[vm] Support shared memory (POSIX)

.github/workflows/demo_test.yml

@@ -495,6 +495,10 @@ jobs:
     - name: Check result
       run: docker exec ${{ github.job }} bash -c "cd /root/occlum/demos/python/python_glibc/occlum_instance; cat smvlight.dat"
 
+    - name: Run python3.10 multiprocessing demo
+      run: docker exec ${{ github.job }} bash -c "cd /root/occlum/demos/python/python_glibc/python3.10-multiprocessing; ./install_python3.10.sh;
+            SGX_MODE=SIM ./run_python3.10_on_occlum.sh"
+
   # Redis test
   Redis_support_test:
     runs-on: ubuntu-20.04

demos/python/python_glibc/.gitignore

@@ -1,3 +1,4 @@
 occlum_instance/
 miniconda/
 Miniconda3*
+python-occlum

demos/python/python_glibc/python3.10-multiprocessing/.gitignore

@@ -0,0 +1,4 @@
+occlum_instance/
+miniconda/
+Miniconda3*
+python-occlum

demos/python/python_glibc/python3.10-multiprocessing/install_python3.10.sh

@@ -0,0 +1,12 @@
+#!/bin/bash
+set -e
+script_dir="$( cd "$( dirname "${BASH_SOURCE[0]}"  )" >/dev/null 2>&1 && pwd )"
+
+# 1. Init occlum workspace
+[ -d occlum_instance ] || occlum new occlum_instance
+
+# 2. Install python and dependencies to specified position
+[ -f Miniconda3-latest-Linux-x86_64.sh ] || wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
+[ -d miniconda ] || bash ./Miniconda3-latest-Linux-x86_64.sh -b -p $script_dir/miniconda
+$script_dir/miniconda/bin/conda create --prefix $script_dir/python-occlum -y \
+python=3.10.0 numpy scipy scikit-learn pandas Cython

demos/python/python_glibc/python3.10-multiprocessing/multiprocessing_demo.py

@@ -0,0 +1,18 @@
+# Python multiprocessing package provides API for process communication and management.
+# This demo demonstrates creating a worker pool and offloading jobs.
+# Processes communicate through shared memory (POSIX).
+import multiprocessing as mp
+import time
+
+def job():
+    print(1)
+
+start = time.time()
+
+if __name__ == '__main__':
+    mp.set_start_method('spawn')
+    pool = mp.Pool(processes=4)
+    for i in range(4):
+        pool.apply(job)
+        print("total time {}".format(time.time() - start))
+

demos/python/python_glibc/python3.10-multiprocessing/python3.10.yaml

@@ -0,0 +1,18 @@
+includes:
+  - base.yaml
+targets:
+  - target: /bin
+    copy:
+      - files:
+          - ../python-occlum/bin/python3.10
+  # python packages
+  - target: /opt
+    copy: 
+      - dirs:
+          - ../python-occlum
+  # below are python code and data
+  - target: /
+    copy:
+      - from: ..
+        files: 
+          - multiprocessing_demo.py

demos/python/python_glibc/python3.10-multiprocessing/run_python3.10_on_occlum.sh

@@ -0,0 +1,28 @@
+#!/bin/bash
+set -e
+
+BLUE='\033[1;34m'
+NC='\033[0m'
+
+script_dir="$( cd "$( dirname "${BASH_SOURCE[0]}"  )" >/dev/null 2>&1 && pwd )"
+python_dir="$script_dir/occlum_instance/image/opt/python-occlum"
+
+rm -rf occlum_instance && occlum new occlum_instance
+
+cd occlum_instance && rm -rf image
+copy_bom -f ../python3.10.yaml --root image --include-dir /opt/occlum/etc/template
+
+if [ ! -d $python_dir ];then
+    echo "Error: cannot stat '$python_dir' directory"
+    exit 1
+fi
+
+new_json="$(jq '.resource_limits.user_space_size = "1000MB" |
+        .resource_limits.kernel_space_heap_size = "300MB" |
+        .env.default += ["PYTHONHOME=/opt/python-occlum", "PATH=/bin"]' Occlum.json)" && \
+echo "${new_json}" > Occlum.json
+occlum build
+
+# Run the python demo
+echo -e "${BLUE}occlum run /bin/python3.10 multiprocessing_demo.py${NC}"
+occlum run /bin/python3.10 multiprocessing_demo.py

src/libos/src/ipc/shm.rs

@@ -212,7 +212,7 @@ impl Drop for ShmSegment {
         assert!(self.process_set.len() == 0);
         USER_SPACE_VM_MANAGER
             .internal()
-            .munmap_chunk(&self.chunk, None);
+            .munmap_chunk(&self.chunk, None, false);
     }
 }
 

src/libos/src/vm/chunk.rs

@@ -6,6 +6,7 @@ use super::vm_perms::VMPerms;
 use super::vm_util::*;
 use crate::process::ProcessRef;
 use crate::process::ThreadRef;
+
 use std::cmp::Ordering;
 use std::collections::HashSet;
 use std::hash::{Hash, Hasher};
@@ -71,10 +72,10 @@ impl Chunk {
         &self.internal
     }
 
-    pub fn get_vma_for_single_vma_chunk(&self) -> VMArea {
+    pub fn get_vma_for_single_vma_chunk(&self) -> SgxMutexGuard<VMArea> {
         match self.internal() {
+            ChunkType::SingleVMA(vma) => return vma.lock().unwrap(),
             ChunkType::MultiVMA(internal_manager) => unreachable!(),
-            ChunkType::SingleVMA(vma) => return vma.lock().unwrap().clone(),
         }
     }
 
@@ -98,7 +99,7 @@ impl Chunk {
             vm_range.clone(),
             *options.perms(),
             options.initializer().backed_file(),
-            DUMMY_CHUNK_PROCESS_ID,
+            current!().process().pid(),
         );
         // Initialize the memory of the new range
         unsafe {
@@ -247,6 +248,13 @@ impl Chunk {
                 .is_free_range(request_range),
         }
     }
+
+    pub fn is_shared(&self) -> bool {
+        match self.internal() {
+            ChunkType::SingleVMA(vma) => vma.lock().unwrap().is_shared(),
+            ChunkType::MultiVMA(_) => false,
+        }
+    }
 }
 
 #[derive(Debug)]

src/libos/src/vm/free_space_manager.rs

@@ -2,10 +2,11 @@
 // Currently only use simple vector as the base structure.
 //
 // Basically use address-ordered first fit to find free ranges.
-use std::cmp::Ordering;
+use super::*;
 
 use super::vm_util::VMMapAddr;
-use super::*;
+
+use std::cmp::Ordering;
 
 const INITIAL_SIZE: usize = 100;
 
@@ -37,7 +38,7 @@ impl VMFreeSpaceManager {
         align: usize,
         addr: VMMapAddr,
     ) -> Result<VMRange> {
-        // Record the minimal free range that satisfies the contraints
+        // Record the minimal free range that satisfies the constraints
         let mut result_free_range: Option<VMRange> = None;
         let mut result_idx: Option<usize> = None;
         let mut free_list = &mut self.free_manager;

src/libos/src/vm/mod.rs

@@ -64,6 +64,7 @@ use std::fmt;
 mod chunk;
 mod free_space_manager;
 mod process_vm;
+mod shm_manager;
 mod user_space_vm;
 mod vm_area;
 mod vm_chunk_manager;

src/libos/src/vm/process_vm.rs

@@ -1,17 +1,18 @@
 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;
 use super::vm_perms::VMPerms;
 use super::vm_util::{
     FileBacked, VMInitializer, VMMapAddr, VMMapOptions, VMMapOptionsBuilder, VMRemapOptions,
 };
+use crate::config;
+use crate::ipc::SHM_MANAGER;
+use crate::process::elf_file::{ElfFile, ProgramHeaderExt};
+use crate::util::sync::rw_lock::RwLockWriteGuard;
+
 use std::collections::HashSet;
-use util::sync::rw_lock::RwLockWriteGuard;
 
