- 14 Jun, 2016 25 commits
-
-
Lars Ellenberg authored
Scenario, starting with normal operation Connected Primary/Secondary UpToDate/UpToDate NetworkFailure Primary/Unknown UpToDate/DUnknown (frozen) ... more failures happen, secondary loses it's disk, but eventually is able to re-establish the replication link ... Connected Primary/Secondary UpToDate/Diskless (resumed; needs to bump uuid!) We used to just resume/resent suspended requests, without bumping the UUID. Which will lead to problems later, when we want to re-attach the disk on the peer, without first disconnecting, or if we experience additional failures, because we now have diverging data without being able to recognize it. Make sure we also bump the current data generation UUID, if we notice "peer disk unknown" -> "peer disk known bad". 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>
-
Lars Ellenberg authored
We already serialize connection state changes, and other, non-connection state changes (role changes) while we are establishing a connection. But if we have an established connection, then trigger a resync handshake (by primary --force or similar), until now we just had to be "lucky". Consider this sequence (e.g. deployment scenario): create-md; up; -> Connected Secondary/Secondary Inconsistent/Inconsistent then do a racy primary --force on both peers. block drbd0: drbd_sync_handshake: block drbd0: self 0000000000000004:0000000000000000:0000000000000000:0000000000000000 bits:25590 flags:0 block drbd0: peer 0000000000000004:0000000000000000:0000000000000000:0000000000000000 bits:25590 flags:0 block drbd0: peer( Unknown -> Secondary ) conn( WFReportParams -> Connected ) pdsk( DUnknown -> Inconsistent ) block drbd0: peer( Secondary -> Primary ) pdsk( Inconsistent -> UpToDate ) *** HERE things go wrong. *** block drbd0: role( Secondary -> Primary ) block drbd0: drbd_sync_handshake: block drbd0: self 0000000000000005:0000000000000000:0000000000000000:0000000000000000 bits:25590 flags:0 block drbd0: peer C90D2FC716D232AB:0000000000000004:0000000000000000:0000000000000000 bits:25590 flags:0 block drbd0: Becoming sync target due to disk states. block drbd0: Writing the whole bitmap, full sync required after drbd_sync_handshake. block drbd0: Remote failed to finish a request within 6007ms > ko-count (2) * timeout (30 * 0.1s) drbd s0: peer( Primary -> Unknown ) conn( Connected -> Timeout ) pdsk( UpToDate -> DUnknown ) The problem here is that the local promotion happens before the sync handshake triggered by the remote promotion was completed. Some assumptions elsewhere become wrong, and when the expected resync handshake is then received and processed, we get stuck in a deadlock, which can only be recovered by reboot :-( Fix: if we know the peer has good data, and our own disk is present, but NOT good, and there is no resync going on yet, we expect a sync handshake to happen "soon". So reject a racy promotion with SS_IN_TRANSIENT_STATE. Result: ... as above ... block drbd0: peer( Secondary -> Primary ) pdsk( Inconsistent -> UpToDate ) *** local promotion being postponed until ... *** block drbd0: drbd_sync_handshake: block drbd0: self 0000000000000004:0000000000000000:0000000000000000:0000000000000000 bits:25590 flags:0 block drbd0: peer 77868BDA836E12A5:0000000000000004:0000000000000000:0000000000000000 bits:25590 flags:0 ... block drbd0: conn( WFBitMapT -> WFSyncUUID ) block drbd0: updated sync uuid 85D06D0E8887AD44:0000000000000000:0000000000000000:0000000000000000 block drbd0: conn( WFSyncUUID -> SyncTarget ) *** ... after the resync handshake *** block drbd0: role( Secondary -> Primary ) Signed-off-by:
-
Lars Ellenberg authored
If in a two-primary scenario, we lost our peer, freeze IO, and are still frozen (no UUID rotation) when the peer comes back as Secondary after a hard crash, we will see identical UUIDs. The "rule_nr = 40" chose to use the "CRASHED_PRIMARY" bit as arbitration, but that would cause the still running (but frozen) Primary to become SyncTarget (which it typically refuses), and the handshake is declined. Fix: check current roles. If we have *one* current primary, the Primary wins. (rule_nr = 41) Since that is a protocol change, use the newly introduced DRBD_FF_WSAME to determine if rule_nr = 41 can be applied. Signed-off-by:
-
Lars Ellenberg authored
We will support WRITE_SAME, if * all peers support WRITE_SAME (both in kernel and DRBD version), * all peer devices support WRITE_SAME * logical_block_size is identical on all peers. We may at some point introduce a fallback on the receiving side for devices/kernels that do not support WRITE_SAME, by open-coding a submit loop. But not yet. Signed-off-by:
-
Lars Ellenberg authored
Signed-off-by:
-
Lars Ellenberg authored
Even if discard_zeroes_data != 0, if discard_zeroes_if_aligned is set, we assume we can reliably zero-out/discard using the drbd_issue_peer_discard() helper. Signed-off-by:
