## Why
An Argon2 verify is **~8.9 ms of frozen world per login attempt** — more than half a 16 ms frame. Failed attempts cost exactly the same as successful ones, by design, so a credential-stuffing flood is a full-cost stall per packet without needing valid credentials. `SetPassword` derives a hash too, so `[password`, the admin gump and account creation each pay the same.
## What the measurement says
Off-loading does not delete the cost, it relocates it. Three things stay on the loop:
| Component | Measured |
|---|---:|
| Inline verify (today) | **8.92 ms** |
| Dispatch to the worker | 210 ns |
| Drain the continuation off `LoopContext` | 13 ns |
| Loop's own work slowed by shared-L3 eviction | **0.05 – 5.44 ms** |
Net gain **3.5 – 8.9 ms** of on-loop time per login. Harness in `ModernUO-Benchmarks` (`Benchmarks/Argon2OffLoop/`): it models the loop as a dependent-load pointer chase swept across working-set sizes, which is an upper bound on cache-latency sensitivity, and copies `EventLoopContext` so the hand-off cost is the real one.
Two results shaped the design:
- **The contention tax peaks in the middle of the working-set range**, not at the top — 5.44 ms at 8 MiB (a quarter of this chip's L3), but 0.76 ms at 30 MiB and 0.10 ms at 256 KiB. A tiny hot set has nothing in L3 to lose; a huge one is already DRAM-bound.
- **Per-login tax falls as concurrency rises** (5.44 → 2.56 → 1.60 ms at 1/2/4 hashers) while *total* loop damage rises. Contention is shared, not additive, so a login rush is not the disaster case — a single login is.
## Why exactly one worker
It is load-bearing three times over, which is also why it must not quietly become a pool:
- **Cost bound.** Off-loop loses to inline only if a hash steals ~82% of the loop's throughput. One hasher contending for one core leaves the loop ~50%. **A single background hasher cannot cost the loop more than the inline verify under any scheduling regime**, which is what lets the measurement hold on hardware we cannot inspect — AMD, VPS, oversubscribed VM. Four hashers drop the loop to ~20% and break it.
- **Memory.** Exactly one hashing arena is live at a time whatever the login volume.
- **Ordering.** Writes apply in dispatch order *only* because a single thread drains FIFO. A second worker would need ordering reintroduced; `WritesApplyInDispatchOrder` fails if that happens.
Throughput is ~110 verifies/sec. Only loop time matters, not login latency, so head-of-line blocking during a rush costs nothing.
## Making every protection safe off-thread
The worker was initially Argon2-only. That was the right call for the wrong reason — it was blamed on Argon2's salt RNG, which is a stateless syscall wrapper and was never a problem. The real blockers were elsewhere, and both are fixed at the source:
| Protection | Was | Now |
|---|---|---|
| MD5/SHA1/SHA2 | shared `HashAlgorithm.ComputeHash`, which carries the running digest across `HashCore`/`HashFinal` through process-wide singletons | static `HashData` into a `stackalloc` span — no state, no allocation, identical bytes |
| PBKDF2 | `Utility.RandomMinMax` → shared `System.Random`, thread-unsafe *and* game state | `RandomNumberGenerator.GetInt32`, matching the salt beside it |
| Argon2 | already safe (`Verify` is static + stackalloc) | unchanged, singleton reused |
Literal digests are pinned in a test **before** the change and still pass after it. These are compared as strings against every account database, so any casing or encoding drift would lock out every SHA and MD5 account at once.
With all three safe, the worker no longer knows which algorithm it runs and the dispatch conditions collapse to "is off-loop available".
## Correctness
- **Phrase derivation** moves to `AccountSecurity.DerivePhrase`, so verification (stored algorithm's rule) and rehash (target algorithm's rule) cannot disagree. Deriving with the wrong one is the shape of the lockout fixed in #2562.
- **Liveness** is checked at dequeue *and* at apply — a connection can drop while queued or while the result sits in the loop queue. A job with no connection attached, such as an admin password change, runs regardless.
- **Queue overflow rejects** a login rather than verifying inline; steering work back onto the loop is what a flood wants. A password change instead falls back to hashing inline, because unlike a login it must not be dropped.
- **Shutdown and crash** both just stop the thread, and pending jobs are dropped. No save is initiated once shutdown begins — saving is the operator's choice up front, via the admin gump's save/no-save variants, and `WaitForWriteCompletion` honours one already in flight — so a write applied during teardown would reach no disk. The crash path needs its own subscription because `HandleClosed` skips `InvokeShutdown` when crashed.
