Xiao Ni [Fri, 7 Apr 2023 00:45:28 +0000 (08:45 +0800)]
Remove the config files in mdcheck_start|continue service
We set MDADM_CHECK_DURATION in the mdcheck_start|continue.service files.
And mdcheck doesn't use any configs from the config file. So we can remove
the dependencies.
Signed-off-by: Xiao Ni <xni@redhat.com> Signed-off-by: Jes Sorensen <jes@trained-monkey.org>
when we excute mdadm --assemble, udev-md-raid-assembly.rules is triggered.
Then we stop array, we found an coredump for mdadm --incremental.func
stack are as follows:
#0 enough (level=10, raid_disks=4, layout=258, clean=1,
avail=avail@entry=0x0) at util.c:555
#1 0x0000562170c26965 in Incremental (devlist=<optimized out>,
c=<optimized out>, st=0x5621729b6dc0) at Incremental.c:514
#2 0x0000562170bfb6ff in main (argc=<optimized out>,
argv=<optimized out>) at mdadm.c:1762
func enough() use array avail,avail allocate space in func count_active,
it may not alloc space, causing a coredump.We fix this coredump.
NeilBrown [Tue, 14 Mar 2023 00:06:25 +0000 (11:06 +1100)]
mdopen: always try create_named_array()
mdopen() will use create_named_array() to ask the kernel to create the
given md array, but only if it is given a number or name.
If it is NOT given a name and is required to choose one itself using
find_free_devnm() it does NOT use create_named_array().
On kernels with CONFIG_BLOCK_LEGACY_AUTOLOAD not set, this can result in
failure to assemble an array. This can particularly seen when the
"name" of the array begins with a host name different to the name of the
host running the command.
NeilBrown [Mon, 13 Mar 2023 03:42:58 +0000 (14:42 +1100)]
mdmon: Improve switchroot interactions.
We need a new mdmon@mdfoo instance to run in the root filesystem after
switch root, as /sys and /dev are removed from the initrd.
systemd will not start a new unit with the same name running while the
old unit is still active, and we want the two mdmon processes to overlap
in time to avoid any risk of deadlock, which can happen when a write is
attempted with no mdmon running.
So we need a different unit name in the initrd than in the root. Apart
from the name, everything else should be the same.
This is easily achieved using a different instance name as the
mdmon@.service unit file already supports multiple instances (for
different arrays).
So start "mdmon@mdfoo.service" from root, but
"mdmon@initrd-mdfoo.service" from the initrd. udev can tell which
circumstance is the case by looking for /etc/initrd-release.
continue_from_systemd() is enhanced so that the "initrd-" prefix can be
requested.
Teach mdmon that a container name like "initrd/foo" should be treated
just like "foo". Note that systemd passes the instance name
"initrd-foo" as "initrd/foo".
We don't need a similar mechanism at shutdown because dracut runs
"mdmon --takeover --all" when appropriate.
NeilBrown [Mon, 13 Mar 2023 03:42:58 +0000 (14:42 +1100)]
mdmon: Remove need for KillMode=none
mdmon needs to keep running during the switchroot out of (at boot) and
then back into (at shutdown) the initrd. It runs until a new mdmon
takes over.
Killmode=none is used to achieve this, with the help of --offroot which
sets argv[0][0] to '@' which systemd understands.
This is needed because mdmon is currently run in system-mdmon.slice
which conflicts with shutdown.target so without Killmode=none mdmon
would get killed early in shutdown when system.mdmon.slice is removed.
As described in systemd.service(5), this conflict with shutdown can be
resolved by explicitly requesting system.slice, which is a natural
counterpart to DefaultDependencies=no.
So add that, and also add IgnoreOnIsolate=true to avoid another possible
source of an early death. With these we no longer need KillMode=none
which the systemd developers have marked as "deprecated".
NeilBrown [Mon, 13 Mar 2023 03:42:58 +0000 (14:42 +1100)]
mdmon: don't test both 'all' and 'container_name'.
If 'all' is not set, then container_name must be NULL, as nothing else
can set it. So simplify the test to ignore container_name.
