This adds a necessary race breaker to these commits:
    drbd: fix for possible deadlock on IO error during resync
    drbd: drop wrong debug asserts, fix recently introduced race

What we do is get a refcount, check the state, then depending on the
state and the requested minimum disk state, either hold it (success),
or give it back immediately (failed "try lock").

Some code paths (flushing of drbd metadata) may still grab and hold a
refcount even if we are D_FAILED (application IO won't).
So even if we hit local_cnt == 0 once after being D_FAILED,
we still need to wait for that again after we changed to D_DISKLESS.
Once local_cnt reaches 0 while we are D_DISKLESS, we can be sure that
no one will look at the protected members anymore, so only then is it
safe to free them.

We cannot easily convert to standard locking primitives here, as we want
to be able to use it in atomic context (we always do a "try lock"),
as well as hold references for a "long time" (from IO submission to
completion callback).
Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

Consolidate the ifdef's for the debug level, accidentally the used both
DEBUG and DRBD_DEBUG_MD_SYNC.  Default to off.

For production, we can safely reduce the grace period for this timer
again the the value we used to have.
Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

 commit 2372c38caadeaebc68a5ee190782c2a0df01edc3
 drbd: fix for possible deadlock on IO error during resync

introduced a new ASSERT, which turns out to be wrong. Drop it.

Also serialize the state change to D_DISKLESS with the after state
change work of the -> D_FAILED transition, don't open a new race.
Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

Scenario:

Something (say, flush-147:0) is in drbd_al_begin_io,
holding a local_cnt, waiting for the resync to make progress.

Disk fails, worker in after_state_ch does drbd_rs_cancel_all,
then waits for local_cnt to drop to zero.

flush-147:0 is woken by drbd_rs_cancel_all, needs to write an AL
transaction, and queues that on the worker.

Deadlock.

Fix: do not wait in the worker, have put_ldev() trigger the
state change D_FAILED -> D_DISKLESS when necessary.
put_ldev() cannot do the state change directly, as it may or may not
already hold various spinlocks. We queue a short work instead.
Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

If it was an "empty" resync, the SyncSource may have already "finished"
the resync and rotated the UUIDs, before noticing the connection loss
(and generating a new uuid, if Primary, rotating again), while the
SyncTarget did not change its uuids at all, or only got to the previous
sync-uuid.
This would then again lead to a full sync on next handshake
(see also Bug #251).

Fix:
Use explicit resync finished notification even for empty resyncs,
do not finish an empty resync implicitly on the SyncSource.
Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

drbd commit 17c854fea474a5eb3cfa12e4fb019e46debbc4ec
drbd: receiving of big packets, for payloads between 64kByte and 4GByte
introduced a new on-the-wire packet header format.  We must no longer
assume either format, but use the result of whatever drbd_recv_header
has decoded.
Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

If we release the page pointed to by md_io_tmpp, we need to zero out the
pointer, too, as that may be used later to decide whether we need to
allocate a new page again.

Impact: a previously freed page may be used and clobbered.  Depending on
what that particular page is being used for meanwhile, this may result
in silent data corruption of completely unrelated things.

Only of concern on devices with logical_block_size != 512 byte,
if you re-attach after becoming diskless once.
Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

There are three ways to get IO suspended:

 * Loss of any access to data
 * Fence-peer-handler running
 * User requested to suspend IO

Track those in different bits, so that one condition clearing its
state bit does not interfere with the other two conditions.

Only when the user resumes IO he overrules all three bits.

The fact is hidden from the user, he sees only a single suspend
bit.
Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

If a synctarget lost connection while being WFSyncUUID,
due to "state sanitizing", the attempted state change to SyncTarget
looked like an "invalidate" to after_state_ch() later,
thus caused a full sync on next handshake (Bug #318).

drbd0: PingAck did not arrive in time.
drbd0: peer( Primary -> Unknown ) conn( WFSyncUUID -> NetworkFailure ) pdsk( UpToDate -> DUnknown )

        from  : { cs:NetworkFailure ro:Secondary/Unknown ds:UpToDate/DUnknown r--- }
        to    : { cs:SyncTarget ro:Secondary/Unknown ds:Inconsistent/DUnknown r--- }
        after sanizising, resulted in
        state: { cs:NetworkFailure ro:Secondary/Unknown ds:Inconsistent/DUnknown r--- }
        drbd0: disk( UpToDate -> Inconsistent )

Fix:
don't mask state transition errors in "sanitizing",
so the requested state change to SyncTarget fails,
instead of being implicitly "remaped" to invalidate.
Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

addendum to baa33ae4eaa4477b60af7c434c0ddd1d182c1ae7

The race:
    drbd_md_sync()
	if (!test_and_clear_bit(MD_DIRTY, &mdev->flags))
		return;
    ==> RACE with drbd_md_mark_dirty() rearming the timer.
	del_timer(&mdev->md_sync_timer);

    Fixed by moving the del_timer before the test_and_clear_bit.

