Files
MetaCrate/docs/inventory.md
Chili Palmer 61d6721287
All checks were successful
Native code generation / deterministic (push) Successful in 13m14s
Imaging and meshing gate / native (push) Successful in 4m13s
Native Rust workspace compile / compile (push) Successful in 4m19s
Implement native inventory manager workflows (#62)
2026-08-10 06:00:28 +00:00

72 lines
4.4 KiB
Markdown

# Inventory models, hierarchy, and cache
The native inventory layer mirrors the value and hierarchy behavior of
LibreMetaverse's `InventoryBase`, `InventoryItem`, `InventoryFolder`,
`InventoryNode`, and `Inventory` types without depending on a network
transport. Concrete item subclasses remain concrete when inserted, queried,
notified, or restored from disk; link detection is based on `AssetType`, not on
the Rust wrapper type. This preserves the C# misclassified-link behavior.
Store mutation is serialized by one inventory state lock. Parent/child and link
indexes and recursive item counts are rebuilt as one transaction, with cycle
and depth bounds. Missing parents become placeholder folders until their real
folder record arrives. Self-loops and longer parent cycles remain queryable but
are not linked into a cyclic ownership graph. Removing a folder removes its
bounded descendant subtree. Root and library roots are tracked independently,
and direct system folders can be found by `FolderType`. Sorting supports the
`ByDate`, `FoldersByName`, and `SystemFoldersToTop` flags.
Added, updated, and removed notifications clone their arguments and subscriber
list before invoking observers. No inventory, node, or subscriber lock is held
during a callback, so observers may safely query or mutate the store. A panic
in one observer is isolated by the shared event registry.
Cache files start with `INVCACHE`, a little-endian format version, owner and
root identities, and tagged concrete records. Reads are limited to 64 MiB,
one million records, one MiB per string, and a 512-level hierarchy. Invalid
magic, versions, UTF-8, enums, duplicate IDs, zero IDs, truncated records, and
trailing data are rejected before the live store is replaced. Synchronous
restore returns `-1` for compatibility; asynchronous restore returns the typed
error. Saves snapshot under the state lock, release it before I/O, write and
sync a same-directory temporary file, then replace the destination. The
replacement includes a rollback path on platforms that cannot rename over an
existing file.
The focused compatibility suites are
`inventory_store_semantics` and `misclassified_link_semantics`; native unit
tests additionally cover concrete cache round trips, permissions, corruption,
unknown versions, system-folder sorting, and callback re-entry.
## UDP inventory manager
`InventoryManager` now owns the live inventory protocol boundary. Login installs
the agent and library skeletons in the transport-free store, while validated
`UpdateCreateInventoryItem`, `BulkUpdateInventory`, `FetchInventoryReply`, and
`InventoryDescendents` replies populate or reconcile it. Create and copy callback
IDs, item fetches, task replies, and inventory offers are correlated independently;
cancelling one waiter removes only that waiter. A cleanup worker bounds abandoned
requests to one minute, and the manager rejects more than 4,096 pending operations.
Outbound create, update, move, copy, remove, give, rez/derez, script-state, and
task-inventory operations build the generated protocol packet types. Generated
multi-block packets use the packet codec's `encode_multiple` path, so large batches
are split according to the UDP transport bounds while retaining the same agent,
session, transaction, callback, owner, and permission data. Capability-backed
inventory fetch and folder-content responses are parsed before becoming visible in
the store. Asset, material, script, notecard, and thumbnail upload bodies remain a
separate asset-pipeline concern.
Legacy task inventory uses a correlated `ReplyTaskInventory` filename followed by
a `RequestXfer` transfer. The implementation confirms packets, handles duplicate
or out-of-order packet numbers deterministically, aborts cancelled transfers, and
rejects announced or accumulated data above 16 MiB. Task text parsing is capped at
100,000 entries and accepts both plain asset IDs and the reference shadow-ID XOR
encoding. The capability path applies the same byte and entry limits.
No store lock is held across an await, a packet send, or an observer callback.
Reply data is decoded and validated before store mutation, and event registries
snapshot subscribers before invoking them. Focused coverage lives in
`inventory_manager_semantics`, `task_inventory_semantics`, and the native
`inventory_manager::tests` fake-packet cases for callback isolation, bulk-copy
completion, stale folder versions, offers, capability events, and bounded Xfer.