This makes the purpose of the code more obvious.
NeilBrown [Mon, 13 Mar 2023 03:42:58 +0000 (14:42 +1100)]
Use existence of /etc/initrd-release to detect initrd.
Since v183, systemd has used the existence of /etc/initrd-release to
detect if it is running in an initrd, rather than looking at the magic
number of the root filesystem's device. It is time for mdadm to do the
same.
Khem Raj [Wed, 18 Jan 2023 08:32:36 +0000 (00:32 -0800)]
Define alignof using _Alignof when using C11 or newer
WG14 N2350 made very clear that it is an UB having type definitions
within "offsetof" [1]. This patch enhances the implementation of macro
alignof_slot to use builtin "_Alignof" to avoid undefined behavior on
when using std=c11 or newer
clang 16+ has started to flag this [2]
Fixes build when using -std >= gnu11 and using clang16+
Older compilers gcc < 4.9 or clang < 8 has buggy _Alignof even though it
may support C11, exclude those compilers too
Logan Gunthorpe [Wed, 1 Mar 2023 20:41:33 +0000 (13:41 -0700)]
mdadm: Add --write-zeros option for Create
Add the --write-zeros option for Create which will send a write zeros
request to all the disks before assembling the array. After zeroing
the array, the disks will be in a known clean state and the initial
sync may be skipped.
Writing zeroes is best used when there is a hardware offload method
to zero the data. But even still, zeroing can take several minutes on
a large device. Because of this, all disks are zeroed in parallel using
their own forked process and a message is printed to the user. The main
process will proceed only after all the zeroing processes have completed
successfully.
Logan Gunthorpe [Wed, 1 Mar 2023 20:41:31 +0000 (13:41 -0700)]
Create: Factor out add_disks() helpers
The Create function is massive with a very large number of variables.
Reading and understanding the function is almost impossible. To help
with this, factor out the two pass loop that adds the disks to the array.
This moves about 160 lines into three new helper functions and removes
a bunch of local variables from the main Create function. The main new
helper function add_disks() does the two pass loop and calls into
add_disk_to_super() and update_metadata(). Factoring out the
latter two helpers also helps to reduce a ton of indentation.
Logan Gunthorpe [Wed, 1 Mar 2023 20:41:30 +0000 (13:41 -0700)]
Create: remove safe_mode_delay local variable
All .getinfo_super() call sets the info.safe_mode_delay variables
to a constant value, so no matter what the current state is
that function will always set it to the same value.
Create() calls .getinfo_super() multiple times while creating the array.
The value is stored in a local variable for every disk in the loop
to add disks (so the last disc call takes precedence). The local
variable is then used in the call to sysfs_set_safemode().
This can be simplified by using info.safe_mode_delay directly. The info
variable had .getinfo_super() called on it early in the function so, by the
reasoning above, it will have the same value as the local variable which
can thus be removed.
Doing this allows for factoring out code from Create() in a subsequent
patch.
Logan Gunthorpe [Wed, 1 Mar 2023 20:41:29 +0000 (13:41 -0700)]
Create: goto abort_locked instead of return 1 in error path
The return 1 after the fstat_is_blkdev() check should be replaced
with an error return that goes through the error path to unlock
resources locked by this function.
Wu Guanghao [Fri, 3 Mar 2023 16:21:33 +0000 (00:21 +0800)]
isuper-intel.c: fix double free in load_imsm_mpb()
In load_imsm_mpb() there is potential double free issue on super->buf.
The first location to free super->buf is from get_super_block() <==
load_and_parse_mpb() <== load_imsm_mpb():
4514 if (posix_memalign(&super->migr_rec_buf, MAX_SECTOR_SIZE,
4515 MIGR_REC_BUF_SECTORS*MAX_SECTOR_SIZE) != 0) {
4516 pr_err("could not allocate migr_rec buffer\n");
4517 free(super->buf);
4518 return 2;
4519 }
If the above error condition happens, super->buf is freed and value 2
is returned to get_super_block() eventually. Then in the following code
block inside load_imsm_mpb(),
5289 error:
5290 if (!err) {
5291 s->next = *super_list;
5292 *super_list = s;
5293 } else {
5294 if (s)
5295 free_imsm(s);
5296 close_fd(&dfd);
5297 }
at line 5295 when free_imsm() is called, super->buf is freed again from
the call chain free_imsm() <== __free_imsm(), in following code block,
4651 if (super->buf) {
4652 free(super->buf);
4653 super->buf = NULL;
4654 }
This patch sets super->buf as NULL after line 4517 in load_imsm_mpb()
to avoid the potential double free().
