* refactor(ecstore): drop the client shim, import rustfs-s3-client directly
Completes the migration window opened by the rustfs-s3-client extraction (rustfs/backlog#1842 PR3): every consumer now imports the client crate directly and the crate::client shim is deleted.
- All in-crate crate::client:: paths (tier warm backends, tier core, lifecycle tier_sweeper, replication storage boundary, set_disk) now import rustfs_s3_client::* directly; crates/ecstore/src/client/mod.rs and the lib.rs mod client declaration are gone.
- The two server-side modules historically misfiled under client/ move to their real homes: object_api_utils.rs to crates/ecstore/src/object_api/ (it builds engine-side object readers/writers), and object_handlers_common.rs to crates/ecstore/src/bucket/lifecycle/ (it is the lifecycle noncurrent-version cleanup helper). The latter now routes its replication calls through the lifecycle replication_sink boundary (schedule_delete wrapper and the sink's ReplicationObjectBridge re-export), as the lifecycle guard requires.
- The ecstore public facade drops api::client: object_api_utils is exposed as api::object_api_utils, and the rustfs crate takes admin_handler_utils (AdminError) from rustfs-s3-client directly (new dependency).
- Guard updates: the migration guard no longer pins mod client in ecstore's lib.rs or the admin_handler_utils facade module (it pins the new api::object_api_utils facade instead), and the module-lint register follows object_api_utils.rs to its new path.
Verification: cargo check -p rustfs-ecstore --all-targets and -p rustfs; cargo fmt --all; tier/transition/lifecycle-focused nextest (626 passed) and the decommission/rebalance/heal families in a filtered run (603 passed; the full-suite parallel run only fails on this machine's known decommission/rebalance baseline flakes, which pass in filtered reruns and fail identically on pristine origin/main); layer/migration/s3s/logging/error-format/doc-path guard scripts all pass.
* docs(architecture): record the S3 client extraction and reword invariant 4 (#6669)
Closes the documentation step of rustfs/backlog#1842. ARCHITECTURE.md invariant 4 now states the serving-vs-consuming distinction the adversarial ruling asked for: ecstore must not serve HTTP/S3 wire types, while consuming remote S3 endpoints is a legitimate engine capability that lives in the extracted rustfs-s3-client crate. The violation note is updated from the pre-extraction snapshot (58 files, embedded client) to the current ratcheted state (shrink-only S3S_ECSTORE_FILES_BASELINE in scripts/check_s3s_footprint.sh, object_lock converted first), and the crate map gains s3-client. ecstore-module-split-plan.md gets the client-directory entry the plan was missing: a Current Shape row and a completed-extraction section describing the pure-move + shim + direct-import sequence and the re-homing of the two misfiled server-side modules.
* refactor(rustfs): carve app/object out of object_usecase.rs — shared, extract, test_support children (backlog#1841 step 1)
Mechanical move-only split of rustfs/src/app/object_usecase.rs (19.7K lines). The file body moves to rustfs/src/app/object/mod.rs, and the first self-contained slices move into children: shared.rs (cross-cutting helpers: quota admission, response checksum injection, object-lock write validation, table-catalog mutation guard, deadlock request guard, proxy passthrough utilities), extract.rs (snowball auto-extract path incl. tar/pax helpers and execute_put_object_extract), and cfg(test) test_support.rs for cross-module test scaffolding. object_usecase.rs stays as a thin pub use facade so every existing crate::app::object_usecase:: path keeps working.
No behavior change: items move verbatim; the only source edits are visibility widenings required by the new module boundaries (private -> pub(super); pub(super) -> pub(crate) for the three helpers multipart_usecase and the app gating tests import). Guard scripts that pinned rustfs/src/app/object_usecase.rs now scan the rustfs/src/app/object tree, and the table_catalog source-text guard test concatenates the split files.
* refactor(rustfs): move the GetObject read path into app/object/get.rs (backlog#1841 step 2)
Move-only continuation of the object_usecase split: cold-fill orchestration, disk-permit admission, streaming readers and resume control, stream-buffer tuning, execute_get_object / execute_get_object_attributes, the GET replication proxy helpers, and their unit tests move from app/object/mod.rs into app/object/get.rs. Items keep their original text; cross-module call sites rely on the visibility widenings introduced in step 1.