## Bounding
`MaxPending` is 4096 — a backstop, not a flood defense. `SentFirstPacket` holds a connection to one pending verify and the engine caps connections at 4096, so the queue is already bounded by construction and this can only trip if that invariant breaks. A cap low enough to blunt an attack would reject real players first; during a mass reconnect they *are* the queue. Flood defense belongs at the connection layer.
The real DoS improvement is elsewhere: today every attempt stalls the world, and after this a flood occupies one core while the loop keeps ticking.
## Gate
Release builds on 4+ cores. Below that there is no spare core to move work to, so off-loading buys nothing by construction; `DEBUG` is excluded because dev boxes and test shards have few logins. Both modes call the same code — the gate only chooses where it runs.
## Engine change
One property, `AccountLoginEventArgs.Deferred`, so a subscriber can say "no verdict yet". `EventSink.AccountLogin` is `Action<...>` with no continuation, and the packet handler replies in the same call. Approved separately since it touches `Projects/Server/`.
## Docs
`dev-docs/threading-model.md` and the threading skill gain a vetted-workers section. The forbidden-patterns table bans `new Thread`, `ConcurrentQueue<T>`, `Interlocked` and `volatile` in `UOContent`, and its exceptions covered only `Projects/Server/` — the existing Advanced Search fan-out already sat outside it. The new section leads with proving the need (measure on-loop time, not wall-clock; gate on core count; record the measurement), keeps game logic on the loop via chunking, and documents the hand-off protocol in both directions.
## Testing
698 UOContent tests, 810 Server tests, Release build clean.
Covered: verify and rehash outcomes, phrase rules for SHA1/SHA2 vs Argon2, stored-format stability for MD5/SHA1/SHA2, jobs with no connection attached, and dispatch ordering through the real queue. The liveness and ordering guards are mutation-verified.
## What
`Dictionary<K,V>.Remove` and `HashSet<T>.Remove` do not bump the collection's version, so removing an entry during a `foreach` does not invalidate the enumerator. A number of loops were still paying for a `PooledRefQueue`/`PooledRefList` to collect keys and drain them in a second pass. This drops those guards.
## Why it's safe
Verified against .NET 10.0.10 rather than taken on trust, since the documented guarantee covers only `Dictionary<TKey,TValue>.Remove` while several of these call sites are `HashSet<T>` or enumerate `.Keys`/`.Values`:
| Case | Result |
|---|---|
| `Dictionary` foreach + `Remove` | safe, all entries visited |
| `Dictionary.Keys` / `.Values` foreach + `Remove` | safe, all entries visited |
| `HashSet` foreach + `Remove` | safe, all entries visited |
| `Dictionary` foreach + `Remove` **then `Add`** | throws `InvalidOperationException` |
Reflection on `_version` confirms the mechanism: neither `Dictionary.Remove` nor `HashSet.Remove` touches it. Because `Remove` never bumps the version, the `Keys` and `Values` enumerators are just as safe as the dictionary's own, even though only `Dictionary.Remove` documents the behaviour. No entries were skipped in any case.
The `HashSet` half is confirmed by [stephentoub on dotnet/dotnet-api-docs#8177](https://github.com/dotnet/dotnet-api-docs/issues/8177#issuecomment-1167251052): *"Both HashSet and Dictionary have been improved to support removal during enumeration. The docs may just benefit from updating."* The gap is in the documentation, not the runtime.
`Remove` followed by `Add` in the same enumeration still throws. That is the line this PR does not cross.
## Guards removed
`VisibilityList`, `ChampionTitleSystem`, `Channel`, `BombingRun`, `Ruleset`, `PuzzleChest`, `RaceChangeGump`, `StepCache`, `PlayerMurderSystem`, `VirtueSystem`, `ProjectedItem`, `StaminaSystem`, `AIGroupMovement`, `PromotedGuard`, `AutoDenylist`, `LoginAllowlist`, `AntiMacroSystem`, `DetectHidden`.
Both collection kinds are covered: `Dictionary` (including loops over `.Keys` and `.Values`) and `HashSet` (`ProjectedItem._active`, `PlayerMurderSystem._contextTerms`, `StaminaSystem._resetHash`). In `StaminaSystem.ResetTimer` the `Count == queue.Count → Clear()` branch goes away with the queue — it only existed to avoid paying for N individual removes.
