git:20261027.697dffc to git:20261027.e88f0f3

---

name: ceph-s3

description: >-

Run Ceph RGW as an S3 endpoint under Rook (CephObjectStore,

CephObjectStoreUser, ObjectBucketClaim, COSI, CephBucketTopic, multisite

realms/zones). Core knowledge: the 19.2.6 / 20.2.4 RGW security fixes and

their fallout (SigV4 hardening rejecting presigned PUTs and RGW's own

multisite forwarding; rgw_sigv4_insecure + rgw_s3_client_max_sig_ver),

AWS SDK default checksums vs RGW (CRC64NVME only in Tentacle), bucket

index sharding and large-omap warnings, Tentacle's two-phase reshard, RGW

accounts (CephObjectStoreAccount, experimental), and Tentacle S3 behaviour

changes (LastModified truncation, tenant IAM deprecation).

when_to_use: >-

Use for RGW / S3-on-Ceph work: "radosgw", "RGW", "ceph s3", "CephObjectStore",

"CephObjectStoreUser", "ObjectBucketClaim", "OBC", "COSI", "bucket

notifications", "CephBucketTopic", "rgw multisite", "zonegroup", "realm",

"sync status", "403 AccessDenied after upgrade", "SignatureDoesNotMatch",

"presigned URL fails", "restic ceph s3 broken", "aws sdk checksum error",

"x-amz-checksum", "CRC64NVME", "InvalidRequest", "bucket index shards",

"large omap objects", "reshard", "radosgw-admin", "lifecycle not running",

"rgw accounts", "IAM on ceph". NOT for generic S3 API usage against

AWS itself; cluster-wide upgrades and CephX rotation are

rook-ceph-best-practices; RADOS-level health is ceph-troubleshooting.

argument-hint: "[security|compat|sharding|multisite|rook] (optional focus area)"

---

# ceph-s3

Focus argument → section: `security` → §19.2.6 / 20.2.4 fixes;

`compat` → §AWS SDK checksums and §Tentacle S3 behaviour changes;

`sharding` → §Bucket index and resharding; `multisite` → SigV4 item 2, §Multisite sync status and

the Multisite row of §Rook object CRDs; `rook` → §Rook object CRDs. No

argument: match the symptom to a section heading.

## 19.2.6 / 20.2.4 RGW security fixes and their fallout

| CVE | What | Fixed |

|---|---|---|

| CVE-2026-54330 | SigV4 did not check that added `x-amz-*` headers on a presigned request were signed | 19.2.6, 20.2.4 |

| CVE-2026-39944 | STS session-token fields (`is_admin`, account type) could be bit-flipped → RGW admin | 19.2.6, 20.2.4 |

| CVE-2025-30156 | CephX (cluster-wide, not RGW-only) — see rook-ceph-best-practices | 19.2.6, 20.2.4 |

The SigV4 fix breaks two things:

1. **Clients sending an unsigned `Content-Type` or `x-amz-content-sha256`**

(boto3 presigned PUTs with a content type, aws-sdk-php, minio-go — so

restic, Thanos, Loki, Mimir): bare `403 AccessDenied`; `debug_rgw` 5

logs `Signature rejected` (trackers 79674, 79803). Fixed in 19.2.7 —

upgrade Squid straight to 19.2.7. Tentacle: no fixed release after

20.2.4 yet — until one, `ceph config set client.rgw rgw_sigv4_insecure true`

(runtime, no restart). Any other unsigned `x-amz-*` header: AWS rejects

it too — fix the client, no workaround.

2. **Multisite.** Pre-fix zones sign forwarded requests improperly: set

`rgw_sigv4_insecure: "true"` in every zone **before** upgrading. If

forwarded requests (non-master → master, e.g. bucket create) still get

`403 AccessDenied`, also set `rgw_s3_client_max_sig_ver: "2"` — Squid

19.2.6 breaks forwarding even between all-19.2.6 zones (tracker 79698 →

74579); fixed in 19.2.7, and in Tentacle since 20.2.2. Under Rook, put

them in the CephObjectStore `gateway.rgwConfig` or `rook-config-override`.

Both workarounds re-open CVE-2026-54330: scope them to the affected

clients/zones and revert (`"false"` / `"-1"`) once everything runs a fixed

release.