Additionally only rearm the timer in drbd_md_mark_dirty, if MD_DIRTY was
not already set, reduce the grace period from five to one second, and
add an ifdef'ed debuging aid to find code paths missing an explicit
drbd_md_sync, if any, as those are the only relevant ones for this race.
Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

The actual race happened int the drbd_start_resync() function. Where
drbd_resync_finished() -> __drbd_set_state() set STOP_SYNC_TIMER and
armed the timer.

If the timer fired before execution reaches the mod_timer statement
at the end of drbd_start_resync() the latter would cause an
unexpected call to w_make_resync_request().

Removed the STOP_SYNC_TIMER bit, and base it on the connection state.

The STOP_SYNC_TIMER bit probably originates probably the time before
the state engine.
Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

The race:
	drbd_md_mark_dirty()
	drbd_md_sync()
		if (!test_and_clear_bit(MD_DIRTY, &mdev->flags))
			return;
		drbd_md_sync_page_io(mdev, mdev->ldev, sector, WRITE)
  ==> RACE
		clear_bit(MD_DIRTY, &mdev->flags); <== spurious

Fixed by removing the spurious clear_bit.
Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

When the complete device is marked as out of sync, we can disable
updates of the on disk AL. Currently AL updates are only disabled
if one uses the "invalidate-remote" command on an unconnected,
primary device, or when at attach time all bits in the bitmap are
set.

As of now, AL updated do not get disabled when a all bits becomes
set due to application writes to an unconnected DRBD device.
While this is a missing feature, it is not considered important,
and might get added later.

BTW, after initializing a "one legged" DRBD device
drbdadm create-md resX
drbdadm -- --force primary resX
AL updates also get disabled, until the first connect.
Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

Now we have multiple BIOs per ee, packets with a 32 bit length field,
it gets time to use these goodies.
Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

We now track the data rate of locally submitted resync related requests,
and can thus detect non-resync activity on the lower level device.

If the current sync rate is above c-min-rate, and the lower level device
appears to be busy, we throttle the resyncer.
Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

The current resync speed as displayed in /proc/drbd fluctuates a lot.
Using an array of rolling marks makes this calculation much more stable.
We used to have this (a long time ago with 0.7), but it got lost somehow.

If "stalled", do not discard the rest of the information, just add a
" (stalled)" tag to the progress line.

This patch also shortens a spinlock critical section somewhat, and
reduces the number of atomic operations in put_ldev.
Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

We may not free tl_hash when IO is suspended, since we can not wait
until ap_bio_cnt reaches zero.

We can do this after susp reched 0, since then tl_clear was called
Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

After disconnect (most likely mdev->net_cnt == 0) and we are
still in an unstable state (!drbd_state_is_stable()). When we
get an IO request in drbd_get_max_buffers() (called from
__inc_ap_bio_cond(), called from inc_ap_bio()) we wake up
misc_wait. Misc_wait is also used in inc_ap_bio() to sleep
until the outcome of __inc_ap_bio_cond() changes. => Busy loop!

Solution: Have a dedicated wait queue for get_net_conf() and
put_net_conf().
Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

When a fencing policy of "resource-and-stonith" is configured,
and DRBD looses connection to it's peer, we can delay the
creation of a new current-UUID until IO gets thawed.

That allows one to deploy fence-peer handlers that actually
commit suicide on the machine they get started.
Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

State transitions in the space of non-allowed states used
to be very noisy. Reduce that, since that has little value
for the majority of the user base.
Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

When no data is accessible (no connection to the peer, nor a local disk)
allow the user to select to freeze all IO operations instead of getting
IO errors.
Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

If IO was frozen for a temporal network outage, resend the
content of the transfer-log into the newly established connection.
Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

With that the drbd_fail_pending_reads() function becomes obsolete.
Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>

It was a now abandoned attempt to throttle resync bandwidth
based on the delay it causes on the bulk data socket.
It has no userbase yet, and has been disabled by
9173465ccb51c09cc3102a10af93e9f469a0af6f already.
This removes the now unused code.

The basic feature, namely using up "idle" bandwith
of network and disk IO subsystem, with minimal impact
to application IO, is being reimplemented differently.
Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>
Cc: stable@kernel.org
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

Signed-off-by: NPhilipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: NLars Ellenberg <lars.ellenberg@linbit.com>
Cc: stable@kernel.org
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>