Do not accept requests until initial sync is finished.
`Apply` is deliberately left out -- we don't want to miss anything new.
Signed-off-by: Dmitrii Stepanov <d.stepanov@yadro.com>
Dynamically growing an unbounded buffers can cause a large
amount of memory to be pinned and never be freed.
Signed-off-by: Dmitrii Stepanov <d.stepanov@yadro.com>
Now it's possible to run evacuate shard in async.
Also only one evacuate process can be in progress.
Signed-off-by: Dmitrii Stepanov <d.stepanov@yadro.com>
If a connection has not been established earlier, it stores `nil` in LRU
cache. Cache eviction tries to close every connection (even a `nil` one) and
panics but not crash the app because we are using pools.
That ugly bug also leads to a deadlock where `Unlock` is not called via
`defer` func (and that is the way I found it).
Signed-off-by: Pavel Karpy <p.karpy@yadro.com>
Doctor RPC performs complex operations on the storage engine.
Currently only duplicate removal is supported.
Signed-off-by: Evgenii Stratonikov <e.stratonikov@yadro.com>
In case of session token (ST) with object IDs search should
return only objects allowed in static session
Signed-off-by: Dmitrii Stepanov <d.stepanov@yadro.com>
Initially it was there to check whether an update is being initiated by
a proper node. It is now obsolete for 2 reasons:
1. Background synchronization fetches all operations from a single node.
2. There are a lot more problems with trust in the tree service, it is
only used in controlled environments.
Signed-off-by: Evgenii Stratonikov <e.stratonikov@yadro.com>
This reverts commit 2567f8020e. It assumes
that assembling logic could break some failover scenarios if request
forwarding is done. However, it also breaks requesting big objects via a
non-container node with TTL=2. Failover has been rechecked without that
commit and no problems were found. Any (if found) other bugs related to
the forwarding and object assembling must be solved more carefully.
Signed-off-by: Pavel Karpy <p.karpy@yadro.com>
Allow replication of any (expired too) locked object. Information about
object locking is considered to be presented on the _container nodes_.
Signed-off-by: Pavel Karpy <p.karpy@yadro.com>
Previously a token could've expired in the middle of an object.PUT
stream, leading to upload being interrupted. This is bad, because user
doesn't always now what is the right values for the session token
lifetime. More than that, setting it to a very high value will
eventually blow up the session token database.
In this commit we read the session token once and reuse it for the whole
stream duration.
Signed-off-by: Evgenii Stratonikov <e.stratonikov@yadro.com>
Currently, under high load clients are blocked on channel send
and the number of goroutines can increase indefinitely.
In this commit we drop replication messages if send/recv queue is full
and rely on a background synchronization.
Signed-off-by: Evgenii Stratonikov <e.stratonikov@yadro.com>
Because synchronization _most likely_ will have apply already existing
operations, it is much faster to check their presence in a read
transaction. However, always doing this will degrade the perfomance
for normal `Apply`. And, let's be honest, it is already not good.
Thus we add a separate parameter which specifies whether this logic is
enabled.
Signed-off-by: Evgenii Stratonikov <e.stratonikov@yadro.com>
`GETRANGEHASH` request spawns `GETRANGE` requests if an object could not be
found locally. If the original request contains session, it can be static
and, therefore, fetching session key can not be performed successfully.
As the best effort a node could request object's range with its own key.
Signed-off-by: Pavel Karpy <p.karpy@yadro.com>
We use cache to avoid policing the same object multiple times in a short
time span (< 30 seconds). If we have 200_000 objects in a blobstor, it is a bit useless
-- if it takes 1 second to process an object and we have `replicator.pool_size: 20`
in config, the next iteration will happen in 10_000 second which is much
larger than 30 second. However we still consume a lot of memory, so it
makes sense to use saner default.
Signed-off-by: Evgenii Stratonikov <e.stratonikov@yadro.com>
Stop child objects collection if the last returned object (the most "left"
object in the collected chain) starts exactly from the `GETRANGE`'s `from`
value.
Signed-off-by: Pavel Karpy <carpawell@nspcc.ru>
That could happen if a node forwards request to a node that closed the
connection during the original object stream.
Signed-off-by: Pavel Karpy <carpawell@nspcc.ru>
```
2022/11/15 08:40:56 worker exits from a panic: runtime error: index out of range [0] with length 0
2022/11/15 08:40:56 worker exits from panic: goroutine 1188 [running]:
github.com/panjf2000/ants/v2.(*goWorker).run.func1.1()
github.com/panjf2000/ants/v2@v2.4.0/worker.go:58 +0x10c
panic({0x1042b60, 0xc0015ae018})
runtime/panic.go:1038 +0x215
github.com/nspcc-dev/neofs-node/pkg/services/policer.(*Policer).shardPolicyWorker.func1()
github.com/nspcc-dev/neofs-node/pkg/services/policer/process.go:65 +0x366
github.com/panjf2000/ants/v2.(*goWorker).run.func1()
github.com/panjf2000/ants/v2@v2.4.0/worker.go:68 +0x97
created by github.com/panjf2000/ants/v2.(*goWorker).run
github.com/panjf2000/ants/v2@v2.4.0/worker.go:48 +0x68
```
Signed-off-by: Evgenii Stratonikov <evgeniy@morphbits.ru>
Includes extending listing methods in the Storage Engine with object types.