* refactor(rustfs): move the PutObject and CopyObject paths into app/object (backlog#1841 step 3)
Move-only continuation: put.rs takes the PUT body admission and timeout readers, zero-copy and eager-commit machinery, execute_put_object, and the PUT unit tests; copy.rs takes the copy namespace/lifecycle lock helpers and execute_copy_object with its tests. Two source edits beyond visibility widenings: PutObjectChecksums fields become pub(super) (read by shared::apply_trailing_checksums across the new module boundary) and one relative super::storage_api call in the copy path becomes crate::app::storage_api since super now resolves to app::object. The table_catalog source-text guard concatenates the new files.
* refactor(rustfs): finish the object_usecase split — delete, head, restore modules (backlog#1841 step 4)
Move-only completion: delete.rs takes the delete helpers, cfg(test) delete hooks, and execute_delete_object/execute_delete_objects; head.rs takes execute_head_object with the HEAD replication proxy helpers; restore.rs takes execute_restore_object. app/object/mod.rs is now just the shared import prelude, module wiring, and the DefaultObjectUsecase struct with its constructors, accessors, and the execute_select_object_content delegation; the emptied tests module is gone. The delete re-export glob is cfg(test)-gated because its only cross-module consumers are the delete test hooks.
The table_catalog source-text guard now isolates the delete entrypoints from app/object/delete.rs, and doc/comment references that pointed at rustfs/src/app/object_usecase.rs internals now point at the per-operation modules.
refactor(protos): move internode compat manifest send-site assertions into owning crates
Promotes the rolling-upgrade dual-write manifest from a test-only constant in rustfs-protos into the public rustfs_protos::compat_manifest module, moves the JSON-encoder send-site assertions into the crates that own the asserted sources (ecstore remote_disk.rs for requests, the rustfs binary node_service/disk.rs for responses), and splits the scanner Phase-0 overlap inventory so its heal- and ecstore-owned halves live in those crates. Adds a cross-crate include_str!/include! guard with fixture self-tests to scripts/check_layer_dependencies.sh so a library crate can never again read another crate's Rust source at compile time, and records the rule in docs/architecture/crate-boundaries.md.
Part of rustfs/backlog#1884.
* feat(madmin): add account and two-factor wire contract
Defines the self-service account and MFA API shapes in one place so the
console and the `rc` CLI decode identical payloads instead of each
carrying its own copy of the contract.
`AccountMutability` is part of the contract on purpose: a client needs to
know whether the server will accept a password change for this identity
before offering the control, rather than discovering it from a rejected
request.
* feat(s3-types): add IAM identity audit events
Adds `iam:Identity:CredentialChanged` and `iam:Identity:AuthChallenge`
so account and authentication activity reaches the audit pipeline in its
own namespace, the way the KMS events already do. Neither is reachable
from a bucket notification config.
Two variants for the whole surface rather than one per operation:
`mask()` gives every variant its own bit in a `u64`, and the budget is
nearly spent (63 of 64 used after this). The per-operation detail lives
in `AuditEntry::api.name` and the `iamOperation` tag, which is what a
SIEM filters on anyway. Splitting these further needs `mask()` widened
first.
* feat(iam): add two-factor authentication primitives
Implements the state machine behind TOTP enrollment and verification in
the IAM domain, so the admin handlers stay HTTP plumbing and the console
and CLI drive identical logic.
* `totp`: RFC 6238 over the workspace's existing hmac/sha1, pinned to the
published Appendix B vectors. SHA-1, 6 digits, 30s: the parameters every
mainstream authenticator app implements. Verification returns the
matched time step so the caller can burn it.
* `recovery`: ten single-use codes, 100 bits each, in a Crockford base32
alphabet without I/L/O/U. Stored as domain-separated SHA-256 digests —
a password KDF would have to run once per stored code on every attempt,
turning each guess into an attacker-controlled cost, and with uniform
100-bit input there is no dictionary for it to defend against.
