Blog 126: Phase 1 Core POSIX gaps -- sessions, fcntl locks, statx, rlimits, /proc

Date: 2026-03-29 Milestone: M10 Alpine Linux -- Phase 1 (Core POSIX Gaps)

Summary

Seven improvements closing fundamental POSIX gaps identified in the Alpine drop-in compatibility audit:

1. statx timestamps fixed -- returns real atime/mtime/ctime from inode
2. File creation timestamps -- ext4 files/dirs now get current time on create
3. Session tracking -- session_id in Process, proper setsid/getsid/TIOCSCTTY
4. fcntl record locks -- F_SETLK/F_GETLK/F_SETLKW with byte-range lock table
5. /proc/[pid]/cwd,root,limits -- three missing per-process proc files
6. /proc/net/tcp,udp real data -- enumerate actual smoltcp sockets
7. setrlimit with per-process storage -- rlimits stored, inherited, enforced

These collectively unblock: SSH daemonization (sessions), sqlite/database ACID (record locks), find/rsync/make (timestamps), and monitoring tools like lsof/ss/top (/proc gaps).

1. statx: real timestamps from inode

The problem

statx(2) returned hardcoded zero timestamps for all fields (atime, mtime, ctime, btime), even though the underlying inode had real values from utimes/utimensat. Also returned hardcoded 1 for nlink and 0 for uid/gid.

The fix

kernel/syscalls/statx.rs: Copy all fields from the Stat struct returned by inode.stat() into the StatxBuf:

#![allow(unused)]
fn main() {
stx_atime: StatxTimestamp { tv_sec: stat.atime.as_isize() as i64, ... },
stx_mtime: StatxTimestamp { tv_sec: stat.mtime.as_isize() as i64, ... },
stx_ctime: StatxTimestamp { tv_sec: stat.ctime.as_isize() as i64, ... },
stx_nlink: stat.nlink.as_usize() as u32,
stx_uid: stat.uid.as_u32(),
stx_gid: stat.gid.as_u32(),
stx_blocks: stat.blocks.as_isize() as u64,
}

Added as_isize(), as_usize() getters to Time, NLink, BlockCount in kevlar_vfs/src/stat.rs.

2. File creation timestamps

The problem

Ext4's create_file and create_dir initialized all timestamps to 0 (epoch 1970-01-01). ls -la showed every file created at the dawn of Unix.

The fix

After create_file/create_dir returns the new inode, the kernel syscall layer calls set_times(now, now) with the current wall clock:

• kernel/syscalls/open.rs (O_CREAT path)
• kernel/syscalls/openat.rs (O_CREAT path)
• kernel/syscalls/mkdir.rs
• kernel/syscalls/mkdirat.rs

This keeps timer dependencies in the kernel crate (ext4 service crate doesn't need to import the clock).

3. Session tracking

The problem

No session concept existed. getsid() returned the process group ID. setsid() created a new process group but never tracked the session. TIOCSCTTY was a no-op. /proc/[pid]/stat reported the PID itself for both pgrp and session fields.

The fix

Added session_id: AtomicI32 to the Process struct:

• Idle thread: session_id = 0
• Init (PID 1): session_id = 1 (session leader)
• fork/vfork/clone: inherit parent's session_id
• setsid(): sets session_id = caller's PID (becomes session leader)
• getsid(): returns actual session_id
• TIOCSCTTY: sets foreground process group to caller's group
• TIOCGSID: returns actual session_id
• /proc/[pid]/stat: fields 5 (pgrp) and 6 (session) now report real values

This unblocks getty, login, SSH daemonization, and proper job control.

4. fcntl record locks (F_SETLK/F_GETLK/F_SETLKW)

The problem

fcntl(2) only supported file descriptor operations (F_DUPFD, F_GETFD, F_SETFD, F_GETFL, F_SETFL). POSIX record locks (F_SETLK/F_GETLK/F_SETLKW) returned ENOSYS. This breaks sqlite WAL mode, postgresql, and any application using lockf().

The fix

Full byte-range record lock implementation in kernel/syscalls/fcntl.rs:

• Lock table: global HashMap<InodeKey, Vec<RecordLock>> keyed by (dev_id, inode_no)
• RecordLock: { start: u64, end: u64, l_type: i16, pid: i32 }
• F_GETLK: checks for conflicts, returns conflicting lock info or F_UNLCK
• F_SETLK: non-blocking acquire -- checks conflicts, splits/merges ranges
• F_SETLKW: returns EAGAIN (no real blocking yet, like flock)
• Conflict rules: write locks conflict with everything; read locks only conflict with write locks; same PID can overlap its own locks
• Range operations: set_lock() properly trims/splits existing locks when a new lock overlaps partial ranges
• Cleanup: release_all_record_locks(pid) called from Process::exit()

Struct flock ABI (x86_64, 32 bytes):

offset 0: l_type (i16)    -- F_RDLCK=0, F_WRLCK=1, F_UNLCK=2
offset 2: l_whence (i16)  -- SEEK_SET/SEEK_CUR/SEEK_END
offset 8: l_start (i64)
offset 16: l_len (i64)    -- 0 means to EOF
offset 24: l_pid (i32)

5. /proc/[pid]/cwd, root, limits

The problem

Tools like lsof, fuser, ps, and top read /proc/[pid]/cwd (current directory symlink), /proc/[pid]/root (root directory symlink), and /proc/[pid]/limits (resource limits). All returned ENOENT.

The fix

Added three entries to ProcPidDir::lookup() in proc_self.rs:

• cwd: symlink resolved from process.root_fs().lock().cwd_path()
• root: symlink always pointing to / (no chroot support yet)
• limits: formatted file matching Linux's /proc/[pid]/limits layout with all 16 RLIMIT_* entries

6. /proc/net/tcp,udp with real socket data

The problem

/proc/net/tcp and /proc/net/udp were static files that returned only the header line. ss, netstat, and monitoring tools saw zero sockets.

The fix

Two new dynamic file types (ProcNetTcpFile, ProcNetUdpFile) in kernel/fs/procfs/system.rs that call helper functions in kernel/net/mod.rs:

• format_proc_net_tcp(): iterates SOCKETS.lock().iter(), matches Socket::Tcp, formats local/remote endpoints as hex + TCP state code
• format_proc_net_udp(): same for Socket::Udp with listen endpoints

TCP state mapping follows Linux conventions (ESTABLISHED=01, SYN_SENT=02, ..., LISTEN=0A, CLOSING=0B).

IP addresses formatted as AABBCCDD:PORT using Ipv4Addr::octets().

7. setrlimit with per-process rlimit storage

The problem

getrlimit() returned hardcoded values. setrlimit() didn't exist. prlimit64() ignored writes. Daemons that set fd limits, stack sizes, or core dump settings had no effect.

The fix

Added rlimits: SpinLock<[[u64; 2]; 16]> to the Process struct:

• 16 resources indexed by RLIMIT_* constants, each with [cur, max]
• Defaults: STACK=8MB/INF, NOFILE=1024/4096, CORE=0/INF, rest=INF
• Inheritance: fork/vfork/clone copy parent's rlimits
• getrlimit: reads from process rlimits table
• setrlimit (syscall 160, new): writes to process rlimits table
• prlimit64: now reads old AND writes new values (was read-only)

Files changed