 // Used for heap and stack start address randomization.
 const RANGE_FOR_RANDOMIZATION: usize = 256 * 4096; // 1M
@@ -50,14 +51,14 @@ impl<'a, 'b> ProcessVMBuilder<'a, 'b> {
     }
 
     // Generate a random address within [0, range]
-    // Note: This function doesn't gurantee alignment
+    // Note: This function doesn't guarantee alignment
     fn get_randomize_offset(range: usize) -> usize {
         if cfg!(debug_assertions) {
             return range;
         }
 
         use crate::misc;
-        trace!("entrophy size = {}", range);
+        trace!("entropy size = {}", range);
         let mut random_buf: [u8; 8] = [0u8; 8]; // same length as usize
         misc::get_random(&mut random_buf).expect("failed to get random number");
         let random_num: usize = u64::from_le_bytes(random_buf) as usize;
@@ -74,7 +75,7 @@ impl<'a, 'b> ProcessVMBuilder<'a, 'b> {
             .stack_size
             .unwrap_or(config::LIBOS_CONFIG.process.default_stack_size);
 
-        // Before allocating memory, let's first calcualte how much memory
+        // Before allocating memory, let's first calculate how much memory
         // we need in total by iterating the memory layouts required by
         // all the memory regions
         let elf_layouts: Vec<VMLayout> = self
@@ -211,7 +212,9 @@ impl<'a, 'b> ProcessVMBuilder<'a, 'b> {
 
     fn handle_error_when_init(&self, chunks: &HashSet<Arc<Chunk>>) {
         chunks.iter().for_each(|chunk| {
-            USER_SPACE_VM_MANAGER.internal().munmap_chunk(chunk, None);
+            USER_SPACE_VM_MANAGER
+                .internal()
+                .munmap_chunk(chunk, None, false);
         });
     }
 
@@ -227,7 +230,7 @@ impl<'a, 'b> ProcessVMBuilder<'a, 'b> {
         let mut empty_start_offset = 0;
         let mut empty_end_offset = 0;
 
-        // Init all loadable segements
+        // Init all loadable segments
         elf_file
             .program_headers()
             .filter(|segment| segment.loadable())
@@ -309,7 +312,9 @@ impl Drop for ProcessVM {
         mem_chunks
             .drain_filter(|chunk| chunk.is_single_vma())
             .for_each(|chunk| {
-                USER_SPACE_VM_MANAGER.internal().munmap_chunk(&chunk, None);
+                USER_SPACE_VM_MANAGER
+                    .internal()
+                    .munmap_chunk(&chunk, None, false);
             });
 
         assert!(mem_chunks.len() == 0);
@@ -351,8 +356,9 @@ impl ProcessVM {
         mem_chunks: &mut RwLockWriteGuard<HashSet<ChunkRef>>,
     ) -> Result<Vec<VMArea>> {
         // Get all single VMA chunks
+        // Shared chunks shouldn't be merged since they are managed by shm manager and shared by multi processes
         let mut single_vma_chunks = mem_chunks
-            .drain_filter(|chunk| chunk.is_single_vma())
+            .drain_filter(|chunk| chunk.is_single_vma() && !chunk.is_shared())
             .collect::<Vec<ChunkRef>>();
         single_vma_chunks.sort_unstable_by(|chunk_a, chunk_b| {
             chunk_a
@@ -395,7 +401,7 @@ impl ProcessVM {
         for chunk in single_vma_chunks.into_iter().filter_map(|chunk| {
             if !chunk.is_single_dummy_vma() {
                 if chunk.is_single_vma_with_conflict_size() {
-                    let new_vma = chunk.get_vma_for_single_vma_chunk();
+                    let new_vma = chunk.get_vma_for_single_vma_chunk().clone();
                     merged_vmas.push(new_vma);
 
                     // Don't insert the merged chunks to mem_chunk list here. It should be updated later.

