فهرست منبع

Delete ACCESS_PROPOSAL.md

Toby Chui 1 هفته پیش
والد
کامیت
745d47c9af
1فایلهای تغییر یافته به همراه0 افزوده شده و 433 حذف شده
  1. 0 433
      src/mod/storage/ACCESS_PROPOSAL.md

+ 0 - 433
src/mod/storage/ACCESS_PROPOSAL.md

@@ -1,433 +0,0 @@
-# Proposal: one access list per drive
-
-Status: implemented on branch `v3.0.5`. Kept as the design record; see "Who can
-use a drive" in the root `CLAUDE.md` for how it works now, and
-[What changed during the work](#what-changed-during-the-work) for where the
-result differs from the text below.
-
-The findings below came from reading the code before the work started. Paths are
-relative to `src/`, and line numbers refer to the code as it was then.
-
-## Summary
-
-Give every drive one access list: a row per permission group, plus an
-"Everyone else" row. That one mechanism covers all four problems.
-
-| Problem | What changes |
-|---|---|
-| 1. share:/ and cluster:/ show as Removed | They are listed as built-in drives, and the server works out each drive's state instead of the page guessing it. |
-| 2. cluster:/ has no access control | It gets an access list like any other drive, kept the same on every cluster node. share:/ gets one too. |
-| 3. The editors are dated | A new Add Drive flow inside the settings page replaces the Edit window. The Bridge window goes away. |
-| 4. Other Permission does not work | It becomes the "Everyone else" row, set per drive and actually enforced. |
-
-**Can bridging go?** Yes, but Other Permission alone cannot replace it, because
-it is one value for everybody outside the group. Bridging means "this one drive,
-for this one group". The access list does both jobs, so bridging can be removed
-once the list exists. See [Can Other Permission replace bridging?](#can-other-permission-replace-bridging)
-
-**Existing installs keep working as they do now.** The default rules reproduce
-today's access, and existing bridges are converted to rows on first boot.
-
-## What the code does today
-
-### Why share:/ and cluster:/ show as Removed
-
-`poolList.html` works out a drive's state by comparing the live pool with the
-config file. A drive that is live, but has no config row and no bridge record,
-is labelled Removed and loses its buttons (lines 286-309).
-
-share:/ and cluster:/ are attached by code (`sharedDriveAttach`,
-`clusterMountDrive`), so they never have a config row. Both drives are working
-normally. Only the label is wrong.
-
-### Why cluster:/ cannot be restricted
-
-`clusterMountDrive` attaches the drive to the base pool (`cluster.go:375`), and
-every user gets the base pool as `HomeDirectories` (`mod/user/user.go:105`).
-Nothing filters a base pool drive by group.
-
-### Why Other Permission does nothing
-
-It is a label with nothing behind it:
-
-- The values are hard-coded when pools load: `readwrite` for system, `denied` for
-  groups (`storage.go:123`, `242`, `256`). No endpoint changes them and they are
-  never saved.
-- `GetHighestAccessRightStoragePool` adds `u.HomeDirectories` even for a group
-  pool (`mod/user/permissionHandler.go:166`). Every drive therefore resolves to
-  the system pool, and line 127, the only line that reads the value, is never
-  reached.
-- `GetAllFileSystemHandler` only walks the base pool and the user's own groups
-  (`mod/user/directoryHandler.go:55`). A non-member never has the drive, so the
-  value could not grant anything even if it were read.
-
-**Do not fix line 166 on its own.** Line 107 compares the pool owner, which is a
-group name, with the username. With only line 166 fixed, group members count as
-"others" on their own pool, get `denied`, and lose their group drives.
-
-### Other bugs found along the way
-
-| Where | Bug | Effect |
-|---|---|---|
-| `storage.pool.go:340`, `375` | Reloading the system pool closes cluster:/, and `LoadBaseStoragePool` only re-adds share:/ | cluster:/ disappears for all users until a restart or a volume change |
-| `storage.go:332`, `mod/filesystem/config.go:46` | Drive IDs are only checked against reserved names, and `GetAllLoadedFsh` keeps one handler per ID | Two groups can both define `data:/`, and a global lookup may serve one group the other group's drive |
-| `mod/filesystem/config.go:46` | `share` is not a reserved ID | A drive with that ID turns off Shared with me, with only a log line |
-| `storage.pool.go:98-108` | Editing skips `ValidateOption` | Edit can save a config that New would refuse |
-| `storage.pool.go:249-268` | The new ID is used to find the old row | Changing a drive's ID keeps the old row and adds a second one |
-| `storage.pool.go:570`, `poolList.html:239` | `listraw` returns the raw config and the page logs it | Network drive passwords and S3 keys reach the browser console |
-| `poolList.html:247` | Any stopped drive marks every config row as found | Offline drives vanish from the list while another drive is stopped |
-| `storage.pool.go:325` | Stop only flips an in-memory flag | A stopped drive comes back after a restart |
-| `fshedit.html:69-79`, `296` | The title elements are inside an HTML comment | New drive still says "Edit File System Handler", as in your screenshot |
-
-## Can Other Permission replace bridging?
