turbonss/docs/architecture.md

Design & constraints
--------------------

To be fast, the user/group database (later: DB) has to be small
([background][data-oriented-design]). It encodes user & group information in a
way that minimizes the DB size, and reduces jumping across the DB ("chasing
pointers and thrashing CPU cache").

To understand how this is done efficiently, let's analyze the
[`getpwnam_r(3)`][getpwnam_r] in high level. This API call accepts a username
and returns the following user information:

```
struct passwd {
   char   *pw_name;       /* username */
   char   *pw_passwd;     /* user password */
   uid_t   pw_uid;        /* user ID */
   gid_t   pw_gid;        /* group ID */
   char   *pw_gecos;      /* user information */
   char   *pw_dir;        /* home directory */
   char   *pw_shell;      /* shell program */
};
```

Turbonss, among others, implements this call, and takes the following steps to
resolve a username to a `struct passwd*`:

- Open the DB (using `mmap`) and interpret it's first 128 bytes as a `*struct
  Header`. The header stores offsets to the sections of the file. This needs to
  be done once, when the NSS library is loaded.
- Hash the username using a perfect hash function. Perfect hash function
  returns a number `n ∈ [0,N-1]`, where N is the total number of users.
- Jump to the `n`'th location in the `idx_name2user` section, which contains
  the index `i` to the user's information.
- Jump to the location `i` of section `Users`, which stores the full user
  information.
- Decode the user information (which is all in a continuous memory block) and
  return it to the caller.

In total, that's one hash for the username (~150ns), two pointer jumps within
the group file (to sections `idx_name2user` and `Users`), and, now that the
user record is found, `memcpy` for each field.

The turbonss DB file is be `mmap`-ed, making it simple to jump across the file
using pointer arithmetic. This also reduces memory usage, as the mmap'ed
regions are shared. Turbonss reads do not consume any heap space.

Tight packing places some constraints on the underlying data:

- Permitted length of username and groupname: 1-32 bytes.
- Permitted length of shell and home: 1-256 bytes.
- Permitted comment ("gecos") length: 0-255 bytes.
- User name, groupname, gecos and shell must be utf8-encoded.
- User and Groups sections are up to 2^35B (~34GB) large. Assuming an "average"
  user record takes 50 bytes, this section would fit ~660M users. The
  worst-case upper bound is left as an exercise to the reader.

Sorting is stable. In v0:
- Groups are sorted by gid, ascending.
- Users are sorted by their name, ascending by the unicode codepoints
  (locale-independent).

remarks on `id(1)`
------------------

A known implementation runs id(1) at ~250 rps sequentially on ~20k users and
~10k groups. Our rps target is much higher.

To better reason about the trade-offs, it is useful to understand how `id(1)`
is implemented, in rough terms:
- lookup user by name ([`getpwent_r(3)`][getpwent]).
- get all gids for the user ([`getgrouplist(3)`][getgrouplist]). Note: it is
  actually using `initgroups_dyn`, accepts a uid, and is very poorly
  documented.
- for each additional gid, get the `struct group*`
  ([`getgrgid_r(3)`][getgrgid_r]).

Assuming a member is in ~100 groups on average, to reach 10k id/s translates to
1M group lookups per second. We need to convert gid to a group index, and group
index to a group gid/name quickly.

Caveat: `struct group` contains an array of pointers to names of group members
(`char **gr_mem`). However, `id` does not use that information, resulting in
read amplification, sometimes by 10-100x. Therefore, if `argv[0] == "id"`, our
implementation of [`getgrid_r(3)`][getgrid] returns the `struct group*` without
the members. This speeds up `id` by about 10x on a known NSS implementation.

Relatedly, because [`getgrid_r(3)`][getgrid] does not need the group members,
the group members are stored in a different DB section, reducing the `Groups`
section and making more of it fit the CPU caches.