* `challenge`: stateless HMAC tokens. A TTL cache would be node-local, so
a cluster without session affinity would issue on one node and verify
on another; nothing here needs replicating.
* `record`: two-phase enrollment, replay high-water mark, and lockout.
Pending enrollment never gates a login, so a mis-scanned QR cannot lock
an operator out, and re-configuring keeps the old factor working until
the new one is confirmed.
* `store`: one object per identity under `config/mfa/`, a sibling of
`config/iam/` so the IAM cache loader's startup walk does not sweep it
up. Optimistic `If-Match` writes; deliberately uncached, because a cache
would need cluster-wide invalidation to keep the replay mark and the
lockout counter honest.
* `qr`: server-side rendering, so neither client needs a QR encoder.
Enrollment is refused without `RUSTFS_IAM_MASTER_KEY`. A TOTP secret is
credential-equivalent, and one written in plaintext could be lifted off a
disk — worse than no second factor, because the user believes they have
one. IAM identities tolerate a missing master key for backward
compatibility; a new feature has no such history to honour.
Also adds `IamSys::revoke_sts_sessions_for_parent`, so a credential
rotation can invalidate the sessions minted under the old secret.
* feat(admin): add self-service account endpoints and the two-factor login gate
Adds the account surface (`/v3/account/*`), the second-factor endpoints,
the administrative reset (`/v3/user/mfa`), and `PUT
/v3/set-user-secret-key`, plus the gate on `AssumeRole`.
What the gate covers, and what it deliberately does not:
* `AssumeRole` is the only interactive login RustFS has, so it is where a
second factor can be enforced. With one enrolled it requires
`TokenCode`; without an enrollment the code path is unchanged, so
existing deployments are untouched.
* A request signed directly with a long-term access key stays ungated.
Gating it would break every script and CLI the moment a human enabled
2FA on their own account, and would add no protection: whoever holds
the secret key already has full access without presenting a code. This
is the division AWS draws; making 2FA meaningful for API access needs an
`aws:MultiFactorAuthPresent` policy condition, tracked separately.
`SerialNumber`/`TokenCode` are STS's own parameters, so an SDK or script
authenticates the same way the console does.
`caller_identity` resolves who a request acts as. The console signs with
a short-lived STS session, so "the caller" is almost never the key that
signed. It reports two separate capabilities: root cannot rotate its
secret (a process-wide `OnceLock` that also derives the internode RPC
secret) but *can* enroll a second factor — conflating the two would leave
the default deployment's console login unprotectable.
The self-service routes carry no admin action. Giving them one would be
wrong in both directions: it would stop an ordinary user from changing
their own password, and let any holder of that action change someone
else's. They gate on possession of the credential plus, for the
mutations, knowledge of the current secret — a signature only proves a
credential was used, so without that a hijacked tab could rewrite the
account's credentials or strip its second factor.
`set-user-secret-key` exists because the only prior way to change a
password was to re-POST the whole user through `add-user`, which rewrote
`status` and dropped the policy field — a password reset that silently
re-enabled a disabled account.
Wrong, replayed and malformed codes are indistinguishable on the wire;
the distinction survives only in the audit trail, where no submitted
value, secret or code is ever recorded.
* test(e2e): cover the two-factor lifecycle and its regressions
Unit tests cover the state machine at its edges; only an end-to-end test
proves the pieces are wired together and that the existing
authentication paths still behave.
Asserts, against a real server: enrollment is refused without a master
key; the full enroll/activate flow works with a genuine RFC 6238 code;
`AssumeRole` refuses without a factor and accepts a valid one; a recovery
code works exactly once; a direct SigV4 admin request keeps working with
a factor enrolled; `AssumeRole` for an unenrolled identity is unchanged;
and a password rotation invalidates the old secret.
The test computes TOTP codes itself rather than calling the server's
implementation — a shared helper could agree with a bug on both sides.
This suite caught a real defect during development: enrollment was
refused for root because its *password* is immutable, which would have
left the default deployment — an administrator signing into the console
as root — unable to protect the one login the feature exists for.
