[libos] Add support for SHM

src/libos/src/ipc/mod.rs

@@ -0,0 +1,7 @@
+use super::*;
+
+mod shm;
+mod syscalls;
+
+pub use self::shm::{key_t, shmids_t, SHM_MANAGER};
+pub use self::syscalls::{do_shmat, do_shmctl, do_shmdt, do_shmget};

src/libos/src/ipc/shm.rs

@@ -0,0 +1,509 @@
+use super::*;
+
+use crate::fs::FileMode;
+use crate::process::{do_getegid, do_geteuid, gid_t, uid_t, ThreadRef};
+use crate::time::{do_gettimeofday, time_t};
+use crate::vm::{
+    ChunkRef, VMInitializer, VMMapOptionsBuilder, VMPerms, VMRange, USER_SPACE_VM_MANAGER,
+};
+use std::collections::{HashMap, HashSet};
+
+#[allow(non_camel_case_types)]
+pub type key_t = u32;
+pub type ShmId = u32;
+pub type CmdId = u32;
+
+// min shared seg size (bytes)
+const SHMMIN: usize = 1;
+// max shared seg size (bytes)
+const SHMMAX: usize = (usize::MAX - (1_usize << 24));
+// max num of segs system wide,
+// also indicates the max shmid - 1 in Occlum
+const SHMMNI: ShmId = 4096;
+
+const IPC_PRIVATE: key_t = 0;
+
+// For cmd in shmctl()
+const IPC_RMID: CmdId = 0;
+const IPC_SET: CmdId = 1;
+const IPC_STAT: CmdId = 2;
+const IPC_INFO: CmdId = 3;
+const SHM_LOCK: CmdId = 11;
+const SHM_UNLOCK: CmdId = 12;
+const SHM_STAT: CmdId = 13;
+const SHM_INFO: CmdId = 14;
+const SHM_STAT_ANY: CmdId = 15;
+
+#[allow(non_camel_case_types)]
+#[derive(Debug)]
+#[repr(C)]
+struct ipc_perm_t {
+    key: key_t,
+    uid: uid_t,
+    gid: gid_t,
+    cuid: uid_t,
+    cgid: gid_t,
+    mode: u16,
+    pad1: u16,
+    seq: u16,
+    pad2: u16,
+    unused1: u64,
+    unused2: u64,
+}
+
+#[allow(non_camel_case_types)]
+#[derive(Debug)]
+#[repr(C)]
+pub struct shmids_t {
+    shm_perm: ipc_perm_t,
+    shm_segsz: size_t,
+    shm_atime: time_t,
+    shm_dtime: time_t,
+    shm_ctime: time_t,
+    shm_cpid: pid_t,
+    shm_lpid: pid_t,
+    shm_nattach: u64,
+    unused1: u64,
+    unused2: u64,
+}
+
+// shared memory segment status
+bitflags! {
+    struct ShmStatus : u16 {
+        // segment will be destroyed on last detach
+        const SHM_DEST = 0o1000;
+        // segment will not be swapped, unused now
+        const SHM_LOCKED = 0o2000;
+    }
+}
+
+bitflags! {
+    pub struct ShmFlags: u32 {
+        const IPC_CREAT = 0o1000;
+        const IPC_EXCL = 0o2000;
+
+        /// read by owner
+        const S_IRUSR = FileMode::S_IRUSR.bits() as u32;
+        /// write by owner
+        const S_IWUSR = FileMode::S_IWUSR.bits() as u32;
+        /// execute/search by owner
+        const S_IXUSR = FileMode::S_IXUSR.bits() as u32;
+        /// read by group
+        const S_IRGRP = FileMode::S_IRGRP.bits() as u32;
+        /// write by group
+        const S_IWGRP = FileMode::S_IWGRP.bits() as u32;
+        /// execute/search by group
+        const S_IXGRP = FileMode::S_IXGRP.bits() as u32;
+        /// read by others
+        const S_IROTH = FileMode::S_IROTH.bits() as u32;
+        /// write by others
+        const S_IWOTH = FileMode::S_IWOTH.bits() as u32;
+        /// execute/search by others
+        const S_IXOTH = FileMode::S_IXOTH.bits() as u32;
+    }
+}
+
+impl ShmFlags {
+    fn to_file_mode(&self) -> FileMode {
+        let mut shmflgs = *self;
+        shmflgs.remove(ShmFlags::IPC_CREAT);
+        shmflgs.remove(ShmFlags::IPC_EXCL);
+        let file_mode = FileMode::from_bits(shmflgs.bits as u16 & FileMode::all().bits()).unwrap();
+        file_mode
+    }
+}
+
+#[derive(Debug)]
+struct ShmSegment {
+    shmid: ShmId,
+    key: key_t,
+
+    uid: uid_t,
+    gid: gid_t,
+    cuid: uid_t,
+    cgid: gid_t,
+    mode: FileMode,
+    status: ShmStatus,
+
+    shm_atime: time_t,
+    shm_dtime: time_t,
+    shm_ctime: time_t,
+
+    shm_cpid: pid_t,
+    shm_lpid: pid_t,
+
+    shm_nattach: u64,
+    chunk: ChunkRef,
+    process_set: HashSet<pid_t>,
+}
+
+impl ShmSegment {
+    fn new(shmid: ShmId, key: key_t, size: size_t, mode: FileMode) -> Result<Self> {
+        let vm_option = VMMapOptionsBuilder::default()
+            .size(size)
+            // Currently, Occlum only support shared memory segment with rw permission
