nimbus-eth1/nimbus/sync/misc/block_queue.nim

673 lines
23 KiB
Nim

# Nimbus
# Copyright (c) 2018-2021 Status Research & Development GmbH
# Licensed and distributed under either of
# * MIT license (license terms in the root directory or at
# https://opensource.org/licenses/MIT).
# * Apache v2 license (license terms in the root directory or at
# https://www.apache.org/licenses/LICENSE-2.0).
# at your option. This file may not be copied, modified, or distributed
# except according to those terms.
## Fetch and queue blocks
## ======================
##
## Worker items state diagram and sketch of sync algorithm:
## ::
## unprocessed | | ready for | store into
## block ranges | peer workers | persistent database | database
## =======================================================================
##
## +------------------------------------------+
## | |
## | +----------------------------+ |
## | | | |
## V v | |
## <unprocessed> ---+---> <worker-0> ---+-----> <staged> -------> OUTPUT
## | |
## +---> <worker-1> ---+
## | |
## +---> <worker-2> ---+
## : :
##
## A work item is created from a range of block numbers extracted from the
## `<unprocessed>` set of block ranges.
##
## A work item consists of a
## * current state `<worker-#>` or `<staged>`
## * given range of consecutive block numbers `[from..to]`
## * sequence of block headers relating to `[from..to]` (to be completed)
## * sequence of block buddies relating to `[from..to]` (to be completed)
##
## Block ranges *may* be recycled back into the `<unprocessed>` set when a
## work item is destroyed. This is supposed to be an exceptional case.
## Typically, a `<staged>` work item is added to the persistent block chain
## database and destroyed without block range recycling.
##
## Beware of `<staged>` overflow
## -----------------------------
## When the `<staged>` queue gets too long in non-backtrack/re-org mode, this
## may be caused by a gap between the least `<unprocessed>` block number and
## the least `<staged>` block number. Then a mechanism is invoked where
## `<unprocessed>` block range is updated.
##
## For backtrack/re-org the system runs in single instance mode tracing
## backvards parent hash references. So updating `<unprocessed>` block numbers
## would have no effect. In that case, the record with the largest block
## numbers are deleted from the `<staged>` list.
##
import
std/[algorithm, options, sequtils, strutils],
chronicles,
chronos,
eth/[common, p2p],
stew/[byteutils, interval_set, sorted_set],
../../db/db_chain,
../../utils/utils,
".."/[protocol, sync_desc, types]
{.push raises:[Defect].}
logScope:
topics = "block-queue"
const
maxStagedWorkItems = 70
## Maximal items in the `staged` list.
stagedWorkItemsTrigger = 50
## Turn on the global `poolMode` if there are more than this many items
## staged.
type
BlockQueueRC* = enum
## Return & error codes
AllSmileOk
EmptyQueue
StagedQueueOverflow
BlockNumberGap
BacktrackDisabled
FetchHeadersError
FetchBodiesError
NoMoreUnprocessed
NoMorePeerBlocks
BlockRangeSetRef = IntervalSetRef[BlockNumber,UInt256]
## Disjunct sets of block number intervals
BlockRange = Interval[BlockNumber,UInt256]
## Block number interval
BlockItemQueue = SortedSet[BlockNumber,BlockItemRef]
## Block intervals sorted by least block number
BlockItemWalkRef = SortedSetWalkRef[BlockNumber,BlockItemRef]