Turbonss header
---------------

The turbonss header looks like this:

```
OFFSET     TYPE     NAME                      DESCRIPTION
   0      [4]u8     magic                     f0 9f a4 b7
   4         u8     version                   0
   5         u8     endian                    0 for little, 1 for big
   6         u8     nblocks_shell_blob        max value: 63
   7         u8     num_shells                max value: 63
   8        u32     num_groups                number of group entries
  12        u32     num_users                 number of passwd entries
  16        u32     nblocks_bdz_gid           bdz_gid section block count
  20        u32     nblocks_bdz_groupname
  24        u32     nblocks_bdz_uid
  28        u32     nblocks_bdz_username
  32        u64     nblocks_groups
  40        u64     nblocks_users
  48        u64     nblocks_groupmembers
  56        u64     nblocks_additional_gids
  64        u64     getgr_bufsize
  72        u64     getpw_bufsize
  80     [48]u8     padding
```

`magic` is 0xf09fa4b7, and `version` must be `0`. All integers are
native-endian. `nblocks_*` is the count of blocks of a particular section; this
helps calculate the offsets to all sections.

Some numbers, like `nblocks_shell_blob`, `num_shells`, would fit to smaller
number of bytes. However, interpreting `[2]u6` with `xxd(1)` is harder than
interpreting `[2]u8`. Therefore we are using the space we have to make these
integers byte-wide.

`getgr_bufsize` and `getpw_bufsize` is a hint for the caller of `getgr*` and
`getpw*`-family calls. This is the recommended size of the buffer, so the
caller does not receive `ENOMEM`.

Primitive types
---------------

`User` and `Group` entries are sorted by the order they were received in the
input file. All entries are aligned to 8 bytes. All `User` and `Group` entries
are referred by their byte offset (shifted by 3 bits due to the 8-byte
alignment) in the `Users` and `Groups` section relative to the beginning of the
section.

```
const PackedGroup = packed struct {
    gid: u32,
    padding: u3,
    groupname_len: u5,
}
```

PackedGroup is followed by the group name (of length `groupname_len`), followed
by a varint-compressed offset to the groupmembers section, followed by padding
upto 8 bytes.

PackedUser is a bit more involved:

```
pub const PackedUser = packed struct {
    uid: u32,
    gid: u32,
    shell_len_or_idx: u8,
    shell_here: bool,
    name_is_a_suffix: bool,
    home_len: u6,
    name_len: u5,
    gecos_len: u11,
}
```

... followed by `userdata: []u8`:
- home.
- name (optional).
- gecos.
- shell (optional).
- `additional_gids_offset`: varint.

First byte of home is stored right after the `gecos_len` field, and its length
is `home_len`. The same logic applies to all the `stringdata` fields: there is
a way to calculate their relative position from the length of the fields before
them.

PackedUser employs two data-oriented compression techniques:
- shells are often shared across different users, see the "Shells" section.
- `name` is frequently a suffix of `home`. For example, `home=/home/vidmantas`
  and `name=vidmantas`. In this case storing both name and home is wasteful.
  Therefore name has two options:
  1. `name_is_a_suffix=true`: name is a suffix of the home dir. Then `name`
  starts at the `home_len - name_len`'th byte of `home`, and ends at the same
  place as `home`.
  2. `name_is_a_suffix=false`: name begins one byte after home, and it's length
  is `name_len`.

The last field `additional_gids_offset: varint` points to the `additional_gids`
section for this user.

Shells
------

Normally there is a limited number of separate shells even in huge user
databases. A few examples: `/bin/bash`, `/usr/bin/nologin`, `/bin/zsh` among
others. Therefore, "shells" have an optimization: they can be pointed by in the
external list, or, if they are unique to the user, reside among the user's
data.

255 most popular shells (i.e. referred to by at least two User entries) are
stored externally in "Shells" area. The less popular ones are stored with
userdata.

Shells section consists of two sub-sections: the index and the blob. The index
is an array of offsets: the i'th shell starts at `offsets[i]` byte, and ends at
`offsets[i+1]` byte. If there is at least one shell in the shell section, the
index contains a sentinel index as the last element, which signifies the position
of the last byte of the shell blob.

`shell_here=true` in the User struct means the shell is stored with userdata,
and it's length is `shell_len_or_idx`. `shell_here=false` means it is stored in
the `Shells` section, and it's index is `shell_len_or_idx` (and the actual
string start and end offsets are resolved as described in the paragraph above).

Variable-length integers (varints)
----------------------------------

Varint is an efficiently encoded integer (packed for small values). Same as
[protocol buffer varints][varint], except the largest possible value is `u64`.
They compress integers well. Varints are stored for group memberships.

Group memberships
-----------------

There are two group memberships at play:

1. Given a group (gid/name), resolve the members' names (e.g. `getgrgid`).
2. Given a username, resolve user's group gids (for `initgroups(3)`).