* docs(operations): document the two-factor authentication model
Records what the second factor protects and what it deliberately does
not, because several of the boundaries look like gaps until the
alternative is spelled out: why direct SigV4 access stays ungated, why
root credentials cannot be rotated at runtime, why secret keys cannot be
hashed in an S3 server, and why at-rest protection is mandatory for a
TOTP secret but optional for an IAM identity.
Also states the limitations plainly, including that GHSA-m77q-r63m-pj89
is unaffected: a holder of the root secret can still forge a session
token, 2FA claim included.
Placed alongside the other authentication and KMS security documents
rather than under a new `docs/security/`, which `.gitignore` excludes.
* fix(admin): route the new account handlers through the admin s3 facade
Two of the guardrails in the CI "Quick Checks" job rejected the previous
commits, so the required check would have gone red as soon as a maintainer
approved the workflow run.
`check_architecture_migration_rules.sh` requires everything under
`rustfs/src/admin` to reach `ECStore` through a domain module rather than
the root of `storage_api`. The MFA handler and the two `AssumeRole`
signatures now use `storage_api::runtime::ECStore`, which is where the
other ten admin handlers already take it from.
`check_s3s_footprint.sh` ratchets two counters that new code may not grow:
files referencing `s3s` and error-macro invocation lines. This branch added
four files and thirty-two lines to them. The ratchet is lower-only and its
header forbids raising a baseline to get green, so the construction moves
behind the facade instead: `storage_api::s3` now re-exports the request and
body types these handlers need and gains an `error` constructor over
`S3Error::with_message`. That is the same constructor the macro expands to
and the one `handlers/mod.rs`, `rebalance_internal_error` and
`invalid_object_lock_configuration` already call, so this is the existing
practice rather than a new one, and it keeps the `s3s` dependency in the
boundary file the s3gate migration replaces.
Every error code and message is carried over unchanged. In `sts.rs` only
the call site this branch added is converted; the sixteen that predate it
are left alone, because rewriting them would put unrelated churn in a
feature PR and push the counter below the baseline it is meant to hold.
* test(e2e): activate data usage regressions
* test(e2e): bind data usage selection to Linux listing
* test(e2e): refresh data usage Darwin selection
---------
Co-authored-by: houseme <housemecn@gmail.com>
* fix(ecstore): persist unresolved decommission entries
* fix(ecstore): type decommission completion result
* fix(ecstore): allow intentional decommission listing signatures under strict clippy
The sftp/swift feature-matrix clippy gates run with -D warnings and
flag the unresolved-entry resolver (large Err payload by design, 8
context parameters) and the decommission listing driver (9 args).
Document why and align with the existing decommission_entry precedent.
Add a read-only Prometheus report helper for backlog#2007 so the 200us vs 50us coalescer delay experiment can capture RPC, batch distribution, stage latency, and host-cost signals with one fixed evidence format.
Co-authored-by: heihutu <heihutu@gmail.com>
The merge of rustfs#6261 lost the last 64 lines of the English
translation: merging main (to pick up rustfs#6258) resolved the
conflict on the renamed file by cutting it mid-table in section 6,
which dropped section 7 (backlog/history index), section 8 (audit
method and limitations) and section 9 (landing results) that the
Chinese counterpart still carries. Restore them verbatim from the
translation commit (0e051602f) so both language versions are complete
568-line mirrors of the full 0-9 baseline, as the PR body promised.
Co-authored-by: heihutu <heihutu@gmail.com>
* docs(operations): land the heal/scanner MinIO audit baseline with closure results
Move the comprehensive heal/scanner vs MinIO analysis (2026-08-16) into
docs/operations/ so it finally enters the tree — the docs/ root is
ignored by the gitignore whitelist, which is why the baseline the audit
issue referenced as "to be merged with a PR" never landed. Append §9
closure results: all 14 backlog sub-issues (#1865-#1878) closed with the
per-item PR map, two further misjudgment corrections (HS-17 was already
implemented; HS-14's MinIO idle semantics drifted upstream), HS-12/HS-18
audit conclusions, and the registered follow-ups.