## Fast traversal descriptor for `BlockItemQueue`
BlockItemRef* = ref object
## Public block items, OUTPUT
blocks*: BlockRange ## Block numbers ranvge covered
topHash*: Option[Hash256] ## Fetched by top hash rather than block
headers*: seq[BlockHeader] ## Block headers received
hashes*: seq[Hash256] ## Hashed from `headers[]` for convenience
bodies*: seq[BlockBody] ## Block bodies received
BlockQueueCtxRef* = ref object
## Globally shared data among `block` instances
backtrack: Option[Hash256] ## Find reverse block after re-org
unprocessed: BlockRangeSetRef ## Block ranges to fetch
staged: BlockItemQueue ## Blocks fetched but not stored yet
topAccepted: BlockNumber ## Up to this block number processed OK
BlockQueueWorkerRef* = ref object
## Local descriptor data extension
global: BlockQueueCtxRef ## Common data
bestNumber: Option[BlockNumber] ## Largest block number reported
ctrl: BuddyCtrlRef ## Control and state settings
peer: Peer ## network peer
BlockQueueStats* = object
## Statistics
topAccepted*: BlockNumber
nextUnprocessed*: Option[BlockNumber]
nextStaged*: Option[BlockNumber]
nStagedQueue*: int
reOrg*: bool
let
highBlockNumber = high(BlockNumber)
highBlockRange = BlockRange.new(highBlockNumber,highBlockNumber)
static:
doAssert stagedWorkItemsTrigger < maxStagedWorkItems
# ------------------------------------------------------------------------------
# Private helpers
# ------------------------------------------------------------------------------
proc `+`(n: BlockNumber; delta: static[int]): BlockNumber =
## Syntactic sugar for expressions like `xxx.toBlockNumber + 1`
n + delta.toBlockNumber
proc `-`(n: BlockNumber; delta: static[int]): BlockNumber =
## Syntactic sugar for expressions like `xxx.toBlockNumber - 1`
n - delta.toBlockNumber
proc merge(ivSet: BlockRangeSetRef; wi: BlockItemRef): Uint256 =
## Syntactic sugar
ivSet.merge(wi.blocks)
proc reduce(ivSet: BlockRangeSetRef; wi: BlockItemRef): Uint256 =
## Syntactic sugar
ivSet.reduce(wi.blocks)
# ---------------
proc pp(n: BlockNumber): string =
## Dedicated pretty printer (`$` is defined elsewhere using `UInt256`)
if n == highBlockNumber: "high" else:"#" & $n
proc `$`(iv: BlockRange): string =
## Needed for macro generated DSL files like `snap.nim` because the
## `distinct` flavour of `NodeTag` is discarded there.
result = "[" & iv.minPt.pp
if iv.minPt != iv.maxPt:
result &= "," & iv.maxPt.pp
result &= "]"
proc `$`(n: Option[BlockRange]): string =
if n.isNone: "n/a" else: $n.get
proc `$`(n: Option[BlockNumber]): string =
if n.isNone: "n/a" else: n.get.pp
proc `$`(brs: BlockRangeSetRef): string =
"{" & toSeq(brs.increasing).mapIt($it).join(",") & "}"
# ------------------------------------------------------------------------------
# Private helpers
# ------------------------------------------------------------------------------
proc nextUnprocessed(ctx: BlockQueueCtxRef): Option[BlockNumber] =
## Pseudo getter
let rc = ctx.unprocessed.ge()
if rc.isOK:
result = some(rc.value.minPt)
proc nextStaged(ctx: BlockQueueCtxRef): Option[BlockRange] =
## Pseudo getter
let rc = ctx.staged.ge(low(BlockNumber))
if rc.isOK:
result = some(rc.value.data.blocks)
template safeTransport(
qd: BlockQueueWorkerRef;
info: static[string];
code: untyped) =
try:
code
except TransportError as e:
error info & ", stop", error=($e.name), msg=e.msg
qd.ctrl.stopped = true
# ------------------------------------------------------------------------------
# Private functions
# ------------------------------------------------------------------------------
proc newWorkItem(qd: BlockQueueWorkerRef): Result[BlockItemRef,BlockQueueRC] =
## Fetch the next unprocessed block range and register it as work item.
##
## This function will grab a block range from the `unprocessed` range set,
## ove it and return it as a `BlockItemRef`. The returned range is registered
## in the `pending` list.
let rc = qd.global.unprocessed.ge()
if rc.isErr:
return err(NoMoreUnprocessed) # no more data for this peer
# Check whether there is somthing to do at all
if qd.bestNumber.isNone or
qd.bestNumber.unsafeGet < rc.value.minPt:
return err(NoMorePeerBlocks) # no more data for this peer
# Compute interval
let iv = BlockRange.new(
rc.value.minPt,
min(rc.value.maxPt,
min(rc.value.minPt + maxHeadersFetch - 1, qd.bestNumber.unsafeGet)))
discard qd.global.unprocessed.reduce(iv)
ok(BlockItemRef(blocks: iv))
proc stageItem(
qd: BlockQueueWorkerRef;
wi: BlockItemRef;
): Result[void,BlockQueueRC] =
## Add work item to the list of staged items
##
## Typically, the function returns `AllSmileOk` unless there is a queue
## oberflow (with return code`StagedQueueOverflow`) which needs to be handled
## in *pool mode* by running `blockQueueGrout()`.
var
error = AllSmileOk
let
peer = qd.peer
rc = qd.global.staged.insert(wi.blocks.minPt)
if rc.isOk:
rc.value.data = wi
# Return `true` if staged queue oberflows (unless backtracking.)
if stagedWorkItemsTrigger < qd.global.staged.len and
qd.global.backtrack.isNone and
wi.topHash.isNone:
debug "Staged queue too long", peer,
staged=qd.global.staged.len, max=stagedWorkItemsTrigger
error = StagedQueueOverflow
# The list size is limited. So cut if necessary and recycle back the block
# range of the discarded item (tough luck if the current work item is the
# one removed from top.)
while maxStagedWorkItems < qd.global.staged.len:
let topValue = qd.global.staged.le(highBlockNumber).value
discard qd.global.unprocessed.merge(topValue.data)
discard qd.global.staged.delete(topValue.key)
else:
# Ooops, duplicates should not exist (but anyway ...)
let wj = block:
let rc = qd.global.staged.eq(wi.blocks.minPt)
doAssert rc.isOk
# Store `wi` and return offending entry
let rcData = rc.value.data
rc.value.data = wi
rcData
# Update `staged` list and `unprocessed` ranges
block:
debug "Replacing dup item in staged list", peer,
range=($wi.blocks), discarded=($wj.blocks)
let rc = wi.blocks - wj.blocks
if rc.isOk:
discard qd.global.unprocessed.merge(rc.value)
if error != AllSmileOk:
return err(error)
ok()
# ------------------------------------------------------------------------------
# Private functions, asynchroneous data network activity
# ------------------------------------------------------------------------------
proc fetchHeaders(
qd: BlockQueueWorkerRef;
wi: BlockItemRef;
): Future[bool]
