| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
Now we have the cached meta_dev_idx member,
we can get rid of a few rcu_read_lock() sections and rcu_dereference().
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
Signed-off-by: Jens Axboe <[email protected]>
|
|
Introduce two new on-disk meta data fields: al_stripes and al_stripe_size_4k
The intended use case is activity log on RAID 0 or similar.
Logically consecutive transactions will advance their on-disk position
by al_stripe_size_4k 4kB (transaction sized) blocks.
Right now, these are still asserted to be the backward compatible
values al_stripes = 1, al_stripe_size_4k = 8 (which amounts to 32kB).
Also introduce a caching member for meta_dev_idx in the in-core
structure: even though it is initially passed in in the rcu-protected
disk_conf structure, it cannot change without a detach/attach cycle.
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
Signed-off-by: Jens Axboe <[email protected]>
|
|
Add a comment about our meta data layout variants,
and rename a few defines (e.g. MD_RESERVED_SECT -> MD_128MB_SECT)
to make it clear that they are short hand for fixed constants,
and not arbitrarily to be redefined as one may see fit.
Properly pad struct meta_data_on_disk to 4kB,
and initialize to zero not only the first 512 Byte,
but all of it in drbd_md_sync().
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
Signed-off-by: Jens Axboe <[email protected]>
|
|
This fixes ASSERT( mdev->state.disk == D_FAILED ) in drivers/block/drbd/drbd_main.c
When we detach from local disk, we let the local refcount hit zero twice.
First, we transition to D_FAILED, so we won't give out new references
to incoming requests; we still may give out *internal* references, though.
Once the refcount hits zero [1] while in D_FAILED, we queue a transition
to D_DISKLESS to our worker. We need to queue it, because we may be in
atomic context when putting the reference.
Once the transition to D_DISKLESS actually happened [2] from worker context,
we don't give out new internal references either.
Between hitting zero the first time [1] and actually transition to
D_DISKLESS [2], there may be a few very short lived internal get/put,
so we may hit zero more than once while being in D_FAILED, or even see a
race where a an internal get_ldev() happened while D_FAILED, but the
corresponding put_ldev() happens just after the transition to D_DISKLESS.
That's why we have the additional test_and_set_bit(GO_DISKLESS,);
and that's why the assert was placed wrong.
Since there was exactly one code path left to drbd_go_diskless(),
and that checks already for D_FAILED, drop that assert,
and fold in the drbd_queue_work().
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
Signed-off-by: Jens Axboe <[email protected]>
|
|
Convert to the much saner new idr interface.
Signed-off-by: Tejun Heo <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
|
|
When we notice a disk failure on the receiving side,
we stop sending it new incoming writes.
Depending on exact timing of various events, the same transfer log epoch
could end up containing both replicated (before we noticed the failure)
and local-only requests (after we noticed the failure).
The sanity checks in tl_release(), called when receiving a
P_BARRIER_ACK, check that the ack'ed transfer log epoch matches
the expected epoch, and the number of contained writes matches
the number of ack'ed writes.
In this case, they counted both replicated and local-only writes,
but the peer only acknowledges those it has seen. We get a mismatch,
resulting in a protocol error and disconnect/reconnect cycle.
Messages logged are
"BAD! BarrierAck #%u received with n_writes=%u, expected n_writes=%u!\n"
A similar issue can also be triggered when starting a resync while
having a healthy replication link, by invalidating one side, forcing a
full sync, or attaching to a diskless node.
Fix this by closing the current epoch if the state changes in a way
that would cause the replication intent of the next write.
Epochs now contain either only non-replicated,
or only replicated writes.
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
We no longer need the connector.
But we need libcrc32c.
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
This was introduces when moving the code over from the 8.3 codebase
with commit 328e0f125bf41f4f33f684db22015f92cb44fe56
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
drbd_set_role(, R_PRIMARY, ) does the state change to Primary,
some more housekeeping, and possibly generates a new UUID set.
All of this holding the "state_mutex".
The connection handshake involves sending of various state information,
including the current data generation UUID set, and two connection
state changes from C_WF_CONNECTION to C_WF_REPORT_PARAMS further to
a number of different outcomes, resync being one of them.