+            .perms(VMPerms::READ | VMPerms::WRITE)
+            .initializer(VMInitializer::FillZeros())
+            .build()?;
+        let chunk = USER_SPACE_VM_MANAGER.internal().mmap_chunk(&vm_option)?;
+
+        Ok(ShmSegment {
+            shmid: shmid,
+            key: key,
+            uid: do_geteuid().unwrap() as u32,
+            cuid: do_geteuid().unwrap() as u32,
+            gid: do_getegid().unwrap() as u32,
+            cgid: do_getegid().unwrap() as u32,
+            mode: mode,
+            status: ShmStatus::empty(),
+            shm_atime: 0,
+            shm_dtime: 0,
+            shm_ctime: ShmManager::current_time(),
+            shm_cpid: current!().process().pid(),
+            shm_lpid: 0,
+            shm_nattach: 0,
+            chunk: chunk,
+            process_set: HashSet::new(),
+        })
+    }
+
+    fn check_perm(&self) -> Result<()> {
+        // TODO: Add permission control
+        Ok(())
+    }
+
+    fn set_destruction(&mut self) {
+        self.status.insert(ShmStatus::SHM_DEST)
+    }
+
+    fn is_destruction(&self) -> bool {
+        self.status.contains(ShmStatus::SHM_DEST)
+    }
+
+    fn shm_start(&self) -> usize {
+        self.chunk.range().start()
+    }
+
+    fn shm_size(&self) -> usize {
+        self.chunk.range().size()
+    }
+
+    fn shm_add_pid(&mut self, pid: &pid_t) -> Result<()> {
+        if self.process_set.contains(pid) {
+            return_errno!(EINVAL, "this pid has been attached to the shm");
+        }
+        self.process_set.insert(*pid);
+        Ok(())
+    }
+
+    fn shm_remove_pid(&mut self, pid: &pid_t) -> Result<()> {
+        if !self.process_set.contains(pid) {
+            return_errno!(EINVAL, " this pid has not been attached to the shm");
+        }
+        self.process_set.remove(&pid);
+        Ok(())
+    }
+}
+
+impl Drop for ShmSegment {
+    fn drop(&mut self) {
+        debug!("drop shm: {:?}", self);
+        assert!(self.shm_nattach == 0);
+        assert!(self.process_set.len() == 0);
+        USER_SPACE_VM_MANAGER
+            .internal()
+            .munmap_chunk(&self.chunk, None);
+    }
+}
+
+#[derive(Debug)]
+struct ShmIdManager {
+    used_id: HashSet<ShmId>,
+    free_num: u32,
+    last_alloc_id: ShmId,
+}
+
+impl ShmIdManager {
+    fn new() -> Self {
+        let used_id = HashSet::new();
+        let free_num = SHMMNI as u32;
+        let last_alloc_id = SHMMNI - 1;
+        ShmIdManager {
+            used_id,
+            free_num,
+            last_alloc_id,
+        }
+    }
+
+    // Always return next free id for shmid
+    fn get_new_shmid(&mut self) -> Result<ShmId> {
+        if self.free_num == 0 {
+            return_errno!(ENOSPC, "all possible shared memory IDs have been taken");
+        } else {
+            self.free_num -= 1;
+        }
+        let mut id = self.last_alloc_id + 1;
+        loop {
+            if id == SHMMNI {
+                id = 0;
+            }
+            if !self.used_id.contains(&id) {
+                break;
+            }
+            id += 1;
+        }
+        self.last_alloc_id = id;
+        Ok(id)
+    }
+
+    fn free_shmid(&mut self, shmid: &ShmId) -> Result<()> {
+        self.free_num += 1;
+        self.used_id.remove(shmid);
+        Ok(())
+    }
+}
+
+lazy_static! {
+    pub static ref SHM_MANAGER: ShmManager = ShmManager::new();
+}
+
+#[derive(Debug)]
+pub struct ShmManager {
+    shm_segments: RwLock<HashMap<ShmId, ShmSegment>>,
+    shmid_manager: RwLock<ShmIdManager>,
+}
+
+impl ShmManager {
+    fn new() -> Self {
+        ShmManager {
+            shm_segments: RwLock::new(HashMap::new()),
+            shmid_manager: RwLock::new(ShmIdManager::new()),
+        }
+    }
+
+    fn current_time() -> time_t {
+        do_gettimeofday().sec()
+    }
+
+    fn get_new_shmid(&self) -> Result<ShmId> {
+        let mut shmid_manager = self.shmid_manager.write().unwrap();
+        shmid_manager.get_new_shmid()
+    }
+