When group's memberships are resolved in (1), the same call also requires other
group information: gid and group name. Therefore it makes sense to store a
pointer to the group members in the group information itself. However, the
memberships are not *always* necessary (see remarks about `id(1)`), therefore
the memberships will be stored separately, outside of the groups section.

Similarly, when user's groups are resolved in (2), they are not always necessary
(i.e. not part of `struct user*`), therefore the memberships themselves are
stored out of bound.

`groupmembers` and `additional_gids` store group and user memberships
respectively. Membership IDs are packed — not necessitating random access, thus
suitable for compression.

- `groupmembers` consists of a number X followed by a list of offsets to User
  records, because `getgr*` returns pointers to membernames, thus a name has to
  be immediately resolvable.
- `additional_gids` is a list of gids, because `initgroups_dyn` (and friends)
  returns an array of gids.

Each entry of `groupmembers` and `additional_gids` starts with a varint N,
which is the number of upcoming elements. Then N delta-compressed varints,
which are:

- **additional_gids** a list of gids.
- **groupmembers** byte-offsets to the User records in the `users` section.

Indices
-------

Now that we've sketched the implementation of `id(3)`, it's clearer to
understand which operations need to be fast; in order of importance:

1. lookup gid -> group info (this is on hot path in id) without members.
2. lookup username -> user's groups.
3. lookup uid -> user.
4. lookup groupname -> group.
5. lookup username -> user.

These indices can use perfect hashing like [bdz from cmph][cmph]: a perfect
hash hashes a list of bytes to a sequential list of integers. Perfect hashing
algorithms require some space, and take some time to calculate ("hashing
duration"). I've tested BDZ, which hashes `[][]u8` to a sequential list of
integers (not preserving order) and CHM, preserves order. BDZ accepts an
optional argument `3 <= b <= 10`.

* BDZ algorithm requires (b=3, 900KB, b=7, 338KB, b=10, 306KB) for 1M values.
* Latency to resolve 1M keys: (170ms, 180ms, 230ms, respectively).
* Packed vs non-packed latency differences are not meaningful.

CHM retains order, however, 1M keys weigh 8MB. 10k keys are ~20x larger with
CHM than with BDZ, eliminating the benefit of preserved ordering: we can just
have a separate index.

None of the tested perfect hashing algorithms makes the distinction between
existing (in the initial dictionary) and new keys. In other words, HASH(value)
will be pointing to a number `n ∈ [0,N-1]`, regardless whether the value was in
the initial dictionary. Therefore one must always confirm, after calculating
the hash, that the key matches what's been hashed.

`idx_*` sections are of type `[]u32` and are pointing from `hash(key)` to the
respective `Groups` and `Users` entries (from the beginning of the respective
section). Since User and Group records are 8-byte aligned, the actual offset to
the record is acquired by right-shifting this value by 3 bits.

Database file structure
-----------------------

Each section is padded to 64 bytes.

```
SECTION               SIZE             DESCRIPTION
header                128              see "Turbonss header" section
bdz_gid               ?                bdz(gid)
bdz_groupname         ?                bdz(groupname)
bdz_uid               ?                bdz(uid)
bdz_username          ?                bdz(username)
idx_gid2group         len(group)*4     bdz->offset Groups
idx_groupname2group   len(group)*4     bdz->offset Groups
idx_uid2user          len(user)*4      bdz->offset Users
idx_name2user         len(user)*4      bdz->offset Users
shell_index           len(shells)*2    shell index array
shell_blob            <= 65280         shell data blob (max 255*256 bytes)
groups                ?                packed Group entries (8b padding)
users                 ?                packed User entries (8b padding)
groupmembers          ?                per-group delta varint memberlist (no padding)
additional_gids       ?                per-user delta varint gidlist (no padding)
```

[cmph]: http://cmph.sourceforge.net/
[id]: https://linux.die.net/man/1/id
[data-oriented-design]: https://media.handmade-seattle.com/practical-data-oriented-design/
[getpwnam_r]: https://linux.die.net/man/3/getpwnam_r
[varint]: https://developers.google.com/protocol-buffers/docs/encoding#varints
[getpwent]: https://www.man7.org/linux/man-pages/man3/getpwent_r.3.html
[getgrouplist]: https://www.man7.org/linux/man-pages/man3/getgrouplist.3.html
[getgrid]: https://www.man7.org/linux/man-pages/man3/getgrid_r.3.html