# Nimbus # Copyright (c) 2018-2024 Status Research & Development GmbH # Licensed under either of # * Apache License, version 2.0, ([LICENSE-APACHE](LICENSE-APACHE) or # http://www.apache.org/licenses/LICENSE-2.0) # * MIT license ([LICENSE-MIT](LICENSE-MIT) or # http://opensource.org/licenses/MIT) # at your option. This file may not be copied, modified, or distributed except # according to those terms. ## TODO: ## ===== ## * Impose a size limit to the bucket database. Which items would be removed? ## ## * There is a conceivable problem with the per-account optimisation. The ## algorithm chooses an account and does not stop packing until all txs ## of the account are packed or the block is full. In the latter case, ## there might be some txs left unpacked from the account which might be ## the most lucrative ones. Should this be tackled (see also next item)? ## ## * The classifier throws out all txs with negative gas tips. This implies ## that all subsequent txs must also be suspended for this account even ## though these following txs might be extraordinarily profitable so that ## packing the whole account might be woth wile. Should this be considered, ## somehow (see also previous item)? ## ## ## Transaction Pool ## ================ ## ## The transaction pool collects transactions and holds them in a database. ## This database consists of the three buckets *pending*, *staged*, and ## *packed* and a *waste basket*. These database entities are discussed in ## more detail, below. ## ## At some point, there will be some transactions in the *staged* bucket. ## Upon request, the pool will pack as many of those transactions as possible ## into to *packed* bucket which will subsequently be used to generate a ## new Ethereum block. ## ## When packing transactions from *staged* into *packed* bucked, the staged ## transactions are sorted by *sender account* and *nonce*. The *sender ## account* values are ordered by a *ranking* function (highest ranking first) ## and the *nonce* values by their natural integer order. Then, transactions ## are greedily picked from the ordered set until there are enough ## transactions in the *packed* bucket. Some boundary condition applies which ## roughly says that for a given account, all the transactions packed must ## leave no gaps between nonce values when sorted. ## ## The rank function applied to the *sender account* sorting is chosen as a ## guess for higher profitability which goes with a higher rank account. ## ## ## Rank calculator ## --------------- ## Let *tx()* denote the mapping ## :: ## tx: (account,nonce) -> tx ## ## from an index pair *(account,nonce)* to a transaction *tx*. Also, for some ## external parameter *baseFee*, let ## :: ## maxProfit: (tx,baseFee) -> tx.effectiveGasTip(baseFee) * tx.gasLimit ## ## be the maximal tip a single transation can achieve (where unit of the ## *effectiveGasTip()* is a *price* and *gasLimit* is a *commodity value*.). ## Then the rank function ## :: ## rank(account) = Σ maxProfit(tx(account,ν),baseFee) / Σ tx(account,ν).gasLimit ## ν ν ## ## is a *price* estimate of the maximal avarage tip per gas unit over all ## transactions for the given account. The nonces `ν` for the summation ## run over all transactions from the *staged* and *packed* bucket. ## ## ## ## ## Pool database: ## -------------- ## :: ## . . ## . . ## . . +----------+ ## add() ----+---------------------------------> | | ## | . +-----------+ . | disposed | ## +-----------> | pending | ------> | | ## . +-----------+ . | | ## . | ^ ^ . | waste | ## . v | | . | basket | ## . +----------+ | . | | ## . | staged | | . | | ## . +----------+ | . | | ## . | | ^ | . | | ## . | v | | . | | ## . | +----------+ . | | ## . | | packed | -------> | | ## . | +----------+ . | | ## . +----------------------> | | ## . . +----------+ ## ## The three columns *Batch queue*, *State bucket*, and *Terminal state* ## represent three different accounting (or database) systems. The pool ## database is continuosly updated while new transactions are added. ## Transactions are bundled with meta data which holds the full datanbase ## state in addition to other cached information like the sender account. ## ## ## New transactions ## ---------------- ## When entering the pool, new transactions are bundled with meta data and ## appended to the batch queue. These bundles are called *items* which are ## forwarded to one of the