-
-Not as it is defined. To give `staff` one `marketing` drive using only Other
-Permission, you would have to open every `marketing` drive to every user.
-
-| Need | Other Permission (per pool) | Bridge | Access list (per drive) |
-|---|---|---|---|
-| Give access to one specific group | No | Yes | Yes |
-| Share one drive of a pool, not all | No | Yes | Yes |
-| Read only for those users, read write for the owners | Yes | No | Yes |
-| Open a drive to all users | Yes | Only by bridging into system | Yes |
-| Restrict share:/ or cluster:/ | No | No | Yes |
-| Survives a pool reload without extra code | Yes | No, needs re-bridging | Yes |
-
-Your instinct is right that the two features overlap. The fix is to merge them:
-Other Permission becomes the "Everyone else" row, and each bridge becomes a group
-row.
-
-**Alternatives I rejected:**
-
-- **Special pools for share and cluster.** Pools are tied one to one with
-  permission groups, so a pool with no group does not fit, and pool-level rules
-  still cannot share a single drive.
-- **Keeping bridging and letting admins bridge cluster:/ into chosen pools.**
-  This gives no per-group read only, no "everyone" default, and keeps the
-  re-bridging logic in every reload path.
-
-## Proposed model
-
-A **drive** is what the code calls a File System Handler. Pools stay: a pool is
-still where a drive's config lives, and its group is the drive's **home group**.
-Pools just stop being the only thing that decides access.
-
-Each drive has an access list with two parts:
-
-- **Group rows:** a group and a level.
-- **Everyone else:** the level for users who match no row. This is what Other
-  Permission was meant to be.
-
-The levels are the existing constants: None (`denied`), Read (`readonly`) and
-Read & write (`readwrite`).
-
-### How a user's level is worked out
-
-The first rule that matches wins:
-
-1. `user:/` and `tmp:/`: Read & write, always.
-2. The user is in the drive's home group: Read & write.
-3. The user is in an admin group, and the drive is built-in or in the system
-   pool: Read & write.
-4. Any of the user's groups has a row: the highest of those rows.
-5. Otherwise: the Everyone else level.
-
-The result is then capped: a read-only drive, and share:/ always, never gives
-more than Read. A level of None hides the drive completely.
-
-Rule 4 taking priority over Everyone else is what lets an admin block one group,
-and it matches how owner, group and other work on Unix.
-
-### Examples
-
-| Goal | Everyone else | Group rows |
-|---|---|---|
-| cluster:/ for `family` only | None | `family`: Read & write |
-| A drive for everyone except `interns` | Read & write | `interns`: None |
-| A `marketing` drive that `staff` can read | None | `staff`: Read |
-
-### Defaults
-
-A drive with no saved rule uses these, which match today's behaviour:
-
-| Drive | Everyone else |
-|---|---|
-| System pool drive | Read & write |
-| Group pool drive | None |
-| share:/ | Read |
-| cluster:/ | Read & write |
-
-### Where rules are stored
-
-Local rules go in `system/storage.access.json`, not in `system/storage/`, because
-files in that folder are named after groups. Only drives with a changed rule
-appear:
-
-```json
-{
-  "version": 1,
-  "drives": {
-    "media": { "other": "denied", "groups": { "family": "readwrite", "guests": "readonly" } },
-    "share": { "other": "readonly", "groups": { "guests": "denied" } }
-  }
-}
-```
-
-Rules are keyed by drive ID, so IDs must be unique across all pools. New and
-edited drives are validated for this. Duplicates that already exist are covered
-in [Decisions for you](#decisions-for-you).
