Guoqing Jiang [Wed, 10 Jun 2015 05:42:11 +0000 (13:42 +0800)]
mdadm: add the ability to change cluster name
To support change the cluster name, the commit do the followings:
1. extend original write_bitmap function for new scenario.
2. add the scenarion to handle the modification of cluster's name
in write_bitmap1.
3. let the cluster name also show in examine_super1 and detail_super1
Guoqing Jiang [Wed, 10 Jun 2015 05:42:09 +0000 (13:42 +0800)]
Convert a bitmap=none device to clustered
This adds the ability to convert a regular md without bitmap
(--bitmap=none) to a clustered device (--bitmap=clustered).
To convert a device with --bitmap=internal or --bitmap=external,
you have to convert to --bitmap=none and then re-execute the
command with --bitmap=clustered.
Guoqing Jiang [Wed, 10 Jun 2015 05:42:08 +0000 (13:42 +0800)]
Add a new clustered disk
A clustered disk is added by the traditional --add sequence.
However, other nodes need to acknowledge that they can "see"
the device. This is done by --cluster-confirm:
--cluster-confirm SLOTNUM:/dev/whatever (if disk is found)
or
--cluster-confirm SLOTNUM:missing (if disk is not found)
The node initiating the --add, has the disk state tagged with
MD_DISK_CLUSTER_ADD and the one confirming tag the disk with
MD_DISK_CANDIDATE.
Guoqing Jiang [Wed, 10 Jun 2015 05:42:06 +0000 (13:42 +0800)]
Set home-cluster while creating an array
The home-cluster is stored in the bitmap super block of the
array. The device can be assembled on a cluster with the
cluster name same as the one recorded in the bitmap.
If home-cluster is not specified, this is auto-detected using
dlopen corosync cmap library.
neilb: allow code to compile when corosync-devel is not installed.
Guoqing Jiang [Wed, 10 Jun 2015 05:42:05 +0000 (13:42 +0800)]
Add nodes option while creating md
Specifies the maximum number of nodes in the cluster that may use
this device simultaneously. This is equivalent to the number of
bitmaps created in the internal superblock (patches to follow).
NeilBrown [Thu, 28 May 2015 06:53:26 +0000 (16:53 +1000)]
test: make 'check wait' more reliable.
'recover' etc doesn't appear in /proc/mdstat immediately.
The "sync" thread must be started first.
But 'sync_action' shows it as soon as MD_RECOVERY_NEEDED is set
in the kernel. So look there too.
Now maybe I can get rid of some of those silly 'sleep' calls.
NeilBrown [Thu, 28 May 2015 06:43:15 +0000 (16:43 +1000)]
Grow: fix problem with --grow --continue
If an array is being reshaped using backup space on a 'spare' device,
then
mdadm --grow --continue
won't find it as by the time it runs, nothing looks like a spare are
more. The spare has been added to the array, but has no data yet.
So allow reshape_prepare_fdlist to find a newly-incorporated spare and
report this so it can be used.
Reported-by: Xiao Ni <xni@redhat.com> Signed-off-by: NeilBrown <neilb@suse.de>
NeilBrown [Mon, 25 May 2015 06:33:45 +0000 (16:33 +1000)]
Grow: another attempt to fix stop-during-reshape race.
When the array is stopped during a critical section, we sometimes
erase the backup, which is bad.
This happens when 'completed' is zero.
This can happen easily when 'stop' freezes reshape.
So try to be more careful and check 'reshape_position'.
Sergey Vidishev [Tue, 19 May 2015 19:02:46 +0000 (22:02 +0300)]
mdadm: monitor: fix nullptr dereference when get_md_name() returns NULL
Function add_new_arrays() expects that function get_md_name() should
return pointer to devname, but also get_md_name() may return NULL. So
check the pointer before use it in add_new_arrays().