Where the collection supports it, `Contains` + `Remove` and `TryGetValue` + `Remove` also collapse into a single lookup (`if (list.Remove(x))`, `if (m_Pending.Remove(ns, out var state))`).
`Utility.Tidy<K,V>` keeps its two branches: when `K` is serializable the value is not inspected, otherwise the value is. Only the serializable side may be cast, so `Dictionary<Mobile, int>` and `Dictionary<Mobile, string>` stay valid.
## Deliberately unchanged
**`BaseCreature.LoyaltyTimer.OnTick`** keeps its deferred-delete queue. Removing from `World.Mobiles` while enumerating it is safe, but `Mobile.Delete()` is not a `Remove` — it runs `OnDelete`/`OnAfterDelete`, the `OnParentDeleted` cascade over the creature's pack, `DropHolding()`, and region and guild callbacks. Anything in that surface that constructs a `Mobile` is an `Add` into the dictionary being enumerated, which does invalidate it. `BaseHire.PayTimer.OnTick` has the same shape and is likewise untouched.
**Spatial-query buffers** — `GuardedRegion.CallGuards`, `Thunderstorm`, `Exorcism`, `LeverPuzzleController`, `BaseCreature.TeleportPets` — are a different hazard. They buffer the result of a range query because the drain moves or harms mobiles, which mutates sectors mid-enumeration.
**Re-entrant drains.** The `_users` sets in `Firebomb` and the explosion, conflagration and confusion-blast potions look like this pattern but are not: the loop collects, `Clear()`s, and only then runs `Target.Cancel` on each, which can re-enter. `AnimalTrainer` enumerates `pm.Stabled` and drains through `RemoveStabled`, which nulls the `Stabled` field once it empties — safe for an in-flight enumerator, which holds the set reference rather than the field, but subtle enough not to be worth inlining on a cold path.
## Verification
`dotnet build` clean with 0 warnings; 810 Server and 684 UOContent tests pass.
## Why
The shard owner, on a Starlink CGNAT address, was blocked by the imported reputation blocklist.
The cause was not CrowdSec. The address was a literal line in `ip-blocklist.txt`, so `BlocklistFilter` denied it at accept and then promoted it — and clearing the CrowdSec decision could not fix it either, because the file entry re-reports within `promoteSuppression` of every reconnect attempt.
This is structural, not a one-off. Reputation feeds list shared consumer address space constantly: on CGNAT one public address fronts many subscribers **at the same time**, so a single abusive customer gets the address listed and everyone else behind it is blocked with them. Where leases rotate, a listing says little about whoever holds the address now. Around 1,000 Starlink addresses sit in the current list.
So exemptions go where they cost nothing, and escalation is driven by what a connection actually does.
## Generator — `tools/Export-IpBlocklist.ps1`
`-AllowlistFile` takes multiple paths, subtracted from the merged set before the output is written. Defaults to every `ip-allowlist*.txt` beside the output, merged into one allow set:
- `ip-allowlist.txt` — operator exemptions, created once and **never rewritten**
- `ip-allowlist-<name>.txt` — a carve-out you built, regenerable and copyable between shards
**Subtraction is range-correct.** An allowlisted address inside a blocked CIDR splits that CIDR around the hole rather than being silently ignored. This also fixes `-ExcludeAnonymizers`, which parsed CIDR entries into `$anonCidr` and then only ever subtracted singles.
**No carve-out ships.** A carve-out names a real network, and which ones a shard should exempt depends on where its players actually are — so publishing one would make that policy call for every shard and put a specific provider's address space in the repo. The script builds them on request instead:
```powershell
.\Export-IpBlocklist.ps1 -AddCarveout starlink -Asn 14593
```
Carve-outs are **discovered, not configured**: every `ip-allowlist*.txt` beside the output is subtracted, by the generator and by the shard, so a file an admin adds needs no config edit and no code change. Each carries an `asn=` marker in its header, which is how `-RefreshCarveouts` rebuilds it without the script keeping a list of anyone's networks; a hand-written allowlist has no marker and is never rewritten.
Prefixes come from **announcements, not ownership records**, because registry data disagrees with what is actually routed and silently caps result sets: ARIN whois returns at most 256 rows and gives per-customer /24s, and `206.83.96.0/19` reads as APNIC in RDAP even though `206.83.96/21` is announced by Starlink.