-
-## How share:/ and cluster:/ are exposed
-
-Both appear in a Built-in section of the page, next to `user:/` and `tmp:/`. They
-cannot be edited, stopped or removed, and the server refuses those calls too. An
-admin only sets who can use them.
-
-### share:/
-
-- The drive is read only, so the page offers just two levels: **Visible** and
-  **Hidden**.
-- Hidden removes the drive and its sidebar entry. It does not revoke any shares,
-  and share links keep opening as before.
-- The rule is local to the node.
-
-### cluster:/
-
-- **The rule is replicated.** A local file would not work: every member serves
-  the same files, so a rule set on node A would leave them open through node B.
-  It is stored with `Submit(KindSetting, ...)`, the way scheduling weights are
-  (`mod/cluster/scheduling/scheduling.go:159`). `clusterMetadata.OnChange` needs
-  to handle `KindSetting` as well as `KindVolume` so that each node refreshes.
-- **Groups match by name**, which is already the cluster's rule ("create the same
-  permission groups on every node"). The editor warns when a row names a group
-  that does not exist on this node.
-- **Jobs follow the rule** because they run as their owner on the executing node
-  (`cluster.jobs.go:43`).
-- **When the drive is not mounted** (not in a cluster, or no volumes yet), the row
-  shows "Not active" with a link to Cluster Settings. Its rule cannot be edited
-  then, because `Submit` returns `ErrNotInCluster`. With `-disable_cluster` the
-  row is hidden.
-
-## What must be fixed for rules to hold
-
-Today, a user can write to anything they can see unless the whole drive is read
-only, so several paths never check. Without these fixes, a Read or None rule would
-look active in the UI but be bypassed. All of this has to ship before the first
-rule can be set.
-
-| Path | Today | Needed |
-|---|---|---|
-| SFTP (`mod/storage/sftpserver/vroot.go`) | Writes straight to the drive, and the drive list is fixed at login | Check the level on every operation, so a rule change applies to open sessions |
-| WebDAV (`mod/storage/webdav/fshAdapter.go`) | `OpenFile`, `Mkdir`, `RemoveAll` and `Rename` have no per-user check | Same as SFTP |
-| Samba (`mod/fileservers/servers/samba/handlers.go:367`) | Shares are created when a user is activated, and the OS enforces them | Honour Read, and update or remove shares when rules change |
-| Global drive lookup (`GetFSHandlerSubpathFromVpath`, `GetFsHandlerByUUID`) | Searches every pool and relies on the caller to check | Audit all callers (details below) |
-| Cluster job datasets (`mod/cluster/jobs/mapreduce.go:102`) | Patterns expand from the metadata index with no user check, so a denied user could still get file paths | Refuse a dataset unless the user can read cluster:/ |
-| Shared with me sidebar (`mod/share/drive.go:315`) | Always available to every user | Hide the entry and return nothing when the level is None |
-| Quota (`mod/user/user.go:112`) | Each user's quota manager keeps the drive list from first use | Clear the buffer when rules change |
-
-**The lookup audit matters most.** A missing check on these lookups was harmless
-for cluster:/ and share:/ while every user could reach them. A rough scan of
-`src/*.go` found 26 functions using them, 8 with no check in the same function:
-
-- 3 are bridge code that gets deleted.
-- 2 look fine: `sharedDriveCopyDenied`, and the admin installer in `module.go`.
-- 3 are file operation helpers whose callers I checked only in part.
-
-The `mod/` packages were not scanned, so this needs a proper pass during the work.
-
-FTP, TFTP and the cluster AGI calls already use `CanRead` / `CanWrite`. AGI
-`filelib` uses them too, but I did not check every function.