If the connection handshake happens between the state change to Primary
and the generation of the new UUIDs, the resync decision based on the
old UUID set may be confused, depending on circumstances.
Make sure that, before we do the handshake, any promotion to Primary
role will either be complete (including the housekeeping stuff), or can
see, and serialize with, the ongoing handshake, based on the
"STATE_SENT" bit, which is set when we start the handshake, and cleared
only when we leave C_WF_REPORT_PARAMS again.
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
We need to propagate the configuration into the flag bits,
or it won't be effective.
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
Smatch complained about it this redundanct check.
The check was introduced in 2006-09-13. On 2007-07-24 the body of the
function was enclosed by get_ldev()/put_ldev() reference counting.
Since then the check is useless and miss leading.
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
Compiling drbd yields:
drivers/block/drbd/drbd_state.c: In function ‘_conn_request_state’:
drivers/block/drbd/drbd_state.c:1804:5: error: macro "wait_event_lock_irq" passed 4 arguments, but takes just 3
drivers/block/drbd/drbd_state.c:1801:3: error: ‘wait_event_lock_irq’ undeclared (first use in this function)
drivers/block/drbd/drbd_state.c:1801:3: note: each undeclared identifier is reported only once for each function it appears in
drivers/block/drbd/drbd_state.c: At top level:
drivers/block/drbd/drbd_state.c:1734:1: warning: ‘_conn_rq_cond’ defined but not used [-Wunused-function]
Due to drbd having copied the MD definition for wait_event_lock_irq()
as well. Kill them.
Signed-off-by: Jens Axboe <[email protected]>
|
|
Use copy_highpage() to copy from one page to another.
Signed-off-by: Akinobu Mita <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
The 8.3.12 commit drbd: Bugfix for the connection behavior fixes a
"wasted established connection", if a former connection attempt failed
during its early stages.
However it opened a window for a regression, if a connection attempt
fails during its last stages. The result was a terminated receiver
thread, that left behind the supposedly transient "C_UNCONNECTED" state.
Any later requests to change the connection state fail, as they wait for
the connection state to "stabilize".
Fix: short circuit and keep retrying to restablish a new connection,
if we don't reach C_WF_REPORT_PARAMS.
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
Signed-off-by: Jing Wang <[email protected]>
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
|
|
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
If the disk has failed already, there is no point trying to change the
bitmap. drbd_set_out_of_sync() already had this safeguard,
time to add it to drbd_set_in_sync() as well.
This also prevents some warning messages, like
FIXME asender in bm_change_bits_to, bitmap locked for 'detach' by worker
if our disk fails during resync, while there are some resync acks queued up.
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
recent commit
drbd: always write bitmap on detach
introduced a bitmap writeout during detach,
which obviously needs some meta data device to write to.
Unfortunately, that same error path may be taken if we fail to attach,
e.g. due to UUID mismatch, after we changed state to D_ATTACHING,
but before the lower level device pointer is even assigned.
We need to test for presence of mdev->ldev.
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
If we detach due to local read-error (which sets a bit in the bitmap),
stay Primary, and then re-attach (which re-reads the bitmap from disk),
we potentially lost the "out-of-sync" (or, "bad block") information in
the bitmap.
Always (try to) write out the changed bitmap pages before going diskless.
That way, we don't lose the bit for the bad block,
the next resync will fetch it from the peer, and rewrite
it locally, which may result in block reallocation in some
lower layer (or the hardware), and thereby "heal" the bad blocks.
If the bitmap writeout errors out as well, we will (again: try to)
mark the "we need a full sync" bit in our super block,
if it was a READ error; writes are covered by the activity log already.
If that superblock does not make it to disk either, we are sorry.
Maybe we just lost an entire disk or controller (or iSCSI connection),
and there actually are no bad blocks at all, so we don't need to
re-fetch from the peer, there is no "auto-healing" necessary.
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
force-detached
The intention of force-detach is to be able to deal with a completely
unresponsive lower level IO stack, which does not even deliver error
completions anymore, but no completion at all.
