filesystems outside of init_user_ns is
s/&init_usre_ns/dquot->dq_sb->s_user_ns/. An actual need for
supporting quotas on filesystems outside of s_user_ns is quite a ways
away and to be done responsibily needs an audit on what can happen
with hostile quota files. Until that audit is complete don't attempt
to support quota files on filesystems outside of s_user_ns.
Cc: Jan Kara <jack@suse.cz>
Acked-by: Seth Forshee <seth.forshee@canonical.com>
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
In Q_XSETQLIMIT use sb->s_user_ns to detect when we are dealing with
the filesystems notion of id 0.
Cc: Jan Kara <jack@suse.cz>
Acked-by: Seth Forshee <seth.forshee@canonical.com>
Inspired-by: Seth Forshee <seth.forshee@canonical.com>
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
Introduce the helper qid_has_mapping and use it to ensure that the
quota system only considers qids that map to the filesystems
s_user_ns.
In practice for quota supporting filesystems today this is the exact
same check as qid_valid. As only 0xffffffff aka (qid_t)-1 does not
map into init_user_ns.
Replace the qid_valid calls with qid_has_mapping as values come in
from userspace. This is harmless today and it prepares the quota
system to work on filesystems with quotas but mounted by unprivileged
users.
Call qid_has_mapping from dqget. This ensures the passed in qid has a
prepresentation on the underlying filesystem. Previously this was
unnecessary as filesystesm never had qids that could not map. With
the introduction of filesystems outside of s_user_ns this will not
remain true.
All of this ensures the quota code never has to deal with qids that
don't map to the underlying filesystem.
Cc: Jan Kara <jack@suse.cz>
Acked-by: Seth Forshee <seth.forshee@canonical.com>
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
It is expected that filesystems can not represent uids and gids from
outside of their user namespace. Keep things simple by not even
trying to create filesystem nodes with non-sense uids and gids.
Acked-by: Seth Forshee <seth.forshee@canonical.com>
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
When a filesystem outside of init_user_ns is mounted it could have
uids and gids stored in it that do not map to init_user_ns.
The plan is to allow those filesystems to set i_uid to INVALID_UID and
i_gid to INVALID_GID for unmapped uids and gids and then to handle
that strange case in the vfs to ensure there is consistent robust
handling of the weirdness.
Upon a careful review of the vfs and filesystems about the only case
where there is any possibility of confusion or trouble is when the
inode is written back to disk. In that case filesystems typically
read the inode->i_uid and inode->i_gid and write them to disk even
when just an inode timestamp is being updated.
Which leads to a rule that is very simple to implement and understand
inodes whose i_uid or i_gid is not valid may not be written.
In dealing with access times this means treat those inodes as if the
inode flag S_NOATIME was set. Reads of the inodes appear safe and
useful, but any write or modification is disallowed. The only inode
write that is allowed is a chown that sets the uid and gid on the
inode to valid values. After such a chown the inode is normal and may
be treated as such.
Denying all writes to inodes with uids or gids unknown to the vfs also
prevents several oddball cases where corruption would have occurred
because the vfs does not have complete information.
One problem case that is prevented is attempting to use the gid of a
directory for new inodes where the directories sgid bit is set but the
directories gid is not mapped.
Another problem case avoided is attempting to update the evm hash
after setxattr, removexattr, and setattr. As the evm hash includeds
the inode->i_uid or inode->i_gid not knowning the uid or gid prevents
a correct evm hash from being computed. evm hash verification also
fails when i_uid or i_gid is unknown but that is essentially harmless
as it does not cause filesystem corruption.
Acked-by: Seth Forshee <seth.forshee@canonical.com>
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
Filesystem uids which don't map into a user namespace may result
in inode->i_uid being INVALID_UID. A symlink and its parent
could have different owners in the filesystem can both get
mapped to INVALID_UID, which may result in following a symlink
when this would not have otherwise been permitted when protected
symlinks are enabled.
Signed-off-by: Seth Forshee <seth.forshee@canonical.com>
Acked-by: Serge Hallyn <serge.hallyn@canonical.com>
Signed-off-by: Eric W. Biederman <ebiederm@xmission.com>
Update posix_acl_valid to verify that an acl is within a user namespace.
Update the callers of posix_acl_valid to pass in an appropriate
user namespace. For posix_acl_xattr_set and v9fs_xattr_set_acl pass in
inode->i_sb->s_user_ns to posix_acl_valid. For md_unpack_acl pass in
&init_user_ns as no inode or superblock is in sight.
Acked-by: Seth Forshee <seth.forshee@canonical.com>
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
Add checks to notify_change to verify that uid and gid changes
will map into the superblock's user namespace. If they do not
fail with -EOVERFLOW.
This is mandatory so that fileystems don't have to even think
of dealing with ia_uid and ia_gid that
--EWB Moved the test from inode_change_ok to notify_change
Signed-off-by: Seth Forshee <seth.forshee@canonical.com>
Acked-by: Serge Hallyn <serge.hallyn@canonical.com>
Signed-off-by: Eric W. Biederman <ebiederm@xmission.com>
If a process gets access to a mount from a different user
namespace, that process should not be able to take advantage of
setuid files or selinux entrypoints from that filesystem. Prevent
this by treating mounts from other mount namespaces and those not
owned by current_user_ns() or an ancestor as nosuid.
This will make it safer to allow more complex filesystems to be
mounted in non-root user namespaces.
This does not remove the need for MNT_LOCK_NOSUID. The setuid,
setgid, and file capability bits can no longer be abused if code in
a user namespace were to clear nosuid on an untrusted filesystem,
but this patch, by itself, is insufficient to protect the system
from abuse of files that, when execed, would increase MAC privilege.
As a more concrete explanation, any task that can manipulate a
vfsmount associated with a given user namespace already has
capabilities in that namespace and all of its descendents. If they
can cause a malicious setuid, setgid, or file-caps executable to
appear in that mount, then that executable will only allow them to
elevate privileges in exactly the set of namespaces in which they
are already privileges.
On the other hand, if they can cause a malicious executable to
appear with a dangerous MAC label, running it could change the
caller's security context in a way that should not have been
possible, even inside the namespace in which the task is confined.
As a hardening measure, this would have made CVE-2014-5207 much
more difficult to exploit.
