The motivation for this recatoring are based on two independent issues:
- @JMoVS found that the changes merged as part of #259 slowed his OS X
based installation down significantly.
Analysis of the zfs command logging introduced in #296 showed that
`zfs holds` took most of the execution time, and they pointed out
that not all of those `zfs holds` invocations were actually necessary.
I.e.: zrepl was inefficient about retrieving information from ZFS.
- @InsanePrawn found that failures on initial replication would lead
to step holds accumulating on the sending side, i.e. they would never
be cleaned up in the HintMostRecentCommonAncestor RPC handler.
That was because we only sent that RPC if there was a most recent
common ancestor detected during replication planning.
@InsanePrawn prototyped an implementation of a `zrepl zfs-abstractions release`
command to mitigate the situation.
As part of that development work and back-and-forth with @problame,
it became evident that the abstractions that #259 built on top of
zfs in package endpoint (step holds, replication cursor,
last-received-hold), were not well-represented for re-use in the
`zrepl zfs-abstractions release` subocommand prototype.
This commit refactors package endpoint to address both of these issues:
- endpoint abstractions now share an interface `Abstraction` that, among
other things, provides a uniform `Destroy()` method.
However, that method should not be destroyed directly but instead
the package-level `BatchDestroy` function should be used in order
to allow for a migration to zfs channel programs in the future.
- endpoint now has a query facitilty (`ListAbstractions`) which is
used to find on-disk
- step holds and bookmarks
- replication cursors (v1, v2)
- last-received-holds
By describing the query in a struct, we can centralized the retrieval
of information via the ZFS CLI and only have to be clever once.
We are "clever" in the following ways:
- When asking for hold-based abstractions, we only run `zfs holds` on
snapshot that have `userrefs` > 0
- To support this functionality, add field `UserRefs` to zfs.FilesystemVersion
and retrieve it anywhere we retrieve zfs.FilesystemVersion from ZFS.
- When asking only for bookmark-based abstractions, we only run
`zfs list -t bookmark`, not with snapshots.
- Currently unused (except for CLI) per-filesystem concurrent lookup
- Option to only include abstractions with CreateTXG in a specified range
- refactor `endpoint`'s various ZFS info retrieval methods to use
`ListAbstractions`
- rename the `zrepl holds list` command to `zrepl zfs-abstractions list`
- make `zrepl zfs-abstractions list` consume endpoint.ListAbstractions
- Add a `ListStale` method which, given a query template,
lists stale holds and bookmarks.
- it uses replication cursor has different modes
- the new `zrepl zfs-abstractions release-{all,stale}` commands can be used
to remove abstractions of package endpoint
- Adjust HintMostRecentCommonAncestor RPC for stale-holds cleanup:
- send it also if no most recent common ancestor exists between sender and receiver
- have the sender clean up its abstractions when it receives the RPC
with no most recent common ancestor, using `ListStale`
- Due to changed semantics, bump the protocol version.
- Adjust HintMostRecentCommonAncestor RPC for performance problems
encountered by @JMoVS
- by default, per (job,fs)-combination, only consider cleaning
step holds in the createtxg range
`[last replication cursor,conservatively-estimated-receive-side-version)`
- this behavior ensures resumability at cost proportional to the
time that replication was donw
- however, as explained in a comment, we might leak holds if
the zrepl daemon stops running
- that trade-off is acceptable because in the presumably rare
this might happen the user has two tools at their hand:
- Tool 1: run `zrepl zfs-abstractions release-stale`
- Tool 2: use env var `ZREPL_ENDPOINT_SENDER_HINT_MOST_RECENT_STEP_HOLD_CLEANUP_MODE`
to adjust the lower bound of the createtxg range (search for it in the code).
The env var can also be used to disable hold-cleanup on the
send-side entirely.
supersedes closes#293
supersedes closes#282fixes#280fixes#278
Additionaly, we fixed a couple of bugs:
- zfs: fix half-nil error reporting of dataset-does-not-exist for ZFSListChan and ZFSBookmark
- endpoint: Sender's `HintMostRecentCommonAncestor` handler would not
check whether access to the specified filesystem was allowed.
- **Resumable Send & Recv Support**
No knobs required, automatically used where supported.
- **Hold-Protected Send & Recv**
Automatic ZFS holds to ensure that we can always resume a replication step.
- **Encrypted Send & Recv Support** for OpenZFS native encryption.
Configurable at the job level, i.e., for all filesystems a job is responsible for.
- **Receive-side hold on last received dataset**
The counterpart to the replication cursor bookmark on the send-side.
Ensures that incremental replication will always be possible between a sender and receiver.