src/libos/src/vm/shm_manager.rs

@@ -0,0 +1,278 @@
+//! Shared memory manager. (POSIX)
+use super::*;
+
+use super::vm_manager::InternalVMManager;
+use super::vm_util::VMMapOptions;
+use crate::process::ThreadStatus;
+
+use rcore_fs::vfs::{FileType, Metadata};
+use std::collections::{HashMap, HashSet};
+use std::sync::{Arc, Weak};
+
+type InodeId = usize;
+
+/// Shared VM manager.
+#[derive(Debug)]
+pub struct ShmManager {
+    // K: Inode id of shared backed file. V: Chunk which is shared by processes.
+    shared_chunks: HashMap<InodeId, ChunkRef>,
+}
+
+/// Result types of `mmap()` with `MAP_SHARED`.
+#[derive(Clone, Debug)]
+pub enum MmapSharedResult {
+    /// Can share successfully
+    Success(usize),
+    /// Need to create a new shared chunk
+    NeedCreate,
+    /// Current shared chunk needs to expand range. (old shared chunk, expand range)
+    NeedExpand(ChunkRef, VMRange),
+    /// Current shared chunk needs to be replaced to satisfy new request
+    NeedReplace(ChunkRef),
+}
+
+/// Result types of unmapping a shared memory chunk.
+/// This could come from `munmap()` or `mremap` request.
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub enum MunmapSharedResult {
+    /// Current shared chunk is still being shared and used by other processes
+    StillInUse,
+    /// All shared processes are detached, the chunk can be freed
+    Freeable,
+}
+
+impl ShmManager {
+    pub fn new() -> Self {
+        Self {
+            shared_chunks: HashMap::new(),
+        }
+    }
+
+    pub fn mmap_shared_chunk(&mut self, options: &VMMapOptions) -> Result<MmapSharedResult> {
+        Self::qualified_for_sharing(options)?;
+
+        let backed_file = options.initializer().backed_file().unwrap();
+        let inode_id = backed_file.metadata().inode;
+        let offset = backed_file.offset();
+
+        let shared_chunk = match self.shared_chunks.get(&inode_id) {
+            Some(shared_chunk) => shared_chunk,
+            None => {
+                return Ok(MmapSharedResult::NeedCreate);
+            }
+        };
+        let mut shared_vma = Self::vma_of(&shared_chunk);
+
+        let current = current!();
+        let current_pid = current.process().pid();
+
+        let contained = shared_vma.belong_to(current_pid);
+        let exclusived = shared_vma.exclusive_by(current_pid);
+
+        let addr = {
+            let sc_addr = shared_vma.start();
+            let sc_size = shared_vma.size();
+            let sc_offset = shared_vma
+                .writeback_file()
+                .map(|(_, offset)| offset)
+                .unwrap();
+            let new_size = *options.size();
+            let mut target_addr = usize::MAX;
+            match *options.addr() {
+                vm_util::VMMapAddr::Any | vm_util::VMMapAddr::Hint(_) => {
+                    if offset == sc_offset && new_size <= sc_size {
+                        target_addr = sc_addr;
+                    } else if offset > sc_offset && offset - sc_offset + new_size <= sc_size {
+                        target_addr = sc_addr + offset - sc_offset;
+                    } else if exclusived {
+                        return Ok(MmapSharedResult::NeedReplace(shared_chunk.clone()));
+                    } else {
+                        return_errno!(EINVAL, "mmap shared chunk failed");
+                    }
+                }
+                vm_util::VMMapAddr::Need(addr) | vm_util::VMMapAddr::Force(addr) => {
+                    if addr == sc_addr && offset == sc_offset && new_size <= sc_size {
+                        target_addr = addr;
+                    } else if Self::can_expand_shared_vma(
+                        &shared_vma,
+                        (
+                            &VMRange::new_with_size(addr, new_size)?,
+                            options
+                                .initializer()
+                                .backed_file()
+                                .unwrap()
+                                .writeback_file()
+                                .unwrap(),
+                        ),
+                    ) {
+                        return Ok(MmapSharedResult::NeedExpand(
+                            shared_chunk.clone(),
+                            VMRange::new_with_size(addr, new_size).unwrap(),
+                        ));
+                    } else if exclusived {
+                        return Ok(MmapSharedResult::NeedReplace(shared_chunk.clone()));
+                    } else {
+                        return_errno!(EINVAL, "mmap shared chunk failed");
+                    }
+                }
+            }
+            target_addr
+        };
+
+        Self::apply_new_perms_if_higher(&mut shared_vma, *options.perms());
+        if !contained {
+            shared_vma.attach_shared_process(current_pid)?;
+            current.vm().add_mem_chunk(shared_chunk.clone());
+        }
+
+        Ok(MmapSharedResult::Success(addr))
+    }
+
+    pub fn munmap_shared_chunk(
+        &mut self,
+        chunk: &ChunkRef,
+        unmap_range: &VMRange,
+        force_unmap: bool,
+    ) -> Result<MunmapSharedResult> {
+        debug_assert!(chunk.is_shared());
+        let mut shared_vma = Self::vma_of(chunk);
+        let shared_range = shared_vma.range();
+        let current_pid = current!().process().pid();
+        let partial_unmap = !unmap_range.is_superset_of(shared_range);
+
+        // Fails when force unmap a partial of shared chunk which is still shared by other process
+        if force_unmap && (partial_unmap || !shared_vma.exclusive_by(current_pid)) {
+            return_errno!(EINVAL, "force unmap shared chunk failed");
+        }
+
+        // Treat partial unmapped shared chunk as still-in-use(do nothing)
+        if partial_unmap {
+            return Ok(MunmapSharedResult::StillInUse);
+        }
+
+        if shared_vma.detach_shared_process(current_pid)? {
+            self.shared_chunks.remove(&Self::inode_id_of(&shared_vma));
+            Ok(MunmapSharedResult::Freeable)
+        } else {
+            Ok(MunmapSharedResult::StillInUse)
+        }
+    }
+
+    pub fn mprotect_shared_chunk(&self, chunk: &ChunkRef, new_perms: VMPerms) -> Result<()> {
+        let mut vma = Self::vma_of(chunk);
+        if !vma.is_shared() {
+            return_errno!(EINVAL, "not a shared chunk");
+        }
+        Self::apply_new_perms_if_higher(&mut vma, new_perms);
+        Ok(())
+    }
+
+    pub fn create_shared_chunk(
+        &mut self,
+        options: &VMMapOptions,
+        new_chunk: ChunkRef,
+    ) -> Result<usize> {
+        let backed_file = options.initializer().backed_file().ok_or(errno!(EINVAL))?;
+        let (inode_id, addr) = {
+            let mut new_vma = Self::vma_of(&new_chunk);
+            new_vma.mark_shared();
+
+            let inode_id = backed_file.metadata().inode;
+            debug_assert_eq!(inode_id, Self::inode_id_of(&new_vma));
+            (inode_id, new_vma.start())
+        };
+
+        self.shared_chunks.insert(inode_id, new_chunk);
+        Ok(addr)
+    }
+
+    pub fn replace_shared_chunk(&mut self, old_shared_chunk: ChunkRef, new_chunk: ChunkRef) {
+        debug_assert!(old_shared_chunk.is_shared() && new_chunk.is_shared());
+        let inode_id = {
+            let mut new_vma = Self::vma_of(&new_chunk);
+            let old_vma = Self::vma_of(&old_shared_chunk);
+            // Inherits access and perms from the old one
+            new_vma.inherits_access_from(&old_vma);
+
+            // Apply higher perms to the whole new range
+            let new_perms = new_vma.perms();
+            let old_perms = old_vma.perms();
+            if new_perms != old_perms {
+                let perms = new_perms | old_perms;
+                VMPerms::apply_perms(new_vma.range(), perms);
+                new_vma.set_perms(perms);
+            }
+
+            let inode_id = Self::inode_id_of(&new_vma);
+            debug_assert_eq!(inode_id, Self::inode_id_of(&old_vma));
+            inode_id
+        };
+
+        let replaced = self.shared_chunks.insert(inode_id, new_chunk).unwrap();
+        debug_assert!(Arc::ptr_eq(&replaced, &old_shared_chunk));
+    }
+
+    // Left: Old shared vma. Right: New vm range, backed file and offset.
+    fn can_expand_shared_vma(lhs: &VMArea, rhs: (&VMRange, (&FileRef, usize))) -> bool {
+        debug_assert!(lhs.is_shared());
+        let (lhs_range, lhs_file, lhs_file_offset) = {
+            let writeback_file = lhs.writeback_file().unwrap();
+            (lhs.range(), writeback_file.0, writeback_file.1)
+        };
+        let (rhs_range, (rhs_file, rhs_file_offset)) = rhs;
+        debug_assert!(lhs_range.end() <= rhs_range.start());
+        if lhs_range.size() == 0 || rhs_range.size() == 0 {
+            return false;
+        }
+
+        // The two vm range must border with each other
+        if lhs_range.end() != rhs_range.start() {
+            return false;
+        }
+
+        Arc::ptr_eq(lhs_file, rhs_file)
+            && rhs_file_offset > lhs_file_offset
+            && rhs_file_offset - lhs_file_offset == lhs_range.size()
+    }
+
+    fn qualified_for_sharing(options: &VMMapOptions) -> Result<()> {
+        if !options.is_shared() {
+            return_errno!(EINVAL, "not a mmap(MAP_SHARED) request");
+        }
+        let backed_file = options.initializer().backed_file().unwrap();
+        if backed_file.metadata().type_ != FileType::File {
+            return_errno!(
+                EINVAL,
+                "unsupported file type when creating shared mappings"
+            );
+        }
+        Ok(())
+    }
+
+    fn vma_of(chunk: &ChunkRef) -> SgxMutexGuard<VMArea> {
+        match chunk.internal() {
+            ChunkType::SingleVMA(vma) => vma.lock().unwrap(),
+            ChunkType::MultiVMA(_) => unreachable!(),
+        }
+    }
+
+    /// Associated functions below only applied to shared vmas.
+
+    fn inode_id_of(vma: &SgxMutexGuard<VMArea>) -> InodeId {
+        debug_assert!(vma.is_shared());
+        vma.writeback_file()
+            .map(|(file, _)| file.metadata().unwrap().inode)
+            .unwrap()
+    }
+
+    fn apply_new_perms_if_higher(vma: &mut SgxMutexGuard<VMArea>, new_perms: VMPerms) {
+        debug_assert!(vma.is_shared());
+        let old_perms = vma.perms();
+        let perms = new_perms | old_perms;
+        if perms == old_perms {
+            return;
+        }
+        VMPerms::apply_perms(vma.range(), perms);
+        vma.set_perms(perms);
+    }
+}

src/libos/src/vm/user_space_vm.rs

@@ -1,10 +1,12 @@
-use super::ipc::SHM_MANAGER;
 use super::*;
+
+use super::vm_manager::VMManager;
+use crate::config::LIBOS_CONFIG;
 use crate::ctor::dtor;
+use crate::ipc::SHM_MANAGER;
 use crate::util::pku_util;
-use config::LIBOS_CONFIG;
+
 use std::ops::{Deref, DerefMut};
-use vm_manager::VMManager;
 
 const RSRV_MEM_PERM: MemPerm =
     MemPerm::from_bits_truncate(MemPerm::READ.bits() | MemPerm::WRITE.bits());
@@ -16,7 +18,7 @@ impl UserSpaceVMManager {
     fn new() -> Result<UserSpaceVMManager> {
         let rsrv_mem_size = LIBOS_CONFIG.resource_limits.user_space_size;
         let vm_range = unsafe {
-            // TODO: Current sgx_alloc_rsrv_mem implmentation will commit all the pages of the desired size, which will consume
+            // TODO: Current sgx_alloc_rsrv_mem implementation will commit all the pages of the desired size, which will consume
             // a lot of time. When EDMM is supported, there is no need to commit all the pages at the initialization stage. A function
             // which reserves memory but not commit pages should be provided then.
             let ptr = sgx_alloc_rsrv_mem(rsrv_mem_size);

src/libos/src/vm/vm_area.rs

@@ -1,19 +1,26 @@
-use std::ops::{Deref, DerefMut};
+use super::*;
 
 use super::vm_perms::VMPerms;
 use super::vm_range::VMRange;
 use super::vm_util::FileBacked;
-use super::*;
 
 use intrusive_collections::rbtree::{Link, RBTree};
 use intrusive_collections::{intrusive_adapter, KeyAdapter};
+use std::collections::HashSet;
+use std::ops::{Deref, DerefMut};
 