## AWS SDK checksums

AWS SDKs released from early 2025 send checksums on every supported request

by default (`x-amz-checksum-*`, often trailing CRC32/CRC64NVME). RGW learned

CRC64NVME only in **Tentacle 20.2.0** (tracker 70040, ceph/ceph#61878); the

Squid backport (tracker 70736) is still **not shipped**.

On Squid, set the client to send checksums only when required:

`AWS_REQUEST_CHECKSUM_CALCULATION=when_required` and

`AWS_RESPONSE_CHECKSUM_VALIDATION=when_required` (env or SDK config), or

pin the SDK/tool to a version before the change.

## Bucket index and resharding

| Option | Default |

|---|---|

| `rgw_dynamic_resharding` | true |

| `rgw_max_objs_per_shard` | 100 000 |

| `rgw_max_dynamic_shards` | 1999 |

| `osd_deep_scrub_large_omap_object_key_threshold` | 200 000 keys |

- `LARGE_OMAP_OBJECTS` on an index pool = a bucket index shard over the

key threshold: `radosgw-admin bucket limit check`, then

`radosgw-admin bucket reshard --bucket <b> --num-shards <n>`. The warning

clears only after the next deep scrub of that PG.

- Squid: resharding a very large bucket blocks writes for minutes (clients

see 5xx). Tentacle's two-phase reshard does the heavy copy before

blocking. On Squid, reshard large buckets in a quiet window or pre-shard

at creation.

- Multisite: never run `radosgw-admin reshard stale-instances rm` — the

docs forbid stale-instance cleanup in multisite deployments.

## Multisite sync status

- `radosgw-admin sync status` on the non-master zone. "behind on N shards"

still needs sync; "recovery" shards hit an error and are retried (usually

transient lock contention). Persisting errors: `radosgw-admin sync error

list`.

- 403 on forwarded requests (bucket create from a secondary) is the SigV4

fallout above, not a sync fault.

- Never promote a zone to metadata master while its `sync status` shows

metadata behind: the unsynced metadata changes are lost.

## Lifecycle and GC

- Lifecycle runs only in `rgw_lifecycle_work_time` (default `00:00-06:00`)

with `rgw_lc_max_worker` 3. "Lifecycle not running" at midday is usually

the window. Check `radosgw-admin lc list`; `radosgw-admin lc process` runs

it now.

- Deleted objects' tail data waits `rgw_gc_obj_min_wait` (2 h) before GC;

pool usage drops late by design. `radosgw-admin gc list --include-all`.

## Tentacle S3 behaviour changes

- `LastModified` is truncated to the second for AWS parity — timestamps on

existing objects can appear to move backwards after the upgrade; sync

tools comparing mtimes may re-copy.

- `GetObjectAttributes` is supported.

- Tenant-level IAM APIs (CreateRole, PutRolePolicy, PutUserPolicy) and

STS Lite / `GetSessionToken` are deprecated in favour of RGW accounts.

- New EC pools default to ISA-L — relevant for EC data pools.

## Rook object CRDs

|---|---|---|

| Bucket for an app | `ObjectBucketClaim` + StorageClass | Stable, documented default |

| Kubernetes-native buckets | COSI (`CephCOSIDriver`, `BucketClaim`) | **Experimental**; the COSI controller repo has moved — take the URL from the current Rook `cosi.md` |

| Users and keys | `CephObjectStoreUser` (`spec.keys` to supply keys) | Stable; since v1.17 undeclared extra keys are purged |

| RGW accounts / IAM | `CephObjectStoreAccount`, `accountRef` on users | **Experimental — only works with `quay.ceph.io/ceph-ci/ceph:main`**, no released Ceph |

| Notifications | `CephBucketTopic` + `CephBucketNotification` | Kafka SASL via `mechanism` + secret refs; topics are owned by their creating user |

| Multisite | `CephObjectRealm` / `ZoneGroup` / `Zone` | Set the multisite SigV4 workaround above before upgrading to 19.2.6+ / 20.2.4+ |

## References

- `references/sources.md` — every source with its verification stamp.

- Sibling skills: `rook-ceph-best-practices`, `ceph-troubleshooting`,

`ceph-performance`.