It allows tuning replication/policer algorithms: container nodes do
not remove `LOCK` objects as redundant and try to fulfill `LOCK` placement
on the ohter container nodes.
Signed-off-by: Pavel Karpy <carpawell@nspcc.ru>
A container node is expected to have full "get" access to assemble the
object.
A non-container node is expected to forward any request to a container node.
Any token is expected to be issued for an original request sender not for a
node so any new request is invalid by design with that token.
Signed-off-by: Pavel Karpy <carpawell@nspcc.ru>
Do not lose meta information of the original requests: cache session and
bearer tokens of the original request b/w a new generated ones. Middle
request wrappers should not contain any meta information, since it is
useless (e.g. ACL service checks only the original tokens).
Signed-off-by: Pavel Karpy <carpawell@nspcc.ru>
After presenting request statuses on the API level, all the errors are
unwrapped before sending to the caller side. It led to a losing invalid
request's context.
Signed-off-by: Pavel Karpy <carpawell@nspcc.ru>
Do not use node's local storage if it is clear that an object will be
removed anyway as a redundant. It requires moving the changing local storage
logic from the validation step to the local target implementation.
It allows performing any relations checks (e.g. object locking) only if a
node is considered as a valid container member and is expected to store
(stored previously) all the helper objects (e.g. `LOCK`, `TOMBSTONE`, etc).
Signed-off-by: Pavel Karpy <carpawell@nspcc.ru>
In previous implementation turning to maintenance mode using NeoFS CLI
required NeoFS API endpoint. This was not convenient from the user
perspective. It's worth to move networks settings' check to the server
side.
Add `force_maintenance` field to `SetNetmapStatusRequest.Body` message
of Control API. Add `force` flag to `neofs-cli control set-status`
command which sets corresponding field in the requests body if status is
`maintenance`. Force flag is ignored for any other status.
Signed-off-by: Leonard Lyubich <ctulhurider@gmail.com>
Add `SynchronizeAllTrees` method of the Tree service. It allows fetching
tree IDs and sync all of them. Share common logic b/w the new method and
the `SynchronizeTree`.
Signed-off-by: Pavel Karpy <carpawell@nspcc.ru>
Add the node position in a container and the container size to the CID
descriptor that is passed to the `TreeApply`. Previously, `checkValid` does
not allow any log operarations.
Signed-off-by: Pavel Karpy <carpawell@nspcc.ru>
Storage node should not provide NeoFS Object API service when it is
under maintenance.
Declare `Common` service that unifies behavior of all object operations.
The implementation pre-checks if node is under maintenance and returns
`apistatus.NodeUnderMaintenance` if so. Use `Common` service as a first
logical processor in object service pipeline.
Signed-off-by: Leonard Lyubich <ctulhurider@gmail.com>
In some scenarios original session can be unrelated to the objects which
are read internally by the node. For example, node requests child
objects when removing the parent one.
Tune internal NeoFS API client used by node's Object API server to
ignore unrelated sessions in `GetObject` / `HeadObject` / `PayloadRange`
ops.
Signed-off-by: Leonard Lyubich <ctulhurider@gmail.com>
Make it store its internal `zap.Logger`'s level. Also, make all the
components to accept internal `logger.Logger` instead of `zap.Logger`; it
will simplify future refactor.
Signed-off-by: Pavel Karpy <carpawell@nspcc.ru>
Current spec allows denying GET_RANGE requests from other storage nodes.
However, GET should always be allowed and it is enough to perform
GET_RANGE locally
Signed-off-by: Evgenii Stratonikov <evgeniy@morphbits.ru>
In previous implementation `ObjectService.Get` RPC handler failed with
`parent address in child object differs` while assembling the "big"
object. This was caused by the child check which required parent
reference to be set in all child objects. The check was impracticable
because not all elements of the split-chain have a link to the parent.
Make `execCtx.isChild` to return `true` if parameterized object has no
parent header in its own header.
Signed-off-by: Leonard Lyubich <ctulhurider@gmail.com>
Node response with `NODE_UNDER_MAINTENANCE` status signals that the node
was switched to maintenance mode. There is a delay between the actual
switch and the reflection in the network map of up to one epoch. To
speed up the reaction to the maintenance, it is required to recognize
such node responses in the Policer.