Signed-off-by: Andy Lutomirski <luto@amacapital.net>
Signed-off-by: Seth Forshee <seth.forshee@canonical.com>
Acked-by: James Morris <james.l.morris@oracle.com>
Acked-by: Serge Hallyn <serge.hallyn@canonical.com>
Signed-off-by: Eric W. Biederman <ebiederm@xmission.com>
Now that SB_I_NODEV controls the nodev behavior devpts can just clear
this flag during mount. Simplifying the code and making it easier
to audit how the code works. While still preserving the invariant
that s_iflags is only modified during mount.
Acked-by: Seth Forshee <seth.forshee@canonical.com>
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
Replace the implict setting of MNT_NODEV on mounts that happen with
just user namespace permissions with an implicit setting of SB_I_NODEV
in s_iflags. The visibility of the implicit MNT_NODEV has caused
problems in the past.
With this change the fragile case where an implicit MNT_NODEV needs to
be preserved in do_remount is removed. Using SB_I_NODEV is much less
fragile as s_iflags are set during the original mount and never
changed.
In do_new_mount with the implicit setting of MNT_NODEV gone, the only
code that can affect mnt_flags is fs_fully_visible so simplify the if
statement and reduce the indentation of the code to make that clear.
Acked-by: Seth Forshee <seth.forshee@canonical.com>
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
Verify all filesystems that we check in mount_too_revealing set
SB_I_NOEXEC and SB_I_NODEV in sb->s_iflags. That is true for today
and it should remain true in the future.
Remove the now unnecessary checks from mnt_already_visibile that
ensure MNT_LOCK_NOSUID, MNT_LOCK_NOEXEC, and MNT_LOCK_NODEV are
preserved. Making the code shorter and easier to read.
Relying on SB_I_NOEXEC and SB_I_NODEV instead of the user visible
MNT_NOSUID, MNT_NOEXEC, and MNT_NODEV ensures the many current
systems where proc and sysfs are mounted with "nosuid, nodev, noexec"
and several slightly buggy container applications don't bother to
set those flags continue to work.
Acked-by: Seth Forshee <seth.forshee@canonical.com>
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
Introduce a function may_open_dev that tests MNT_NODEV and a new
superblock flab SB_I_NODEV. Use this new function in all of the
places where MNT_NODEV was previously tested.
Add the new SB_I_NODEV s_iflag to proc, sysfs, and mqueuefs as those
filesystems should never support device nodes, and a simple superblock
flags makes that very hard to get wrong. With SB_I_NODEV set if any
device nodes somehow manage to show up on on a filesystem those
device nodes will be unopenable.
Acked-by: Seth Forshee <seth.forshee@canonical.com>
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
The cgroup filesystem is in the same boat as sysfs. No one ever
permits executables of any kind on the cgroup filesystem, and there is
no reasonable future case to support executables in the future.
Therefore move the setting of SB_I_NOEXEC which makes the code proof
against future mistakes of accidentally creating executables from
sysfs to kernfs itself. Making the code simpler and covering the
sysfs, cgroup, and cgroup2 filesystems.
Acked-by: Seth Forshee <seth.forshee@canonical.com>
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
Allowing a filesystem to be mounted by other than root in the initial
user namespace is a filesystem property not a mount namespace property
and as such should be checked in filesystem specific code. Move the
FS_USERNS_MOUNT test into super.c:sget_userns().
Acked-by: Seth Forshee <seth.forshee@canonical.com>
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
Start marking filesystems with a user namespace owner, s_user_ns. In
this change this is only used for permission checks of who may mount a
filesystem. Ultimately s_user_ns will be used for translating ids and
checking capabilities for filesystems mounted from user namespaces.
The default policy for setting s_user_ns is implemented in sget(),
which arranges for s_user_ns to be set to current_user_ns() and to
ensure that the mounter of the filesystem has CAP_SYS_ADMIN in that
user_ns.
The guts of sget are split out into another function sget_userns().
The function sget_userns calls alloc_super with the specified user
namespace or it verifies the existing superblock that was found
has the expected user namespace, and fails with EBUSY when it is not.
This failing prevents users with the wrong privileges mounting a
filesystem.
The reason for the split of sget_userns from sget is that in some
cases such as mount_ns and kernfs_mount_ns a different policy for
permission checking of mounts and setting s_user_ns is necessary, and
the existence of sget_userns() allows those policies to be
implemented.
The helper mount_ns is expected to be used for filesystems such as
proc and mqueuefs which present per namespace information. The
function mount_ns is modified to call sget_userns instead of sget to
ensure the user namespace owner of the namespace whose information is
presented by the filesystem is used on the superblock.
For sysfs and cgroup the appropriate permission checks are already in
place, and kernfs_mount_ns is modified to call sget_userns so that
the init_user_ns is the only user namespace used.
For the cgroup filesystem cgroup namespace mounts are bind mounts of a
subset of the full cgroup filesystem and as such s_user_ns must be the
same for all of them as there is only a single superblock.
Mounts of sysfs that vary based on the network namespace could in principle
change s_user_ns but it keeps the analysis and implementation of kernfs
simpler if that is not supported, and at present there appear to be no
benefits from supporting a different s_user_ns on any sysfs mount.
Getting the details of setting s_user_ns correct has been
a long process. Thanks to Pavel Tikhorirorv who spotted a leak
in sget_userns. Thanks to Seth Forshee who has kept the work alive.
Thanks-to: Seth Forshee <seth.forshee@canonical.com>
Thanks-to: Pavel Tikhomirov <ptikhomirov@virtuozzo.com>
Acked-by: Seth Forshee <seth.forshee@canonical.com>
Signed-off-by: Eric W. Biederman <ebiederm@xmission.com>
Move the call of get_pid_ns, the call of proc_parse_options, and
the setting of s_iflags into proc_fill_super so that mount_ns
can be used.
Convert proc_mount to call mount_ns and remove the now unnecessary
code.
Acked-by: Seth Forshee <seth.forshee@canonical.com>
Reviewed-by: Djalal Harouni <tixxdz@gmail.com>
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
Today what is normally called data (the mount options) is not passed
to fill_super through mount_ns.
Pass the mount options and the namespace separately to mount_ns so
that filesystems such as proc that have mount options, can use
mount_ns.
Pass the user namespace to mount_ns so that the standard permission
check that verifies the mounter has permissions over the namespace can
be performed in mount_ns instead of in each filesystems .mount method.
