Turbo NSS

Turbonss is a plugin for GNU Name Service Switch (NSS) functionality of GNU C Library (glibc). Turbonss implements lookup for user and passwd database entries (i.e. system users, groups, and group memberships). It's main goal is performance, with focus on making id(1) run as fast as possible.

Turbonss is optimized for reading. If the data changes in any way, the whole file will need to be regenerated (and tooling only supports only full generation). It was created, and best suited, for environments that have a central user & group database which then needs to be distributed to many servers/services, and the data does not change very often (e.g. hourly).

To understand more about name service switch, start with nsswitch.conf(5).

Design & constraints

To be fast, the user/group database (later: DB) has to be small (background). 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) 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 64 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:

• Maximum database size: 4GB.
• Permitted length of username and groupname: 1-32 bytes.
• Permitted length of shell and home: 1-64 bytes.
• Permitted comment ("gecos") length: 0-255 bytes.
• User name, groupname, gecos and shell must be utf8-encoded.

Checking out and building

$ git clone --recursive https://git.sr.ht/~motiejus/turbonss

Alternatively, if you forgot --recursive:

$ git submodule update --init

And run tests:

$ zig build test

Other commands will be documented as they are implemented.

This project uses git subtrac for managing dependencies. They work just like regular submodules, except all the refs of the submodules are in this repository. Repeat after me: all the submodules are in this repository. So if you have a copy of this repo, dependencies will not disappear.

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)).
• get all gids for the user (getgrouplist(3)). 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                         always 0xf09fa4b7
   4         u8     version                       now `0`
   5        u16     bom                           0x1234
             u8     num_shells                    max value: 63.
   8        u32     num_users                     number of passwd entries
  12        u32     num_groups                    number of group entries
  16        u32     offset_bdz_uid2user
  24        u32     offset_bdz_name2user
  20        u32     offset_bdz_groupname2group
  28        u32     offset_idx                    offset to the first idx_ section
  32        u32     offset_groups
  36        u32     offset_users
  40        u32     offset_groupmembers
  44        u32     offset_additional_gids

magic is 0xf09fa4b7, and version must be 0. All integers are native-endian. bom is a byte-order-mark. It must resolve to 0x1234 (4460). If that's not true, the file is consumed in a different endianness than it was created at. Turbonss files cannot be moved across different-endianness computers. If that happens, turbonss will refuse to read the file.

Offsets are indices to further sections of the file, with zero being the first block (pointing to the magic field). As all sections are aligned to 64 bytes, the offsets are always pointing to the beginning of an 64-byte "block". Therefore, all offset_* values could be u26. As u32 is easier to visualize with xxd, and the header block fits to 64 bytes anyway, we are keeping them as u32 now.

Sections whose lengths can be calculated do not have a corresponding offset_* header field. For example, bdz_gid2group comes immediately after the header, and idx_groupname2group comes after idx_gid2group, whose offset is offset_idx, and size can be calculated.

num_shells would fit to u6; however, we would need 2 bits of padding (all other fields are byte-aligned). If we instead do u2 followed by u6, the byte would look very unusual on a little-endian architecture. Therefore we will just reject the DB if the number of shells exceeds 63.

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 in the Users and Groups section relative to the beginning of the section.

const PackedGroup = packed struct {
    gid: u32,
    // index to a separate structure with a list of members. The memberlist is
    // 2^5-byte aligned (32b), this is an index there.
    members_offset: u27,
    groupname_len: u5,
    // a groupname_len-sized string
    groupname []u8;
}

pub const PackedUser = packed struct {
    uid: u32,
    gid: u32,
    additional_gids_offset: u29,
    shell_here: bool,
    shell_len_or_idx: u6,
    home_len: u6,
    name_is_a_suffix: bool,
    name_len: u5,
    gecos_len: u8,
    // pseudocode: variable-sized array that will be stored immediately after
    // this struct.
    stringdata []u8;
}

stringdata contains a few string entries:

• home.
• name (optional).
• gecos.
• shell (optional).

First byte of home is stored right after the gecos_len field, and it's 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.

Additionally, there are two "easy" optimizations:

• shells are often shared across different users, see the "Shells" section.
• name is frequently a suffix of home. For example, /home/motiejus and motiejus. 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.

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.

63 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 a list of structs which point to a location in the "blob" area:

const ShellIndex = struct {
    offset: u10,
    len: u6,
};

In the user's struct shell_here=true signifies that 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.

Variable-length integers (varints)

Varint is an efficiently encoded integer (packed for small values). Same as protocol buffer varints, 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 Username2gids store group and user memberships respectively. Membership IDs are used in their entirety — not necessitating random access, thus suitable for tight packing and varint encoding.

• For each group — a list of pointers (offsets) to User records, because getgr*_r returns pointers to membernames.
• For each user — a list of gids, because initgroups_dyn (and friends) returns an array of gids.

An entry of Groupmembers and Username2gids looks like this piece of pseudo-code:

const PackedList = struct {
    Length: varint,
    Members: [Length]varint,
}
const Groupmembers = PackedList;
const Username2gids  = PackedList;

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: 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 []PackedIntArray(u29) and are pointing to the respective Groups and Users entries (from the beginning of the respective section). Since User and Group records are 8-byte aligned, u29 is used.

Complete file structure

Each section is padded to 64 bytes.

STATUS  SECTION               SIZE             DESCRIPTION
✅      Header                48               see "Turbonss header" section
✅      bdz_gid               ?                bdz(gid)
✅      bdz_groupname         ?                bdz(groupname)
✅      bdz_uid               ?                bdz(uid)
✅      bdz_name              ?                bdz(username)
        idx_gid2group         len(group)*29/8  bdz->offset Groups
        idx_groupname2group   len(group)*29/8  bdz->offset Groups
        idx_uid2user          len(user)*29/8   bdz->offset Users
        idx_name2user         len(user)*29/8   bdz->offset Users
        idx_username2gids     len(user)*29/8   bdz->offset Username2gids
✅      ShellIndex            len(shells)*2    Shell index array
✅      ShellBlob             <= 4032          Shell data blob (max 63*64 bytes)
        Groups                ?                packed Group entries (8b padding)
✅      Users                 ?                packed User entries (8b padding)
        Groupmembers          ?                per-group memberlist (32b padding)
        Username2gids         ?                Per-user gidlist entries (8b padding)