Backlog issue: rustfs/backlog#1862
Co-Authored-By: heihutu <heihutu@gmail.com>
* docs(operations): add an English counterpart of the audit baseline
Rename the Chinese analysis to *_zh.md (matching the repo's bilingual
convention of scanner-excess-alerts.md / _zh.md) and add a full English
translation at the original path, cross-linked at the top of both files.
Co-Authored-By: heihutu <heihutu@gmail.com>
---------
Signed-off-by: houseme <housemecn@gmail.com>
Co-authored-by: heihutu <heihutu@gmail.com>
docs(operations): land the heal/scanner MinIO audit baseline with closure results
Move the comprehensive heal/scanner vs MinIO analysis (2026-08-16) into
docs/operations/ so it finally enters the tree — the docs/ root is
ignored by the gitignore whitelist, which is why the baseline the audit
issue referenced as "to be merged with a PR" never landed. Append §9
closure results: all 14 backlog sub-issues (#1865-#1878) closed with the
per-item PR map, two further misjudgment corrections (HS-17 was already
implemented; HS-14's MinIO idle semantics drifted upstream), HS-12/HS-18
audit conclusions, and the registered follow-ups.
Backlog issue: rustfs/backlog#1862
Co-authored-by: heihutu <heihutu@gmail.com>
* refactor(scanner): drop the always-None single-disk default cycle hook
single_disk_default_cycle_secs returned None for every maintenance
feature combination, so the single-disk startup path already resolved
its default cycle from the speed preset (60s at 'default'). Remove the
never-wired hook and its pin tests, keep the explicit reset, and record
the decision: no special single-disk cycle override without measured
cold-start ILM latency evidence; clean-idle backoff already stretches
idle cadence (backlog#1878 HS-16).
Co-Authored-By: heihutu <heihutu@gmail.com>
* docs(operations): add heal/scanner MinIO parity decision notes
Document the HS-14/16/18 decision batch from backlog#1878: the scanner
idle throttling semantics matrix (RUSTFS_SCANNER_IDLE_MODE x speed
preset x foreground read backoff) side by side with MinIO's current
static idle_speed switch as verified against upstream master, the
migration warnings for env names and value vocabularies, the bitrot
cycle default divergence (30d vs off), the stale-multipart / tmp / trash
three-stage cleanup comparison with the crash-residue window grading,
and the single-disk default cycle decision.
Co-Authored-By: heihutu <heihutu@gmail.com>
---------
Co-authored-by: heihutu <heihutu@gmail.com>
test(kms): move the Vault KV2 Transit-wrapping doc guard into check_fips_wording.sh
`test_vault_kv2_sources_do_not_claim_transit_wrapping` asserted that four
`include_str!`-pinned files never describe the Vault KV2 backend as wrapping key
material through Vault's Transit engine. The invariant is a documentation-claim
invariant with no behavioral twin by construction, and the test form was weak in
both directions: it saw only four files (the same prose in a fifth file passed
silently) and it stopped compiling — rather than reporting a violation — as soon
as one of them was renamed.
Move the four literals verbatim into `scripts/check_fips_wording.sh`, which
already guards the adjacent cryptographic over-claim class (unsupported FIPS
validation wording) and is anchored to the same policy document. The guard now
greps every file under `crates/kms` for the same four case-sensitive literals and
separately reports a moved pinned source instead of failing to build.
`check_fips_wording.sh` previously ran only in `make pre-commit` / `pre-pr`, so
wire it into the Quick Checks job of both CI workflows to keep the invariant's
failure visibility at least as strong as the deleted test's.
Neither rustfs/src/storage/backpressure.rs nor rustfs/src/storage/lock_optimizer.rs had a production caller: their only non-self references were the pub mod lines in storage/mod.rs and a cfg(test) module, so the object transfer path never applied this backpressure and never took these lock shortcuts.
The six removed tests in concurrent_fix_test.rs duplicated tests that lived inside the deleted files; the shared primitives they shadowed keep their own coverage in rustfs-io-core.