Thus removing the duplication between mqueuefs and proc in terms of
permission checks. The extra permission check does not currently
affect the rpc_pipefs filesystem and the nfsd filesystem as those
filesystems do not currently allow unprivileged mounts. Without
unpvileged mounts it is guaranteed that the caller has already passed
capable(CAP_SYS_ADMIN) which guarantees extra permission check will
pass.
Update rpc_pipefs and the nfsd filesystem to ensure that the network
namespace reference is always taken in fill_super and always put in kill_sb
so that the logic is simpler and so that errors originating inside of
fill_super do not cause a network namespace leak.
Acked-by: Seth Forshee <seth.forshee@canonical.com>
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
Replace the call of fs_fully_visible in do_new_mount from before the
new superblock is allocated with a call of mount_too_revealing after
the superblock is allocated. This winds up being a much better location
for maintainability of the code.
The first change this enables is the replacement of FS_USERNS_VISIBLE
with SB_I_USERNS_VISIBLE. Moving the flag from struct filesystem_type
to sb_iflags on the superblock.
Unfortunately mount_too_revealing fundamentally needs to touch
mnt_flags adding several MNT_LOCKED_XXX flags at the appropriate
times. If the mnt_flags did not need to be touched the code
could be easily moved into the filesystem specific mount code.
Acked-by: Seth Forshee <seth.forshee@canonical.com>
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
In rare cases it is possible for s_flags & MS_RDONLY to be set but
MNT_READONLY to be clear. This starting combination can cause
fs_fully_visible to fail to ensure that the new mount is readonly.
Therefore force MNT_LOCK_READONLY in the new mount if MS_RDONLY
is set on the source filesystem of the mount.
In general both MS_RDONLY and MNT_READONLY are set at the same for
mounts so I don't expect any programs to care. Nor do I expect
MS_RDONLY to be set on proc or sysfs in the initial user namespace,
which further decreases the likelyhood of problems.
Which means this change should only affect system configurations by
paranoid sysadmins who should welcome the additional protection
as it keeps people from wriggling out of their policies.
Cc: stable@vger.kernel.org
Fixes: 8c6cf9cc82 ("mnt: Modify fs_fully_visible to deal with locked ro nodev and atime")
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
MNT_LOCKED implies on a child mount implies the child is locked to the
parent. So while looping through the children the children should be
tested (not their parent).
Typically an unshare of a mount namespace locks all mounts together
making both the parent and the slave as locked but there are a few
corner cases where other things work.
Cc: stable@vger.kernel.org
Fixes: ceeb0e5d39 ("vfs: Ignore unlocked mounts in fs_fully_visible")
Reported-by: Seth Forshee <seth.forshee@canonical.com>
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
Add this trivial missing error handling.
Cc: stable@vger.kernel.org
Fixes: 1b852bceb0 ("mnt: Refactor the logic for mounting sysfs and proc in a user namespace")
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
The /dev/ptmx device node is changed to lookup the directory entry "pts"
in the same directory as the /dev/ptmx device node was opened in. If
there is a "pts" entry and that entry is a devpts filesystem /dev/ptmx
uses that filesystem. Otherwise the open of /dev/ptmx fails.
The DEVPTS_MULTIPLE_INSTANCES configuration option is removed, so that
userspace can now safely depend on each mount of devpts creating a new
instance of the filesystem.
Each mount of devpts is now a separate and equal filesystem.
Reserved ttys are now available to all instances of devpts where the
mounter is in the initial mount namespace.
A new vfs helper path_pts is introduced that finds a directory entry
named "pts" in the directory of the passed in path, and changes the
passed in path to point to it. The helper path_pts uses a function
path_parent_directory that was factored out of follow_dotdot.
In the implementation of devpts:
- devpts_mnt is killed as it is no longer meaningful if all mounts of
devpts are equal.
- pts_sb_from_inode is replaced by just inode->i_sb as all cached
inodes in the tty layer are now from the devpts filesystem.
- devpts_add_ref is rolled into the new function devpts_ptmx. And the
unnecessary inode hold is removed.
- devpts_del_ref is renamed devpts_release and reduced to just a
deacrivate_super.
- The newinstance mount option continues to be accepted but is now
ignored.
In devpts_fs.h definitions for when !CONFIG_UNIX98_PTYS are removed as
they are never used.
Documentation/filesystems/devices.txt is updated to describe the current
situation.
This has been verified to work properly on openwrt-15.05, centos5,
centos6, centos7, debian-6.0.2, debian-7.9, debian-8.2, ubuntu-14.04.3,
ubuntu-15.10, fedora23, magia-5, mint-17.3, opensuse-42.1,
slackware-14.1, gentoo-20151225 (13.0?), archlinux-2015-12-01. With the
caveat that on centos6 and on slackware-14.1 that there wind up being
two instances of the devpts filesystem mounted on /dev/pts, the lower
copy does not end up getting used.
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
Cc: Greg KH <greg@kroah.com>
Cc: Peter Hurley <peter@hurleysoftware.com>
Cc: Peter Anvin <hpa@zytor.com>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Serge Hallyn <serge.hallyn@ubuntu.com>
Cc: Willy Tarreau <w@1wt.eu>
Cc: Aurelien Jarno <aurelien@aurel32.net>
Cc: One Thousand Gnomes <gnomes@lxorguk.ukuu.org.uk>
Cc: Jann Horn <jann@thejh.net>
Cc: Jiri Slaby <jslaby@suse.com>
Cc: Florian Weimer <fw@deneb.enyo.de>
Cc: Konstantin Khlebnikov <koct9i@gmail.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Pull btrfs fixes from Chris Mason:
"The important part of this pull is Filipe's set of fixes for btrfs
device replacement. Filipe fixed a few issues seen on the list and a
number he found on his own"
* 'for-linus-4.7' of git://git.kernel.org/pub/scm/linux/kernel/git/mason/linux-btrfs:
Btrfs: deal with duplciates during extent_map insertion in btrfs_get_extent
Btrfs: fix race between device replace and read repair
Btrfs: fix race between device replace and discard
Btrfs: fix race between device replace and chunk allocation
Btrfs: fix race setting block group back to RW mode during device replace
Btrfs: fix unprotected assignment of the left cursor for device replace
Btrfs: fix race setting block group readonly during device replace
Btrfs: fix race between device replace and block group removal
Btrfs: fix race between readahead and device replace/removal
When dealing with inline extents, btrfs_get_extent will incorrectly try
to insert a duplicate extent_map. The dup hits -EEXIST from
add_extent_map, but then we try to merge with the existing one and end
up trying to insert a zero length extent_map.