+    fn free_shmid(&self, shmid: &ShmId) -> Result<()> {
+        let mut shmid_manager = self.shmid_manager.write().unwrap();
+        shmid_manager.free_shmid(&shmid)
+    }
+
+    fn shmctl_rmshm(&self, shmid: ShmId) -> Result<()> {
+        let mut shm_segments = self.shm_segments.write().unwrap();
+        let shm = shm_segments.get_mut(&shmid);
+
+        if let Some(shm) = shm {
+            shm.shm_ctime = ShmManager::current_time();
+            shm.set_destruction();
+            if shm.shm_nattach == 0 {
+                let shmid = shm.shmid;
+                self.free_shmid(&shmid)?;
+                shm_segments.remove(&shmid);
+            }
+        } else {
+            return_errno!(EINVAL, "cannot find shm by shmid");
+        }
+        Ok(())
+    }
+
+    fn shmctl_ipcstat(&self, shmid: ShmId, buf: Option<&mut shmids_t>) -> Result<()> {
+        let shm_segments = self.shm_segments.read().unwrap();
+        let shm = shm_segments.get(&shmid);
+        if let Some(shm) = shm {
+            shm.check_perm()?;
+            let mut buf = match buf {
+                Some(buf) => buf,
+                None => return_errno!(EFAULT, "buf is empty"),
+            };
+            let shm_perm = ipc_perm_t {
+                key: shm.key,
+                uid: shm.uid,
+                gid: shm.uid,
+                cuid: shm.cuid,
+                cgid: shm.cgid,
+                mode: shm.mode.bits() | shm.status.bits(),
+                pad1: 0,
+                seq: 0,
+                pad2: 0,
+                unused1: 0,
+                unused2: 0,
+            };
+            let shmids = shmids_t {
+                shm_perm: shm_perm,
+                shm_segsz: shm.shm_size(),
+                shm_atime: shm.shm_atime,
+                shm_dtime: shm.shm_dtime,
+                shm_ctime: shm.shm_ctime,
+                shm_cpid: shm.shm_cpid,
+                shm_lpid: shm.shm_lpid,
+                shm_nattach: shm.shm_nattach,
+                unused1: 0,
+                unused2: 0,
+            };
+            *buf = shmids;
+        } else {
+            return_errno!(EINVAL, "cannot find shm by shmid");
+        }
+        Ok(())
+    }
+
+    pub fn do_shmget(&self, key: key_t, size: usize, shmflg: ShmFlags) -> Result<ShmId> {
+        debug!(
+            "do_shmget: key: {:?}, size: {:?}, shmflg: {:?}",
+            key, size, shmflg
+        );
+
+        // Check the size from user
+        if size < SHMMIN || size > SHMMAX {
+            return_errno!(EINVAL, "invalid size");
+        }
+
+        let mut mode = shmflg.to_file_mode();
+        // The creator and user must have the read and write permission to the segment
+        let read_per = mode.contains(FileMode::S_IRUSR)
+            || mode.contains(FileMode::S_IRGRP)
+            || mode.contains(FileMode::S_IROTH);
+        let write_per = mode.contains(FileMode::S_IWUSR)
+            || mode.contains(FileMode::S_IWGRP)
+            || mode.contains(FileMode::S_IWOTH);
+        if !(read_per && write_per) {
+            return_errno!(
+                EINVAL,
+                "shared memory segement in occlum should have rw permission now"
+            );
+        }
+
+        let mut shm_segments = self.shm_segments.write().unwrap();
+        let shmid = if key == IPC_PRIVATE {
+            let shmid = self.get_new_shmid()?;
+            let shm = ShmSegment::new(shmid, key, size, mode)?;
+            shm_segments.insert(shm.shmid, shm);
+            shmid
+        } else {
+            // Get the shm from key if the segment is not marked to be destroyed
+            let shm = shm_segments
+                .values()
+                .find(|&shm| !shm.is_destruction() && shm.key == key);
+            let shmid = if let Some(shm) = shm {
+                if shmflg.contains(ShmFlags::IPC_CREAT) && shmflg.contains(ShmFlags::IPC_EXCL) {
+                    return_errno!(
+                        EEXIST,
+                        "the shared memory segment already exists for given key"