following entites: ## ## * the *staged* bucket if the transaction is valid and match some constraints ## on expected minimum mining fees (or a semblance of that for *non-PoW* ## networks) ## * the *pending* bucket if the transaction is valid but is not subject to be ## held in the *staged* bucket ## * the *waste basket* if the transaction is invalid ## ## If a valid transaction item supersedes an existing one, the existing ## item is moved to the waste basket and the new transaction replaces the ## existing one in the current bucket if the gas price of the transaction is ## at least `priceBump` per cent higher (see adjustable parameters, below.) ## ## Status buckets ## -------------- ## The term *bucket* is a nickname for a set of *items* (i.e. transactions ## bundled with meta data as mentioned earlier) all labelled with the same ## `status` symbol and not marked *waste*. In particular, bucket membership ## for an item is encoded as ## ## * the `status` field indicates the particular *bucket* membership ## * the `reject` field is reset/unset and has zero-equivalent value ## ## The following boundary conditions hold for the union of all buckets: ## ## * *Unique index:* ## Let **T** be the union of all buckets and **Q** be the ## set of *(sender,nonce)* pairs derived from the items of **T**. Then ## **T** and **Q** are isomorphic, i.e. for each pair *(sender,nonce)* ## from **Q** there is exactly one item from **T**, and vice versa. ## ## * *Consecutive nonces:* ## For each *(sender0,nonce0)* of **Q**, either ## *(sender0,nonce0-1)* is in **Q** or *nonce0* is the current nonce as ## registered with the *sender account* (implied by the block chain), ## ## The *consecutive nonces* requirement involves the *sender account* ## which depends on the current state of the block chain as represented by the ## internally cached head (i.e. insertion point where a new block is to be ## appended.) ## ## The following notation describes sets of *(sender,nonce)* pairs for ## per-bucket items. It will be used for boundary conditions similar to the ## ones above. ## ## * **Pending** denotes the set of *(sender,nonce)* pairs for the ## *pending* bucket ## ## * **Staged** denotes the set of *(sender,nonce)* pairs for the ## *staged* bucket ## ## * **Packed** denotes the set of *(sender,nonce)* pairs for the ## *packed* bucket ## ## The pending bucket ## ^^^^^^^^^^^^^^^^^^ ## Items in this bucket hold valid transactions that are not in any of the ## other buckets. All itmes might be promoted form here into other buckets if ## the current state of the block chain as represented by the internally cached ## head changes. ## ## The staged bucket ## ^^^^^^^^^^^^^^^^^ ## Items in this bucket are ready to be added to a new block. They typycally ## imply some expected minimum reward when mined on PoW networks. Some ## boundary condition holds: ## ## * *Consecutive nonces:* ## For any *(sender0,nonce0)* pair from **Staged**, the pair ## *(sender0,nonce0-1)* is not in **Pending**. ## ## Considering the respective boundary condition on the union of buckets ## **T**, this condition here implies that a *staged* per sender nonce has a ## predecessor in the *staged* or *packed* bucket or is a nonce as registered ## with the *sender account*. ## ## The packed bucket ## ^^^^^^^^^^^^^^^^^ ## All items from this bucket have been selected from the *staged* bucket, the ## transactions of which (i.e. unwrapped items) can go right away into a new ## ethernet block. How these items are selected was described at the beginning ## of this chapter. The following boundary conditions holds: ## ## * *Consecutive nonces:* ## For any *(sender0,nonce0)* pair from **Packed**, the pair ## *(sender0,nonce0-1)* is neither in **Pending**, nor in **Staged**. ## ## Considering the respective boundary condition on the union of buckets ## **T**, this condition here implies that a *packed* per-sender nonce has a ## predecessor in the very *packed* bucket or is a nonce as registered with the ## *sender account*. ## ## ## Terminal state ## -------------- ## After use, items are disposed into a waste basket *FIFO* queue which has a ## maximal length. If the length is exceeded, the oldest items are deleted. ## The waste basket is used as a cache for discarded transactions that need to ## re-enter the system. Recovering from the waste basket saves the effort of ## recovering the sender account from