This actually works most of the time, except when there are extent maps
past the end of the inline extent. rocksdb will trigger this sometimes
because it preallocates an extent and then truncates down.
Josef made a script to trigger with xfs_io:
#!/bin/bash
xfs_io -f -c "pwrite 0 1000" inline
xfs_io -c "falloc -k 4k 1M" inline
xfs_io -c "pread 0 1000" -c "fadvise -d 0 1000" -c "pread 0 1000" inline
xfs_io -c "fadvise -d 0 1000" inline
cat inline
You'll get EIOs trying to read inline after this because add_extent_map
is returning EEXIST
Signed-off-by: Chris Mason <clm@fb.com>
There are several issues in fscache revalidation code.
- In ceph_revalidate_work(), fscache_invalidate() is called when
fscache_check_consistency() return 0. This is complete wrong
because 0 means cache is valid.
- Handle_cap_grant() calls ceph_queue_revalidate() if client
already has CAP_FILE_CACHE. This code is confusing. Client
should revalidate the cache each time it got CAP_FILE_CACHE
anew.
- In Handle_cap_grant(), fscache_invalidate() is called if MDS
revokes CAP_FILE_CACHE. This is inconsistency with the case
that inode get evicted. In the later case, the cache is not
discarded. Client may use the cache when inode is reloaded.
This patch moves the fscache revalidation into ceph_get_caps().
Client revalidates the cache after it gets CAP_FILE_CACHE.
i_rdcache_gen should keep constance while CAP_FILE_CACHE is
used. If i_fscache_gen is not equal to i_rdcache_gen, client
needs to check cache's consistency.
Signed-off-by: Yan, Zheng <zyan@redhat.com>
All other filesystems do not add dirty pages to fscache. They all
disable fscache when inode is opened for write. Only ceph adds
dirty pages to fscache, but the code is buggy.
Signed-off-by: Yan, Zheng <zyan@redhat.com>
__fscache_check_consistency() calls check_consistency() callback
and return the callback's return value. But the return type of
check_consistency() is bool. So __fscache_check_consistency()
return 1 if the cache is inconsistent. This is inconsistent with
the document.
Signed-off-by: Yan, Zheng <zyan@redhat.com>
Acked-by: David Howells <dhowells@redhat.com>
While we are finishing a device replace operation we can have a concurrent
task trying to do a read repair operation, in which case it will call
btrfs_map_block() to get a struct btrfs_bio which can have a stripe that
points to the source device of the device replace operation. This allows
for the read repair task to dereference the stripe's device pointer after
the device replace operation has freed the source device, resulting in
an invalid memory access. This is similar to the problem solved by my
previous patch in the same series and named "Btrfs: fix race between
device replace and discard".
So fix this by surrounding the call to btrfs_map_block() and the code
that uses the returned struct btrfs_bio with calls to
btrfs_bio_counter_inc_blocked() and btrfs_bio_counter_dec(), giving the
proper serialization with the finishing phase of the device replace
operation.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: Josef Bacik <jbacik@fb.com>
While we are finishing a device replace operation, we can make a discard
operation (fs mounted with -o discard) do an invalid memory access like
the one reported by the following trace:
[ 3206.384654] general protection fault: 0000 [#1] PREEMPT SMP
[ 3206.387520] Modules linked in: dm_mod btrfs crc32c_generic xor raid6_pq acpi_cpufreq tpm_tis psmouse tpm ppdev sg parport_pc evdev i2c_piix4 parport
processor serio_raw i2c_core pcspkr button loop autofs4 ext4 crc16 jbd2 mbcache sr_mod cdrom ata_generic sd_mod virtio_scsi ata_piix libata virtio_pci
virtio_ring scsi_mod e1000 virtio floppy [last unloaded: btrfs]
[ 3206.388595] CPU: 14 PID: 29194 Comm: fsstress Not tainted 4.6.0-rc7-btrfs-next-29+ #1
[ 3206.388595] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS by qemu-project.org 04/01/2014
[ 3206.388595] task: ffff88017ace0100 ti: ffff880171b98000 task.ti: ffff880171b98000
[ 3206.388595] RIP: 0010:[<ffffffff8124d233>] [<ffffffff8124d233>] blkdev_issue_discard+0x5c/0x2a7
[ 3206.388595] RSP: 0018:ffff880171b9bb80 EFLAGS: 00010246
[ 3206.388595] RAX: ffff880171b9bc28 RBX: 000000000090d000 RCX: 0000000000000000
[ 3206.388595] RDX: ffffffff82fa1b48 RSI: ffffffff8179f46c RDI: ffffffff82fa1b48
[ 3206.388595] RBP: ffff880171b9bcc0 R08: 0000000000000000 R09: 0000000000000001
[ 3206.388595] R10: ffff880171b9bce0 R11: 000000000090f000 R12: ffff880171b9bbe8
[ 3206.388595] R13: 0000000000000010 R14: 0000000000004868 R15: 6b6b6b6b6b6b6b6b
[ 3206.388595] FS: 00007f6182e4e700(0000) GS:ffff88023fdc0000(0000) knlGS:0000000000000000
[ 3206.388595] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[ 3206.388595] CR2: 00007f617c2bbb18 CR3: 000000017ad9c000 CR4: 00000000000006e0
[ 3206.388595] Stack:
[ 3206.388595] 0000000000004878 0000000000000000 0000000002400040 0000000000000000
[ 3206.388595] 0000000000000000 ffff880171b9bbe8 ffff880171b9bbb0 ffff880171b9bbb0
[ 3206.388595] ffff880171b9bbc0 ffff880171b9bbc0 ffff880171b9bbd0 ffff880171b9bbd0
[ 3206.388595] Call Trace:
[ 3206.388595] [<ffffffffa042899e>] btrfs_issue_discard+0x12f/0x143 [btrfs]
[ 3206.388595] [<ffffffffa042899e>] ? btrfs_issue_discard+0x12f/0x143 [btrfs]
[ 3206.388595] [<ffffffffa042e862>] btrfs_discard_extent+0x87/0xde [btrfs]
[ 3206.388595] [<ffffffffa04303b5>] btrfs_finish_extent_commit+0xb2/0x1df [btrfs]
[ 3206.388595] [<ffffffff8149c246>] ? __mutex_unlock_slowpath+0x150/0x15b
[ 3206.388595] [<ffffffffa04464c4>] btrfs_commit_transaction+0x7fc/0x980 [btrfs]
[ 3206.388595] [<ffffffff8149c246>] ? __mutex_unlock_slowpath+0x150/0x15b
[ 3206.388595] [<ffffffffa0459af6>] btrfs_sync_file+0x38f/0x428 [btrfs]
[ 3206.388595] [<ffffffff811a8292>] vfs_fsync_range+0x8c/0x9e
[ 3206.388595] [<ffffffff811a82c0>] vfs_fsync+0x1c/0x1e
[ 3206.388595] [<ffffffff811a8417>] do_fsync+0x31/0x4a
[ 3206.388595] [<ffffffff811a8637>] SyS_fsync+0x10/0x14
[ 3206.388595] [<ffffffff8149e025>] entry_SYSCALL_64_fastpath+0x18/0xa8
[ 3206.388595] [<ffffffff81100c6b>] ? time_hardirqs_off+0x9/0x14
[ 3206.388595] [<ffffffff8108e87d>] ? trace_hardirqs_off_caller+0x1f/0xaa
This happens because when we call btrfs_map_block() from
btrfs_discard_extent() to get a btrfs_bio structure, the device replace
operation has not finished yet, but before we use the device of one of the
stripes from the returned btrfs_bio structure, the device object is freed.