NeilBrown [Fri, 15 May 2015 05:11:48 +0000 (15:11 +1000)]
Grow: be even more careful about handing a '0' completed value.
Some old kernels set 'completed' to '0' too soon.
But modern kernels don't.
And when 'mdadm --stop' freezes and resume the grow,
'completed' goes back to zero briefly, which can confuse this
logic.
So only think '0' might be wrong from an old kernel when
the reshape has gone idle.
NeilBrown [Thu, 14 May 2015 01:17:39 +0000 (11:17 +1000)]
Grow: only warn about incompatible metadata when no fallback available.
We might be trying to set_new_data_offset() for RAID10, when it is
a necessary requirement, or for RAID5 where it is optional.
In the latter case, a message about metadata versions is no helpful.
NeilBrown [Wed, 13 May 2015 04:08:41 +0000 (14:08 +1000)]
Manage: when re-adding, do check avail size if ->sb cannot be found.
avail_size1 requires ->sb, so we must only call it if ->sb
was loaded.
If ->sb wasn't loaded, then we are only proceding on the basis that
the kernel might be able to work something out - we don't need to
do any tests on size.
Martin Wilck [Mon, 11 May 2015 14:09:44 +0000 (16:09 +0200)]
DDF: _write_super_to_disk: fix anchor header type
Since commit 30bee0201, the anchor is updated from the active
DDF header. This requires fixing the header type before the
anchor is written.
The LSI Software RAID code will reject DDF meta data with wrong
anchor type and will erase all meta data when it encounters
such a broken anchor. Thus starting Linux md once on a system
with LSI RAID BIOS may cause the meta data to get destroyed.
NeilBrown [Wed, 6 May 2015 05:03:50 +0000 (15:03 +1000)]
Manage: fix test for 'is array failed'.
We 'active_disks' does not count spares, so if array is rebuilding,
this will not necessarily find all devices, so may report an array
as failed when it isn't.
Active arrays with IMSM metadata are counted per hba so far.
This is bad due to new functionality of orom shared between multiple
controllers i.e. more arrays can be created than is supported by orom.
This patch changes the way of counting arrays, so the result will be
sum of arrays under every hba supported by specific orom.
Assemble/force: make it possible to "force" a new device in a reshape.
Normally we do not "force"-assemble devices which are in the
middle of recovery, as they are unlikely to have useful data.
However, when a reshape increases the number of devices,
the newly added devices appear to be recovering because they
do not have complete data on them yet, but then they aren't expected
to until the reshape completes.
So in this case, it can be appropriate to force-assemble them.
allow new_offset to be set, but don't then allow a RAID5
to be reshaped to change that offset.
Due to selective backports, this includes the SLES11-SP3 kernel.
It is quite easy to handle this case in mdadm, so we do.
Specifically: if the reshape with data-offset fails with EINVAL,
abort the data-offset change and try the "old" way.
Pawel Baldysiak [Fri, 27 Feb 2015 14:47:54 +0000 (15:47 +0100)]
IncRemove: Set "auto-read" only after successful excl open.
"mdadm -If" - triggered from udev rules when disk is removed from OS -
tries to set array in auto-read-only mode. This can interrupt rebuild
process which is started automatically, e.g. if array is mounted and
spare disk is available (I/O error is detected faster than removing
failed disk by mdadm).
This patch prevents "mdadm -If" from setting array into "auto-read-only",
by requiring exclusive open to succeed.
NeilBrown [Thu, 12 Feb 2015 02:46:53 +0000 (13:46 +1100)]
Don't break long strings onto multiple lines.
It is best to keep strings all together so that they
are easier to search for in the source code.
If a string is so long that it looks ugly one line,
them maybe it should be broken into multiple lines
for display too.
Only strings which contain a newline can be broken
into multiple lines:
NeilBrown [Thu, 12 Feb 2015 02:21:17 +0000 (13:21 +1100)]
Consistently print program Name and __func__ in debug messages.
make dprintf() print program name and __func__, so that
this messaging is consistent.