-
Lars Ellenberg authored
When re-attaching the local backend device to a C_STANDALONE D_DISKLESS R_PRIMARY with OND_SUSPEND_IO, we may only resume IO if we recognize the backend that is being attached as D_UP_TO_DATE. Signed-off-by:
-
Lars Ellenberg authored
If DRBD lost all path to good data, and the on-no-data-accessible policy is OND_SUSPEND_IO, all pending and new IO requests are suspended (will block). If that setting is OND_IO_ERROR, IO will still be completed. READ to "clean" areas (e.g. on an D_INCONSISTENT device, and bitmap indicates a block is already in sync) will succeed. READ to "unclean" areas (bitmap indicates block is out-of-sync), will return EIO. If we are already D_DISKLESS (or D_FAILED), we also return EIO. Unfortunately, on a former R_PRIMARY C_SYNC_TARGET D_INCONSISTENT, after replication link loss, new WRITE requests still went through OK. The would also set the "out-of-sync" bit on their way, so READ after WRITE would still return EIO. Also, the data generation UUIDs had not been bumped, we would cause data divergence, without being able to detect it on the next sync handshake, given the right sequence of events in a multiple error scenario and "improper" order of recovery actions. The right thing to do is to return EIO for all new writes, unless we have access to good, current, D_UP_TO_DATE data. The "established best practices" way to avoid these situations in the first place is to set OND_SUSPEND_IO, or even do a hard-reset from the pri-on-incon-degr policy helper hook. Signed-off-by:
-
Lars Ellenberg authored
Possibly sequence of events: SyncTarget is made Primary, then loses replication link (only path to good data on SyncSource). Behavior is then controlled by the on-no-data-accessible policy, which defaults to OND_IO_ERROR (may be set to OND_SUSPEND_IO). If OND_IO_ERROR is in fact the current policy, we clear the susp_fen (IO suspended due to fencing policy) flag, do NOT set the susp_nod (IO suspended due to no data) flag. But we forgot to call the IO error completion for all pending, suspended, requests. While at it, also add a race check for a theoretically possible race with a new handshake (network hickup), we may be able to re-send requests, and can avoid passing IO errors up the stack. Signed-off-by:
-
Lars Ellenberg authored
When resync is finished, we already call the "after-resync-target" handler (on the former sync target, obviously), once per volume. Paired with the before-resync-target handler, you can create snapshots, before the resync causes the volumes to become inconsistent, and discard those snapshots again, once they are no longer needed. It was also overloaded to be paired with the "fence-peer" handler, to "unfence" once the volumes are up-to-date and known good. This has some disadvantages, though: we call "fence-peer" for the whole connection (once for the group of volumes), but would call unfence as side-effect of after-resync-target once for each volume. Also, we fence on a (current, or about to become) Primary, which will later become the sync-source. Calling unfence only as a side effect of the after-resync-target handler opens a race window, between a new fence on the Primary (SyncTarget) and the unfence on the SyncTarget, which is difficult to close without some kind of "cluster wide lock" in those handlers. We would not need those handlers if we could still communicate. Which makes trying to aquire a cluster wide lock from those handlers seem like a very bad idea. This introduces the "unfence-peer" handler, which will be called per connection (once for the group of volumes), just like the fence handler, only once all volumes are back in sync, and on the SyncSource. Which is expected to be the node that previously called "fence", the node that is currently allowed to be Primary, and thus the only node that could trigger a new "fence" that could race with this unfence. Which makes us not need any cluster wide synchronization here, serializing two scripts running on the same node is trivial. Signed-off-by:
-
Lars Ellenberg authored
If the replication link breaks exactly during "resync finished" detection, finishing too early on the sync source could again lead to UUIDs rotated too fast, and potentially a spurious full resync on next handshake. Always wait for explicit resync finished state change notification from the sync target. Signed-off-by:
-
Lars Ellenberg authored
Make sure we have at least 67 (> AL_UPDATES_PER_TRANSACTION) al-extents available, and allow up to half of that to be discarded in one bio. Signed-off-by:
-
Lars Ellenberg authored
For consistency, also zero-out partial unaligned chunks of discard requests on the local backend. Signed-off-by:
-
Lars Ellenberg authored