+                    );
+                }
+                // The size from user must be less than the actual size
+                if size > shm.shm_size() {
+                    return_errno!(EINVAL, "the size from user is too large");
+                }
+                // Check the permission
+                shm.check_perm()?;
+                shm.shmid
+            } else {
+                if !shmflg.contains(ShmFlags::IPC_CREAT) {
+                    return_errno!(ENOENT, "no segment exists for given key");
+                }
+                let shmid = self.get_new_shmid()?;
+                let shm = ShmSegment::new(shmid, key, size, mode)?;
+                shm_segments.insert(shm.shmid, shm);
+                shmid
+            };
+            shmid
+        };
+        Ok(shmid)
+    }
+
+    pub fn do_shmat(&self, shmid: ShmId, addr: usize, shmflg: ShmFlags) -> Result<usize> {
+        debug!(
+            "do_shmat: shmid: {:?}, addr: {:?}, shmflg: {:?}",
+            shmid, addr, shmflg
+        );
+        let pid = current!().process().pid();
+        let mut shm_segments = self.shm_segments.write().unwrap();
+        let shm = shm_segments.get_mut(&shmid);
+        let addr = if let Some(shm) = shm {
+            if addr != 0 && addr != shm.shm_start() {
+                return_errno!(EINVAL, "invalid addr");
+            }
+            // Check the permission
+            shm.check_perm()?;
+
+            shm.shm_nattach += 1;
+            shm.shm_atime = ShmManager::current_time();
+            shm.shm_add_pid(&pid)?;
+            shm.shm_lpid = pid;
+            shm.shm_start()
+        } else {
+            return_errno!(EINVAL, "cannot find shm by shmid");
+        };
+        Ok(addr)
+    }
+
+    pub fn do_shmdt(&self, addr: usize) -> Result<()> {
+        debug!("do_shmdt: addr: {:?}", addr);
+        let pid = current!().process().pid();
+        let mut shm_segments = self.shm_segments.write().unwrap();
+        let shm = shm_segments
+            .values_mut()
+            .find(|shm| shm.shm_start() == addr);
+
+        if let Some(shm) = shm {
+            shm.shm_dtime = ShmManager::current_time();
+            shm.shm_lpid = pid;
+            shm.shm_remove_pid(&pid)?;
+            shm.shm_nattach -= 1;
+            if shm.is_destruction() && shm.shm_nattach == 0 {
+                let shmid = shm.shmid;
+                self.free_shmid(&shmid);
+                shm_segments.remove(&shmid);
+            }
+        } else {
+            return_errno!(EINVAL, "cannot find shm by addr");
+        }
+        Ok(())
+    }
+
+    pub fn do_shmctl(&self, shmid: ShmId, cmd: CmdId, buf: Option<&mut shmids_t>) -> Result<()> {
+        debug!(
+            "do_shmctl: shmid: {:?}, cmd: {:?}, buf: {:?}",
+            shmid, cmd, buf
+        );
+        match cmd {
+            IPC_RMID => self.shmctl_rmshm(shmid),
+            IPC_STAT => self.shmctl_ipcstat(shmid, buf),
+            _ => return_errno!(EINVAL, "unimplemented cmd"),
+        }
+    }
+
+    pub fn detach_shm_when_process_exit(&self, thread: &ThreadRef) {
+        let pid = &thread.process().pid();
+        let mut shm_segments = self.shm_segments.write().unwrap();
+        shm_segments.drain_filter(|shmid, shm| match shm.shm_remove_pid(&pid) {
+            // There exists a shm that has been attached to the current process
+            Ok(_) => {
+                shm.shm_nattach -= 1;
+                if shm.is_destruction() && shm.shm_nattach == 0 {
+                    self.free_shmid(&shmid);
+                    true
+                } else {
+                    false
+                }
+            }
+            Err(_) => false,
+        });
+    }
+
+    pub fn clean_when_libos_exit(&self) {
+        let mut shm_segments = self.shm_segments.write().unwrap();
+        for (_, shm) in shm_segments.drain() {
+            debug!("clean shm: {:?}", shm);
+            self.free_shmid(&shm.shmid);
+        }
+    }
+}