Make `Policer.processNodes` to exclude elements with shortage decreasing
on `NODE_UNDER_MAINTENANCE` status response.
Signed-off-by: Leonard Lyubich <ctulhurider@gmail.com>
Nodes under maintenance SHOULD not respond to object requests. Based on
this, storage node's Policer SHOULD consider such nodes as problem ones.
However, to prevent spam with the new replicas, on the contrary, Policer
should consider them normal.
Make `Policer.processNodes` to exclude elements if `IsMaintenance()`
with shortage decreasing.
Signed-off-by: Leonard Lyubich <ctulhurider@gmail.com>
Make `replicator.TaskResult` to accept `netmap.NodeInfo` type instead of
uint64 in order to clarify the meaning and prevent passing the random
numbers.
Signed-off-by: Leonard Lyubich <ctulhurider@gmail.com>
It doesn't make sense to check object relation in session check of
`ObjectService.Put` RPC which has been spawned by `ObjectService.Delete`
with session. Session issuer can't predict identifier of the tombstone
object to be created.
Signed-off-by: Leonard Lyubich <ctulhurider@gmail.com>
In previous implementation of `neofs-node` app object session was not
checked for substitution of the object related to it. Also, for access
checks, the session object was substituted instead of the one from the
request. This, on the one hand, made it possible to inherit the session
from the parent object for authorization for certain actions. On the
other hand, it covered the mentioned object substitution, which is a
critical vulnerability.
Next changes are applied to processing of all Object service requests:
- check if object session relates to the requested object
- use requested object in access checks.
Disclosed problem of object context inheritance will be solved within
Signed-off-by: Leonard Lyubich <ctulhurider@gmail.com>
In previous implementation node blocked any operation of local object
storage in maintenance mode. There is a need to perform some storage
operations like data evacuation or restoration.
Do not call block storage engine in maintenance mode. Make all Object
service operations to return `apistatus.NodeUnderMaintenance` error from
each local op.
Signed-off-by: Leonard Lyubich <ctulhurider@gmail.com>
Replace `ProcessCurrentNetMap` method of `NodeState` interface with
`ReadCurrentNetMap` one with two changes:
* Replace network map type from NeoFS SDK package with the
protocol-generated message. This replaces all the business logic to
the application layer.
* Support error return. This allows to cover problem node states.
Return an error from `NodeState.ReadCurrentNetMap` method implemeted
through `atomic.Value` if `Store` method has not been called yet.
Signed-off-by: Leonard Lyubich <leonard@nspcc.ru>
Use a simple loop at the callsite. This way we remove as much as we can.
Also, `Delete` metrics is more meaningful now.
Signed-off-by: Evgenii Stratonikov <evgeniy@morphbits.ru>
Do not call `CalculateAction` for the eACL checks since it requires object
headers that are meaningless in the tree context.
Signed-off-by: Pavel Karpy <carpawell@nspcc.ru>
1. Do not require a request to be signed by the container owner if a
bearer token is missing
2. Do not check the system role since public requests are not expected to
be signed by IR or a container node (unlike the object requests)
Signed-off-by: Pavel Karpy <carpawell@nspcc.ru>
Previously, the depth was restricted because with BFS the amount of
nodes we have in memory blows up exponentially. With DFS is is linear,
so we can process trees of arbitrary depth.
Signed-off-by: Evgenii Stratonikov <evgeniy@morphbits.ru>
Also, try to fetch object header info from the local storage to find as much
object info as possible for the requests which do not assume returning
object header as a response.
Signed-off-by: Pavel Karpy <carpawell@nspcc.ru>
`Degraded` mode can be set by the administrator if needed.
Modifying operations in this mode can lead node into an inconsistent state
because metabase checks such as lock checking are not performed.
Signed-off-by: Evgenii Stratonikov <evgeniy@nspcc.ru>
Do not check that a node indeed belongs to the container, because the
synchronization will fail in this case anyway.
Signed-off-by: Evgenii Stratonikov <evgeniy@nspcc.ru>
Do not return backend type from the service for now, because memory
backend is expected to vanish.
Signed-off-by: Evgenii Stratonikov <evgeniy@nspcc.ru>
In case node is down or failing for some reason, we can expect `Dial` to
fail. In case we actively try to replicate and `Dial` always takes 2
seconds, replication-related channels quickly become full. That affects
latency of all other write operations.
Signed-off-by: Evgenii Stratonikov <evgeniy@nspcc.ru>
Before this commit the replication channel was quickly filled under
heavy load. This lead to the continuously increasing latency for all
write operations. Now it looks better.
Signed-off-by: Evgenii Stratonikov <evgeniy@nspcc.ru>