This is illustrated by the following diagram.
CPU 1 CPU 2
btrfs_dev_replace_start()
(...)
btrfs_dev_replace_finishing()
btrfs_start_transaction()
btrfs_commit_transaction()
(...)
btrfs_sync_file()
btrfs_start_transaction()
(...)
btrfs_commit_transaction()
btrfs_finish_extent_commit()
btrfs_discard_extent()
btrfs_map_block()
--> returns a struct btrfs_bio
with a stripe that has a
device field pointing to
source device of the replace
operation (the device that
is being replaced)
mutex_lock(&uuid_mutex)
mutex_lock(&fs_info->fs_devices->device_list_mutex)
mutex_lock(&fs_info->chunk_mutex)
btrfs_dev_replace_update_device_in_mapping_tree()
--> iterates the mapping tree and for each
extent map that has a stripe pointing to
the source device, it updates the stripe
to point to the target device instead
btrfs_rm_dev_replace_blocked()
--> waits for fs_info->bio_counter to go down to 0
btrfs_rm_dev_replace_remove_srcdev()
--> removes source device from the list of devices
mutex_unlock(&fs_info->chunk_mutex)
mutex_unlock(&fs_info->fs_devices->device_list_mutex)
mutex_unlock(&uuid_mutex)
btrfs_rm_dev_replace_free_srcdev()
--> frees the source device
--> iterates over all stripes
of the returned struct
btrfs_bio
--> for each stripe it
dereferences its device
pointer
--> it ends up finding a
pointer to the device
used as the source
device for the replace
operation and that was
already freed
So fix this by surrounding the call to btrfs_map_block(), and the code
that uses the returned struct btrfs_bio, with calls to
btrfs_bio_counter_inc_blocked() and btrfs_bio_counter_dec(), so that
the finishing phase of the device replace operation blocks until the
the bio counter decreases to zero before it frees the source device.
This is the same approach we do at btrfs_map_bio() for example.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: Josef Bacik <jbacik@fb.com>
For the benefit of every single caller, take osdc instead of map.
Also, now that osdc->osdmap can't ever be NULL, drop the check.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
While iterating and copying extents from the source device, the device
replace code keeps adjusting a left cursor that is used to make sure that
once we finish processing a device extent, any future writes to extents
from the corresponding block group will get into both the source and
target devices. This left cursor is also used for resuming the device
replace operation at mount time.
However using this left cursor to decide whether writes go into both
devices or only the source device is not enough to guarantee we don't
miss copying extents into the target device. There are two cases where
the current approach fails. The first one is related to when there are
holes in the device and they get allocated for new block groups while
the device replace operation is iterating the device extents (more on
this explained below). The second one is that when that loop over the
device extents finishes, we start dellaloc, wait for all ordered extents
and then commit the current transaction, we might have got new block
groups allocated that are now using a device extent that has an offset
greater then or equals to the value of the left cursor, in which case
writes to extents belonging to these new block groups will get issued
only to the source device.
For the first case where the current approach of using a left cursor
fails, consider the source device currently has the following layout:
[ extent bg A ] [ hole, unallocated space ] [extent bg B ]
3Gb 4Gb 5Gb
While we are iterating the device extents from the source device using
the commit root of the device tree, the following happens:
CPU 1 CPU 2
<we are at transaction N>
scrub_enumerate_chunks()
--> searches the device tree for
extents belonging to the source
device using the device tree's
commit root
--> 1st iteration finds extent belonging to
block group A
--> sets block group A to RO mode
(btrfs_inc_block_group_ro)
--> sets cursor left to found_key.offset
which is 3Gb
--> scrub_chunk() starts
copies all allocated extents from
block group's A stripe at source
device into target device
btrfs_alloc_chunk()
--> allocates device extent
in the range [4Gb, 5Gb[
from the source device for
a new block group C
extent allocated from block
group C for a direct IO,
buffered write or btree node/leaf
extent is written to, perhaps
in response to a writepages()
call from the VM or directly
through direct IO
the write is made only against
the source device and not against
the target device because the
extent's offset is in the interval
[4Gb, 5Gb[ which is larger then
the value of cursor_left (3Gb)
--> scrub_chunks() finishes
--> updates left cursor from 3Gb to
4Gb
--> btrfs_dec_block_group_ro() sets
block group A back to RW mode
<we are still at transaction N>
--> 2nd iteration finds extent belonging to
block group B - it did not find the new
extent in the range [4Gb, 5Gb[ for block
group C because we are using the device
tree's commit root or even because the
block group's items are not all yet
inserted in the respective btrees, that is,
the block group is still attached to some
transaction handle's new_bgs list and
btrfs_create_pending_block_groups() was
not called yet against that transaction
handle, so the device extent items were
not yet inserted into the devices tree
<we are still at transaction N>
--> so we end not copying anything from the newly
allocated device extent from the source device
to the target device
So fix this by making __btrfs_map_block() always redirect writes to the
target device as well, independently of the left cursor's value. With
this change the left cursor is now used only for the purpose of tracking
progress and allow a mount operation to resume a device replace.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: Josef Bacik <jbacik@fb.com>
After it finishes processing a device extent, the device replace code sets
back the block group to RW mode and then after that it sets the left cursor
to match the logical end address of the block group, so that future writes
into extents belonging to the block group go both the source (old) and
target (new) devices. However from the moment we turn the block group
back to RW mode we have a short time window, that lasts until we update
the left cursor's value, where extents can be allocated from the block
group and written to, in which case they will not be copied/written to
the target (new) device. Fix this by updating the left cursor's value
before turning the block group back to RW mode.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: Josef Bacik <jbacik@fb.com>
We were assigning new values to fields of the device replace object
without holding the respective lock after processing each device extent.