Editing an allowlist bypasses `-MinInterval`, so a just-added exemption isn't indistinguishable from the allowlist not working. A Starlink carve-out, if you build one, costs **~4,300 IPs + ~144 CIDRs of 4.2M (0.10%)**.
## Allowlists
**`FileAllowlist`** reads the same files the generator subtracts, so an operator entry means "leave this address alone" for real. Subtraction alone only covers being *blocked*; behavioural detections never consult the blocklist, so without this a carve-out was quietly routed around — one scanner behind a shared address was enough to get everyone behind it contributed and firewalled, with nothing in the shard's own config explaining why. Reading the files also means an entry applies on the next reload rather than the next regeneration, which is what matters when someone is complaining now.
**`LoginAllowlist`** is earned by authenticating, with a 90-day TTL because an address that logged in years ago is a stranger. Its own store rather than `Account.LoginIPs`, which has no timestamps and cannot be backfilled. An entry is evidence rather than a licence: 10 suppressed contributions in an hour revokes it, and a fresh login forgives the tally.
Both are consulted **only after the blocklist has already matched**, so a normal accept pays nothing for them and the accept gate stays allowlist-free. `BanExemptions` combines them behind `BanChannel.IsExempt` and suppresses escalation only — every local defence still applies.
Two limits, both deliberate and documented in the class: `LoginAllowlist` **cannot bootstrap** (an entry is only earned by getting in, so it never repairs an existing false positive), and it is weakest on rotating CGNAT. That is why `FileAllowlist` is the fix for those, and why it is manual.
## Behavioural detection
| Reason | Trigger |
|---|---|
| `silent-connect` | Reaped after 5s having sent **zero bytes** |
| `invalid-seed` | Opened with a zero seed |
| `foreign-protocol` | Positively identified as HTTP, TLS or SSH |
**`ForeignProtocol` inverts the test.** Asking "is this a good UO client?" cannot work: `LoginEncryption.ClientDecrypt` is a byte-for-byte stream XOR, so a legitimate client with encryption enabled when the shard expects none sends a structurally perfect connection whose payload is noise. "Speaks HTTP" is safe where "unreadable" is not — however misconfigured a UO client is, it never sends `GET / HTTP/1.1`.
Nothing assumes arrival framing. TCP has no message boundaries, so a rule of the form "these bytes must arrive together" is broken by construction and drops real players on poor links. A prefix match with too few bytes to confirm waits for more. A four-byte seed can legitimately spell `GET ` (the address 71.69.84.32) or `0x16 0x03 0x0?` (22.3.x.x), so confirmation requires the request line to continue in printable ASCII or an actual ClientHello inside a plausible record — a real client's fifth byte is a packet id (`0x80`, `0x91`, `0xEF`), none of them printable, so those collisions fall through.
Everything is keyed on **bytes-received rather than elapsed time**. A connection that sent something and ran out of time is far more likely a slow link than an attack, and banning those produces the worst failure mode available: the player retries, trips the rate limiter, and compounds a bad connection into hours of being firewalled off.
## `AutoDenylist`
A short-lived local hold (15m) on behavioural detections, as `IConnectionFilter` + `IBanReporter` over one store so the engine detection sites never reach into content.
This closes the gap where a flood pays for a socket, buffer and `NetState` slot per connection while waiting for the OS bouncer — the verdicts that matter most are reachable only *after* reading bytes — and it is the entire defence on a shard running no bouncer, which is the default config. Not persisted: a holding pen that survives restarts is a ban without a ban's review.
Cost: one dictionary lookup on a usually-empty dict per accept.
## `BanReasons`
Centralises the reason slugs. `IsBehavioral` is an **opt-in** set, not "everything except manual", so a future reason escalates normally instead of silently inheriting an exemption or entering a local denylist.
This caught a real bug during review: the first cut of the exemption swallowed `manual` admin bans (`Commands.cs`, three sites in `AdminGump`) for any allowlisted address.
## Fixes found in review
- **`BanConfiguration.Settings` was null until `Configure()` ran**, while the reap path dereferences it every `Slice()`. A harness driving `NetState.Slice()` directly hit an NRE that presented as flaky because it depended on whether an earlier test had already called `Configure()` — which is why it failed on some CI platforms and not others. Now starts at the record's defaults, with idempotency tracked by a flag; this also removes the same latent NRE from the pre-existing rate-limit path.