Design Doc
----------
`replication/design.md` doc describes how we use ZFS holds and bookmarks to ensure that a single replication step is always resumable.
The replication algorithm described in the design doc introduces the notion of job IDs (please read the details on this design doc).
We reuse the job names for job IDs and use `JobID` type to ensure that a job name can be embedded into hold tags, bookmark names, etc.
This might BREAK CONFIG on upgrade.
Protocol Version Bump
---------------------
This commit makes backwards-incompatible changes to the replication/pdu protobufs.
Thus, bump the version number used in the protocol handshake.
Replication Cursor Format Change
--------------------------------
The new replication cursor bookmark format is: `#zrepl_CURSOR_G_${this.GUID}_J_${jobid}`
Including the GUID enables transaction-safe moving-forward of the cursor.
Including the job id enables that multiple sending jobs can send the same filesystem without interfering.
The `zrepl migrate replication-cursor:v1-v2` subcommand can be used to safely destroy old-format cursors once zrepl has created new-format cursors.
Changes in This Commit
----------------------
- package zfs
- infrastructure for holds
- infrastructure for resume token decoding
- implement a variant of OpenZFS's `entity_namecheck` and use it for validation in new code
- ZFSSendArgs to specify a ZFS send operation
- validation code protects against malicious resume tokens by checking that the token encodes the same send parameters that the send-side would use if no resume token were available (i.e. same filesystem, `fromguid`, `toguid`)
- RecvOptions support for `recv -s` flag
- convert a bunch of ZFS operations to be idempotent
- achieved through more differentiated error message scraping / additional pre-/post-checks
- package replication/pdu
- add field for encryption to send request messages
- add fields for resume handling to send & recv request messages
- receive requests now contain `FilesystemVersion To` in addition to the filesystem into which the stream should be `recv`d into
- can use `zfs recv $root_fs/$client_id/path/to/dataset@${To.Name}`, which enables additional validation after recv (i.e. whether `To.Guid` matched what we received in the stream)
- used to set `last-received-hold`
- package replication/logic
- introduce `PlannerPolicy` struct, currently only used to configure whether encrypted sends should be requested from the sender
- integrate encryption and resume token support into `Step` struct
- package endpoint
- move the concepts that endpoint builds on top of ZFS to a single file `endpoint/endpoint_zfs.go`
- step-holds + step-bookmarks
- last-received-hold
- new replication cursor + old replication cursor compat code
- adjust `endpoint/endpoint.go` handlers for
- encryption
- resumability
- new replication cursor
- last-received-hold
- client subcommand `zrepl holds list`: list all holds and hold-like bookmarks that zrepl thinks belong to it
- client subcommand `zrepl migrate replication-cursor:v1-v2`
Symptom: zrepl log message:
rpc error: code = Unavailable desc = transport is closing
Underlying Problem:
* rpc.NewServer was not using grpchelper.NewServer and not setting Server KeepaliveParams by itself
* and even grpchelper.NewServer didn't set a KeepaliveEnforcementPolicy
* However, KeepaliveEnforcementPolicy is necessary if the client keepalive is configured with non-default values
* .. which grpchelper.ClientConn does, and that method is used by rpc.NewClient
* rpc.Client was sending pings
* lacking server-side KeepaliveEnforcementPolicy caused grpc-hard-coded `pingStrikes` counter to go past limit of 2:
021bd5734e/internal/transport/http2_server.go (L726)
How was this debugged:
* GRPC_GO_LOG_VERBOSITY_LEVEL=99 GRPC_GO_LOG_SEVERITY_LEVEL=info PATH=/root/mockpath:$PATH zrepl daemon
* with a patch on grpc package to get more log messages on pingStrikes increases:
diff --git a/internal/transport/http2_server.go b/internal/transport/http2_server.go
index 8b04b039..f68f55ea 100644
--- a/internal/transport/http2_server.go
+++ b/internal/transport/http2_server.go
@@ -214,6 +214,7 @@ func newHTTP2Server(conn net.Conn, config *ServerConfig) (_ ServerTransport, err
if kep.MinTime == 0 {
kep.MinTime = defaultKeepalivePolicyMinTime
}
+ errorf("effective keepalive enforcement policy: %#v", kep)
done := make(chan struct{})
t := &http2Server{
ctx: context.Background(),
@@ -696,6 +697,7 @@ func (t *http2Server) handlePing(f *http2.PingFrame) {
t.controlBuf.put(pingAck)
now := time.Now()
+ errorf("transport:ping handlePing, last ping %s ago", now.Sub(t.lastPingAt))
defer func() {
t.lastPingAt = now
}()
@@ -713,11 +715,13 @@ func (t *http2Server) handlePing(f *http2.PingFrame) {
// Keepalive shouldn't be active thus, this new ping should
// have come after at least defaultPingTimeout.