the signature. An item is identified ## *waste* if ## ## * the `reject` field is explicitely set and has a value different ## from a zero-equivalent. ## ## So a *waste* item is clearly distinguishable from any active one as a ## member of one of the *status buckets*. ## ## ## ## Pool coding ## =========== ## A piece of code using this pool architecture could look like as follows: ## :: ## # see also unit test examples, e.g. "Block packer tests" ## var db: CoreDbRef # to be initialised ## var txs: seq[Transaction] # to be initialised ## ## proc mineThatBlock(blk: EthBlock) # external function ## ## .. ## ## var xq = TxPoolRef.new(db) # initialise tx-pool ## .. ## ## xq.add(txs) # add transactions .. ## .. # .. into the buckets ## ## let newBlock = xq.assembleBlock # fetch current mining block ## ## .. ## mineThatBlock(newBlock) ... # external mining & signing process ## .. ## ## xp.smartHead(newBlock.header) # update pool, new insertion point ## ## ## Discussion of example ## --------------------- ## In the example, transactions are processed into buckets via `add()`. ## ## The `ethBlock()` directive assembles and retrieves a new block for mining ## derived from the current pool state. It invokes the block packer which ## accumulates txs from the `pending` buscket into the `packed` bucket which ## then go into the block. ## ## Then mining and signing takes place ... ## ## After mining and signing, the view of the block chain as seen by the pool ## must be updated to be ready for a new mining process. In the best case, the ## canonical head is just moved to the currently mined block which would imply ## just to discard the contents of the *packed* bucket with some additional ## transactions from the *staged* bucket. A more general block chain state ## head update would be more complex, though. ## ## In the most complex case, the newly mined block was added to some block ## chain branch which has become an uncle to the new canonical head retrieved ## by `getCanonicalHead()`. In order to update the pool to the very state ## one would have arrived if worked on the retrieved canonical head branch ## in the first place, the directive `smartHead()` calculates the actions of ## what is needed to get just there from the locally cached head state of the ## pool. These actions are applied by `smartHead()` after the internal head ## position was moved. ## ## The *setter* behind the internal head position adjustment also caches ## updated internal parameters as base fee, state, fork, etc. ## ## ## Adjustable Parameters ## --------------------- ## ## flags ## The `flags` parameter holds a set of strategy symbols for how to process ## items and buckets. ## ## *autoUpdateBucketsDB* ## Automatically update the state buckets after running batch jobs if the ## `dirtyBuckets` flag is also set. ## ## *autoZombifyUnpacked* ## Automatically dispose *pending* or *staged* tx items that were added to ## the state buckets database at least `lifeTime` ago. ## ## lifeTime ## Txs that stay longer in one of the buckets will be moved to a waste ## basket. From there they will be eventually deleted oldest first when ## the maximum size would be exceeded. ## ## priceBump ## There can be only one transaction in the database for the same `sender` ## account and `nonce` value. When adding a transaction with the same ## (`sender`, `nonce`) pair, the new transaction will replace the current one ## if it has a gas price which is at least `priceBump` per cent higher. ## ## ## Read-Only Parameters ## -------------------- ## head ## Cached block chain insertion point, not necessarily the same header as ## retrieved by the `getCanonicalHead()`. This insertion point can be ## adjusted with the `smartHead()` function. import std/[sequtils, tables], ./tx_pool/[tx_packer, tx_desc, tx_info, tx_item], ./tx_pool/tx_tabs, ./tx_pool/tx_tasks/[ tx_add, tx_bucket, tx_head, tx_dispose], chronicles, stew/keyed_queue, results, ../common/common, ./chain/forked_chain, ./casper export TxItemRef, TxItemStatus, TxPoolFlags, TxPoolRef, TxTabsItemsCount, results, tx_desc.startDate, tx_info, tx_item.effectiveGasTip, tx_item.info, tx_item.itemID, tx_item.sender, tx_item.status, tx_item.timeStamp, tx_item.tx, tx_desc.head {.push raises: [].} logScope: topics = "tx-pool" # ------------------------------------------------------------------------------ # Private functions: tasks processor # ------------------------------------------------------------------------------ proc maintenanceProcessing(xp: TxPoolRef) {.gcsafe,raises: [CatchableError].