-
-## Backend changes
-
-| Area | Change |
-|---|---|
-| `mod/storage/access` (new, with tests) | Rule type, the level calculation, and a file store with an in-memory copy and atomic writes |
-| `mod/user/directoryHandler.go` | `GetAllFileSystemHandler` walks every pool and keeps drives where the user's level is not None |
-| `mod/user/permissionHandler.go` | `GetPathAccessPermission` uses the new calculation; `GetHighestAccessRightStoragePool` is deleted |
-| `mod/storage/storage.go` | `OtherPermission` and `HasHigherOrEqualPermissionThan` are deleted |
-| `storage.go`, `cluster.go`, `file_system.share.go` | One `attachBuiltinDrives` step used by both boot and reload, which also fixes the lost cluster:/ |
-| `storage.pool.go` | New endpoints below; validation on edit; built-in drives refused by edit, stop and remove |
-| `mod/filesystem/config.go` | Reserve `share`; check that IDs are unique across pools |
-| `mod/permission/group.go` | Deleting a group removes its rows, so a new group with the same name does not inherit them |
-
-New endpoints, all on the admin router:
-
-| Endpoint | Purpose |
-|---|---|
-| `GET /system/storage/drives` | Every drive with its pool, kind, state and access summary, computed on the server and with no credentials. Replaces `listraw` and the per-drive `checkBridge` calls. |
-| `GET` / `POST /system/storage/access` | Read or set one drive's access list |
-| `GET /system/storage/access/explain?user=&drive=` | Shows which rule decided a user's level, like `/system/cluster/sched/explain` |
-
-Saving a drive mounts it straight away with `AttachFsHandler`, so most changes no
-longer need a pool reload. The "THIS WILL CRASH THE SYSTEM" reload moves to an
-advanced menu.
-
-**Deleted with bridging:** `storage.bridge.go`, `mod/storage/bridge/`,
-`fshBridge.html`, the `bridge` and `checkBridge` endpoints, and the re-bridge logic
-in `HandleStoragePoolReload`.
-
-## UI
-
-The page becomes a proper settings fragment using the `--ss-*` tokens and `ss-*`
-components, so it follows the theme. It also gets a locale file,
-`system/locale/disk/storagepool.json`, in the six File Manager languages. Today
-the page is hard-coded English, as your first screenshot shows. "File System
-Handler" and "FSH" disappear from the UI in favour of "drive".
-
-### Drives view
-
-```
-Manage where files are stored and who can use them.     [ Drives | Who can access ]
-                                                                     [ + Add drive ]
-BUILT-IN
-  user:/      Home folders       Online       Everyone: Read & write       (fixed)
-  tmp:/       Temporary files    Online       Everyone: Read & write       (fixed)
-  share:/     Shared with me     Online       Everyone: Visible            [Access]
-  cluster:/   Cluster            Online       family: Read & write         [Access]
-
-SYSTEM POOL
-  media:/     Media    Local     Online       Everyone: Read & write       [Access] [...]
-  test:/      test     Local     Offline      Everyone: Read & write       [Access] [...]
-
-GROUP POOL: FAMILY
-  photos:/    Photos   SMB       Online       family, guests (Read)        [Access] [...]
-```
-
-The `[...]` menu holds Edit, Stop and Remove.
-
-### Who can access view
-
-This is the one-screen answer to "which group can use which drive". Each cell is a
-dropdown, and changes save as they are made.
-
-```
-              administrator   default        family         guests     Everyone else
-share:/       Visible*        Visible        Visible        Hidden     Visible
-cluster:/     Read & write*   None           Read & write   None       None
-media:/       Read & write*   Read & write   Read & write   Read       Read & write
-photos:/      None            None           Home group*    Read       None
-
-* fixed by rule: admin access or home group
-```
-
-On a phone, the matrix becomes a group picker with a list of drives below it.
-
-### Add and edit drive
-
-This opens as a panel inside the page instead of a float window, in three steps:
-
-1. **What are you adding?** A folder on this server, a physical disk, or a network
-   drive.
-2. **Details.** Only the fields for that choice.
-3. **Who can use it?** The same access editor as above.
-
-| Today | Proposed |
-|---|---|
-| Every field shown at once, with type defaulting to NTFS | Step 1 decides which fields appear; a plain folder needs no filesystem type |
-| UUID typed by hand, editable later | "Drive ID" filled in from the name, shown as `photos:/`, locked when editing |
-| Hierarchy: Isolated User Folders / Public Access Folders | "Folder layout": one shared folder, or a private folder per user, each with a one-line explanation |
-| Access Permission dropdown | "Read only drive" switch |
-| Mount device, partition UUID, mount point and automount always visible | Physical disk only; Browse fills them in |
-| Confirm, then a Reload Now / Later page | Save mounts the drive straight away |
-| `alert()` errors | Errors next to the field, plus a Test button for paths and network logins |
-| Fixed white float windows | Follows the theme and language of System Settings |
-
-## Migration
-
-No rule file means defaults, so only bridges need converting. On first boot, for
-each record in `system/bridge.json`:
-
-- Write a row giving `SPOwner` Read & write on that drive, which is the access it
-  has today.