{.async.} =
## Get the work item with the least interval and complete it. The function
## returns `true` if bodies were fetched and there were no inconsistencies.
if 0 < wi.hashes.len:
return true
let peer = qd.peer
var hdrReq: BlocksRequest
if wi.topHash.isNone:
hdrReq = BlocksRequest(
startBlock: HashOrNum(
isHash: false,
number: wi.blocks.minPt),
maxResults: wi.blocks.len.truncate(uint),
skip: 0,
reverse: false)
trace trEthSendSendingGetBlockHeaders, peer,
blocks=($wi.blocks)
else:
hdrReq = BlocksRequest(
startBlock: HashOrNum(
isHash: true,
hash: wi.topHash.get),
maxResults: maxHeadersFetch,
skip: 0,
reverse: true)
trace trEthSendSendingGetBlockHeaders & " reverse", peer,
topHash=hdrReq.startBlock.hash, reqLen=hdrReq.maxResults
# Fetch headers from peer
var hdrResp: Option[blockHeadersObj]
block:
let reqLen = hdrReq.maxResults
qd.safeTransport("Error fetching block headers"):
hdrResp = await peer.getBlockHeaders(hdrReq)
# Beware of peer terminating the session
if qd.ctrl.stopped:
return false
if hdrResp.isNone:
trace trEthRecvReceivedBlockHeaders, peer, reqLen, respose="n/a"
return false
let hdrRespLen = hdrResp.get.headers.len
trace trEthRecvReceivedBlockHeaders, peer, reqLen, hdrRespLen
if hdrRespLen == 0:
qd.ctrl.stopped = true
return false
# Update block range for reverse search
if wi.topHash.isSome:
# Headers are in reversed order
wi.headers = hdrResp.get.headers.reversed
wi.blocks = BlockRange.new(
wi.headers[0].blockNumber, wi.headers[^1].blockNumber)
discard qd.global.unprocessed.reduce(wi)
trace "Updated reverse header range", peer, range=($wi.blocks)
# Verify start block number
elif hdrResp.get.headers[0].blockNumber != wi.blocks.minPt:
trace "Header range starts with wrong block number", peer,
startBlock=hdrResp.get.headers[0].blockNumber,
requestedBlock=wi.blocks.minPt
qd.ctrl.zombie = true
return false
# Import into `wi.headers`
else:
wi.headers.shallowCopy(hdrResp.get.headers)
# Calculate block header hashes and verify it against parent links. If
# necessary, cut off some offending block headers tail.
wi.hashes.setLen(wi.headers.len)
wi.hashes[0] = wi.headers[0].hash
for n in 1 ..< wi.headers.len:
if wi.headers[n-1].blockNumber + 1 != wi.headers[n].blockNumber:
trace "Non-consecutive block numbers in header list response", peer
qd.ctrl.zombie = true
return false
if wi.hashes[n-1] != wi.headers[n].parentHash:
# Oops, cul-de-sac after block chain re-org?
trace "Dangling parent link in header list response. Re-org?", peer
wi.headers.setLen(n)
wi.hashes.setLen(n)
break
wi.hashes[n] = wi.headers[n].hash
# Adjust range length if necessary
if wi.headers[^1].blockNumber < wi.blocks.maxPt:
let redRng = BlockRange.new(
wi.headers[0].blockNumber, wi.headers[^1].blockNumber)
trace "Adjusting block range", peer, range=($wi.blocks), reduced=($redRng)
discard qd.global.unprocessed.merge(redRng.maxPt + 1, wi.blocks.maxPt)
wi.blocks = redRng
return true
proc fetchBodies(
qd: BlockQueueWorkerRef;
wi: BlockItemRef
): Future[bool]
{.async.} =
## Get the work item with the least interval and complete it. The function
## returns `true` if bodies were fetched and there were no inconsistencies.
let peer = qd.peer
# Complete group of bodies
qd.safeTransport("Error fetching block bodies"):
while wi.bodies.len < wi.hashes.len:
let
start = wi.bodies.len
reqLen = min(wi.hashes.len - wi.bodies.len, maxBodiesFetch)
top = start + reqLen
hashes = wi.hashes[start ..< top]
trace trEthSendSendingGetBlockBodies, peer, reqLen
# Append bodies from peer to `wi.bodies`
block:
let bdyResp = await peer.getBlockBodies(hashes)
# Beware of peer terminating the session
if qd.ctrl.stopped:
return false
if bdyResp.isNone:
trace trEthRecvReceivedBlockBodies, peer, reqLen, respose="n/a"
qd.ctrl.zombie = true
return false
let bdyRespLen = bdyResp.get.blocks.len
trace trEthRecvReceivedBlockBodies, peer, reqLen, bdyRespLen
if bdyRespLen == 0 or reqLen < bdyRespLen:
qd.ctrl.zombie = true
return false
wi.bodies.add bdyResp.get.blocks
return true
# ------------------------------------------------------------------------------
# Public functions, constructor
# ------------------------------------------------------------------------------
proc init*(
T: type BlockQueueCtxRef; ## Global data descriptor type
firstBlockNumber = 0.toBlockNumber; ## Of first block to fetch from network
): T =
## Global constructor, shared data
result = T(
unprocessed: BlockRangeSetRef.init())
result.staged.init()
result.topAccepted = max(firstBlockNumber,1.toBlockNumber) - 1
discard result.unprocessed.merge(result.topAccepted + 1, highBlockNumber)
proc init*(
T: type BlockQueueWorkerRef; ## Global data descriptor type
ctx: BlockQueueCtxRef; ## Global data descriptor
ctrl: BuddyCtrlRef; ## Control and state settings
peer: Peer; ## For fetching data from network
): T =
## Buddy/local constructor
T(global: ctx,
peer: peer,
ctrl: ctrl)
# ------------------------------------------------------------------------------
# Public functions -- synchronous
# ------------------------------------------------------------------------------
proc blockQueueFetchStaged*(
qd: BlockQueueWorkerRef;