Mateusz Grzonka [Thu, 2 Feb 2023 11:27:01 +0000 (12:27 +0100)]
Mdmonitor: Add helper functions
Add functions:
- is_email_event(),
- get_syslog_event_priority(),
- sprint_event_message(),
with kernel style comments containing more detailed descriptions.
Also update event syslog priorities to be consistent with man. MoveSpare event was described in man as priority info, while implemented as warning. Move event data into a struct, so that it is passed between different functions if needed.
Sort function declarations alphabetically and remove redundant alert() declaration.
Heming Zhao [Thu, 23 Feb 2023 14:39:39 +0000 (22:39 +0800)]
Grow: fix can't change bitmap type from none to clustered.
Commit a042210648ed ("disallow create or grow clustered bitmap with
writemostly set") introduced this bug. We should use 'true' logic not
'== 0' to deny setting up clustered array under WRITEMOSTLY condition.
Kinga Tanska [Fri, 28 Oct 2022 02:51:17 +0000 (04:51 +0200)]
super-intel: make freesize not required for chunk size migration
Freesize is needed to be set for migrations where size of RAID could
be changed - expand. It tells how many free space is determined for
members. In chunk size migartion freesize is not needed to be set,
pointer shouldn't be checked if exists. This commit moves check to
condition which contains size calculations, instead of checking it
always at the first step.
Fix return value when superblock is not set.
Kinga Tanska [Tue, 27 Dec 2022 05:50:43 +0000 (06:50 +0100)]
incremental, manage: do not verify if remove is safe
Function is_remove_safe() was introduced to verify if removing
member device won't cause failed state of the array. This
verification should be used only with set-faulty command. Add
special mode indicating that Incremental removal was executed.
If this mode is used do not execute is_remove_safe() routine.
Kinga Tanska [Tue, 27 Dec 2022 05:50:42 +0000 (06:50 +0100)]
Manage: do not check array state when drive is removed
Array state doesn't need to be checked when drive is
removed, but until now clean state was required. Result
of the is_remove_safe() function will be independent
from array state.
Xiao Ni [Wed, 4 Jan 2023 16:29:20 +0000 (00:29 +0800)]
mdadm/udev: Don't handle change event on raw devices
The raw devices are ready when add event happpens and the raid
can be assembled. So there is no need to handle change events.
And it can cause some inconvenient problems.
For example, the OS is installed on md0(/root) and md1(/home).
md0 and md1 are created on partitions. When it wants to re-install
OS, anaconda can't clear the storage configure. It deletes one
partition and does some jobs. The change event happens. Now
the raid device is assembled again. It can't delete the other
partitions.
So in this patch, we don't handle change event on raw devices
anymore.
Signed-off-by: Xiao Ni <xni@redhat.com> Signed-off-by: Jes Sorensen <jes@trained-monkey.org>
Mateusz Kusiak [Mon, 2 Jan 2023 08:35:24 +0000 (09:35 +0100)]
Change char* to enum in context->update & refactor code
Storing update option in string is bad for frequent comparisons and
error prone.
Replace char array with enum so already existing enum is passed around
instead of string.
Adapt code to changes.
Mateusz Kusiak [Mon, 2 Jan 2023 08:35:23 +0000 (09:35 +0100)]
Manage&Incremental: code refactor, string to enum
Prepare Manage and Incremental for later changing context->update to enum.
Change update from string to enum in multiple functions and pass enum
where already possible.
Mateusz Kusiak [Mon, 2 Jan 2023 08:35:22 +0000 (09:35 +0100)]
Change update to enum in update_super and update_subarray
Use already existing enum, change update_super and update_subarray
update to enum globally.
