The btrfs tool is changed in order to support command line parameters
to configure the IO priority of the scrub tasks. Also the default is
changed. The default IO priority for scrub is the idle class now.
The behavior is the same as when one would type
'ionice ... btrfs scrub start ...' or 'ionice ... btrfs scrub resume ...'
(without this patch applied).
The only reason for adding this to the btrfs tool is that it was not
documented and not obvious that it worked like this, that all internal
scrub tasks inherited the IO priority values of the btrfs tool that is
starting or resuming the scrub operation.
Note that after applying the patch it is no longer possible to set
the IO priority using ionice since the btrfs tool always configures
the priority in order to run in the idle class by default.
Some basic performance measurements have been done with the goal to
measure which IO priority for scrub gives the best overall disk data
throughput. The kernel was configured to use the CFQ IO scheduler
with default configuration and without support for throttling. The
summary is, that the more the disk head movements are avoided, the
faster the overall disk transfer capacity is, which is not really a
big surprise. Therefore it makes sense that the best data throughput
was measured setting the scrub IO priority and the scrub readahead
IO priority to the idle class priority. Running with idle class IO
priority means that scrub and scrub readahead IO is paused while
other tasks access the disk. Doing the tasks one after the other
instead of concurrently avoids many disk head movements. The
overall data throughput of rotating disks is improved this way.
However, if it is desired to have the scrub task done within a
reasonable time, and if at the same time the filesystem is heavily
loaded, the idle IO priority should be avoided. Otherwise the scrub
operation will never take place and thus never terminate.
The best effort IO priority class with the subclass 7 (the lowest
one in the best effort class) is recommended in the case of always
heavily loaded hard disks. If the filesystem is not loaded all the
time and leaves some idle slots for scrub, the idle class IO priority
is recommended. The idle class now is the default if the scrub
operation is started with the btrfs-progs tools.
Note that the patch that sets the scrub readahead IO priority to the
idle class is a seperate patch, this needs to be done in the kernel.
Signed-off-by: Stefan Behrens <sbehrens@giantdisaster.de>
get_fs_info() has been silently switching from a device to a mounted
path as needed; the caller's filehandle was unexpectedly closed &
reopened outside the caller's scope. Not so great.
The callers do want "fdmnt" to be the filehandle for the mount point
in all cases, though - the various ioctls act on this (not on an fd
for the device). But switching it in the local scope of get_fs_info
is incorrect; it just so happens that *usually* the fd number is
unchanged.
So - use the new helpers to detect when an argument is a block
device, and open the the mounted path more obviously / explicitly
for ioctl use, storing the filehandle in fdmnt.
Then, in get_fs_info, ignore the fd completely, and use the path on
the argument to determine if the caller wanted to act on just that
device, or on all devices for the filesystem.
Affects those commands which are documented to accept either
a block device or a path:
* btrfs device stats
* btrfs replace start
* btrfs scrub start
* btrfs scrub status
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
cmd_scrub_cancel had its own mountpoint discovery routine;
just use open_path_or_dev_mnt() for that now.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
consolidate error handling to ensure that peer_fd
is closed on error paths. Add a couple comments
to the error handling after the thread is complete.
Note that scrub_progress_cycle returns negative
errnos on any error.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
The two sigint handlers issue ioctls to clean up, but if
they fail, noone would know. I'm not sure there is
any other error handling to be done at this point, but a
notification seems wise.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
If we request scrub cancel on an unmounted or
non-btrfs device, we still get a "scrub canceled"
success message:
# btrfs scrub cancel /dev/loop1
scrub cancelled
# blkid /dev/loop1
/dev/loop1: UUID="7f586941-1d5e-4ba7-9caa-b35934849957" TYPE="xfs"
Fix this so that if check_mounted_where returns 0
we don't report success.
While we're at it, use perror to report the reason for an open
failure, if we get one.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
If we retry opening the mountpoint and fail, we'll call
close on a filehandle w/ value -1. Rearrange so the
retry uses the same open and same error handling.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
In the case that btrfs scrub cancel is given a device name,
we close the file handle, and then pass it to check_mounted_where()
which eventually preads from that (now closed) fd. Fix the logic
so that we close & re-open the discovered mountpoint properly.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: Zach Brown <zab@redhat.com>
If scrub start discovers that scrub is already running,
we need to set prg_fd to -1 before goto out, or we'll
try to close it again in the error path.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: Zach Brown <zab@redhat.com>
If connection fails the socket is leaked when the status file is used
instead. Close it to trivially cut down on fd use and to bring down the
noise in static code analysis.
Signed-off-by: Zach Brown <zab@redhat.com>
Two convenient utility functions that have so far been local to scrub are
moved to utils.c.
They will be used in the device stats code in a following commit.
Signed-off-by: Stefan Behrens <sbehrens@giantdisaster.de>
Scrub can be invoked to scrub only a single device of a (mounted) filesystem.
The code determines whether the given path is a mountpoint of a filesystem
by issueing a btrfs-specific ioctl to it. Only in case of EINVAL it assumed
it may be a device, all other errnos just caused it fail, but some devices
(correctly) return ENOTTY. This patch adds this to the error check.
Signed-off-by: Arne Jansen <sensille@gmx.net>
There, 'char' is unsigned, so once assigned '-1' from getopt, it gets
the value 255. Then, it compared to '-1' gives false.
Signed-off-by: Lluis Batlle i Rossell <viric@viric.name>
The kernel uses unsigned long long for u64, but PPC64 uses unsigned
long by default. This results in compilation warnings such as:
print-tree.c:333: warning: format '%llu' expects type 'long long
unsigned int', but argument 4 has type 'u64'
To fix this, the macro __KERNEL__ needs to be defined before including
the file <asm/types.h>. This can be done by defining the macro in
"kerncompat.h" and making it the first included file in the relevant
header files; this fixes the compiler warnings on PPC64.
Reviewed-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Wade Cline <clinew@linux.vnet.ibm.com>
The function scrub_fs_info( ) closes and reopen a file handle
passed as argument, when a caller uses the file handle even after the
call.
The function scrub_fs_info( ) is updated to remove the file handle
argument, and instead uses a private own file handle.
The callers are updated to not pass the argument.
The new infrastructure offloads checking number of arguments passed to a
command to individual command handlers. Fix them up accordingly.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Separate every command group into its own file (cmds_<group>.c) and
rearrange includes. Remove btrfs_cmds.c.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Anand Jain
Stefan Behrens
Eric Sandeen
Zach Brown
Arne Jansen
Lluis Batlle i Rossell
Wade Cline
Goffredo Baroncelli
Ilya Dryomov