Keep a tenant's accounts' disk quotas within its own, together (inbuxa-admin#257) #497

Merged
jcoffey-dev merged 1 commits from fix/tenant-disk-quota-total into main 2026-11-20 14:42:49 +00:00
Owner

#494 stopped any one account's disk quota going past its tenant's, but
not several adding up to more: a 5 GB tenant could still give four
accounts 5 GB each. BulTester spotted it on the issue.

  • x:Account/set (JMAP and SCIM) now refuses a disk quota that doesn't
    fit in what the tenant has left to give: its maxDiskQuota less the
    quotas of its other accounts, users and groups alike. The refusal
    names the limit, what's already given out and what's left. Accounts
    with no quota are always allowed and share what's left, which JMAP,
    IMAP and DAV already report. Only a write that grows the total is
    checked (a new quota, a raised one, an account joining the tenant),
    so lowering or clearing one always saves, even in a tenant already
    over a lowered limit (MT-18). The server administrator is held to it
    too.
  • Restoring a deleted account is checked as well. It skipped the check
    before, so a tenant could delete an account, give its space away and
    restore it. The refusal tells the restorer to lower other quotas
    first, since a kept account's quota can't be edited.
  • Moving a domain into a tenant counts the disk quotas of the accounts
    it brings (Refusal::DiskQuota, overQuota).
  • A write that could grow the total takes a per-tenant lock in the
    in-memory store (so across a cluster) before its checks and releases
    it once saved; one left behind expires after 30 seconds. Without it,
    two writes at the same moment could each fit the space left and
    together pass it. A write that can't get the lock in five seconds is
    refused with forbidden, "try again".

Checked live against a debug build (5 GB tenant): four 5 GB accounts,
one created; a batch of two 2 GB with 3 GB left, one created; raising
past what's left refused, lowering saved; no-quota account created;
the server admin refused too; 8 parallel 3 GB creates, 1 created;
delete, give the space away, restore refused; a domain whose account
has 2 GB moved into 1 GB left refused; the tenant admin can't raise its
own tenant's limit. tenant_tests gains the sum and lowering cases;
tenant, undelete, scim and legal_hold suites pass.

Follows up #494. Refs inbuxa/inbuxa-admin#257.

#494 stopped any one account's disk quota going past its tenant's, but not several adding up to more: a 5 GB tenant could still give four accounts 5 GB each. BulTester spotted it on the issue. - x:Account/set (JMAP and SCIM) now refuses a disk quota that doesn't fit in what the tenant has left to give: its maxDiskQuota less the quotas of its other accounts, users and groups alike. The refusal names the limit, what's already given out and what's left. Accounts with no quota are always allowed and share what's left, which JMAP, IMAP and DAV already report. Only a write that grows the total is checked (a new quota, a raised one, an account joining the tenant), so lowering or clearing one always saves, even in a tenant already over a lowered limit (MT-18). The server administrator is held to it too. - Restoring a deleted account is checked as well. It skipped the check before, so a tenant could delete an account, give its space away and restore it. The refusal tells the restorer to lower other quotas first, since a kept account's quota can't be edited. - Moving a domain into a tenant counts the disk quotas of the accounts it brings (Refusal::DiskQuota, overQuota). - A write that could grow the total takes a per-tenant lock in the in-memory store (so across a cluster) before its checks and releases it once saved; one left behind expires after 30 seconds. Without it, two writes at the same moment could each fit the space left and together pass it. A write that can't get the lock in five seconds is refused with forbidden, "try again". Checked live against a debug build (5 GB tenant): four 5 GB accounts, one created; a batch of two 2 GB with 3 GB left, one created; raising past what's left refused, lowering saved; no-quota account created; the server admin refused too; 8 parallel 3 GB creates, 1 created; delete, give the space away, restore refused; a domain whose account has 2 GB moved into 1 GB left refused; the tenant admin can't raise its own tenant's limit. tenant_tests gains the sum and lowering cases; tenant, undelete, scim and legal_hold suites pass. Follows up #494. Refs inbuxa/inbuxa-admin#257.
jcoffey-dev added this to the 27.1 milestone 2026-11-20 14:28:13 +00:00
jcoffey-dev added 1 commit 2026-11-20 14:28:13 +00:00
Keep a tenant's accounts' disk quotas within its own, together (inbuxa-admin#257)
ci / fork-checks (pull_request) Skipped
ci / build (pull_request) Skipped
github/ci (branch) GitHub Actions
ci / github (pull_request) Successful in 13m27s
3b6b9b699f
#494 stopped any one account's disk quota going past its tenant's, but
not several adding up to more: a 5 GB tenant could still give four
accounts 5 GB each. BulTester spotted it on the issue.

- x:Account/set (JMAP and SCIM) now refuses a disk quota that doesn't
  fit in what the tenant has left to give: its maxDiskQuota less the
  quotas of its other accounts, users and groups alike. The refusal
  names the limit, what's already given out and what's left. Accounts
  with no quota are always allowed and share what's left, which JMAP,
  IMAP and DAV already report. Only a write that grows the total is
  checked (a new quota, a raised one, an account joining the tenant),
  so lowering or clearing one always saves, even in a tenant already
  over a lowered limit (MT-18). The server administrator is held to it
  too.
- Restoring a deleted account is checked as well. It skipped the check
  before, so a tenant could delete an account, give its space away and
  restore it. The refusal tells the restorer to lower other quotas
  first, since a kept account's quota can't be edited.
- Moving a domain into a tenant counts the disk quotas of the accounts
  it brings (Refusal::DiskQuota, overQuota).
- A write that could grow the total takes a per-tenant lock in the
  in-memory store (so across a cluster) before its checks and releases
  it once saved; one left behind expires after 30 seconds. Without it,
  two writes at the same moment could each fit the space left and
  together pass it. A write that can't get the lock in five seconds is
  refused with forbidden, "try again".

Checked live against a debug build (5 GB tenant): four 5 GB accounts,
one created; a batch of two 2 GB with 3 GB left, one created; raising
past what's left refused, lowering saved; no-quota account created;
the server admin refused too; 8 parallel 3 GB creates, 1 created;
delete, give the space away, restore refused; a domain whose account
has 2 GB moved into 1 GB left refused; the tenant admin can't raise its
own tenant's limit. tenant_tests gains the sum and lowering cases;
tenant, undelete, scim and legal_hold suites pass.
jcoffey-dev merged commit 16fc3714b5 into main 2026-11-20 14:42:49 +00:00
jcoffey-dev deleted branch fix/tenant-disk-quota-total 2026-11-20 14:42:49 +00:00
Sign in to join this conversation.