A connection timeout affects all volumes of a resource!
Under the following conditions:
A resource with multiple volumes
AND
ko-count >=1
AND
a write request triggers the timeout (ko-count * timeout)
DRBD's internal state gets confused. That in turn may
lead to very miss leading follow up failures. E.g.
"BUG: scheduling while atomic"
CC: stable@kernel.org # v3.17
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
This was not noticed for many years. Affects operation if
md raid is used a backing device for DRBD.
CC: stable@kernel.org # v3.2+
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
request_timer_fn() takes resource->req_lock via the device and releases it via
the connection. Avoid this as it is confusing static code checkers.
Reported-by: "Dan Carpenter" <dan.carpenter@oracle.com>
Signed-off-by: Andreas Gruenbacher <agruen@linbit.com>
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
During resync, if we need to block some specific incoming write because
of active resync requests to that same range, we potentially caused
*all* new application writes (to "cold" activity log extents) to block
until this one request has been processed.
Improve the do_submit() logic to
* grab all incoming requests to some "incoming" list
* process this list
- move aside requests that are blocked by resync
- prepare activity log transactions,
- commit transactions and submit corresponding requests
- if there are remaining requests that only wait for
activity log extents to become free, stop the fast path
(mark activity log as "starving")
- iterate until no more requests are waiting for the activity log,
but all potentially remaining requests are only blocked by resync
* only then grab new incoming requests
That way, very busy IO on currently "hot" activity log extents cannot
starve scattered IO to "cold" extents. And blocked-by-resync requests
are processed once resync traffic on the affected region has ceased,
without blocking anything else.
The only blocking mode left is when we cannot start requests to "cold"
extents because all currently "hot" extents are actually used.
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Background resynchronisation does some "side-stepping", or throttles
itself, if it detects application IO activity, and the current resync
rate estimate is above the configured "cmin-rate".
What was not detected: if there is no application IO,
because it blocks on activity log transactions.
Introduce a new atomic_t ap_actlog_cnt, tracking such blocked requests,
and count non-zero as application IO activity.
This counter is exposed at proc_details level 2 and above.
Also make sure to release the currently locked resync extent
if we side-step due to such voluntary throttling.
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
A request that is to be shipped to the peer goes through a few stages:
- queued
- sent, waiting for ack
- ack received, waiting for "barrier ack", which is re-order epoch being
closed on the peer by acknowledging a "cache flush" equivalent
on the lower level device.
In the later two stages, depending on protocol, we may have already
completed this request to the upper layers, so it won't be found anymore
on device->pending_master_completion[] lists.
Track the oldest request yet to be sent (req_next), the oldest not yet
acknowledged (req_ack_pending) and the oldest "still waiting for
something from the peer" (req_not_net_done), doing short list walks on
the transfer log to find the next pending one whenever such a request
makes progress.
Now we have a fast way to look up the oldest requests,
don't do a transfer log walk every time.
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Adding requests to per-device fifo lists as soon as possible after
allocating them leaves a simple list_first_entry_or_null() to find the
oldest request, regardless what it is still waiting for.
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Record (in jiffies) how much time a request spends in which stages.
Followup commits will use and present this additional timing information
so we can better locate and tackle the root causes of latency spikes,
or present the backlog for asynchronous replication.
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
If we already "pulled ahead", we can short-circuit,
and avoid logging the same messages over and over again.
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
If "dirty" blocks are written to during resync,
that brings them in-sync.
By explicitly requesting write-acks during resync even in protocol != C,
we now can actually respect this.
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
The last user was al_write_transaction, if called with "delegate",
and the last user to call it with "delegate = true" was the receiver
thread, which has no need to delegate, but can call it himself.
Finally drop the delegate parameter, drop the extra
w_al_write_transaction callback, and drop drbd_queue_work_front.
Do not (yet) change dequeue_work_item to dequeue_work_batch, though.
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Reduce the number of calls to first_peer_device(). Instead, call
first_peer_device() just once to assign a local variable peer_device.
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Instead of dropping and re-aquiring the spinlock around the submit,
just remember that we want to submit, and do that only once we have
dropped the spinlock for good.
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Just because it is the oldest not yet completed request
does not make it the oldest request waiting for disk.
Or waiting for the peer.
And we completely missed already completed requests
that would still hold references to activity log extents,
waiting only for the barrier ack.
Find two oldest not yet completely processed requests,
one that is still waiting for local completion,
and one that is still waiting for some response from the peer.
These may or may not be the same request object.
Then separately apply the network and disk timeouts, respectively.
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
When batching more updates to the activity log into single transactions,
we lost the ability for new requests to force themselves into the active
set: all preparation steps became non-blocking, and if all currently
hot extents keep busy, they could starve out new incoming requests
to cold extents for quite a while.
This can only happen if your IO backend accepts more IO operations per
average DRBD replication round trip time than you have al-extents
configured.
If we have incoming requests to cold extents,
at least do one blocking update per transaction.
In an artificial worst-case workload on SSD with an asynchronous 600 ms
replication link, with al-extents = 7 (the minimum we allow), and
concurrent full resynch, without this patch, some write requests have
been observed to be starved for 40 seconds.
With this patch, application observed a worst case latency of twice the
replication round trip time.
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Note that I do NOT call __drbd_chk_io_error for failed REQ_DISCARD.
That may be wrong, though, or needs to differ between EOPNOTSUPP and
other errors...
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
drbd_device_work is a work item that has a reference to a device,
while drbd_work is a more generic work item that does not carry
a reference to a device.
All callbacks get a pointer to a drbd_work instance, those callbacks
that expect a drbd_device_work use the container_of macro to get it.