- **`-AllowlistFile` was typed `[string]`** while documented and used as a list, so passing two paths would have collapsed them into one string.
## Layout and docs
Content network code moves out of `Misc/` into `UOContent/Network/`, one concern per folder — `AutoDenylist/`, `Blocklist/`, `CrowdSec/`, `Firewall/`, `LoginAllowlist/`, `Packets/`. **Namespaces are untouched**, so these are pure file moves (git tracks all 16 as renames).
`dev-docs/ip-bans-and-allowlists.md` documents the subsystem, leading with the operator process for unblocking a player — including the three things that look sufficient and are not: deleting the CrowdSec decision alone, editing `ip-blocklist.txt` by hand, and `cscli allowlists` alone. `.gitignore` covers the new config files.
## Testing
Build clean. **Server.Tests 810 passed**, **UOContent.Tests 637 passed**, zero warnings. This branch adds 38 tests; the rest of the delta is main's, since this is rebased on current `main`.
New coverage: TTL boundary and renewal, private-address exclusion, manual-ban-never-exempt, unopted-reason-never-exempt, strike revocation, quiet-window reset, login forgiveness, file-allowlist CIDR coverage, file-allowlist not spending the earned list's strikes, denylist expiry-on-read, cap enforcement, lapsed-entry reclaim, HTTP/TLS/SSH identification, seed-collision fall-through, and encrypted-login-is-not-foreign.
Generator verified end-to-end against live feeds: a clean run ships no carve-out, `-AddCarveout starlink -Asn 14593` fetches and collapses 213 prefixes to 115 ranges in 0.1s over 4.2M entries, `-RefreshCarveouts` rediscovers it by its `asn=` marker, a hand-written allowlist is left untouched, and deleting a carve-out drops it rather than having it rewritten. CIDR splitting verified exhaustively: a single-IP hole in a /24 leaves exactly 255 of 256 addresses blocked.
## Operator note
Existing installs are unaffected until the generator next runs, which creates `ip-allowlist.txt` and nothing else. To unblock someone: add the address to that file and delete any live CrowdSec decision — the existing ban outlives the config change. The shard picks the entry up on its next reload, so re-running the generator is optional.
A shard whose players are on CGNAT (satellite, mobile, or an ISP short on IPv4) will likely also want `-AddCarveout`; see `dev-docs/ip-bans-and-allowlists.md`.
## Also included: a latent CI failure this PR surfaced
`fix(tests): serialize test classes that rent through STArrayPool` touches a property-list test file that has nothing to do with this feature. It is here because it was failing macOS CI, and it is trivially cherry-pickable out if you would rather it went to `main` on its own — **which may be the better call, since it is failing `main` today.**
CI has since gone green with it applied.
`STArrayPool` is single-threaded by design and its bucket cache is a plain `static`, not `[ThreadStatic]`, with a check-then-act initialize in `Return()`:
```csharp
var cacheBuckets = _cacheBuckets ?? InitializeBuckets();
```
Two threads both see null, both initialize, and the loser trips `Debug.Assert(_cacheBuckets is null)`. Anything renting from it has to stay off parallel test threads — which is what the `DisableParallelization` collections are for.
- `ObjectPropertyListReentrancyTests` and `ObjectPropertyListNestedBuildTests` (added in #2555) build property lists, which rent the interpolation buffer, but were not in the sequential collection — unlike `PropertyListInvalidationDuringBuildTests` in the same file. This is a **latent failure already on `main`**; it is timing-dependent, so it shows on some platforms and not others.
- `AutoDenylistTests` (added here) has the same exposure: its cap tests reach `AutoDenylist.Sweep`, which rents a `PooledRefList` without `mt`. The blocklist tests need no marking because `BlocklistSnapshot.Build` asks for the `mt` pool explicitly.
No production change — `STArrayPool` is the right pool on the game loop, where both `Sweep` and the property list actually run.
## Deliberately not included
Waiting for a fragmented four-byte seed at `AwaitingSeed`. It looked like a bug but the disconnect is a deliberate defence: only pre-0xEF clients reach it (0xEF goes through `HandlePacket`, which already waits for its 21 bytes), and waiting converts an instant drop into a full 5s slot hold for a client sending one or two bytes, or a loris dribbling a byte every few seconds. Against a fixed 4096-entry `MaxConnections` table that trades capacity that matters for a fragmentation case a reconnect already fixes.
Kamron Batman