In all other cases, we must still wait for the meta data IO completion.
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
This has not yet been observed, but conceivably, when using GFP_KERNEL
allocations from drbd_md_sync(), drbd_flush_after_epoch() or
receive_SyncParam(), we could trigger additional IO to our own device,
or an other device in a criss-cross setup, and end up in a local
deadlock, or potentially a distributed deadlock in a criss-cross setup
involving the peer blocked in a similar way waiting for us to make
progress.
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
The former comment arguing that GFP_KERNEL was good enough was wrong: it
did not take resize into account at all, and assumed the only path
leading here was the normal attach on a still secondary device, so no
deadlock would be possible.
Both resize on a Primary, or attach on a diskless Primary,
could potentially deadlock.
drbd_bm_resize() is called while IO to the respective device is
suspended, so we must use GFP_NOIO to avoid potential deadlock.
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
Using list_move_tail() instead of list_del() + list_add_tail().
spatch with a semantic match is used to found this problem.
(http://coccinelle.lip6.fr/)
Signed-off-by: Wei Yongjun <[email protected]>
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
"aborting" requests, or force-detaching the disk, is intended for
completely blocked/hung local backing devices which do no longer
complete requests at all, not even do error completions. In this
situation, usually a hard-reset and failover is the only way out.
By "aborting", basically faking a local error-completion,
we allow for a more graceful swichover by cleanly migrating services.
Still the affected node has to be rebooted "soon".
By completing these requests, we allow the upper layers to re-use
the associated data pages.
If later the local backing device "recovers", and now DMAs some data
from disk into the original request pages, in the best case it will
just put random data into unused pages; but typically it will corrupt
meanwhile completely unrelated data, causing all sorts of damage.
Which means delayed successful completion,
especially for READ requests,
is a reason to panic().
We assume that a delayed *error* completion is OK,
though we still will complain noisily about it.
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
is_valid_transition() might return SS_NOTHING_TO_DO.
The condition function _req_st_cond() returned SS_NOTHING_TO_DO, which
caused the wait_event to abort too early. Therefore drbd_req_state()
did not consume the next CL_ST_CHG_SUCCESS or SS_CW_FAILED_BY_PEER
causing serve disruption of the state machine logic...
Detaching from a single volue was one way to trigger this bug.
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
We use the RQ_POSTPONED flag to mark a request for several reasons.
It may be a conflicting request in a dual-primary setup,
where conflict detection and resolution on the peer decided that
this request needs to be re-submitted, it needs to re-enter
drbd_make_request() to fix the data divergence caused by these
conflicting, partially overlapping, quasi-simultaneous requests.
In this case we need to mark the corresponding area as out-of-sync,
before we call drbd_al_complete_io().
We also use the RQ_POSTPONED flag to just "push back" a request,
before even processing it, if IO is suspended for some reason.
In this case, as this request was neither submitted nor sent yet,
we must not touch the bitmap.
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
A postponed request might has RQ_IN_ACT_LOG already set, but
is POSTPONED before it gets something in the RQ_LOCAL_MASK
set. Up to now this caused a left-over active extent.
Fix that by only testing for the RQ_IN_ACT_LOG bit in drbd_req_destroy()
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
Without this, the meta-data gets updates after 5 seconds by the
md_sync_timer. Better to do it immeditaly after a state change.
If the asender detects a network failure, it may take a bit until
the worker processes the according after-conn-state-change work item.
The worker might be blocked in sending something, i.e. it
takes until it gets into its timeout. That is 6 seconds by
default which is longer than the 5 seconds of the md_sync_timer.
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
* Postponed requests should not set or clear out-of-sync marks
* When a request gets postponed we need to drop its reference
mdev->local_cnt (put_ldev()).
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
With merging the commit
'drbd: Delay/reject other state changes while establishing a connection'
the condition check for clearing the flag was wrong.
Move the bit clearing to the __drbd_set_state() function
in order to have it already cleared for the other parts of
the function. I.e. clearing the susp_fen in the after_state_ch() function.
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
_conn_requests_state()
When _conn_requests_state() is used to change other parts of the state
than the connection, do not check for a valid connection transition.