-- If `SPOwner` is `system`, set Everyone else to Read & write instead, because
-  bridging into the system pool gave the drive to every user.
-- Keep rows for drives that are offline at that moment.
-
-The file is then renamed to `system/bridge.json.migrated`, which keeps a backup
-and stops the step running twice.
-
-## Delivery order
-
-Each step can ship on its own:
-
-1. **Built-in drives show correctly (problem 1).** The `drives` endpoint, no more
-   Removed label, reload keeps cluster:/, ID validation, and no credentials in the
-   browser. Access does not change.
-2. **Access engine and enforcement.** The new calculation and every fix in
-   [What must be fixed](#what-must-be-fixed-for-rules-to-hold). Users see no
-   change, because no rules can be set yet.
-3. **Access UI and cluster replication (problems 2 and 4).** Both views and the
-   replicated cluster rule.
-4. **Remove bridging.** Migration, then delete the code and the window.
-5. **New Add Drive flow (problem 3).** Independent of steps 2 to 4, so it can
-   happen at any point after step 1.
-
-**Testing:**
-
-- Table-driven tests for the level calculation.
-- A test proving the defaults give the same result as today across a matrix of
-  users and drives. This is the main safety net for existing installs.
-- Denied write tests for SFTP and WebDAV.
-- A migration test, including a bridge into the system pool.
-- A live two-node test: restrict cluster:/ on one node and confirm the other node
-  denies access.
-- A pass through the real UI on a scratch instance.
-
-## Decisions for you
-
-1. **Can a group row say None to block a group?** Recommended: yes. Without it,
-   "everyone except interns" cannot be expressed. The alternative is grants only,
-   which is slightly simpler.
-   => YES
-2. **Should admins always keep access to built-in and system drives?** Recommended:
-   yes, so an admin cannot lock themselves out of cluster:/ in the File Manager.
-   Group pool drives stay hidden from admins who are not members, as they are now.
-   => YES
-3. **Should a share still reach someone denied the drive it came from?** For
-   example, Alice shares `cluster:/report.pdf` with Bob, who has None on cluster:/.
-   Recommended: yes, because a share is a deliberate exception for one file. If
-   Alice later loses access to the drive, her shares stop working
-   (`mod/share/drive.go:257`). A per-drive "no sharing" switch could come later.
-   => YES
-4. **Should cluster:/ have one rule for the whole cluster?** Recommended: yes.
-   Per-node rules would let the same file be allowed on one node and denied on
-   another.
-   => YES
-5. **What should happen to duplicate drive IDs that already exist?** Recommended:
-   log a warning and flag them on the page, but keep them loaded. Changing an ID
-   automatically would break share links and saved paths. Until the admin renames
-   one, a rule on that ID applies to both drives.
-   => SOUNDS GOOD
-
-## What changed during the work
-
-- **More write paths needed a check.** The table in
-  [What must be fixed](#what-must-be-fixed-for-rules-to-hold) missed these, which
-  wrote to a user's drive without `CanWrite`: AGI `ffmpeg`, `imagelib` resize,
-  crop and RAW conversion, and SQLite (a user who may only read now gets a read
-  only connection). Emptying the trash also skips drives the user may only read.
-- **`/system/storage/pool/list` went too.** Like `listraw` it sent stored logins
-  to the browser, through each drive's start options.
-- **SFTP reads** opened files read write, which a Read level would have refused,
-  so reading now opens them read only.
-- **"Everyone else" is called "Default access" in the UI**, and a group that
-  follows it says "Use default". It sits in the first column of the matrix: the
-  groups read as exceptions to it, and it stays in view when there are more
-  groups than fit. In the rule file the level is still the `other` field.
-- **`clusterMetadata.OnChange` is not what keeps nodes up to date.** Each access
-  check reads the replicated value, so a rule applies as soon as it arrives. The
-  hook only clears the quota managers and updates Samba.
-- **Samba** needed `read list` rather than `read only`, since one share can have
-  users at both levels. `RemoveSambaShareConfig` also had `/etc/samba/smb.conf`
-  hard-coded and now uses the configured path.
-- **Stop** still lasts until the next restart or reload. The menu says so.
-
-## Out of scope
-
-- Rules on individual folders inside a drive.
-- Rows for individual users rather than groups.
-- Private per-user folders on cluster:/.
-- Uploading into editable shares from share:/.
-- Encrypting network drive passwords, which sit in plain text in the storage
-  config files.