 #[derive(Clone, Debug, Default)]
 pub struct VMArea {
     range: VMRange,
     perms: VMPerms,
     file_backed: Option<FileBacked>,
-    pid: pid_t,
+    access: VMAccess,
+}
+
+#[derive(Clone, Debug, Eq, PartialEq)]
+pub enum VMAccess {
+    Private(pid_t),
+    Shared(HashSet<pid_t>),
 }
 
 impl VMArea {
@@ -27,7 +34,7 @@ impl VMArea {
             range,
             perms,
             file_backed,
-            pid,
+            access: VMAccess::Private(pid),
         }
     }
 
@@ -37,7 +44,7 @@ impl VMArea {
         vma: &VMArea,
         new_range: VMRange,
         new_perms: VMPerms,
-        pid: pid_t,
+        access: VMAccess,
     ) -> Self {
         let new_backed_file = vma.file_backed.as_ref().map(|file| {
             let mut new_file = file.clone();
@@ -57,7 +64,12 @@ impl VMArea {
             new_file
         });
 
-        Self::new(new_range, new_perms, new_backed_file, pid)
+        Self {
+            range: new_range,
+            perms: new_perms,
+            file_backed: new_backed_file,
+            access,
+        }
     }
 
     pub fn perms(&self) -> VMPerms {
@@ -68,8 +80,22 @@ impl VMArea {
         &self.range
     }
 
-    pub fn pid(&self) -> pid_t {
-        self.pid
+    pub fn access(&self) -> &VMAccess {
+        &self.access
+    }
+
+    pub fn belong_to(&self, target_pid: pid_t) -> bool {
+        match &self.access {
+            VMAccess::Private(pid) => *pid == target_pid,
+            VMAccess::Shared(pid_set) => pid_set.contains(&target_pid),
+        }
+    }
+
+    pub fn exclusive_by(&self, target_pid: pid_t) -> bool {
+        match &self.access {
+            VMAccess::Private(pid) => *pid == target_pid,
+            VMAccess::Shared(pid_set) => pid_set.len() == 1 && pid_set.contains(&target_pid),
+        }
     }
 
     pub fn init_file(&self) -> Option<(&FileRef, usize)> {
@@ -96,7 +122,7 @@ impl VMArea {
         self.deref()
             .subtract(other)
             .into_iter()
-            .map(|range| Self::inherits_file_from(self, range, self.perms(), self.pid()))
+            .map(|range| Self::inherits_file_from(self, range, self.perms(), self.access().clone()))
             .collect()
     }
 
@@ -109,7 +135,8 @@ impl VMArea {
             }
             new_range.unwrap()
         };
-        let new_vma = VMArea::inherits_file_from(self, new_range, self.perms(), self.pid());
+        let new_vma =
+            VMArea::inherits_file_from(self, new_range, self.perms(), self.access().clone());
         Some(new_vma)
     }
 
@@ -139,7 +166,7 @@ impl VMArea {
     }
 
     pub fn is_the_same_to(&self, other: &VMArea) -> bool {
-        if self.pid() != other.pid() {
+        if self.access() != other.access() {
             return false;
         }
 
@@ -175,7 +202,7 @@ impl VMArea {
             return false;
         }
         // The two VMAs must be owned by the same process
-        if left.pid() != right.pid() {
+        if left.access() != right.access() {
             return false;
         }
         // The two VMAs must border with each other
@@ -202,6 +229,78 @@ impl VMArea {
             }
         }
     }
+
+    /// Flush a file-backed VMA to its file. This has no effect on anonymous VMA.
+    pub fn flush_backed_file(&self) {
+        self.flush_backed_file_with_cond(|_| true)
+    }
+
+    /// Same as `flush_backed_file()`, except that an extra condition on the file needs to satisfy.
+    pub fn flush_backed_file_with_cond<F: Fn(&FileRef) -> bool>(&self, cond_fn: F) {
+        let (file, file_offset) = match self.writeback_file() {
+            None => return,
+            Some((file_and_offset)) => file_and_offset,
+        };
+        let file_writable = file
+            .access_mode()
+            .map(|ac| ac.writable())
+            .unwrap_or_default();
+        if !file_writable {
+            return;
+        }
+        if !cond_fn(file) {
+            return;
+        }
+        file.write_at(file_offset, unsafe { self.as_slice() });
+    }
+
+    pub fn is_shared(&self) -> bool {
+        match self.access {
+            VMAccess::Private(_) => false,
+            VMAccess::Shared(_) => true,
+        }
+    }
+
+    pub fn mark_shared(&mut self) {
+        let access = match self.access {
+            VMAccess::Private(pid) => VMAccess::Shared(HashSet::from([pid])),
+            VMAccess::Shared(_) => {
+                return;
+            }
+        };
+        self.access = access;
+    }
+
+    pub fn shared_process_set(&self) -> Result<&HashSet<pid_t>> {
+        match &self.access {
+            VMAccess::Private(_) => Err(errno!(EINVAL, "not a shared vma")),
+            VMAccess::Shared(pid_set) => Ok(pid_set),
+        }
+    }
+
+    pub fn attach_shared_process(&mut self, pid: pid_t) -> Result<()> {
+        match &mut self.access {
+            VMAccess::Private(_) => Err(errno!(EINVAL, "not a shared vma")),
+            VMAccess::Shared(pid_set) => {
+                pid_set.insert(pid);
+                Ok(())
+            }
+        }
+    }
+
+    pub fn detach_shared_process(&mut self, pid: pid_t) -> Result<bool> {
+        match &mut self.access {
+            VMAccess::Private(_) => Err(errno!(EINVAL, "not a shared vma")),
+            VMAccess::Shared(pid_set) => {
+                pid_set.remove(&pid);
+                Ok(pid_set.is_empty())
+            }
+        }
+    }
+
+    pub fn inherits_access_from(&mut self, vma: &VMArea) {
+        self.access = vma.access().clone()
+    }
 }
 
 impl Deref for VMArea {
@@ -212,6 +311,12 @@ impl Deref for VMArea {
     }
 }
 
+impl Default for VMAccess {
+    fn default() -> Self {
+        Self::Private(0)
+    }
+}
+
 #[derive(Clone)]
 pub struct VMAObj {
     link: Link,

src/libos/src/vm/vm_chunk_manager.rs

@@ -4,12 +4,12 @@ use super::free_space_manager::VMFreeSpaceManager as FreeRangeManager;
 use super::vm_area::*;
 use super::vm_perms::VMPerms;
 use super::vm_util::*;
-use std::collections::BTreeSet;
 
 use intrusive_collections::rbtree::{Link, RBTree};
 use intrusive_collections::Bound;
 use intrusive_collections::RBTreeLink;
 use intrusive_collections::{intrusive_adapter, KeyAdapter};
+use std::collections::BTreeSet;
 