): Result[BlockItemRef,BlockQueueRC]=
## Fetch the next item from the staged block queue. This item will be removed
## from the staged queue and must be recycled if it cannot be processed.
##
## On error, the function returns `EmptyQueue` if the queue was empty and
## `BlockNumberGap` if processing this item would result in a gap between the
## last accepted block number and the fitsr block number of the next queue
## item.
##
## This gap might appear if another function processes the in-beween block
## in paralell or if something went wrong, see `blockQueueGrout()`, below.
let rc = qd.global.staged.ge(low(BlockNumber))
if rc.isErr:
# No more items in the database
return err(EmptyQueue)
let
peer = qd.peer
wi = rc.value.data
topAccepted = qd.global.topAccepted
startNumber = wi.headers[0].blockNumber
# Check whether this record of blocks can be stored, at all
if topAccepted + 1 < startNumber:
trace "Staged work item postponed", peer, topAccepted,
range=($wi.blocks), staged=qd.global.staged.len
return err(BlockNumberGap)
# Ok, store into the block chain database
trace "Staged work item", peer,
topAccepted, range=($wi.blocks)
# Remove from staged DB
discard qd.global.staged.delete(wi.blocks.minPt)
ok(wi)
proc blockQueueAccept*(qd: BlockQueueWorkerRef; wi: BlockItemRef) =
## Mark this argument item `wi` to be the item with the topmost block number
## accepted. This statement comes tyipcally after the successful processing
## and storage of the work item fetched by `blockQueueFetchStaged()`.
qd.global.topAccepted = wi.blocks.maxPt
proc blockQueueGrout*(qd: BlockQueueWorkerRef) =
## Fill the gap unprocessed and staged block numbers. If there is such a gap
## (which should not at all), the `blockQueueFetchStaged()` will always fail
## with a `true` error code because there is no next work item.
##
## To close the gap and avoid double processing, all other workers should
## have finished their tasks while this function is run. A way to achive that
## is to run this function in *pool mode* once.
# Mind the gap, fill in if necessary
let covered = min(
qd.global.nextUnprocessed.get(otherwise = highBlockNumber),
qd.global.nextStaged.get(otherwise = highBlockRange).minPt)
if qd.global.topAccepted + 1 < covered:
discard qd.global.unprocessed.merge(qd.global.topAccepted + 1, covered - 1)
proc blockQueueRecycle*(qd: BlockQueueWorkerRef; wi: BlockItemRef) =
## Put back and destroy the `wi` argument item. The covered block range needs
## to be re-fetched from the network. This statement is typically used instead
## of `blockQueueAccept()` after a failure tpo process and store the work item
## fetched by `blockQueueFetchStaged()`.
discard qd.global.unprocessed.merge(wi.blocks)