Refactor function references also.
Remove code specific options from update_options.
Mateusz Kusiak [Mon, 2 Jan 2023 08:35:16 +0000 (09:35 +0100)]
Fix --update-subarray on active volume
Options: bitmap, ppl and name should not be updated when array is active.
Those features are mutually exclusive and share the same data area in IMSM (danger of overwriting by kernel).
Remove check for active subarrays from super-intel.
Since ddf is not supported, apply it globally for all options.
Mateusz Kusiak [Mon, 2 Jan 2023 08:35:15 +0000 (09:35 +0100)]
mdadm: Add option validation for --update-subarray
Subset of options available for "--update" is not same as for "--update-subarray".
Define maps and enum for update options and use them instead of direct comparisons.
Add proper error message.
Fix segfault commit [1] introduced check whether given device is
mddevice, but it happend to terminate Mdmonitor if at least one of given
devices didn't fulfill that condition. In result Mdmonitor service was
no longer started on boot (with --scan option) when config contained some
non-existent array entry.
This commit introduces ommiting non-md devices so scan option can still
be used when config is wrong and allow Mdmonitor service to run on boot.
Giving a list of devices to monitor containing non-existing or
non-md devices will result in monitoring only confirmed mddevices.
Mateusz Kusiak [Thu, 18 Aug 2022 09:47:21 +0000 (11:47 +0200)]
Manage: Block unsafe member failing
Kernel may or may not block mdadm from removing member device if it
will cause arrays failed state. It depends on raid personality
implementation in kernel.
Add verification on requested removal path (#mdadm --set-faulty
command).
NeilBrown [Thu, 25 Aug 2022 22:55:56 +0000 (08:55 +1000)]
super1: report truncated device
When the metadata is at the start of the device, it is possible that it
describes a device large than the one it is actually stored on. When
this happens, report it loudly in --examine.
....
Unused Space : before=1968 sectors, after=-2047 sectors DEVICE TOO SMALL
State : clean TRUNCATED DEVICE
....
Also report in --assemble so that the failure which the kernel will
report will be explained.
mdadm: Device /dev/sdb is not large enough for data described in superblock
mdadm: no RAID superblock on /dev/sdb
mdadm: /dev/sdb has no superblock - assembly aborted
Scenario can be demonstrated as follows:
mdadm: Note: this array has metadata at the start and
may not be suitable as a boot device. If you plan to
store '/boot' on this device please ensure that
your boot-loader understands md/v1.x metadata, or use
--metadata=0.90
mdadm: Defaulting to version 1.2 metadata
mdadm: array /dev/md/test started.
mdadm: stopped /dev/md/test
Unused Space : before=1968 sectors, after=-2047 sectors DEVICE TOO SMALL
State : clean TRUNCATED DEVICE
Unused Space : before=1968 sectors, after=-2047 sectors DEVICE TOO SMALL
State : clean TRUNCATED DEVICE
Kinga Tanska [Fri, 19 Aug 2022 00:55:46 +0000 (02:55 +0200)]
Assemble: check if device is container before scheduling force-clean update
Up to now using assemble with force flag making each array as clean.
Force-clean should not be done for the container. This commit add
check if device is different than container before cleaning.
Grow_reshape should be split into helper functions given its size.
- Add helper function for preparing reshape on external metadata.
- Close cfd file descriptor.
mdadm: Don't open md device for CREATE and ASSEMBLE
The mdadm command tries to open the md device for most modes, first
thing, no matter what. When running to create or assemble an array,
in most cases, the md device will not exist, the open call will fail
and everything will proceed correctly.
However, when running tests, a create or assembly command may be run
shortly after stopping an array and the old md device file may still
be around. Then, if create_on_open is set in the kernel, a new md
device will be created when mdadm does its initial open.
When mdadm gets around to creating the new device with the new_array
parameter it issues this error:
mdadm: Fail to create md0 when using
/sys/module/md_mod/parameters/new_array, fallback to creation via node
This is because an mddev was already created by the kernel with the
earlier open() call and thus the new one being created will fail with
EEXIST. The mdadm command will still successfully be created due to
falling back to the node creation method. However, the error message
itself will fail any test that's running it.