 /// Memory chunk manager.
 ///
@@ -77,13 +77,13 @@ impl ChunkManager {
         vmas_cursor.move_next(); // move to the first element of the tree
         while !vmas_cursor.is_null() {
             let vma = vmas_cursor.get().unwrap().vma();
-            if vma.pid() != pid || vma.size() == 0 {
+            if !vma.belong_to(pid) || vma.size() == 0 {
                 // Skip vmas which doesn't belong to this process
                 vmas_cursor.move_next();
                 continue;
             }
 
-            Self::flush_file_vma(vma);
+            vma.flush_backed_file();
 
             if !vma.perms().is_default() {
                 VMPerms::apply_perms(vma, VMPerms::default());
@@ -156,7 +156,7 @@ impl ChunkManager {
         while !vmas_cursor.is_null() && vmas_cursor.get().unwrap().vma().start() <= range.end() {
             let vma = &vmas_cursor.get().unwrap().vma();
             trace!("munmap related vma = {:?}", vma);
-            if vma.size() == 0 || current_pid != vma.pid() {
+            if vma.size() == 0 || !vma.belong_to(current_pid) {
                 vmas_cursor.move_next();
                 continue;
             }
@@ -169,7 +169,7 @@ impl ChunkManager {
             };
 
             // File-backed VMA needs to be flushed upon munmap
-            Self::flush_file_vma(&intersection_vma);
+            intersection_vma.flush_backed_file();
             if !&intersection_vma.perms().is_default() {
                 VMPerms::apply_perms(&intersection_vma, VMPerms::default());
             }
@@ -254,7 +254,7 @@ impl ChunkManager {
             {
                 let vma = &vmas_cursor.get().unwrap().vma();
                 // The old range must be contained in one single VMA
-                if vma.pid() == current_pid && vma.is_superset_of(&old_range) {
+                if vma.belong_to(current_pid) && vma.is_superset_of(&old_range) {
                     break;
                 } else {
                     vmas_cursor.move_next();
@@ -284,7 +284,7 @@ impl ChunkManager {
             && containing_vmas.get().unwrap().vma().start() <= protect_range.end()
         {
             let mut containing_vma = containing_vmas.get().unwrap().vma().clone();
-            if containing_vma.pid() != current_pid {
+            if !containing_vma.belong_to(current_pid) {
                 containing_vmas.move_next();
                 continue;
             }
@@ -332,7 +332,7 @@ impl ChunkManager {
                             &containing_vma,
                             protect_range,
                             new_perms,
-                            current_pid,
+                            VMAccess::Private(current_pid),
                         );
                         VMPerms::apply_perms(&new_vma, new_vma.perms());
                         let new_vma = VMAObj::new_vma_obj(new_vma);
@@ -344,7 +344,7 @@ impl ChunkManager {
                                 &containing_vma,
                                 range,
                                 old_perms,
-                                current_pid,
+                                VMAccess::Private(current_pid),
                             );
                             VMAObj::new_vma_obj(new_vma)
                         };
@@ -369,7 +369,7 @@ impl ChunkManager {
                             &containing_vma,
                             intersection_vma.range().clone(),
                             new_perms,
-                            current_pid,
+                            VMAccess::Private(current_pid),
                         );
                         VMPerms::apply_perms(&new_vma, new_vma.perms());
 
@@ -399,7 +399,7 @@ impl ChunkManager {
                 None => continue,
                 Some(vma) => vma,
             };
-            Self::flush_file_vma(&vma);
+            vma.flush_backed_file();
         }
         Ok(())
     }
@@ -409,41 +409,17 @@ impl ChunkManager {
     pub fn msync_by_file(&mut self, sync_file: &FileRef) {
         for vma_obj in &self.vmas {
             let is_same_file = |file: &FileRef| -> bool { Arc::ptr_eq(&file, &sync_file) };
-            Self::flush_file_vma_with_cond(&vma_obj.vma(), is_same_file);
+            vma_obj.vma().flush_backed_file_with_cond(is_same_file);
         }
     }
 
-    /// Flush a file-backed VMA to its file. This has no effect on anonymous VMA.
-    pub fn flush_file_vma(vma: &VMArea) {
-        Self::flush_file_vma_with_cond(vma, |_| true)
-    }
-
-    /// Same as flush_vma, except that an extra condition on the file needs to satisfy.
-    pub fn flush_file_vma_with_cond<F: Fn(&FileRef) -> bool>(vma: &VMArea, cond_fn: F) {
-        let (file, file_offset) = match vma.writeback_file() {
-            None => return,
-            Some((file_and_offset)) => file_and_offset,
-        };
-        let file_writable = file
-            .access_mode()
-            .map(|ac| ac.writable())
-            .unwrap_or_default();
-        if !file_writable {
-            return;
-        }
-        if !cond_fn(file) {
-            return;
-        }
-        file.write_at(file_offset, unsafe { vma.as_slice() });
-    }
-
     pub fn find_mmap_region(&self, addr: usize) -> Result<VMRange> {
         let vma = self.vmas.upper_bound(Bound::Included(&addr));
         if vma.is_null() {
             return_errno!(ESRCH, "no mmap regions that contains the address");
         }
         let vma = vma.get().unwrap().vma();
-        if vma.pid() != current!().process().pid() || !vma.contains(addr) {
+        if !vma.belong_to(current!().process().pid()) || !vma.contains(addr) {
             return_errno!(ESRCH, "no mmap regions that contains the address");
         }
 
@@ -451,13 +427,13 @@ impl ChunkManager {
     }
 
     pub fn usage_percentage(&self) -> f32 {
-        let totol_size = self.range.size();
+        let total_size = self.range.size();
         let mut used_size = 0;
         self.vmas
             .iter()
             .for_each(|vma_obj| used_size += vma_obj.vma().size());
 
-        return used_size as f32 / totol_size as f32;
+        return used_size as f32 / total_size as f32;
     }
 
     fn merge_all_vmas(&mut self) {

src/libos/src/vm/vm_manager.rs

@@ -4,15 +4,15 @@ use super::chunk::{
     Chunk, ChunkID, ChunkRef, ChunkType, CHUNK_DEFAULT_SIZE, DUMMY_CHUNK_PROCESS_ID,
 };
 use super::free_space_manager::VMFreeSpaceManager;
-use super::vm_area::VMArea;
+use super::shm_manager::{MmapSharedResult, MunmapSharedResult, ShmManager};
+use super::vm_area::{VMAccess, VMArea};
 use super::vm_chunk_manager::ChunkManager;
 use super::vm_perms::VMPerms;
 use super::vm_util::*;
 use crate::process::{ThreadRef, ThreadStatus};
-use std::ops::Bound::{Excluded, Included};
 
-use crate::util::sync::rw_lock;
-use std::collections::{BTreeSet, HashSet};
+use std::collections::BTreeSet;
+use std::ops::Bound::{Excluded, Included};
 
 // Incorrect order of locks could cause deadlock easily.
 // Don't hold a low-order lock and then try to get a high-order lock.
@@ -65,6 +65,24 @@ impl VMManager {
     }
 