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
The previous way of doing the state change was also okay since the
state change on the susp flag gets propagated from the mdev
to the tconn.
Fortunately all this goes away in drbd-9.0
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
fence handler
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
If the md_sync_timer triggers a second time,
while the work queued during the first time is still pending,
this could result in list_add() of an already added item,
and corrupt the work item list.
This likely only triggered because of the erroneous
batch-dequeueing of work items fixed with
drbd: dequeue single work items in wait_for_work()
Still, skip queueing if md_sync_work is already queued.
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
As long as we still use drbd_queue_work_front(),
we must only dequeue the single first item during normal operation.
The comment in drbd_worker() even says so,
but bc8a5a1 drbd: remove struct drbd_tl_epoch objects (barrier works)
introduced the batch dequeueing again via list_splice_init() in
wait_for_work().
Change back to list_move() of the first item, if any.
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
Documentation of mutex_unlock says
we must not use it in interrupt context.
So do not call it while holding the spin_lock_irq,
but give up the spinlock temporarily.
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
If the preconditions for a state change change after the wait_event() we
might hit the BUG() statement in conn_set_state().
With holding the spin_lock while evaluating the condition AND until the
actual state change we ensure the the preconditions can not change anymore.
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
drbd_adm_disk_opts() does
wait_event(mdev->al_wait, lc_try_lock(mdev->act_log));
drbd_al_shrink(mdev);
If the device is very busy, this can take a very long time to succeed.
Fix this by temporarily suspending IO,
then quickly change the settings, and resume.
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
We must only send P_BARRIER for epochs we actually sent P_DATA in.
If we (re-)establish a connection, we reinitialized the
send.current_epoch_nr, but forgot to reset send.current_epoch_writes.
This could result in a spurious P_BARRIER with stale epoch information,
and a disconnect/reconnect cycle once the then "unexpected"
P_BARRIER_ACK is received:
BAD! BarrierAck #28823 received, expected #28829!
Introduce re_init_if_first_write() and maybe_send_barrier() helpers,
and call them appropriately for read/write/set-out-of-sync requests.
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
drbd_disconnected() is supposed to clear the resync lru cache,
by calling drbd_rs_cancel_all().
We must do so before we call drbd_flush_workqueue(), as at least the
callback w_restart_disk_io() may wait for resync progres, and would
otherwise deadlock.
drbd_finish_peer_reqs() may again populate that cache, which will
then potentially be stale after the next resync handshake and bitmap
exchange, we have to do it again after that.
A stale resync lru cache causes no harm but ugly messages like this:
BAD! sector=196608s enr=6 rs_left=-256 rs_failed=0 count=256 cstate=SyncTarget
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
Disconnecting is a cluster wide state change. In case the peer node agrees
to the state transition, it sends back the fact on the meta-data connection
and closes both sockets.
In case the node node that initiated the state transfer sees the closing
action on the data-socket, before the P_STATE_CHG_REPLY packet, it was
going into one of the network failure states.
At least with the fencing option set to something else thatn "dont-care",
the unclean shutdown of the connection causes a short IO freeze or
a fence operation.
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
There is at least the worker context, the receiver context, the context of
receiving netlink packts.
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
We need to write the whole bitmap after we moved the meta data
due to an online resize operation.
With the support for one peta byte devices bitmap IO was optimized
to only write out touched pages. This optimization must be turned
off when writing the bitmap after an online resize.
This issue was introduced with drbd-8.3.10.
The impact of this bug is that after an online resize, the next
resync could become larger than expected.
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|
|
In various places (E.g. CONNECTION_LOST_WHILE_PENDING) the
RQ_COMPLETION_SUSP mask is passed in the clear set to mod_rq_state().
The issue was that it tried to clear the RQ_COMPLETION_SUSP bit
out of the state mask first, and eventuelly set it afterwards,
in the drbd_req_put_completion_ref() function.
Fixed that by moving the reference getting out of
drbd_req_put_completion_ref() into the mod_rq_state(), before the place
where the extra reference might be put.
Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Lars Ellenberg <[email protected]>
|