} = ## Tasks to be done after add/del txs processing # Purge expired items if autoZombifyUnpacked in xp.pFlags: # Move transactions older than `xp.lifeTime` to the waste basket. xp.disposeExpiredItems # Update buckets if autoUpdateBucketsDB in xp.pFlags: if xp.pDirtyBuckets: # For all items, re-calculate item status values (aka bucket labels). # If the `force` flag is set, re-calculation is done even though the # change flag has remained unset. discard xp.bucketUpdateAll xp.pDirtyBuckets = false proc setHead(xp: TxPoolRef; val: Header) {.gcsafe,raises: [CatchableError].} = ## Update cached block chain insertion point. This will also update the ## internally cached `baseFee` (depends on the block chain state.) if xp.head != val: xp.head = val # calculates the new baseFee xp.txDB.baseFee = xp.baseFee xp.pDirtyBuckets = true xp.bucketFlushPacked # ------------------------------------------------------------------------------ # Public constructor/destructor # ------------------------------------------------------------------------------ proc new*(T: type TxPoolRef; com: CommonRef): T {.gcsafe,raises: [CatchableError].} = ## Constructor, returns a new tx-pool descriptor. new result result.init(com) # ------------------------------------------------------------------------------ # Public functions, task manager, pool actions serialiser # ------------------------------------------------------------------------------ # core/tx_pool.go(848): func (pool *TxPool) AddLocals(txs [].. # core/tx_pool.go(864): func (pool *TxPool) AddRemotes(txs [].. proc add*(xp: TxPoolRef; txs: openArray[PooledTransaction]; info = "") {.gcsafe,raises: [CatchableError].} = ## Add a list of transactions to be processed and added to the buckets ## database. It is OK pass an empty list in which case some maintenance ## check can be forced. ## ## The argument Transactions `txs` may come in any order, they will be ## sorted by `` before adding to the database with the ## least nonce first. For this reason, it is suggested to pass transactions ## in larger groups. Calling single transaction jobs, they must strictly be ## passed *smaller nonce* before *larger nonce*. xp.pDoubleCheckAdd xp.addTxs(txs, info).topItems xp.maintenanceProcessing # core/tx_pool.go(854): func (pool *TxPool) AddLocals(txs [].. # core/tx_pool.go(883): func (pool *TxPool) AddRemotes(txs [].. proc add*(xp: TxPoolRef; tx: PooledTransaction; info = "") {.gcsafe,raises: [CatchableError].} = ## Variant of `add()` for a single transaction. xp.add(@[tx], info) proc smartHead*(xp: TxPoolRef; pos: Header, chain: ForkedChainRef): bool {.gcsafe,raises: [CatchableError].} = ## This function moves the internal head cache (i.e. tx insertion point, ## vmState) and ponts it to a now block on the chain. ## ## it calculates the ## txs that need to be added or deleted after moving the insertion point ## head so that the tx-pool will not fail to re-insert quered txs that are ## on the chain, already. Neither will it loose any txs. After updating the ## the internal head cache, the previously calculated actions will be ## applied. ## let rcDiff = xp.headDiff(pos, chain) if rcDiff.isOk: let changes = rcDiff.value # Need to move head before adding txs which may rightly be rejected in # `addTxs()` otherwise. xp.setHead(pos) # Delete already *mined* transactions if 0 < changes.remTxs.len: debug "queuing delta txs", mode = "remove", num = changes.remTxs.len xp.disposeById(toSeq(changes.remTxs.keys), txInfoChainHeadUpdate) xp.maintenanceProcessing return true # ------------------------------------------------------------------------------ # Public functions, getters # ------------------------------------------------------------------------------ func com*(xp: TxPoolRef): CommonRef = ## Getter xp.vmState.com type AssembledBlock* = object blk*: EthBlock blobsBundle*: Opt[BlobsBundle] blockValue*: UInt256 executionRequests*: Opt[array[3, seq[byte]]] proc assembleBlock*( xp: TxPoolRef, someBaseFee: bool = false ): Result[AssembledBlock, string] {.gcsafe,raises: [CatchableError].