     pub fn mmap(&self, options: &VMMapOptions) -> Result<usize> {
+        if options.is_shared() {
+            let res = self.internal().mmap_shared_chunk(options);
+            match res {
+                Ok(addr) => {
+                    trace!(
+                        "mmap_shared_chunk success: addr = 0x{:X}, pid = {}",
+                        res.as_ref().unwrap(),
+                        current!().process().pid()
+                    );
+                    return Ok(addr);
+                }
+                Err(e) => {
+                    warn!("mmap_shared_chunk failed: {:?}", e);
+                    // Do not return when `mmap_shared_chunk()` fails. Try mmap as a regular chunk as below.
+                }
+            }
+        }
+
         let addr = *options.addr();
         let size = *options.size();
         let align = *options.align();
@@ -198,7 +216,7 @@ impl VMManager {
             for chunk in overlapping_chunks.iter() {
                 match chunk.internal() {
                     ChunkType::SingleVMA(_) => {
-                        internal_manager.munmap_chunk(chunk, Some(&munmap_range))?
+                        internal_manager.munmap_chunk(chunk, Some(&munmap_range), false)?
                     }
                     ChunkType::MultiVMA(manager) => manager
                         .lock()
@@ -232,7 +250,11 @@ impl VMManager {
                         .map(|chunk| chunk.clone())
                 };
                 if let Some(overlapping_chunk) = overlapping_chunk {
-                    return internal_manager.munmap_chunk(&overlapping_chunk, Some(&munmap_range));
+                    return internal_manager.munmap_chunk(
+                        &overlapping_chunk,
+                        Some(&munmap_range),
+                        false,
+                    );
                 } else {
                     warn!("no overlapping chunks anymore");
                     return Ok(());
@@ -354,7 +376,7 @@ impl VMManager {
             }
             ChunkType::SingleVMA(vma) => {
                 let vma = vma.lock().unwrap();
-                ChunkManager::flush_file_vma(&vma);
+                vma.flush_backed_file();
             }
         }
         Ok(())
@@ -375,7 +397,9 @@ impl VMManager {
                         .msync_by_file(sync_file);
                 }
                 ChunkType::SingleVMA(vma) => {
-                    ChunkManager::flush_file_vma_with_cond(&vma.lock().unwrap(), is_same_file);
+                    vma.lock()
+                        .unwrap()
+                        .flush_backed_file_with_cond(is_same_file);
                 }
             });
     }
@@ -437,7 +461,7 @@ impl VMManager {
         let ret_addr = if let Some(mmap_options) = remap_result_option.mmap_options() {
             let mmap_addr = self.mmap(mmap_options);
 
-            // FIXME: For MRemapFlags::MayMove flag, we checked if the prefered range is free when parsing the options.
+            // FIXME: For MRemapFlags::MayMove flag, we checked if the preferred range is free when parsing the options.
             // But there is no lock after the checking, thus the mmap might fail. In this case, we should try mmap again.
             if mmap_addr.is_err() && remap_result_option.may_move() == true {
                 return_errno!(
@@ -457,7 +481,7 @@ impl VMManager {
 
         if let Some((munmap_addr, munmap_size)) = remap_result_option.munmap_args() {
             self.munmap(*munmap_addr, *munmap_size)
-                .expect("Shouln't fail");
+                .expect("Shouldn't fail");
         }
 
         return Ok(ret_addr);
@@ -479,7 +503,7 @@ impl VMManager {
         let mut mem_chunks = thread.vm().mem_chunks().write().unwrap();
 
         mem_chunks.iter().for_each(|chunk| {
-            internal_manager.munmap_chunk(&chunk, None);
+            internal_manager.munmap_chunk(&chunk, None, false);
         });
         mem_chunks.clear();
 
@@ -487,13 +511,14 @@ impl VMManager {
     }
 }
 
-// Modification on this structure must aquire the global lock.
+// Modification on this structure must acquire the global lock.
 // TODO: Enable fast_default_chunks for faster chunk allocation
 #[derive(Debug)]
 pub struct InternalVMManager {
     chunks: BTreeSet<ChunkRef>, // track in-use chunks, use B-Tree for better performance and simplicity (compared with red-black tree)
     fast_default_chunks: Vec<ChunkRef>, // empty default chunks
     free_manager: VMFreeSpaceManager,
+    shm_manager: ShmManager, // track chunks which are shared by processes
 }
 
 impl InternalVMManager {
@@ -501,10 +526,12 @@ impl InternalVMManager {
         let chunks = BTreeSet::new();
         let fast_default_chunks = Vec::new();
         let free_manager = VMFreeSpaceManager::new(vm_range);
+        let shm_manager = ShmManager::new();
         Self {
             chunks,
             fast_default_chunks,
             free_manager,
+            shm_manager,
         }
     }
 
@@ -522,26 +549,38 @@ impl InternalVMManager {
 
     // Allocate a chunk with single vma
     pub fn mmap_chunk(&mut self, options: &VMMapOptions) -> Result<ChunkRef> {
+        let new_chunk = self
+            .new_chunk_with_options(options)
+            .map_err(|e| errno!(e.errno(), "mmap_chunk failure"))?;
+        trace!("allocate a new single vma chunk: {:?}", new_chunk);
+        self.chunks.insert(new_chunk.clone());
+        Ok(new_chunk)
+    }
+
+    fn new_chunk_with_options(&mut self, options: &VMMapOptions) -> Result<ChunkRef> {
         let addr = *options.addr();
         let size = *options.size();
         let align = *options.align();
         let free_range = self.find_free_gaps(size, align, addr)?;
-        let free_chunk = Chunk::new_single_vma_chunk(&free_range, options);
-        if let Err(e) = free_chunk {
+        let free_chunk = Chunk::new_single_vma_chunk(&free_range, options).map_err(|e| {
             // Error when creating chunks. Must return the free space before returning error
             self.free_manager
                 .add_range_back_to_free_manager(&free_range);
-            return_errno!(e.errno(), "mmap_chunk failure");
-        }
-        let chunk = Arc::new(free_chunk.unwrap());
-        trace!("allocate a new single vma chunk: {:?}", chunk);
-        self.chunks.insert(chunk.clone());
-        Ok(chunk)
+            e
+        })?;
+        Ok(Arc::new(free_chunk))
     }
 
     // Munmap a chunk
     // For Single VMA chunk, a part of the chunk could be munmapped if munmap_range is specified.
-    pub fn munmap_chunk(&mut self, chunk: &ChunkRef, munmap_range: Option<&VMRange>) -> Result<()> {
+    // `force_unmap` indicates whether a unmap request came from a (re)map request with `MAP_FIXED`,
+    // the chunk would end differently when it is being shared.
+    pub fn munmap_chunk(
+        &mut self,
+        chunk: &ChunkRef,
+        munmap_range: Option<&VMRange>,
+        force_unmap: bool,
+    ) -> Result<()> {
         trace!(
             "munmap_chunk range = {:?}, munmap_range = {:?}",
             chunk.range(),
@@ -569,6 +608,15 @@ impl InternalVMManager {
             }
         };
 
+        if chunk.is_shared() {
+            trace!(
+                "munmap_shared_chunk, chunk_range: {:?}, munmap_range = {:?}",
+                chunk.range(),
+                munmap_range,
+            );
+            return self.munmap_shared_chunk(chunk, munmap_range, force_unmap);
+        }
+
         // Either the munmap range is a subset of the chunk range or the munmap range is
         // a superset of the chunk range. We can handle both cases.
 
@@ -580,7 +628,7 @@ impl InternalVMManager {
         };
 
         // File-backed VMA needs to be flushed upon munmap
-        ChunkManager::flush_file_vma(&intersection_vma);
+        intersection_vma.flush_backed_file();
 
         // Reset memory permissions
         if !&intersection_vma.perms().is_default() {
@@ -625,7 +673,9 @@ impl InternalVMManager {
                 let new_vma = new_vmas.pop().unwrap();
                 let new_vma_chunk = Arc::new(Chunk::new_chunk_with_vma(new_vma));
                 self.chunks.insert(new_vma_chunk.clone());
+                if current.status() != ThreadStatus::Exited {
                     current.vm().add_mem_chunk(new_vma_chunk);
+                }
 
                 let updated_vma = new_vmas.pop().unwrap();
                 self.update_single_vma_chunk(&current, &chunk, updated_vma);
@@ -635,18 +685,119 @@ impl InternalVMManager {
         Ok(())
     }
 