if t.lastPingAt.Add(defaultPingTimeout).After(now) {
+ errorf("transport:ping strike ns < 1 && !t.kep.PermitWithoutStream")
t.pingStrikes++
}
} else {
// Check if keepalive policy is respected.
if t.lastPingAt.Add(t.kep.MinTime).After(now) {
+ errorf("transport:ping strike !(ns < 1 && !t.kep.PermitWithoutStream) kep.MinTime=%s ns=%d", t.kep.MinTime, ns)
t.pingStrikes++
}
}
fixes#181
Verified once again that grpc.DialContext is indeed non-blocking.
However, it checks in a defer stmt that the passed dial is not ctx.Done().
That is highly unusual if the dial is non-blocking.
But it might still happen, maybe because of machine suspend during the function call and before the defer stmt is executed.
panic:
context deadline exceeded
goroutine 49 [running]:
github.com/zrepl/zrepl/rpc/grpcclientidentity/grpchelper.ClientConn(0x1906ea0, 0xc0003ea1e0, 0x1921620, 0xc0002da660, 0x0)
/gopath/src/github.com/zrepl/zrepl/rpc/grpcclientidentity/grpchelper/authlistener_grpc_adaptor_wrapper.go:49 +0x38c
github.com/zrepl/zrepl/rpc.NewClient(0x1906f00, 0xc0002d60f0, 0x1921620, 0xc0002da640, 0x1921620, 0xc0002da660, 0x1921620, 0xc0002da6a0, 0x1921620)
/gopath/src/github.com/zrepl/zrepl/rpc/rpc_client.go:53 +0x199
github.com/zrepl/zrepl/daemon/job.(*modePush).ConnectEndpoints(0xc0000d1e90, 0x1921620, 0xc0002da640, 0x1921620, 0xc0002da660, 0x1921620, 0xc0002da6a0, 0x1906f00, 0xc0002d60f0)
/gopath/src/github.com/zrepl/zrepl/daemon/job/active.go:105 +0x15d
github.com/zrepl/zrepl/daemon/job.(*ActiveSide).do(0xc0000d6120, 0x1918720, 0xc00020f170)
/gopath/src/github.com/zrepl/zrepl/daemon/job/active.go:356 +0x236
github.com/zrepl/zrepl/daemon/job.(*ActiveSide).Run(0xc0000d6120, 0x1918720, 0xc00009c660)
/gopath/src/github.com/zrepl/zrepl/daemon/job/active.go:347 +0x289
github.com/zrepl/zrepl/daemon.(*jobs).start.func1(0xc0000fc880, 0x1921620, 0xc0002da120, 0x191a320, 0xc0000d6120, 0x1918720, 0xc0002d6a80)
* Add Close() in closeState to identify the first closer
* Non-first closers get an error
* Reads and Writes from the Conn get an error if the conn was closed
during the Read / Write was running
* The first closer starts _separate_ goroutine draining the c.frameReads channel
* The first closer then waits for the goroutine that fills c.frameReads
to exit
refs 3bfe0c16d0fixes#174
readFrames would block on `reads <-`
but only after that would stream.Conn.readFrames close c.waitReadFramesDone
which was too late because stream.Conn.Close would wait for c.waitReadFramesDone to be closed before draining the channel
^^^^^^ (not frameconn.Conn, that closed successfully)
195 @ 0x1032ae0 0x1006cab 0x1006c81 0x1006a65 0x15505be 0x155163e 0x1060bc1
0x15505bd github.com/zrepl/zrepl/rpc/dataconn/stream.readFrames+0x16d github.com/zrepl/zrepl/rpc/dataconn/stream/stream.go:220
0x155163d github.com/zrepl/zrepl/rpc/dataconn/stream.(*Conn).readFrames+0xbd github.com/zrepl/zrepl/rpc/dataconn/stream/stream_conn.go:71
195 @ 0x1032ae0 0x10078c8 0x100789e 0x100758b 0x1552678 0x1557a4b 0x1556aec 0x1060bc1
0x1552677 github.com/zrepl/zrepl/rpc/dataconn/stream.(*Conn).Close+0x77 github.com/zrepl/zrepl/rpc/dataconn/stream/stream_conn.go:191
0x1557a4a github.com/zrepl/zrepl/rpc/dataconn.(*Server).serveConn.func1+0x5a github.com/zrepl/zrepl/rpc/dataconn/dataconn_server.go:93
0x1556aeb github.com/zrepl/zrepl/rpc/dataconn.(*Server).serveConn+0x87b github.com/zrepl/zrepl/rpc/dataconn/dataconn_server.go:176
treat handshake errors as permanent on the client
The issue was observed by 100% CPU usage due to lack ofrate-limiting in
dataconn.ReqPing retries=> safeguard that
Summary:
* Logging is still bad
* test output in a lot of placed
* FIXMEs every where
Test Plan: None, just review
Differential Revision: https://phabricator.cschwarz.com/D2
Tear down occurs on each protocol level, stack-wise.