} = ## Getter, retrieves a packed block ready for mining and signing depending ## on the internally cached block chain head, the txs in the pool and some ## tuning parameters. The following block header fields are left ## uninitialised: ## ## * *extraData*: Blob ## * *mixHash*: Hash32 ## * *nonce*: BlockNonce ## ## Note that this getter runs *ad hoc* all the txs through the VM in ## order to build the block. let pst = xp.packerVmExec().valueOr: # updates vmState return err(error) var blk = EthBlock( header: pst.assembleHeader # uses updated vmState ) var blobsBundle: BlobsBundle for _, nonceList in xp.txDB.packingOrderAccounts(txItemPacked): for item in nonceList.incNonce: let tx = item.pooledTx blk.txs.add tx.tx if tx.networkPayload != nil: for k in tx.networkPayload.commitments: blobsBundle.commitments.add k for p in tx.networkPayload.proofs: blobsBundle.proofs.add p for blob in tx.networkPayload.blobs: blobsBundle.blobs.add blob blk.header.transactionsRoot = calcTxRoot(blk.txs) let com = xp.vmState.com if com.isShanghaiOrLater(blk.header.timestamp): blk.withdrawals = Opt.some(com.pos.withdrawals) if not com.isCancunOrLater(blk.header.timestamp) and blobsBundle.commitments.len > 0: return err("PooledTransaction contains blobs prior to Cancun") let blobsBundleOpt = if com.isCancunOrLater(blk.header.timestamp): doAssert blobsBundle.commitments.len == blobsBundle.blobs.len doAssert blobsBundle.proofs.len == blobsBundle.blobs.len Opt.some blobsBundle else: Opt.none BlobsBundle if someBaseFee: # make sure baseFee always has something blk.header.baseFeePerGas = Opt.some(blk.header.baseFeePerGas.get(0.u256)) let executionRequestsOpt = if com.isPragueOrLater(blk.header.timestamp): Opt.some(pst.executionRequests) else: Opt.none(array[3, seq[byte]]) ok AssembledBlock( blk: blk, blobsBundle: blobsBundleOpt, blockValue: pst.blockValue, executionRequests: executionRequestsOpt) # core/tx_pool.go(474): func (pool SetGasPrice,*TxPool) Stats() (int, int) { # core/tx_pool.go(1728): func (t *txLookup) Count() int { # core/tx_pool.go(1737): func (t *txLookup) LocalCount() int { # core/tx_pool.go(1745): func (t *txLookup) RemoteCount() int { func nItems*(xp: TxPoolRef): TxTabsItemsCount = ## Getter, retrieves the current number of items per bucket and ## some totals. xp.txDB.nItems # ------------------------------------------------------------------------------ # Public functions, per-tx-item operations # ------------------------------------------------------------------------------ # core/tx_pool.go(979): func (pool *TxPool) Get(hash common.Hash) .. # core/tx_pool.go(985): func (pool *TxPool) Has(hash common.Hash) bool { func getItem*(xp: TxPoolRef; hash: Hash32): Result[TxItemRef,void] = ## Returns a transaction if it is contained in the pool. xp.txDB.byItemID.eq(hash) func disposeItems*(xp: TxPoolRef; item: TxItemRef; reason = txInfoExplicitDisposal; otherReason = txInfoImpliedDisposal): int {.discardable,gcsafe,raises: [CatchableError].} = ## Move item to wastebasket. All items for the same sender with nonces ## greater than the current one are deleted, as well. The function returns ## the number of items eventally removed. xp.disposeItemAndHigherNonces(item, reason, otherReason) iterator txHashes*(xp: TxPoolRef): Hash32 = for txHash in nextKeys(xp.txDB.byItemID): yield txHash iterator okPairs*(xp: TxPoolRef): (Hash32, TxItemRef) = for x in nextPairs(xp.txDB.byItemID): if x.data.reject == txInfoOk: yield (x.key, x.data) func numTxs*(xp: TxPoolRef): int = xp.txDB.byItemID.len func disposeAll*(xp: TxPoolRef) {.raises: [CatchableError].} = let numTx = xp.numTxs var list = newSeqOfCap[TxItemRef](numTx) for x in nextPairs(xp.txDB.byItemID): list.add x.data for x in list: xp.disposeItems(x) # ------------------------------------------------------------------------------ # Public functions, local/remote accounts # ------------------------------------------------------------------------------ func inPoolAndOk*(xp: TxPoolRef; txHash: Hash32): bool = let res = xp.getItem(txHash) if res.isErr: return false res.get().reject == txInfoOk func inPoolAndReason*(xp: TxPoolRef; txHash: Hash32): Result[void, string] = let res = xp.getItem(txHash) if res.isErr: # try to look in rejecteds let r = xp.txDB.byRejects.eq(txHash) if r.isErr: return err("cannot find tx in txpool") else: return err(r.get().rejectInfo) let item = res.get() if item.reject == txInfoOk: return ok() else: return err(item.rejectInfo) # ------------------------------------------------------------------------------ # End # ------------------------------------------------------------------------------