src/libos/src/ipc/syscalls.rs

@@ -0,0 +1,36 @@
+use super::*;
+
+use util::mem_util::from_user;
+
+use super::shm::{shmids_t, CmdId, ShmFlags, ShmId, SHM_MANAGER};
+
+pub fn do_shmget(key: key_t, size: size_t, shmflg: i32) -> Result<isize> {
+    let shmflg =
+        ShmFlags::from_bits(shmflg as u32).ok_or_else(|| errno!(EINVAL, "invalid flags"))?;
+    let shmid = SHM_MANAGER.do_shmget(key, size, shmflg)?;
+    Ok(shmid as isize)
+}
+
+pub fn do_shmat(shmid: i32, shmaddr: usize, shmflg: i32) -> Result<isize> {
+    let shmflg =
+        ShmFlags::from_bits(shmflg as u32).ok_or_else(|| errno!(EINVAL, "invalid flags"))?;
+    let addr = SHM_MANAGER.do_shmat(shmid as ShmId, shmaddr, shmflg)?;
+    Ok(addr as isize)
+}
+
+pub fn do_shmdt(shmaddr: usize) -> Result<isize> {
+    SHM_MANAGER.do_shmdt(shmaddr)?;
+    Ok(0)
+}
+
+pub fn do_shmctl(shmid: i32, cmd: i32, buf_u: *mut shmids_t) -> Result<isize> {
+    let buf = if !buf_u.is_null() {
+        from_user::check_mut_ptr(buf_u)?;
+        let mut buf = unsafe { &mut *buf_u };
+        Some(buf)
+    } else {
+        None
+    };
+    SHM_MANAGER.do_shmctl(shmid as ShmId, cmd as CmdId, buf)?;
+    Ok(0)
+}

src/libos/src/lib.rs

@@ -81,6 +81,7 @@ mod events;
 mod exception;
 mod fs;
 mod interrupt;
+mod ipc;
 mod misc;
 mod net;
 mod process;

src/libos/src/process/do_exit.rs

@@ -6,6 +6,7 @@ use super::do_vfork::{is_vforked_child_process, vfork_return_to_parent};
 use super::pgrp::clean_pgrp_when_exit;
 use super::process::{Process, ProcessFilter};
 use super::{table, ProcessRef, TermStatus, ThreadRef, ThreadStatus};
+use crate::ipc::SHM_MANAGER;
 use crate::prelude::*;
 use crate::signal::{KernelSignal, SigNum};
 use crate::syscall::CpuContext;
@@ -107,6 +108,7 @@ fn exit_process(thread: &ThreadRef, term_status: TermStatus) {
     let mut process_inner = process.inner();
     // Clean used VM
     USER_SPACE_VM_MANAGER.free_chunks_when_exit(thread);
+    SHM_MANAGER.detach_shm_when_process_exit(thread);
 