* test(heal): add concurrency invariants for heal vs delete/overwrite races (HS-12)
Audit conclusion for backlog#1874: RustFS does not need a persistent
object-level healing marker (MinIO x-minio-healing) because every path
that can touch the same (bucket, object) commit surface serializes on
the same namespace write lock, and the heal lock guard spans the whole
rename commit including the HEAL_RENAME_INCOMPLETE partial path.
Lock the conclusion in with two race regression tests:
- heal_racing_version_delete_never_resurrects_the_deleted_version:
shard damage is injected on the doomed version so a Deep heal has real
reconstruction work while a versioned DELETE runs concurrently; the
deleted version must stay deleted and the survivor intact.
- heal_racing_unversioned_overwrites_preserves_the_last_commit:
unversioned overwrites (activating the post-commit tail that deletes
the replaced data dir without the ns lock) race a Deep heal in a
loop; the final current version must be exactly the last commit.
Also adds docs/operations/heal-concurrency-safety-notes-zh.md with the
full intersection matrix (17 intersections), lock-coverage argument,
and the residual-window classification (commit tail races are
fail-into-retry safe; bare prefix delete has zero production callers;
admin no_lock is an explicit operator opt-in).
Co-Authored-By: heihutu <heihutu@gmail.com>
* test: remove redundant heal etag clone
Co-Authored-By: heihutu <heihutu@gmail.com>
---------
Co-authored-by: heihutu <heihutu@gmail.com>
* feat(scanner): emit excess alerts as S3 notification events
The excess-versions / excess-version-size / excess-folders alerts were
metrics-and-logs only; consoles and external auditors had no way to hear
them (rustfs/backlog#1868, HS-04). MinIO emits s3:ObjectManyVersions /
s3:ObjectLargeVersions / s3:PrefixManyFolders for the same conditions —
RustFS carries those as EventName::Scanner* with s3:Scanner:* wire names
that already existed unpublished.
The three alert sites now also dispatch through the standard event
pipeline (send_event via the storage_api owner facade), carrying the
actual values and thresholds in req_params and UserAgent "Scanner".
Without a cooldown a single over-threshold object would re-emit on every
~60s scan cycle, so emissions are edge-held per (kind, bucket, object)
for 24h (RUSTFS_SCANNER_ALERT_COOLDOWN_SECS, 0 = every cycle), backed by
a process-global map with a 4096-key hard cap that clears rather than
grows. Metrics and structured logs stay level-triggered every cycle;
only the notification events are held back. A restart resets the
cooldown deliberately: one re-emission per still-hot key buys back
visibility after the restarts that accompany incident response.
Tests pin the edge-hold semantics (first fires, immediate re-check held,
independent keys, cooldown expiry re-fires, zero cooldown always emits,
hard bound) in one sequential test for the process-global map, and pin
the emitted wire names against EventName's canonical string forms so a
subscribed bucket notification can never silently stop matching.
docs/operations/scanner-excess-alerts.md documents the three events,
the metric-vs-event cadence difference, and the HS-15 threshold deltas
(alert_excess_folders 65538 vs MinIO 50000 is deliberate: Proxmox
Backup Server chunk layout compatibility).
Closesrustfs/backlog#1868.
Co-Authored-By: heihutu <heihutu@gmail.com>
* docs(operations): split scanner excess alerts into English and Chinese pages
The page shipped Chinese-only; keep it as scanner-excess-alerts_zh.md and
add a faithful English translation at the original path, cross-linked at
the top of both.
Co-Authored-By: heihutu <heihutu@gmail.com>
---------
Co-authored-by: heihutu <heihutu@gmail.com>
* fix(storage): restore multipart disk compression and make the legacy decompressor resumable
Multipart uploads have bypassed disk compression since #5169 removed the session marker as a stopgap for mid-stream GET failures. The actual root cause was never the multipart layout: the legacy DecompressReader reset its payload consumption state on every poll re-entry, so a Poll::Pending in the middle of a block payload (routine under the erasure duplex) desynchronized the block framing and surfaced as LZ4 frameType errors. This rewrites the decoder as a resumable state machine, restores the multipart session compression marker, reports logical part sizes in ListParts, and makes the rebalance migration read raw stored bytes so compressed and encrypted objects survive migration verbatim.