This issue is a race condition that is very rare, but a recent change
in the kernel caused this to happen more frequently: about 1 in 50
times.
To fix this, don't bother trying to open the md device for CREATE,
ASSEMBLE and BUILD commands, as the file descriptor will never be used
anyway even if it is successfully openned. The mdfd has not been used
for these commands since:
7f91af49ad09 ("Delay creation of array devices for assemble/build/create")
The checks that were done on the open device can be changed to being
done with stat.
Side note: it would be nice to disable create_on_open as well to help
solve this, but it seems the work for this was never finished. By default,
mdadm will create using the old node interface when a name is specified
unless the user specifically puts names=yes in a config file, which
doesn't seem to be common or desirable to require this..
Current behavior is not documented and tested. This test is a base for
future improvements. It is enough to test it only with native metadata,
because it is generic code. Generated properties are passed to metadata
handler.
Mateusz Grzonka [Fri, 12 Aug 2022 14:36:02 +0000 (16:36 +0200)]
mdadm: Replace obsolete usleep with nanosleep
According to POSIX.1-2001, usleep is considered obsolete.
Replace it with a wrapper that uses nanosleep, as recommended in man.
Add handy macros for conversions between msec, usec and nsec.
Logan Gunthorpe [Wed, 22 Jun 2022 20:25:19 +0000 (14:25 -0600)]
tests: Add broken files for all broken tests
Each broken file contains the rough frequency of brokeness as well
as a brief explanation of what happens when it breaks. Estimates
of failure rates are not statistically significant and can vary
run to run.
This is really just a view from my window. Tests were done on a
small VM with the default loop devices, not real hardware. We've
seen different kernel configurations can cause bugs to appear as well
(ie. different block schedulers). It may also be that different race
conditions will be seen on machines with different performance
characteristics.
These annotations were done with the kernel currently in md/md-next:
facef3b96c5b ("md: Notify sysfs sync_completed in md_reap_sync_thread()")
Logan Gunthorpe [Wed, 22 Jun 2022 20:25:18 +0000 (14:25 -0600)]
mdadm/test: Mark and ignore broken test failures
Add functionality to continue if a test marked as broken fails.
To mark a test as broken, a file with the same name but with the suffix
'.broken' should exist. The first line in the file will be printed with
a KNOWN BROKEN message; the rest of the file can describe the how the
test is broken.
Also adds --skip-broken and --skip-always-broken to skip all the tests
that have a .broken file or to skip all tests whose .broken file's first
line contains the keyword always.
tests/04update-metadata: avoid passing chunk size to raid1
'04update-metadata' test fails with error, "specifying chunk size is
forbidden for this level" added by commit, 5b30a34aa4b5e. Hence,
correcting the test to ignore passing chunk size to raid1.
Some of the test cases fail because raid0 creation fails with the error,
"0.90 metadata does not support layouts for RAID0" added by commit, 329dfc28debb. Fix some of the test cases by switching from raid0 to
linear level for 0.9 metadata where possible.
Logan Gunthorpe [Wed, 22 Jun 2022 20:25:11 +0000 (14:25 -0600)]
mdadm: Fix mdadm -r remove option regression
The commit noted below globally adds a parameter to the -r option but missed
the fact that -r is used for another purpose: --remove.
After that commit, a command such as:
mdadm /dev/md0 -r /dev/loop0
will do nothing seeing the device parameter will be consumed as a
argument to the -r option; thus, there will only be one device
seen one the command line, devs_found will only be 1 and nothing will
happen.
This caused the 01r5integ and 01raid6integ tests to hang indefinitely
as mdadm did not remove the failed device. With the device not removed,
it would not be readded. Then the loop waiting for the array status to
change would loop forever.
This commit was recently reverted, but the legitimate fix for the
monitor operations was still not fixed. So add specific monitor
short ops to re-fix the --monitor -r option.