     // The parent is the idle process
     if parent_inner.is_none() {

src/libos/src/process/mod.rs

@@ -68,6 +68,8 @@ pub mod task;
 pub type pid_t = u32;
 #[allow(non_camel_case_types)]
 pub type uid_t = u32;
+#[allow(non_camel_case_types)]
+pub type gid_t = u32;
 
 pub type ProcessRef = Arc<Process>;
 pub type ThreadRef = Arc<Thread>;

src/libos/src/syscall/mod.rs

@@ -35,6 +35,7 @@ use crate::fs::{
     FileRef, HostStdioFds, Stat, Statfs,
 };
 use crate::interrupt::{do_handle_interrupt, sgx_interrupt_info_t};
+use crate::ipc::{do_shmat, do_shmctl, do_shmdt, do_shmget, key_t, shmids_t};
 use crate::misc::{resource_t, rlimit_t, sysinfo_t, utsname_t, RandFlags};
 use crate::net::{
     do_accept, do_accept4, do_bind, do_connect, do_epoll_create, do_epoll_create1, do_epoll_ctl,
@@ -118,9 +119,9 @@ macro_rules! process_syscall_table_with_callback {
             (Msync = 26) => do_msync(addr: usize, size: usize, flags: u32),
             (Mincore = 27) => handle_unsupported(),
             (Madvise = 28) => handle_unsupported(),
-            (Shmget = 29) => handle_unsupported(),
+            (Shmget = 29) => do_shmget(key: key_t, size: size_t, shmflg: i32),
-            (Shmat = 30) => handle_unsupported(),
+            (Shmat = 30) => do_shmat(shmid: i32, shmaddr: usize, shmflg: i32),
-            (Shmctl = 31) => handle_unsupported(),
+            (Shmctl = 31) => do_shmctl(shmid: i32, cmd: i32, buf: *mut shmids_t),
             (Dup = 32) => do_dup(old_fd: FileDesc),
             (Dup2 = 33) => do_dup2(old_fd: FileDesc, new_fd: FileDesc),
             (Pause = 34) => handle_unsupported(),
@@ -156,7 +157,7 @@ macro_rules! process_syscall_table_with_callback {
             (Semget = 64) => handle_unsupported(),
             (Semop = 65) => handle_unsupported(),
             (Semctl = 66) => handle_unsupported(),
-            (Shmdt = 67) => handle_unsupported(),
+            (Shmdt = 67) => do_shmdt(shmaddr: usize),
             (Msgget = 68) => handle_unsupported(),
             (Msgsnd = 69) => handle_unsupported(),
             (Msgrcv = 70) => handle_unsupported(),

src/libos/src/vm/mod.rs

@@ -75,10 +75,12 @@ mod vm_util;
 
 use self::vm_layout::VMLayout;
 
+pub use self::chunk::ChunkRef;
 pub use self::process_vm::{MMapFlags, MRemapFlags, MSyncFlags, ProcessVM, ProcessVMBuilder};
 pub use self::user_space_vm::USER_SPACE_VM_MANAGER;
 pub use self::vm_perms::VMPerms;
 pub use self::vm_range::VMRange;
+pub use self::vm_util::{VMInitializer, VMMapOptionsBuilder};
 
 pub fn do_mmap(
     addr: usize,

src/libos/src/vm/process_vm.rs

@@ -2,6 +2,7 @@ use super::*;
 
 use super::chunk::*;
 use super::config;
+use super::ipc::SHM_MANAGER;
 use super::process::elf_file::{ElfFile, ProgramHeaderExt};
 use super::user_space_vm::USER_SPACE_VM_MANAGER;
 use super::vm_area::VMArea;

src/libos/src/vm/user_space_vm.rs

@@ -1,3 +1,4 @@
+use super::ipc::SHM_MANAGER;
 use super::*;
 use crate::ctor::dtor;
 use config::LIBOS_CONFIG;
@@ -46,6 +47,7 @@ impl UserSpaceVMManager {
 // be called after the main function. Static variables are still safe to visit at this time.
 #[dtor]
 fn free_user_space() {
+    SHM_MANAGER.clean_when_libos_exit();
     let range = USER_SPACE_VM_MANAGER.range();
     assert!(USER_SPACE_VM_MANAGER.verified_clean_when_exit());
     let addr = range.start() as *const c_void;