Fixes#5957. Internal tracking: backlog#1848, backlog#1850.
* feat(storage): stage multipart compression behind RUSTFS_COMPRESSION_MULTIPART_ENABLED
Review follow-up: a rolling-upgrade window must not create new compressed multipart objects while pre-fix nodes (whose decompressor is not resumable) may still serve reads. The session marker is now additionally gated on RUSTFS_COMPRESSION_MULTIPART_ENABLED, default off, so the restored capability stays dark until the operator confirms fleet convergence. The default flips per the multipart-compression-default-off-window entry in docs/architecture/compat-cleanup-register.md once the minimum supported direct-upgrade release ships the resumable decoder.
* chore(compat): satisfy the cleanup-register guard for the multipart compression switch
The architecture guard requires every backticked identifier in a register entry to carry a RUSTFS_COMPAT_TODO source marker: keep only the entry slug in backticks, and add the marker (with its literal Remove-after condition) at the switch definition.
* chore(rio): drop a dead store in the poison guard and note the end-block branch
Review follow-up: the poison gate re-assigned an already-true flag, and the COMPRESS_TYPE_END branch reads as dead without stating that the writer never emits an end block — that absence is exactly what lets concatenated per-part streams decode as one.
* fix(s3): report empty compressed multipart part size
* fix(s3): report empty encrypted multipart part size
* fix(site-replication): route every state RMW through the locked transaction
P1-15 PR2 (rustfs/backlog#1796, batch B2 of rustfs/backlog#1675), the
follow-up promised by rustfs/rustfs#5882.
PR1 left ~26 read-modify-write call sites on
config/site-replication/state.json in the pre-transaction shape: a
process-local mutex around load / mutate / save, each IO taking its own
object lock. Nothing held a distributed lock across the whole sequence, so
two nodes of one site still lost each other's updates, and the transitional
mutex kept the old shape available to copy.
Every remaining RMW now runs inside update_site_replication_state;
read-only sites use load_site_replication_state, whose object read comes
with the object-level read lock. SITE_REPLICATION_STATE_LOCK and its owner
helper are gone, together with their architecture-guard allowlist entry and
inventory row.
The multi-stage flows (add / edit / peer join / peer edit / remove / rotate)
keep their updated_at and pending-id CAS, but the CAS now runs inside the
transaction that writes, against the state that transaction loaded. Peer
probes, IAM work and fan-outs run between transactions and hold no lock at
all — the add no longer blocks every writer of the site across its peer join
round trips, and it re-checks the precondition right after the capability
probes so the common race is rejected before any IAM write or remote join.
When the add's commit CAS still fails, the error says the peers may already
be joined and that re-running the add reconverges. The add adopts only the
fields it computed (exhaustive destructure — adding a state field is a
compile error until classified); fields owned by writers that do not bump
updated_at keep their freshly loaded values.
Ordering of peer-edit deliveries now rests on the generation fence landed in
PR1 rather than on a guard that could never order two nodes: the add's
finalize fan-out carries the generation allocated in its commit. An accepted
peer join PRESERVES the applied-generation high-water marks — join fan-outs
are routine (adds and rotations both deliver SRPeerJoin to existing peers),
so wiping them would let stalled older edits land after any join; the
unilateral-removal rejoin misfence that a wipe would have patched is
pre-existing since the fence landed and needs an epoch in the fence instead.
The rotation handler now takes the lifecycle guard: the background
service-account reconciler runs its repair under a lifecycle try-acquire,
and its pending-rotation precheck is only sound if a rotation cannot start
mid-repair — an exclusion the removed process mutex used to provide as a
side effect.
update_site_replication_state_when_changed adds persist-or-skip so ack
markers and pending-clearing paths stop rewriting the object on a miss —
load-bearing, because the shared persist helper clears the whole object for
a ≤1-peer pending-free state — and save_site_replication_state is now
cfg(test): the pre-P1-15 shape can no longer be written in production code.