This is important for the left cursor field which can be accessed by a
concurrent task running __btrfs_map_block (which, correctly, takes the
device replace lock).
So change these fields while holding the device replace lock.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: Josef Bacik <jbacik@fb.com>
When we do a device replace, for each device extent we find from the
source device, we set the corresponding block group to readonly mode to
prevent writes into it from happening while we are copying the device
extent from the source to the target device. However just before we set
the block group to readonly mode some concurrent task might have already
allocated an extent from it or decided it could perform a nocow write
into one of its extents, which can make the device replace process to
miss copying an extent since it uses the extent tree's commit root to
search for extents and only once it finishes searching for all extents
belonging to the block group it does set the left cursor to the logical
end address of the block group - this is a problem if the respective
ordered extents finish while we are searching for extents using the
extent tree's commit root and no transaction commit happens while we
are iterating the tree, since it's the delayed references created by the
ordered extents (when they complete) that insert the extent items into
the extent tree (using the non-commit root of course).
Example:
CPU 1 CPU 2
btrfs_dev_replace_start()
btrfs_scrub_dev()
scrub_enumerate_chunks()
--> finds device extent belonging
to block group X
<transaction N starts>
starts buffered write
against some inode
writepages is run against
that inode forcing dellaloc
to run
btrfs_writepages()
extent_writepages()
extent_write_cache_pages()
__extent_writepage()
writepage_delalloc()
run_delalloc_range()
cow_file_range()
btrfs_reserve_extent()
--> allocates an extent
from block group X
(which is not yet
in RO mode)
btrfs_add_ordered_extent()
--> creates ordered extent Y
flush_epd_write_bio()
--> bio against the extent from
block group X is submitted
btrfs_inc_block_group_ro(bg X)
--> sets block group X to readonly
scrub_chunk(bg X)
scrub_stripe(device extent from srcdev)
--> keeps searching for extent items
belonging to the block group using
the extent tree's commit root
--> it never blocks due to
fs_info->scrub_pause_req as no
one tries to commit transaction N
--> copies all extents found from the
source device into the target device
--> finishes search loop
bio completes
ordered extent Y completes
and creates delayed data
reference which will add an
extent item to the extent
tree when run (typically
at transaction commit time)
--> so the task doing the
scrub/device replace
at CPU 1 misses this
and does not copy this
extent into the new/target
device
btrfs_dec_block_group_ro(bg X)
--> turns block group X back to RW mode
dev_replace->cursor_left is set to the
logical end offset of block group X
So fix this by waiting for all cow and nocow writes after setting a block
group to readonly mode.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: Josef Bacik <jbacik@fb.com>
When it's finishing, the device replace code iterates all extent maps
representing block group and for each one that has a stripe that refers
to the source device, it replaces its device with the target device.
However when it replaces the source device with the target device it,
the target device still has an ID of 0ULL (BTRFS_DEV_REPLACE_DEVID),
only after its ID is changed to match the one from the source device.
This leads to races with the chunk removal code that can temporarly see
a device with an ID of 0ULL and then attempt to use that ID to remove
items from the device tree and fail, causing a transaction abort:
[ 9238.594364] BTRFS info (device sdf): dev_replace from /dev/sdf (devid 3) to /dev/sde finished
[ 9238.594377] ------------[ cut here ]------------
[ 9238.594402] WARNING: CPU: 14 PID: 21566 at fs/btrfs/volumes.c:2771 btrfs_remove_chunk+0x2e5/0x793 [btrfs]
[ 9238.594403] BTRFS: Transaction aborted (error 1)
[ 9238.594416] Modules linked in: btrfs crc32c_generic acpi_cpufreq xor tpm_tis tpm raid6_pq ppdev parport_pc processor psmouse parport i2c_piix4 evdev sg i2c_core se
rio_raw pcspkr button loop autofs4 ext4 crc16 jbd2 mbcache sr_mod cdrom sd_mod ata_generic virtio_scsi ata_piix virtio_pci libata virtio_ring virtio e1000 scsi_mod fl
oppy [last unloaded: btrfs]
[ 9238.594418] CPU: 14 PID: 21566 Comm: btrfs-cleaner Not tainted 4.6.0-rc7-btrfs-next-29+ #1
[ 9238.594419] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS by qemu-project.org 04/01/2014
[ 9238.594421] 0000000000000000 ffff88017f1dbc60 ffffffff8126b42c ffff88017f1dbcb0
[ 9238.594422] 0000000000000000 ffff88017f1dbca0 ffffffff81052b14 00000ad37f1dbd18
[ 9238.594423] 0000000000000001 ffff88018068a558 ffff88005c4b9c00 ffff880233f60db0
[ 9238.594424] Call Trace:
[ 9238.594428] [<ffffffff8126b42c>] dump_stack+0x67/0x90
[ 9238.594430] [<ffffffff81052b14>] __warn+0xc2/0xdd
[ 9238.594432] [<ffffffff81052b7a>] warn_slowpath_fmt+0x4b/0x53
[ 9238.594434] [<ffffffff8116c311>] ? kmem_cache_free+0x128/0x188
[ 9238.594450] [<ffffffffa04d43f5>] btrfs_remove_chunk+0x2e5/0x793 [btrfs]
[ 9238.594452] [<ffffffff8108e456>] ? arch_local_irq_save+0x9/0xc
[ 9238.594464] [<ffffffffa04a26fa>] btrfs_delete_unused_bgs+0x317/0x382 [btrfs]
[ 9238.594476] [<ffffffffa04a961d>] cleaner_kthread+0x1ad/0x1c7 [btrfs]
[ 9238.594489] [<ffffffffa04a9470>] ? btree_invalidatepage+0x8e/0x8e [btrfs]
[ 9238.594490] [<ffffffff8106f403>] kthread+0xd4/0xdc
[ 9238.594494] [<ffffffff8149e242>] ret_from_fork+0x22/0x40
[ 9238.594495] [<ffffffff8106f32f>] ? kthread_stop+0x286/0x286
[ 9238.594496] ---[ end trace 183efbe50275f059 ]---
The sequence of steps leading to this is like the following:
CPU 1 CPU 2
btrfs_dev_replace_finishing()
at this point
dev_replace->tgtdev->devid ==
BTRFS_DEV_REPLACE_DEVID (0ULL)
...