Also remove all __func__ messages from pr_err(). We shouldn't
leak that internal data in error message.
If we really want function name there, we new pr_XXX might
be wanted.
Pawel Baldysiak [Wed, 11 Feb 2015 21:25:03 +0000 (22:25 +0100)]
Change way of printing name of a process
Sometimes mdadm prints messages with wrong name "mdmon",
and vice versa.
This patch solves this problem by changing method of determining
process name.
Now "Name" will be set in const at start of a program,
previously was hardcoded as #define.
Monitor: fix for regression with container devices
This patch fixes 2 problems introduced by commit 9a518d8: not closing a
file descriptor and ignoring container devices. Array state is always
"inactive" for containers, so we make sure that the device is not a
container by reading also the "level" sysfs entry.
NeilBrown [Tue, 3 Feb 2015 22:06:47 +0000 (09:06 +1100)]
mdcheck: be careful when sourcing the output of "mdadm --detail --export"
The output of "mdadm --detail --export" isn't quoted properly so
fields that contain spaces can be a problem.
We only want the MD_UUID field, and it has a very well defined
format with no spaces.
So use 'grep' to limit the output to just that.
Pawel Baldysiak [Tue, 20 Jan 2015 12:52:25 +0000 (13:52 +0100)]
IMSM: Clear migration record on disks more often
Migration record is not always cleared after successful migration. This can
block another reshape from being started. Migration will not be continued via
systemd service due to error in verifying reshape position. This patch added
clearing migration record when disk is added to container, and after successful
migration.
Pawel Baldysiak [Thu, 27 Nov 2014 11:35:24 +0000 (12:35 +0100)]
Grow: Fix wrong 'goto' in set_new_data_offset
Commit a821c95f114724b38df1ea99b2858178e0ed28ce
besides introducing additional message, also changed
direct return to "goto" instruction.
'goto release' will cause routine to return with '-1',
when previously '1' was returned.
Described behaviour breaks e.g. IMSM reshape process.
This patch fixes this issue by changing 'goto' to proper one -
the one that returns '1'.
NeilBrown [Tue, 28 Oct 2014 21:48:02 +0000 (08:48 +1100)]
Monitor: don't open md array that doesn't exist.
Opening a block-special-device for an array that doesn't
exist causes that array to be instantiated (as an empty array).
Races at array shutdown can cause the array to spontaneously
re-appear if some deamon notices a 'change' event and goes
to investigate.
Teach "mdadm --monitor" to avoid this race by checking the
"array_state" before opening the device.
Reported-by: Francis Moreau <francis.moro@gmail.com> Signed-off-by: NeilBrown <neilb@suse.de>
Print platform details per OROM, not per controller, differentiate
RST(e) platforms from legacy IMSM, print NVMe device paths, adjust port
printing to newer sysfs path.
imsm: support for second and combined AHCI controllers in UEFI mode
Grantly platform introduces a second AHCI controller (sSATA) and two new
UEFI variables for the RSTe firmware. This patch adds support for those
variables in order to correctly determine IMSM platform capabilities in
UEFI mode.
Signed-off-by: Artur Paszkiewicz <artur.paszkiewicz@intel.com> Signed-off-by: NeilBrown <neilb@suse.de>
The IMSM platform code was based on an assumption that the OROM or UEFI
capability structure (represented by struct imsm_orom) always belongs to
only one HBA. This assumption is no longer valid, because of newer
platforms with dual AHCI HBAs. Each HBA can have a separate OROM, but
some versions have a combined OROM for both HBAs.
This patch implements this HBA-OROM relationship in struct orom_entry,
which matches an OROM with a list of HBA PCI ids. All the detected
orom_entries are stored and retrieved using a global array and the
functions add_orom(), add_orom_device_id() and get_orom_by_device_id().