Logan Gunthorpe [Wed, 22 Jun 2022 20:25:10 +0000 (14:25 -0600)]
monitor: Avoid segfault when calling NULL get_bad_blocks
Not all struct superswitch implement a get_bad_blocks() function,
yet mdmon seems to call it without checking for NULL and thus
occasionally segfaults in the test 10ddf-geometry.
Logan Gunthorpe [Wed, 22 Jun 2022 20:25:09 +0000 (14:25 -0600)]
mdadm/Grow: Fix use after close bug by closing after fork
The test 07reshape-grow fails most of the time. But it succeeds around
1 in 5 times. When it does succeed, it causes the tests to die because
mdadm has segfaulted.
The segfault was caused by mdadm attempting to repoen a file
descriptor that was already closed. The backtrace of the segfault
was:
#0 __strncmp_avx2 () at ../sysdeps/x86_64/multiarch/strcmp-avx2.S:101
#1 0x000056146e31d44b in devnm2devid (devnm=0x0) at util.c:956
#2 0x000056146e31dab4 in open_dev_flags (devnm=0x0, flags=0)
at util.c:1072
#3 0x000056146e31db22 in open_dev (devnm=0x0) at util.c:1079
#4 0x000056146e3202e8 in reopen_mddev (mdfd=4) at util.c:2244
#5 0x000056146e329f36 in start_array (mdfd=4,
mddev=0x7ffc55342450 "/dev/md0", content=0x7ffc55342860,
st=0x56146fc78660, ident=0x7ffc55342f70, best=0x56146fc6f5d0,
bestcnt=10, chosen_drive=0, devices=0x56146fc706b0, okcnt=5,
sparecnt=0, rebuilding_cnt=0, journalcnt=0, c=0x7ffc55342e90,
clean=1, avail=0x56146fc78720 "\001\001\001\001\001",
start_partial_ok=0, err_ok=0, was_forced=0)
at Assemble.c:1206
#6 0x000056146e32c36e in Assemble (st=0x56146fc78660,
mddev=0x7ffc55342450 "/dev/md0", ident=0x7ffc55342f70,
devlist=0x56146fc6e2d0, c=0x7ffc55342e90)
at Assemble.c:1914
#7 0x000056146e312ac9 in main (argc=11, argv=0x7ffc55343238)
at mdadm.c:1510
The file descriptor was closed early in Grow_continue(). The noted commit
moved the close() call to close the fd above the fork which caused the
parent process to return with a closed fd.
This meant reshape_array() and Grow_continue() would return in the parent
with the fd forked. The fd would eventually be passed to reopen_mddev()
which returned an unhandled NULL from fd2devnm() which would then be
dereferenced in devnm2devid.
Fix this by moving the close() call below the fork. This appears to
fix the 07revert-grow test. While we're at it, switch to using
close_fd() to invalidate the file descriptor.
Fixes: 77b72fa82813 ("mdadm/Grow: prevent md's fd from being occupied during delayed time") Cc: Alex Wu <alexwu@synology.com> Cc: BingJing Chang <bingjingc@synology.com> Cc: Danny Shih <dannyshih@synology.com> Cc: ChangSyun Peng <allenpeng@synology.com> Signed-off-by: Logan Gunthorpe <logang@deltatee.com> Acked-by: Mariusz Tkaczyk <mariusz.tkaczyk@linux.intel.com> Signed-off-by: Jes Sorensen <jes@trained-monkey.org>
Logan Gunthorpe [Wed, 22 Jun 2022 20:25:06 +0000 (14:25 -0600)]
Makefile: Don't build static build with everything and everything-test
Running the test suite requires building everything, but it seems to be
difficult to build the static version of mdadm now seeing there
is no readily available static udev library.
The test suite doesn't need the static binary so just don't build it
with the everything or everything-test targets.
Leave the mdadm.static and install-static targets in place in case
someone still has a use case for the static binary.
Monitor: use devname as char array instead of pointer
Device name wasn't filled properly due to incorrect use of strcpy.