btrfs_start_transaction()
btrfs_commit_transaction()
btrfs_delete_unused_bgs()
btrfs_remove_chunk()
looks up for the extent map
corresponding to the chunk
lock_chunks() (chunk_mutex)
check_system_chunk()
unlock_chunks() (chunk_mutex)
locks fs_info->chunk_mutex
btrfs_dev_replace_update_device_in_mapping_tree()
--> iterates fs_info->mapping_tree and
replaces the device in every extent
map's map->stripes[] with
dev_replace->tgtdev, which still has
an id of 0ULL (BTRFS_DEV_REPLACE_DEVID)
iterates over all stripes from
the extent map
--> calls btrfs_free_dev_extent()
passing it the target device
that still has an ID of 0ULL
--> btrfs_free_dev_extent() fails
--> aborts current transaction
finishes setting up the target device,
namely it sets tgtdev->devid to the value
of srcdev->devid (which is necessarily > 0)
frees the srcdev
unlocks fs_info->chunk_mutex
So fix this by taking the device list mutex while processing the stripes
for the chunk's extent map. This is similar to the race between device
replace and block group creation that was fixed by commit 50460e3718
("Btrfs: fix race when finishing dev replace leading to transaction abort").
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: Josef Bacik <jbacik@fb.com>
The self-test was updated to cover zero-length strings; the function
needs to be updated, too.
Reported-by: Geert Uytterhoeven <geert@linux-m68k.org>
Signed-off-by: George Spelvin <linux@sciencehorizons.net>
Fixes: fcfd2fbf22 ("fs/namei.c: Add hashlen_string() function")
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The original name was simply hash_string(), but that conflicted with a
function with that name in drivers/base/power/trace.c, and I decided
that calling it "hashlen_" was better anyway.
But you have to do it in two places.
[ This caused build errors for architectures that don't define
CONFIG_DCACHE_WORD_ACCESS - Linus ]
Signed-off-by: George Spelvin <linux@sciencehorizons.net>
Reported-by: Guenter Roeck <linux@roeck-us.net>
Fixes: fcfd2fbf22 ("fs/namei.c: Add hashlen_string() function")
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The HPFS filesystem used generic_show_options to produce string that is
displayed in /proc/mounts. However, there is a problem that the options
may disappear after remount. If we mount the filesystem with option1
and then remount it with option2, /proc/mounts should show both option1
and option2, however it only shows option2 because the whole option
string is replaced with replace_mount_options in hpfs_remount_fs.
To fix this bug, implement the hpfs_show_options function that prints
options that are currently selected.
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Cc: stable@vger.kernel.org
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Commit c8f33d0bec ("affs: kstrdup() memory handling") checks if the
kstrdup function returns NULL due to out-of-memory condition.
However, if we are remounting a filesystem with no change to
filesystem-specific options, the parameter data is NULL. In this case,
kstrdup returns NULL (because it was passed NULL parameter), although no
out of memory condition exists. The mount syscall then fails with
ENOMEM.
This patch fixes the bug. We fail with ENOMEM only if data is non-NULL.
The patch also changes the call to replace_mount_options - if we didn't
pass any filesystem-specific options, we don't call
replace_mount_options (thus we don't erase existing reported options).
Fixes: c8f33d0bec ("affs: kstrdup() memory handling")
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Cc: stable@vger.kernel.org # v4.1+
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Commit ce657611ba ("hpfs: kstrdup() out of memory handling") checks if
the kstrdup function returns NULL due to out-of-memory condition.
However, if we are remounting a filesystem with no change to
filesystem-specific options, the parameter data is NULL. In this case,
kstrdup returns NULL (because it was passed NULL parameter), although no
out of memory condition exists. The mount syscall then fails with
ENOMEM.
This patch fixes the bug. We fail with ENOMEM only if data is non-NULL.
The patch also changes the call to replace_mount_options - if we didn't
pass any filesystem-specific options, we don't call
replace_mount_options (thus we don't erase existing reported options).
Fixes: ce657611ba ("hpfs: kstrdup() out of memory handling")
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Cc: stable@vger.kernel.org
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Various builds (such as i386:allmodconfig) fail with
fs/binfmt_aout.c:133:2: error: expected identifier or '(' before 'return'
fs/binfmt_aout.c:134:1: error: expected identifier or '(' before '}' token
[ Oops. My bad, I had stupidly thought that "allmodconfig" covered this
on x86-64 too, but it obviously doesn't. Egg on my face. - Linus ]
Fixes: 5d22fc25d4 ("mm: remove more IS_ERR_VALUE abuses")
Signed-off-by: Guenter Roeck <linux@roeck-us.net>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Pull string hash improvements from George Spelvin:
"This series does several related things:
- Makes the dcache hash (fs/namei.c) useful for general kernel use.
(Thanks to Bruce for noticing the zero-length corner case)
- Converts the string hashes in <linux/sunrpc/svcauth.h> to use the
above.
- Avoids 64-bit multiplies in hash_64() on 32-bit platforms. Two
32-bit multiplies will do well enough.
- Rids the world of the bad hash multipliers in hash_32.
This finishes the job started in commit 689de1d6ca ("Minimal
fix-up of bad hashing behavior of hash_64()")
The vast majority of Linux architectures have hardware support for
32x32-bit multiply and so derive no benefit from "simplified"
multipliers.
The few processors that do not (68000, h8/300 and some models of
Microblaze) have arch-specific implementations added. Those
patches are last in the series.
- Overhauls the dcache hash mixing.
The patch in commit 0fed3ac866 ("namei: Improve hash mixing if
CONFIG_DCACHE_WORD_ACCESS") was an off-the-cuff suggestion.