+    pub fn mmap_shared_chunk(&mut self, options: &VMMapOptions) -> Result<usize> {
+        match self.shm_manager.mmap_shared_chunk(options)? {
+            MmapSharedResult::Success(addr) => Ok(addr),
+            MmapSharedResult::NeedCreate => {
+                let new_chunk = self.mmap_chunk(options)?;
+                current!().vm().add_mem_chunk(new_chunk.clone());
+                self.shm_manager
+                    .create_shared_chunk(options, new_chunk.clone())
+                    .map_err(|e| {
+                        let vma = new_chunk.get_vma_for_single_vma_chunk();
+                        // Reset memory permissions
+                        if !vma.perms().is_default() {
+                            VMPerms::apply_perms(&vma, VMPerms::default());
+                        }
+                        // Reset memory contents
+                        unsafe {
+                            let buf = vma.as_slice_mut();
+                            buf.iter_mut().for_each(|b| *b = 0)
+                        }
+                        drop(vma);
+
+                        self.free_chunk(&new_chunk);
+                        current!().vm().remove_mem_chunk(&new_chunk);
+                        e
+                    })
+            }
+            MmapSharedResult::NeedExpand(old_shared_chunk, expand_range) => {
+                let new_chunk = {
+                    let new_chunk = self.new_chunk_with_options(options)?;
+                    self.merge_two_single_vma_chunks(&old_shared_chunk, &new_chunk)
+                };
+                let new_range = *new_chunk.range();
+                debug_assert_eq!(new_range, expand_range);
+                self.shm_manager
+                    .replace_shared_chunk(old_shared_chunk, new_chunk);
+                Ok(new_range.start())
+            }
+            MmapSharedResult::NeedReplace(_) => {
+                return_errno!(EINVAL, "mmap shared chunk failed");
+                // TODO: Support replace shared chunk when necessary,
+                // e.g., Current shared chunk is exclusived and `remap()` by same process
+            }
+        }
+    }
+
+    pub fn munmap_shared_chunk(
+        &mut self,
+        chunk: &ChunkRef,
+        munmap_range: &VMRange,
+        force_unmap: bool,
+    ) -> Result<()> {
+        if !chunk.is_shared() {
+            return_errno!(EINVAL, "not a shared chunk");
+        }
+        if !chunk.range().overlap_with(munmap_range) {
+            return Ok(());
+        }
+
+        if self
+            .shm_manager
+            .munmap_shared_chunk(chunk, munmap_range, force_unmap)?
+            == MunmapSharedResult::Freeable
+        {
+            let vma = chunk.get_vma_for_single_vma_chunk();
+            // Flush memory contents to backed file
+            vma.flush_backed_file();
+            // Reset memory permissions
+            if !vma.perms().is_default() {
+                VMPerms::apply_perms(&vma, VMPerms::default());
+            }
+            // Reset memory contents
+            unsafe {
+                let buf = vma.as_slice_mut();
+                buf.iter_mut().for_each(|b| *b = 0)
+            }
+            drop(vma);
+
+            self.free_chunk(chunk);
+            let current = current!();
+            if current.status() != ThreadStatus::Exited {
+                current.vm().remove_mem_chunk(&chunk);
+            }
+        }
+        Ok(())
+    }
+
     fn update_single_vma_chunk(
         &mut self,
         current_thread: &ThreadRef,
         old_chunk: &ChunkRef,
         new_vma: VMArea,
-    ) {
+    ) -> ChunkRef {
         let new_chunk = Arc::new(Chunk::new_chunk_with_vma(new_vma));
         current_thread
             .vm()
             .replace_mem_chunk(old_chunk, new_chunk.clone());
         self.chunks.remove(old_chunk);
-        self.chunks.insert(new_chunk);
+        self.chunks.insert(new_chunk.clone());
+        new_chunk
+    }
+
+    fn merge_two_single_vma_chunks(&mut self, lhs: &ChunkRef, rhs: &ChunkRef) -> ChunkRef {
+        let mut new_vma = {
+            let lhs_vma = lhs.get_vma_for_single_vma_chunk();
+            let rhs_vma = rhs.get_vma_for_single_vma_chunk();
+            debug_assert_eq!(lhs_vma.end(), rhs_vma.start());
+
+            let mut new_vma = rhs_vma.clone();
+            new_vma.set_start(lhs_vma.start());
+            new_vma
+        };
+
+        self.update_single_vma_chunk(&current!(), lhs, new_vma)
     }
 
     // protect_range should a sub-range of the chunk range
@@ -657,6 +808,17 @@ impl InternalVMManager {
         new_perms: VMPerms,
     ) -> Result<()> {
         debug_assert!(chunk.range().is_superset_of(&protect_range));
+        if chunk.is_shared() {
+            trace!(
+                "mprotect_shared_chunk, chunk_range: {:?}, mprotect_range = {:?}",
+                chunk.range(),
+                protect_range,
+            );
+            // When protect range hits a shared chunk, new perms are
+            // applied in a all-or-nothing mannner of the whole vma
+            return self.shm_manager.mprotect_shared_chunk(chunk, new_perms);
+        }
+
         let vma = match chunk.internal() {
             ChunkType::MultiVMA(_) => {
                 unreachable!();
@@ -677,6 +839,7 @@ impl InternalVMManager {
                 return Ok(());
             }
 
+            let current_pid = current!().process().pid();
             let same_start = protect_range.start() == containing_vma.start();
             let same_end = protect_range.end() == containing_vma.end();
             match (same_start, same_end) {
@@ -701,7 +864,7 @@ impl InternalVMManager {
                         &containing_vma,
                         protect_range,
                         new_perms,
-                        DUMMY_CHUNK_PROCESS_ID,
+                        VMAccess::Private(current_pid),
                     );
                     VMPerms::apply_perms(&new_vma, new_vma.perms());
 
@@ -711,7 +874,7 @@ impl InternalVMManager {
                             &containing_vma,
                             range,
                             old_perms,
-                            DUMMY_CHUNK_PROCESS_ID,
+                            VMAccess::Private(current_pid),
                         )
                     };
 
@@ -732,7 +895,7 @@ impl InternalVMManager {
                         &containing_vma,
                         protect_range,
                         new_perms,
-                        DUMMY_CHUNK_PROCESS_ID,
+                        VMAccess::Private(current_pid),
                     );
                     VMPerms::apply_perms(&new_vma, new_vma.perms());
 
@@ -800,7 +963,7 @@ impl InternalVMManager {
     //
     // Previously, this method is implemented for VMManager and only acquire the internal lock when needed. However, this will make mmap non-atomic.
     // In multi-thread applications, for example, when a thread calls mmap with MAP_FIXED flag, and that desired memory is mapped already, the libos
-    // will first munmap the corresponding memory and then do a mmap with desired address. If the lock is not aquired during the whole process,
+    // will first munmap the corresponding memory and then do a mmap with desired address. If the lock is not acquired during the whole process,
     // the unmapped memory might be allocated by other thread who is waiting and acquiring the lock.
     // Thus, in current code, this method is implemented for InternalManager and holds the lock during the whole process.
     // Below method "force_mmap_across_multiple_chunks" is the same.
@@ -853,7 +1016,7 @@ impl InternalVMManager {
                             }
 
                             // Munmap the corresponding single vma chunk
-                            self.munmap_chunk(&chunk, Some(&target_range))?;
+                            self.munmap_chunk(&chunk, Some(&target_range), true)?;
                         }
                         VMMapAddr::Any => unreachable!(),
                     }

src/libos/src/vm/vm_range.rs

@@ -89,7 +89,7 @@ impl VMRange {
         self.start() <= addr && addr < self.end()
     }
 
-    // Returns whether two ranges have non-empty interesection.
+    // Returns whether two ranges have non-empty intersection.
     pub fn overlap_with(&self, other: &VMRange) -> bool {
         let intersection_start = self.start().max(other.start());
         let intersection_end = self.end().min(other.end());

src/libos/src/vm/vm_util.rs

@@ -3,12 +3,12 @@ use super::*;
 use super::vm_area::*;
 use super::vm_perms::VMPerms;
 use crate::fs::FileMode;
-use std::collections::BTreeSet;
 
 use intrusive_collections::rbtree::{Link, RBTree};
 use intrusive_collections::Bound;
 use intrusive_collections::RBTreeLink;
 use intrusive_collections::{intrusive_adapter, KeyAdapter};
+use rcore_fs::vfs::Metadata;
 