Strcpy was used twice. Firstly to fill devname with "/dev/md/"
and then to add chosen name. First strcpy result was overwritten by
second one (as a result <device_name> instead of "/dev/md/<device_name>"
was assigned). This commit changes this implementation to use snprintf
and devname with fixed size.
imsm_create_metadata_update_for_size_change() that returns u_size value
could return 0 in the past. As its behavior changed, and returned value
is always the size of imsm_update_size_change structure, check for
u_size is no longer needed.
Mariusz Tkaczyk [Mon, 20 Jun 2022 16:10:43 +0000 (00:10 +0800)]
imsm: block changing slots during creation
If user specifies drives for array creation, then slot order across
volumes is not preserved.
Ideally, it should be checked in validate_geometry() but it is not
possible in current implementation (order is determined later).
Add verification in add_to_super_imsm_volume() and throw error if
mismatch is detected.
IMSM allows to use only same members within container.
This is not hardware dependency but metadata limitation.
Therefore, 09-imsm-overlap test is removed. Testing it is pointless.
After this patch, creation in this scenario is blocked. Offset
verification is covered in other tests.
Mariusz Tkaczyk [Mon, 20 Jun 2022 16:10:42 +0000 (00:10 +0800)]
imsm: use same slot across container
Autolayout relies on drives order on super->disks list, but
it is not quaranted by readdir() in sysfs_read(). As a result
drive could be put in different slot in second volume.
Make it consistent by reffering to first volume, if exists.
Heming Zhao [Mon, 20 Jun 2022 16:10:40 +0000 (00:10 +0800)]
mdadm/super1: restore commit 45a87c2f31335 to fix clustered slot issue
Commit 9d67f6496c71 ("mdadm:check the nodes when operate clustered
array") modified assignment logic for st->nodes in write_bitmap1(),
which introduced bitmap slot issue:
load_super1 didn't set up supertype.nodes, which made spare disk only
have one slot info. Then it triggered kernel md_bitmap_load_sb to get
wrong bitmap slot data.
For fixing this issue, there are two methods:
1> revert the related code of commit 9d67f6496c71. and restore the code
from former commit 45a87c2f31335 ("super1: add more checks for
NodeNumUpdate option").
st->nodes value would be 0 & 1 under current code logic. i.e.
When adding a spare disk, there is no place to init st->nodes, and
the value is ZERO.
2> keep 9d67f6496c71, add additional ->nodes handling in load_super1(),
let load_super1 to set st->nodes when bitmap is BITMAP_MAJOR_CLUSTERED.
Under current mdadm code logic, load_super1 will be called many
times, any new code in load_super1 will cost mdadm running more time.
And more reason is I prefer as much as possible to limit clustered
code spreading in every corner.
the output of current "mdadm -X /dev/sdc":
(there should be (by default) 4 slot info for correct output)
```
Filename : /dev/sdc
Magic : 6d746962
Version : 5
UUID : a74642f8:a6b1fba8:58e1f8db:cfe7b082
Events : 29
Events Cleared : 0
State : OK
Chunksize : 64 MB
Daemon : 5s flush period
Write Mode : Normal
Sync Size : 306176 (299.00 MiB 313.52 MB)
Bitmap : 5 bits (chunks), 5 dirty (100.0%)
```
And mdadm later operations will trigger kernel output error message:
(triggered by "mdadm -A /dev/md0 /dev/sdb /dev/sdc")
```
kernel: md0: invalid bitmap file superblock: bad magic
kernel: md_bitmap_copy_from_slot can't get bitmap from slot 1
kernel: md-cluster: Could not gather bitmaps from slot 1
kernel: md0: invalid bitmap file superblock: bad magic
kernel: md_bitmap_copy_from_slot can't get bitmap from slot 2
kernel: md-cluster: Could not gather bitmaps from slot 2
kernel: md0: invalid bitmap file superblock: bad magic
kernel: md_bitmap_copy_from_slot can't get bitmap from slot 3
kernel: md-cluster: Could not gather bitmaps from slot 3
kernel: md-cluster: failed to gather all resyn infos
kernel: md0: detected capacity change from 0 to 612352
```
projects / thirdparty / mdadm.git / log