This replaces the arrays: imsm_orom, populated_orom, imsm_efi,
populated_efi.
The scan() function is extended to find all HBAs for an OROM. The list
of their device ids is retrieved from the PCI Expansion ROM Data
Structure, hence the additional field devListOffset in struct
pciExpDataStructFormat.
In UEFI mode we can't read the PCI Expansion ROM Data Structure and the
imsm_orom structures are stored in UEFI variables. They do not provide a
similar device id list, so we also check the HBA PCI class to make sure
that the HBA has RAID mode enabled.
In super-intel.c there are changes which allow spanning of IMSM
containers over HBAs of the same type, but only if the HBAs share the
same OROM. This is done by comparing imsm_orom pointers, which (outside
of platform-intel.c) always point to the global array containing all the
detected oroms. Additional warnings are added to
validate_container_imsm() to warn about potentially dangerous operations
in all the possible cases, e.g. when an array is assembled using disks
attached to HBAs with separate OROMs.
Signed-off-by: Artur Paszkiewicz <artur.paszkiewicz@intel.com> Signed-off-by: NeilBrown <neilb@suse.de>
NeilBrown [Wed, 5 Nov 2014 05:21:42 +0000 (16:21 +1100)]
Incremental: don't be distracted by partition table when calling try_spare.
Currently a partition table on a device makes "mdadm -I" think
the array has a particular metadata type and so will only
add it to an array of that (partition table) type .. which doesn't
make any sense.
So tell guess_super to only look for 'array' metadata.
NeilBrown [Mon, 3 Nov 2014 22:35:20 +0000 (09:35 +1100)]
Detail: fix handling of 'disks' array.
Since the introduction of replacement devices, we reserve
to places in the "disks" array for each raid disk.
That means we should allocate to twice "max_disk" as the array
could have that many raid_disks (though that would limit the
number of replacements).
A couple of other places need to use "max_disks*2" instead of
"max_disks" to co-ordinate with this.
Reported-by: Or Sagi <ors@reduxio.com> Signed-off-by: NeilBrown <neilb@suse.de>
NeilBrown [Mon, 3 Nov 2014 01:49:05 +0000 (12:49 +1100)]
Rebuildmap: strip local host name from device name.
When /run/mdadm/map is being rebuilt, e.g. by "mdadm -Ir",
if the device doesn't exist in /dev, we have to choose
a name.
Currently we don't strip the hostname which is wrong if
it is the local host.
Reported-by: Stephen Kent <smkent@smkent.net> Signed-off-by: NeilBrown <neilb@suse.de>
Justin Maggard [Sat, 25 Oct 2014 00:55:02 +0000 (17:55 -0700)]
Grow: fix resize of array component size to > 32bits
If the request --size to --grow an array to is larger
than 32bits, then mdadm may make the wrong choice and
use ioctl instead of setting component_size via sysfs
and the change is ignored.
Instead of using casts to check for a 32-bit overflow,
just check for set bits outside of INT32_MAX.
mdmon: already read sysfs files once after opening.
seq_file in the kernel will allocate a read buffer on
first read. We want this to happen under the managemon thread,
not the 'monitor' thread, as the latter is not allow to allocate
memory (might deadlock).
So do a first read after opening.
Samuli Suominen [Thu, 21 Aug 2014 03:56:48 +0000 (06:56 +0300)]
Fix parallel make problem.
When make is called with, for example,
"make -j9 install install-system"
i.e. both install and install-systemd targets at the same
line and with high -j value,
then the same install.tmp file was used, and udev rules
ends up in systemd service files, or otherway around.
For more information, see:
http://www.spinics.net/lists/raid/msg46782.html
http://bugs.gentoo.org/show_bug.cgi?id=517218
NeilBrown [Fri, 15 Aug 2014 05:45:54 +0000 (15:45 +1000)]
super1: don't allow adding a bitmap if there is no space.