Open RWC
Open ML (with NewMessageLayer)
Open RPC (with NewServer/ NewClient)
Close RPC (with Close() from Client())
Close ML
* in Server: after error / receive of Close request
* in Client: after getting ACK for Close request from Server
Close RWC
To achieve this, a DataType for RPC control messages was added, which
has a separate set of endpoints. Not exactly pretty, but works for now.
The necessity of the RST frame remains to be determined. However, it is
nice to have a way to signal the other side something went terribly
wrong in the middle of an operation. Example: A frameBridingWriter fails
to read the next chunk of a file it is supposed to send, it can just
send an RST frame to signal this operation failed... Wouldn't trailers
make sense then?
The existing ByteStreamRPC requires writing RPC stub + server code
for each RPC endpoint. Does not scale well.
Goal: adding a new RPC call should
- not require writing an RPC stub / handler
- not require modifications to the RPC lib
The wire format is inspired by HTTP2, the API by net/rpc.
Frames are used for framing messages, i.e. a message is made of multiple
frames which are glued together using a frame-bridging reader / writer.
This roughly corresponds to HTTP2 streams, although we're happy with
just one stream at any time and the resulting non-need for flow control,
etc.
Frames are typed using a header. The two most important types are
'Header' and 'Data'.
The RPC protocol is built on top of this:
- Client sends a header => multiple frames of type 'header'
- Client sends request body => mulitiple frames of type 'data'
- Server reads a header => multiple frames of type 'header'
- Server reads request body => mulitiple frames of type 'data'
- Server sends response header => ...
- Server sends response body => ...
An RPC header is serialized JSON and always the same structure.
The body is of the type specified in the header.
The RPC server and client use some semi-fancy reflection tequniques to
automatically infer the data type of the request/response body based on
the method signature of the server handler; or the client parameters,
respectively.
This boils down to a special-case for io.Reader, which are just dumped
into a series of data frames as efficiently as possible.
All other types are (de)serialized using encoding/json.
The RPC layer and Frame Layer log some arbitrary messages that proved
useful during debugging. By default, they log to a non-logger, which
should not have a big impact on performance.
pprof analysis shows the implementation spends its CPU time
60% waiting for syscalls
30% in memmove
10% ...
On a Intel(R) Core(TM) i7-6600U CPU @ 2.60GHz CPU, Linux 4.12, the
implementation achieved ~3.6GiB/s.
Future optimization may include spice(2) / vmspice(2) on Linux, although
this doesn't fit so well with the heavy use of io.Reader / io.Writer
throughout the codebase.
The existing hackaround for local calls was re-implemented to fit the
new interface of PRCServer and RPCClient.
The 'R'PC method invocation is a bit slower because reflection is
involved inbetween, but otherwise performance should be no different.
The RPC code currently does not support multipart requests and thus does
not support the equivalent of a POST.
Thus, the switch to the new rpc code had the following fallout:
- Move request objects + constants from rpc package to main app code
- Sacrifice the hacky 'push = pull me' way of doing push
-> need to further extend RPC to support multipart requests or
something to implement this properly with additional interfaces
-> should be done after replication is abstracted better than separate
algorithms for doPull() and doPush()
Pushing is achieved by inverting the roles on the established
connection, i.e. the client tells the server what data it should pull
from the client (PullMeRequest).
Role inversion is achieved by moving the server loop to the serverLoop
function of ByteStreamRPC, which can be called from both the Listen()
function (server-side) and the PullMeRequest() client-side function.
A donwside of this PullMe approach is that the replication policies
become part of the rpc, because the puller must follow the policy.
JSONDecoder was buffering more of connection data than just the JSON.
=> Unchunker didn't bother and just started unchunking.
While chaining JSONDecoder.Buffered() and the connection using
ChainedReader works, it's still not a clean architecture.
=> Every JSON message is now wrapped in a chunked stream
(chunked and unchunked)
=> no special-cases
=> Keep ChainedReader, might be useful later on...