Replaced with a much more careful design that's simultaneously
faster and better. (My own invention, as there was noting suitable
in the literature I could find. Comments welcome!)
- Modify the hash_name() loop to skip the initial HASH_MIX(). This
would let us salt the hash if we ever wanted to.
- Sort out partial_name_hash().
The hash function is declared as using a long state, even though
it's truncated to 32 bits at the end and the extra internal state
contributes nothing to the result. And some callers do odd things:
- fs/hfs/string.c only allocates 32 bits of state
- fs/hfsplus/unicode.c uses it to hash 16-bit unicode symbols not bytes
- Modify bytemask_from_count to handle inputs of 1..sizeof(long)
rather than 0..sizeof(long)-1. This would simplify users other
than full_name_hash"
Special thanks to Bruce Fields for testing and finding bugs in v1. (I
learned some humbling lessons about "obviously correct" code.)
On the arch-specific front, the m68k assembly has been tested in a
standalone test harness, I've been in contact with the Microblaze
maintainers who mostly don't care, as the hardware multiplier is never
omitted in real-world applications, and I haven't heard anything from
the H8/300 world"
* 'hash' of git://ftp.sciencehorizons.net/linux:
h8300: Add <asm/hash.h>
microblaze: Add <asm/hash.h>
m68k: Add <asm/hash.h>
<linux/hash.h>: Add support for architecture-specific functions
fs/namei.c: Improve dcache hash function
Eliminate bad hash multipliers from hash_32() and hash_64()
Change hash_64() return value to 32 bits
<linux/sunrpc/svcauth.h>: Define hash_str() in terms of hashlen_string()
fs/namei.c: Add hashlen_string() function
Pull out string hash to <linux/stringhash.h>
This is just the infrastructure; there are no users yet.
This is modelled on CONFIG_ARCH_RANDOM; a CONFIG_ symbol declares
the existence of <asm/hash.h>.
That file may define its own versions of various functions, and define
HAVE_* symbols (no CONFIG_ prefix!) to suppress the generic ones.
Included is a self-test (in lib/test_hash.c) that verifies the basics.
It is NOT in general required that the arch-specific functions compute
the same thing as the generic, but if a HAVE_* symbol is defined with
the value 1, then equality is tested.
Signed-off-by: George Spelvin <linux@sciencehorizons.net>
Cc: Geert Uytterhoeven <geert@linux-m68k.org>
Cc: Greg Ungerer <gerg@linux-m68k.org>
Cc: Andreas Schwab <schwab@linux-m68k.org>
Cc: Philippe De Muyter <phdm@macq.eu>
Cc: linux-m68k@lists.linux-m68k.org
Cc: Alistair Francis <alistai@xilinx.com>
Cc: Michal Simek <michal.simek@xilinx.com>
Cc: Yoshinori Sato <ysato@users.sourceforge.jp>
Cc: uclinux-h8-devel@lists.sourceforge.jp
Patch 0fed3ac866 improved the hash mixing, but the function is slower
than necessary; there's a 7-instruction dependency chain (10 on x86)
each loop iteration.
Word-at-a-time access is a very tight loop (which is good, because
link_path_walk() is one of the hottest code paths in the entire kernel),
and the hash mixing function must not have a longer latency to avoid
slowing it down.
There do not appear to be any published fast hash functions that:
1) Operate on the input a word at a time, and
2) Don't need to know the length of the input beforehand, and
3) Have a single iterated mixing function, not needing conditional
branches or unrolling to distinguish different loop iterations.
One of the algorithms which comes closest is Yann Collet's xxHash, but
that's two dependent multiplies per word, which is too much.
The key insights in this design are:
1) Barring expensive ops like multiplies, to diffuse one input bit
across 64 bits of hash state takes at least log2(64) = 6 sequentially
dependent instructions. That is more cycles than we'd like.
2) An operation like "hash ^= hash << 13" requires a second temporary
register anyway, and on a 2-operand machine like x86, it's three
instructions.
3) A better use of a second register is to hold a two-word hash state.
With careful design, no temporaries are needed at all, so it doesn't
increase register pressure. And this gets rid of register copying
on 2-operand machines, so the code is smaller and faster.
4) Using two words of state weakens the requirement for one-round mixing;
we now have two rounds of mixing before cancellation is possible.
5) A two-word hash state also allows operations on both halves to be
done in parallel, so on a superscalar processor we get more mixing
in fewer cycles.
I ended up using a mixing function inspired by the ChaCha and Speck
round functions. It is 6 simple instructions and 3 cycles per iteration
(assuming multiply by 9 can be done by an "lea" instruction):
x ^= *input++;
y ^= x; x = ROL(x, K1);
x += y; y = ROL(y, K2);
y *= 9;
Not only is this reversible, two consecutive rounds are reversible:
if you are given the initial and final states, but not the intermediate
state, it is possible to compute both input words. This means that at
least 3 words of input are required to create a collision.
(It also has the property, used by hash_name() to avoid a branch, that
it hashes all-zero to all-zero.)
The rotate constants K1 and K2 were found by experiment. The search took
a sample of random initial states (I used 1023) and considered the effect
of flipping each of the 64 input bits on each of the 128 output bits two
rounds later. Each of the 8192 pairs can be considered a biased coin, and
adding up the Shannon entropy of all of them produces a score.
The best-scoring shifts also did well in other tests (flipping bits in y,
trying 3 or 4 rounds of mixing, flipping all 64*63/2 pairs of input bits),
so the choice was made with the additional constraint that the sum of the
shifts is odd and not too close to the word size.
The final state is then folded into a 32-bit hash value by a less carefully
optimized multiply-based scheme. This also has to be fast, as pathname
components tend to be short (the most common case is one iteration!), but
there's some room for latency, as there is a fair bit of intervening logic
before the hash value is used for anything.
(Performance verified with "bonnie++ -s 0 -n 1536:-2" on tmpfs. I need
a better benchmark; the numbers seem to show a slight dip in performance
between 4.6.0 and this patch, but they're too noisy to quote.)
Special thanks to Bruce fields for diligent testing which uncovered a
nasty fencepost error in an earlier version of this patch.
[checkpatch.pl formatting complaints noted and respectfully disagreed with.]
Signed-off-by: George Spelvin <linux@sciencehorizons.net>
Tested-by: J. Bruce Fields <bfields@redhat.com>