Now that we have the discard_zeroes_if_aligned setting, we should also check it when setting up our queue parameters on the primary, not only on the receiving side. We announce discard support, UNLESS * we are connected to a peer that does not support TRIM on the DRBD protocol level. Otherwise, it would either discard, or do a fallback to zero-out, depending on its backend and configuration. * our local backend does not support discards, or (discard_zeroes_data=0 AND discard_zeroes_if_aligned=no). Signed-off-by:
-
Lars Ellenberg authored
We can avoid spurious data divergence caused by partially-ignored discards on certain backends with discard_zeroes_data=0, if we translate partial unaligned discard requests into explicit zero-out. The relevant use case is LVM/DM thin. If on different nodes, DRBD is backed by devices with differing discard characteristics, discards may lead to data divergence (old data or garbage left over on one backend, zeroes due to unmapped areas on the other backend). Online verify would now potentially report tons of spurious differences. While probably harmless for most use cases (fstrim on a file system), DRBD cannot have that, it would violate our promise to upper layers that our data instances on the nodes are identical. To be correct and play safe (make sure data is identical on both copies), we would have to disable discard support, if our local backend (on a Primary) does not support "discard_zeroes_data=true". We'd also have to translate discards to explicit zero-out on the receiving (typically: Secondary) side, unless the receiving side supports "discard_zeroes_data=true". Which both would allocate those blocks, instead of unmapping them, in contrast with expectations. LVM/DM thin does set discard_zeroes_data=0, because it silently ignores discards to partial chunks. We can work around this by checking the alignment first. For unaligned (wrt. alignment and granularity) or too small discards, we zero-out the initial (and/or) trailing unaligned partial chunks, but discard all the aligned full chunks. At least for LVM/DM thin, the result is effectively "discard_zeroes_data=1". Arguably it should behave this way internally, by default, and we'll try to make that happen. But our workaround is still valid for already deployed setups, and for other devices that may behave this way. Setting discard-zeroes-if-aligned=yes will allow DRBD to use discards, and to announce discard_zeroes_data=true, even on backends that announce discard_zeroes_data=false. Setting discard-zeroes-if-aligned=no will cause DRBD to always fall-back to zero-out on the receiving side, and to not even announce discard capabilities on the Primary, if the respective backend announces discard_zeroes_data=false. We used to ignore the discard_zeroes_data setting completely. To not break established and expected behaviour, and suddenly cause fstrim on thin-provisioned LVs to run out-of-space, instead of freeing up space, the default value is "yes". Signed-off-by:
-
Lars Ellenberg authored
To maintain write-order fidelity accros all volumes in a DRBD resource, the receiver of a P_BARRIER needs to issue flushes to all volumes. We used to do this by calling blkdev_issue_flush(), synchronously, one volume at a time. We now submit all flushes to all volumes in parallel, then wait for all completions, to reduce worst-case latencies on multi-volume resources. Signed-off-by:
-
Lars Ellenberg authored
The command line parameter the kernel module uses to communicate the device minor to userland helper is flawed in a way that the device indentifier "minor-%d" is being truncated to minors with a maximum of 5 digits. But DRBD 8.4 allows 2^20 == 1048576 minors, thus a minimum of 7 digits must be supported. Reported by Veit Wahlich on drbd-dev. Signed-off-by:
-
Lars Ellenberg authored
Regression introduced with 8.4.5 drbd: application writes may set-in-sync in protocol != C Overwriting the same block (LBA) while a former version is still "in-flight" to the peer (to be exact: we did not receive the P_BARRIER_ACK for its epoch yet) would wait for the full epoch of that former version to be acknowledged by the peer. In synchronous and quasi-synchronous protocols C and B, this may double the latency on overwrites. With protocol A, which is supposed to be asynchronous and only wait for local completion, it is even worse: it would make overwrites quasi-synchronous, they would be hit by the full RTT, which protocol A was specifically meant to avoid, and possibly the additional time it takes to drain the buffers first. Particularly bad for databases, or anything else that does frequent updates to the same blocks (various file system meta data). No impact if >= rtt passes between updates to the same block. Signed-off-by:
-
Lars Ellenberg authored
Signed-off-by:
-
Philipp Reisner authored
If thinly provisioned volumes are used, during a resync the sync source tries to find out if a block is deallocated. If it is deallocated, then the resync target uses block_dev_issue_zeroout() on the range in question. Signed-off-by:
-
Philipp Reisner authored
As long as the value is 0 the feature is disabled. With setting it to a positive value, DRBD limits and aligns its resync requests to the rs-discard-granularity setting. If the sync source detects all zeros in such a block, the resync target discards the range on disk. Signed-off-by:
-
Philipp Reisner authored
If during resync we read only zeroes for a range of sectors assume that these secotors can be discarded on the sync target node. Signed-off-by:
-
Philipp Reisner authored
Signed-off-by:
-
Lars Ellenberg authored
When leaving resync states because of disconnect, do the bitmap write-out synchronously in the drbd_disconnected() path. When leaving resync states because we go back to AHEAD/BEHIND, or because resync actually finished, or some disk was lost during resync, trigger the write-out from after_state_ch(). The bitmap write-out for resync -> ahead/behind was missing completely before. Note that this is all only an optimization to avoid double-resyncs of already completed blocks in case this node crashes. Signed-off-by:
-
Lars Ellenberg authored
The intention was to only suspend IO if some normal bitmap operation is supposed to be locked out, not always. If the bulk operation is flaged as BM_LOCKED_CHANGE_ALLOWED, we do not need to suspend IO. Signed-off-by:
-
- 12 Jun, 2016 9 commits
-
-
Christoph Hellwig authored
We want to apply this to Fabrics drivers as well, so move it to common code. Reviewed-by:
Jay Freyensee <james.p.freyensee@intel.com> Reviewed-by:
Sagi Grimberg <sagi@grimberg.me> Tested-by:
Ming Lin <ming.l@ssi.samsung.com> Signed-off-by:
Christoph Hellwig <hch@lst.de> Reviewed-by:
Keith Busch <keith.busch@intel.com> Signed-off-by:
-
Christoph Hellwig authored
Centralize the check if a given NVMe command reads or writes data. Reviewed-by:
-
Christoph Hellwig authored
Some transport drivers may have a lower transfer size than the controller. So allow the transport to set it in the controller max_hw_sectors. Reviewed-by:
-
James Smart authored
Signed-off-by:
James Smart <james.smart@broadcom.com> Reviewed-by:
-
Christoph Hellwig authored
Reviewed-by:
-
Christoph Hellwig authored
Reviewed-by:
-
Armen Baloyan authored
Add get_log_page command structure and a corresponding entry in nvme_command union Signed-off-by:
Armen Baloyan <armenx.baloyan@intel.com> Reviewed-by:
-
Christoph Hellwig authored
These have been added in NVMe 1.2 and we'll need at least oaes for the NVMe target driver. Reviewed-by:
-
Bhaktipriya Shridhar authored
alloc_workqueue replaces deprecated create_workqueue(). Dedicated workqueues have been used since bcache_wq and moving_gc_wq are workqueues for writes and are being used on a memory reclaim path. WQ_MEM_RECLAIM has been set to ensure forward progress under memory pressure. Since there are only a fixed number of work items, explicit concurrency limit is unnecessary here. Signed-off-by:
Bhaktipriya Shridhar <bhaktipriya96@gmail.com> Acked-by:
Tejun Heo <tj@kernel.org> Signed-off-by:
-
- 09 Jun, 2016 1 commit
-
-
Christoph Hellwig authored
Instead of overloading the discard support with the REQ_SECURE flag. Use the opportunity to rename the queue flag as well, and remove the dead checks for this flag in the RAID 1 and RAID 10 drivers that don't claim support for secure erase. Signed-off-by:
-
- 08 Jun, 2016 3 commits
-
-
Mike Christie authored
Because we define WRITE/READ as REQ_OPs, we cannot do switch (rq_data_dir(request)) case READ .... case WRITE ... without getting warnings about handling other REQ_OPs. This just has mq_disk do a if/else like it does in other places. Signed-off-by:
Mike Christie <mchristi@redhat.com> Signed-off-by:
-
Sunad Bhandary authored
In case of the active namespace list scanning method, a namespace that is detached is not removed from the host if it was the last entry in the list. Fix this by adding a scan to validate namespaces greater than the value of prev. This also handles the case of removing namespaces whose value exceed the device's reported number of namespaces. Signed-off-by:
Sunad Bhandary S <sunad.s@samsung.com> Reviewed-by:
-
Minfei Huang authored
It's more elegant to use UINT_MAX to represent the max value of type unsigned int. So replace the actual value by using this define. Signed-off-by:
Minfei Huang <mnghuan@gmail.com> Signed-off-by:
-
- 07 Jun, 2016 2 commits
-
-
Ming Lin authored
So it can be used by fabrics driver also. Signed-off-by:
Ming Lin <ming.l@samsung.com> Reviewed-by:
Johannes Thumshirn <jthumshirn@suse.de> Reviewed-by:
Keith Busch <keith.bsuch@intel.com> Signed-off-by:
-
Ming Lin authored
nvme_cancel_io is a bit confusing (given the distinction of io/admin), so rename it to nvme_cancel_request. And update it a bit to pass in struct nvme_ctrl, so it can be used by Fabrics driver also. Signed-off-by:
-