If the data is too close to the superblock there may be
no space for a bitmap.
If that happens, fail the adding of the bitmap rather than
corrupt data.
Reported-by: Lars Wijtemans <rhelbugzilla@lars.wijtemans.nl>
Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=922944
NeilBrown [Mon, 11 Aug 2014 00:30:42 +0000 (10:30 +1000)]
Manage: fix removal of non-existent devices.
"--remove detached" and others stopped working a while
back when I refactored some code.
For 'remove' and 'fail', the device may not exist so
if it is "MM:mm", (e.g. added by "detached"), just parse
out the numbers.
Reported-by: Killian De Volder <killian.de.volder@megasoft.be> Signed-off-by: NeilBrown <neilb@suse.de>
NeilBrown [Mon, 11 Aug 2014 00:22:24 +0000 (10:22 +1000)]
Manage: simplify `rdev` handling in Manage_subdevs.
The only use 'struct stat stb' to get the 'rdev', and sometimes
we don't even use 'stat'.
So make 'rdev' a stand-alone variable, and only declare stb'
when we actually need it.
NeilBrown [Wed, 6 Aug 2014 05:56:12 +0000 (15:56 +1000)]
Detail: Avoid dereferencing some NULL pointers.
dm devices which only have a single underlying md device
will respond to md ioctls as though they were that md device.
This can confuse mdadm and lead it to violating its segments.
So add tests for NULL where appropriate. You might not get exactly
the right answer when you "mdadm -D" a dm device, but at least it won't
crash now.
Guy Menanteau [Mon, 4 Aug 2014 14:53:03 +0000 (16:53 +0200)]
DDF: cast print arguments in super-ddf.c
mdadm fails to build on ppc64 and ppc64le architectures.
===
super-ddf.c: In function '_set_config_size':
super-ddf.c:2849:4: error: format '%llx' expects argument of type 'long long unsigned int', but argument 6 has type '__u64' [-Werror=format=]
pr_err("%s: %x:%x: workspace size 0x%llx too big, ignoring\n",
^
super-ddf.c:2855:2: error: format '%llx' expects argument of type 'long long unsigned int', but argument 6 has type '__u64' [-Werror=format=]
dprintf("%s: %x:%x config_size %llx, DDF structure is %llx blocks\n",
^
cc1: all warnings being treated as errors
<builtin>: recipe for target 'super-ddf.o' failed
===
Grow: fix that preventing resize of array to 32bit size.
If the request --size to --grow an array to is 32bits
(i.e. msb in bit 32) then mdadm make wrong choice and
uses ioctl instead of setting component_size via sysfs
and the change is ignored.
This is fixed by using correct casts.
Reported-and-tested-by: Killian De Volder <killian.de.volder@megasoft.be> Signed-off-by: NeilBrown <neilb@suse.de>
Grow process did not check if reshape is already started
when deciding about restarting.
Sync_action should be checked in this case, and if
reshape is running - restart flag should not be set.
Otherwise, Grow process will fail to write data to
sysfs, and reshape will not be continued.
DDF: validate metadata_update size before using it.
process_update already checks update->len, for all but
the 'magic', prepare_update doesn't at all.
So add tests to prepare_update that we don't exceed the buffer.
This will consequently protect process_update from looking
for a 'magic' which isn't there.
Reported-by: Vincent Berg <vberg@ioactive.com> Signed-off-by: NeilBrown <neilb@suse.de>
If 'prepare_update' fails for some reason there is little
point continuing on to 'process_update'.
For now only malloc failures are caught, but other failures
will be considered in future.
Pawel Baldysiak [Mon, 30 Jun 2014 12:22:22 +0000 (12:22 +0000)]
IMSM: Add warning message when assemble spanned container
Due to several changes in code assemble with disks
spanned between different controllers can be obtained
in some cases. After IMSM container will be assembled, check HBA of
disks, and print proper warning if mismatch is detected.