Signed-off-by: Andreas Gruenbacher <agruen@linbit.com>
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
The implicit dependency on a variable inside the macro is problematic.
Signed-off-by: Andreas Gruenbacher <agruen@linbit.com>
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
DRBD was using dev_err() and similar all over the code; instead of having to
write dev_err(disk_to_dev(device->vdisk), ...) to convert a drbd_device into a
kernel device, a DEV macro was used which implicitly references the device
variable. This is terrible; introduce separate drbd_err() and similar macros
with an explicit device parameter instead.
Signed-off-by: Andreas Gruenbacher <agruen@linbit.com>
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
In a setup where a device (aka volume) can replicate to multiple peers and one
connection can be shared between multiple devices, we need separate objects to
represent devices on peer nodes and network connections.
As a first step to introduce multiple connections per device, give each
drbd_device object a single drbd_peer_device object which connects it to a
drbd_connection object.
Signed-off-by: Andreas Gruenbacher <agruen@linbit.com>
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
sed -i -e 's:all_tconn:connections:g' -e 's:tconn:connection:g'
Signed-off-by: Andreas Gruenbacher <agruen@linbit.com>
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
sed -i -e 's:\<drbd_conf\>:drbd_device:g'
Signed-off-by: Andreas Gruenbacher <agruen@linbit.com>
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Mark functions drbd_request_prepare() and find_oldest_request() as
static in drbd/drbd_req.c because they are not used outside this file.
This eliminates the following warnings in drbd/drbd_req.c:
drivers/block/drbd/drbd_req.c:1037:1: warning: no previous prototype for ‘drbd_request_prepare’ [-Wmissing-prototypes]
drivers/block/drbd/drbd_req.c:1323:22: warning: no previous prototype for ‘find_oldest_request’ [-Wmissing-prototypes]
Signed-off-by: Rashika Kheria <rashika.kheria@gmail.com>
Reviewed-by: Josh Triplett <josh@joshtriplett.org>
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
For a long time, the receiving side has spread "too large" incoming
requests over multiple bios. No need to shrink our max_bio_size
(max_hw_sectors) if the peer is reconfigured to use a different storage.
The problem manifests itself if we are not the top of the device stack
(DRBD is used a LVM PV).
A hardware reconfiguration on the peer may cause the supported
max_bio_size to shrink, and the connection handshake would now
unnecessarily shrink the max_bio_size on the active node.
There is no way to notify upper layers that they have to "re-stack"
their limits. So they won't notice at all, and may keep submitting bios
that are suddenly considered "too large for device".
We already check for compatibility and ignore changes on the peer,
the code only was masked out unless we have a fully established connection.
We just need to allow it a bit earlier during the handshake.
Also consider max_hw_sectors in our merge bvec function, just in case.
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The sanity check when receiving P_BARRIER_ACK does expect all write
requests with a given req->epoch to have been either all replicated,
or all not replicated.
Because req->epoch was assigned before calling maybe_pull_ahead(),
this expectation was not met, leading to an off-by-one in the sanity
check, and further to a "Protocol Error".
Fix: move the call to maybe_pull_ahead() a few lines up,
and assign req->epoch only after that.
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We completed empty flushes (blkdev_issue_flush()) with IO error
if we lost the local disk, even if we still have an established
replication link to a healthy remote disk.
Fix this to only report errors to upper layers,
if neither local nor remote data is reachable.
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
There may have been more incoming requests while we where preparing
the current transaction. Try to consolidate more updates into this
transaction until we make no more progres.
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The IO accounting of the drbd "queue depth" was misleading.
We only started IO accounting once we already wrote the activity log.
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Depending on current IO depth, try to consolidate as many updates
as possible into one activity log transaction.
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This is in preparation to be able to defer requests that need to wait
for an activity log transaction to a submitter workqueue.
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
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 <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
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 <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
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 <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
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 <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
* 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 <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
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 <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
To avoid confusion with REQ_DISCARD aka TRIM, rename our
"discard concurrent write acks" from P_DISCARD_WRITE to P_SUPERSEDED.
At the same time, rename the drbd request event DISCARD_WRITE
to CONFLICT_RESOLVED. It already triggers both successful completion
or restart of the request, depending on our RQ_POSTPONED flag.
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Don't drop a request from the transfer log just because it was NEG_ACKED.
We need it around to be able to verify P_BARRIER_ACKs against the
transver log.
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Almost all code paths calling start_new_tl_epoch() guarded it with
if (... current_tle_writes > 0 ... ).
Just move that inside start_new_tl_epoch().
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Requests of an acked epoch are stored on the barrier_acked_requests list. In
case the private bio of such a request completes while IO on the drbd device
is suspended [req_mod(completed_ok)] then the request stays there.
When thawing IO because the fence_peer handler returned, then we use
tl_clear() to apply the connection_lost_while_pending event to all requests
on the transfer-log and the barrier_acked_requests list.
Up to now the connection_lost_while_pending event was not applied
on requests on the barrier_acked_requests list. Fixed that.
I.e. now the connection_lost_while_pending and resend events are
applied to requests on the barrier_acked_requests list. For that
it is necessary that the resend event finishes (local only)
READS correctly.
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
DRBD has a concept of request epochs or reorder-domains,
which are separated on the wire by P_BARRIER packets.
Older DRBD is not able to handle zero-sized requests at all,
so we need to map empty flushes to these drbd barriers.
These are the equivalent of empty flushes, and
by default trigger flushes on the receiving side anyways
(unless not supported or explicitly disabled),
so there is no need to handle this differently in newer drbd either.
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Gu Zheng