proc blockQueueRecycleStaged*(qd: BlockQueueWorkerRef) =
## Similar to `blockQueueRecycle()`, recycle all items from the staged queue.
# using fast traversal
let
walk = BlockItemWalkRef.init(qd.global.staged)
var
rc = walk.first()
while rc.isOk:
# Store back into `unprocessed` ranges set
discard qd.global.unprocessed.merge(rc.value.data)
rc = walk.next()
# optional clean up, see comments on the destroy() directive
walk.destroy()
qd.global.staged.clear()
proc blockQueueBacktrackFrom*(qd: BlockQueueWorkerRef; wi: BlockItemRef) =
## Set backtrack mode starting with the blocks before the argument work
## item `wi`.
qd.global.backtrack = some(wi.headers[0].parentHash)
proc blockQueueBacktrackOk*(qd: BlockQueueWorkerRef): bool =
## Returns `true` if the queue is in backtrack mode.
qd.global.backtrack.isSome
proc blockQueueStats*(ctx: BlockQueueCtxRef; stats: var BlockQueueStats) =
## Collect global statistics
stats.topAccepted = ctx.topAccepted
stats.nextUnprocessed = ctx.nextUnprocessed
stats.nStagedQueue = ctx.staged.len
stats.reOrg = ctx.backtrack.isSome
stats.nextStaged =
if ctx.nextStaged.isSome: some(ctx.nextStaged.unsafeGet.minPt)
else: none(BlockNumber)
# ------------------------------------------------------------------------------
# Public functions -- asynchronous
# ------------------------------------------------------------------------------
proc blockQueueBacktrackWorker*(
qd: BlockQueueWorkerRef;
): Future[Result[void,BlockQueueRC]]
{.async.} =
## This function does some backtrack processing on the queue. Backtracking
## is single threaded due to the fact that the next block is identified by
## the hash of the parent header. So this function needs to run in *single
## mode*.
##
## If backtracking is enabled, this function fetches the next parent work
## item from the network and makes it available on the staged queue to be
## retrieved with `blockQueueFetchStaged()`. In that case, the function
## succeeds and `blockQueueBacktrackOk()` will return `false`.
##
## In all other cases, the function returns an error code.
var error = BacktrackDisabled
if qd.global.backtrack.isSome:
let
peer = qd.peer
wi = BlockItemRef(
# This dummy interval can savely merged back without any effect
blocks: highBlockRange,
# Enable backtrack
topHash: some(qd.global.backtrack.unsafeGet))
# Fetch headers and bodies for the current work item
trace "Single mode worker, re-org backtracking", peer
if not await qd.fetchHeaders(wi):
error = FetchHeadersError
elif not await qd.fetchBodies(wi):
error = FetchBodiesError
else:
qd.global.backtrack = none(Hash256)
discard qd.stageItem(wi)
return ok()
# This work item failed, nothing to do anymore.
discard qd.global.unprocessed.merge(wi)
return err(error)
proc blockQueueWorker*(
qd: BlockQueueWorkerRef;
): Future[Result[void,BlockQueueRC]]
{.async.} =
## Normal worker function used to stage another work item be retrieved by
## `blockQueueFetchStaged()`. This function may run in *multi mode*. Not
## until retrieving work items the queue will be synchronised in a way that
## after the next item can be retrieved the queue will be blocked by a
## *gap* until the item is commited by `blockQueueAccept()`.
##
## On error, with most error codes there is not much that can be done. The
## one remarcable error code is `StagedQueueOverflow` which pops up if there
## is a gap between unprocessed and staged block numbers. One of the actions
## to be considered here is to run `blockQueueGrout()` in *pool mode*.
## Otherwise, the `StagedQueueOverflow` can be treated as a success would be.
##
# Fetch work item
let wi = block:
let rc = qd.newWorkItem()
if rc.isErr:
# No way, end of capacity for this peer => re-calibrate
qd.bestNumber = none(BlockNumber)
return err(rc.error)
rc.value
# Fetch headers and bodies for the current work item
var error = AllSmileOk
if not await qd.fetchHeaders(wi):
error = FetchHeadersError
elif not await qd.fetchBodies(wi):
error = FetchBodiesError
else:
return qd.stageItem(wi)
# This work item failed
discard qd.global.unprocessed.merge(wi)
return err(error)
# ------------------------------------------------------------------------------
# End
# ------------------------------------------------------------------------------