No on-disk format change.
Verification: cargo nextest run -p rustfs -E
'test(/admin::handlers::site_replication::/)' (181 passed); site-replication
dual/three-node e2e (13 passed); cargo clippy -p rustfs --all-targets -D
warnings; make pre-commit. Mutation checks: dropping the state-object lock
from the boundary reds the separate-node concurrency tests; flipping a
persist-or-skip miss to a persist reds
test_missed_pending_clear_must_not_rewrite_the_state_object. Reviewed by
three independent adversarial passes (correctness/concurrency,
security/compatibility, simplicity/test-coverage); their confirmed findings
are folded in.
* fix(site-replication): serialize peer-join admission around its IAM write
Review follow-up (overtrue): two joins accepted by the same node could
interleave as "A checks a stale snapshot and pauses reading its body, B
applies secret B and commits, A resumes, overwrites IAM with secret A, and
A's commit is refused as superseded" — the persisted state advertised B's
contract while IAM only accepted A's secret, failing every peer
control-plane call. The pre-P1-15 process mutex serialized same-node joins
end to end; removing it dropped that exclusion.
admit_peer_join now runs the staleness check, the IAM upsert and the state
commit under the lifecycle guard, with the authoritative pre-check taken
against a load under that guard BEFORE IAM changes anything. The closing
transaction still re-checks staleness: the guard is process-local (exactly
as far as the old mutex reached) and the state-object lock arbitrates joins
accepted by different nodes. The body is fully read before the guard so a
stalling sender cannot block add/remove/rotate/reconciler.
The IAM step is injected, and the gated-body regression test reproduces the
review's ordering: join A is held mid-IAM while a newer join B arrives; B
must wait at the guard, and both IAM order and the final persisted state end
on B. Mutation-verified: removing the lifecycle guard from admit_peer_join
turns the test red.
Verification: cargo nextest run -p rustfs -E
'test(/admin::handlers::site_replication::/)' (182 passed);
site-replication dual/three-node e2e (13 passed); cargo clippy -p rustfs
--all-targets -D warnings; make pre-commit.
* fix(site-replication): fence peer-join admission across nodes
Review follow-up (overtrue, round 2): the lifecycle guard only serializes
joins within one process. Node A could pass the staleness check for an
older T1, node B write secret B to IAM and commit a newer T2, and node A
then overwrite IAM with secret A while its own state commit is refused as
superseded — state advertising T2's contract while IAM only accepts A's
secret.
The admission (staleness check -> IAM upsert -> state commit) now also runs
under a distributed join-admission lock, a namespace-lock key with no
backing object, following the repair execution lock's pattern — including
its nesting of config-object locks (admission -> state), and delegating
crash safety to the lock subsystem's lease expiry instead of a hand-rolled
TTL. The staleness check runs against a load taken inside the lock, before
IAM changes anything, so a superseded join exits without touching IAM. The
closing transaction keeps its re-check for defence in depth and for
old-version nodes that do not take the admission lock during a rolling
upgrade (that mixed-version window keeps today's behavior and closes when
the upgrade completes).
admit_peer_join_across_nodes is the admission minus the process-local
lifecycle guard — exactly what a second node runs — and the new
separate-nodes regression test drives it directly with join A gated
mid-IAM: join B must wait at the distributed lock, and both the IAM write
order and the final persisted state end on B. Mutation-verified: removing
the admission lock turns the test red while the same-node test (which
drives the full admit_peer_join) stays green.
Verification: cargo nextest run -p rustfs -E
'test(/admin::handlers::site_replication::/)' (183 passed);
site-replication dual/three-node e2e (13 passed); cargo clippy -p rustfs
--all-targets -D warnings; make pre-commit.
Stream materialized GET bodies by moving the buffered Bytes once instead of wrapping the stream in an extra bytes_stream layer.
Add an operations runbook for object I/O tuning A/B sweeps.
Co-authored-by: heihutu <heihutu@gmail.com>
Zhengchao An