 #[derive(Clone, Debug)]
 pub enum VMInitializer {
@@ -20,7 +20,7 @@ pub enum VMInitializer {
     FileBacked {
         file: FileBacked,
     },
-    // For ELF files, there is specical handling to not copy all the contents of the file. This is only used for tracking.
+    // For ELF files, there is special handling to not copy all the contents of the file. This is only used for tracking.
     ElfSpecific {
         elf_file: FileRef,
     },
@@ -150,6 +150,10 @@ impl FileBacked {
             None
         }
     }
+
+    pub fn metadata(&self) -> Metadata {
+        self.file.metadata().unwrap()
+    }
 }
 
 #[derive(Clone, Copy, Debug, PartialEq)]
@@ -258,6 +262,13 @@ impl VMMapOptions {
     pub fn initializer(&self) -> &VMInitializer {
         &self.initializer
     }
+
+    pub fn is_shared(&self) -> bool {
+        if let Some(backed_file) = self.initializer.backed_file() {
+            return backed_file.need_write_back();
+        }
+        false
+    }
 }
 
 #[derive(Clone, Copy, PartialEq)]
@@ -407,7 +418,7 @@ pub trait VMRemapParser {
         // For Linux, writing to either memory vma or the file will update the other two equally. But we won't be able to support this before
         // we really have paging. Thus, if the old_range is not equal to a recorded vma, we will just return with error.
         if writeback_file.is_some() && &old_range != vma.range() {
-            return_errno!(EINVAL, "Known limition")
+            return_errno!(EINVAL, "Known limitation")
         }
 
         // Implement mremap as one optional mmap followed by one optional munmap.
@@ -449,13 +460,13 @@ pub trait VMRemapParser {
                 (Some(mmap_opts), ret_addr)
             }
             (MRemapFlags::MayMove, VMRemapSizeType::Growing, None) => {
-                let prefered_new_range =
+                let preferred_new_range =
                     VMRange::new_with_size(old_addr + old_size, new_size - old_size)?;
-                if self.is_free_range(&prefered_new_range) {
+                if self.is_free_range(&preferred_new_range) {
                     // Don't need to move the old range
                     let mmap_ops = VMMapOptionsBuilder::default()
-                        .size(prefered_new_range.size())
-                        .addr(VMMapAddr::Need(prefered_new_range.start()))
+                        .size(preferred_new_range.size())
+                        .addr(VMMapAddr::Need(preferred_new_range.start()))
                         .perms(perms)
                         .initializer(VMInitializer::DoNothing())
                         .build()?;
@@ -475,9 +486,9 @@ pub trait VMRemapParser {
                 }
             }
             (MRemapFlags::MayMove, VMRemapSizeType::Growing, Some((backed_file, offset))) => {
-                let prefered_new_range =
+                let preferred_new_range =
                     VMRange::new_with_size(old_addr + old_size, new_size - old_size)?;
-                if self.is_free_range(&prefered_new_range) {
+                if self.is_free_range(&preferred_new_range) {
                     // Don't need to move the old range
                     let vm_initializer_for_new_range = VMInitializer::FileBacked {
                         file: FileBacked::new(
@@ -487,8 +498,8 @@ pub trait VMRemapParser {
                         ),
                     };
                     let mmap_ops = VMMapOptionsBuilder::default()
-                        .size(prefered_new_range.size())
-                        .addr(VMMapAddr::Need(prefered_new_range.start()))
+                        .size(preferred_new_range.size())
+                        .addr(VMMapAddr::Need(preferred_new_range.start()))
                         .perms(perms)
                         .initializer(vm_initializer_for_new_range)
                         .build()?;

test/Makefile

@@ -22,7 +22,7 @@ TESTS ?= env empty hello_world malloc mmap file fs_perms getpid spawn sched pipe
 	truncate readdir mkdir open stat link symlink chmod chown tls pthread system_info rlimit \
 	server server_epoll unix_socket cout hostfs cpuid rdtsc device sleep exit_group posix_flock \
 	ioctl fcntl eventfd emulate_syscall access signal sysinfo prctl rename procfs wait \
-	spawn_attribute exec statfs random umask pgrp vfork mount flock utimes shm epoll brk
+	spawn_attribute exec statfs random umask pgrp vfork mount flock utimes shm epoll brk posix_shm
 # Benchmarks: need to be compiled and run by bench-% target
 BENCHES := spawn_and_exit_latency pipe_throughput unix_socket_throughput
 

test/posix_shm/Makefile

@@ -0,0 +1,10 @@
+include ../test_common.mk
+
+EXTRA_C_FLAGS := -Wno-stringop-truncation -Wno-nonnull
+EXTRA_LINK_FLAGS := -lrt
+BIN_ARGS :=
+
+CUSTOM_PRE_BUILD := 1
+
+custom_pre_build:
+	@cp /lib/x86_64-linux-gnu/librt.so.1 $(BUILD_DIR)/test/image/opt/occlum/glibc/lib/

test/posix_shm/main.c

@@ -0,0 +1,170 @@
+// C test program illustrating POSIX shared-memory API.
+#define _GNU_SOURCE
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <fcntl.h>
+#include <sys/shm.h>
+#include <sys/stat.h>
+#include <sys/mman.h>
+#include "test.h"
+
+// ============================================================================
+// Helper macros
+// ============================================================================
+
+#define SHM_OBJ "shm_test" // Name of shared memory object
+#define SHM_SIZE 0x1000 // Size (in bytes) of shared memory object
+
+/* Messages to communicate through shared memory */
+#define MSG0 "1st Hello"
+#define MSG1 "2nd Hello"
+#define MSG2 "3rd Hello"
+#define MSG3 "4th Hello"
+#define MSG_SIZE strlen(MSG0)
+
+// ============================================================================
+// Test cases
+// ============================================================================
+
+int producer_process() {
+    // Shared memory file descriptor
+    int shm_fd;
+    // Shared memory buffer
+    void *shm_buf;
+
+    // Create the shared memory object
+    shm_fd = shm_open(SHM_OBJ, O_CREAT | O_RDWR, 0666);
+    if (shm_fd < 0) {
+        THROW_ERROR("shm_open failed");
+    }
+
+    // Configure the size of the shared memory object
+    if (ftruncate(shm_fd, SHM_SIZE) < 0) {
+        THROW_ERROR("ftruncate error");
+    }
+
+    // Memory map the shared memory object
+    shm_buf = mmap(NULL, SHM_SIZE, PROT_READ | PROT_WRITE, MAP_SHARED, shm_fd, 0);
+    if (shm_buf == MAP_FAILED) {
+        THROW_ERROR("mmap(MAP_SHARED) failed");
+    }
+
+    // Fork a child and launch consumer process
+    pid_t child_pid = vfork();
+    if (child_pid < 0) {
+        // THROW_ERROR("Spawn a child process failed");
+        perror("Spawn a child process failed");
+        return -1;
+    } else if (child_pid == 0) {
+        execl("/bin/posix_shm", NULL, NULL);
+        THROW_ERROR("exec failed");
+    }
+
+    // Communicate through shared memory
+    strncpy(shm_buf, MSG0, MSG_SIZE);
+    printf("[Producer] send %s\n", MSG0);
+    while (1) {
+        if (strncmp(shm_buf, MSG1, MSG_SIZE) != 0) {
+            sleep(1);
+            continue;
+        }
+        printf("[Producer] receive %s\n", MSG1);
+        strncpy(shm_buf, MSG2, MSG_SIZE);
+        printf("[Producer] send %s\n", MSG2);
+        while (1) {
+            if (strncmp(shm_buf, MSG3, MSG_SIZE) != 0) {
+                sleep(1);
+                continue;
+            }
+            printf("[Producer] receive %s\n", MSG3);
+            break;
+        }
+        break;
+    }
+
+    // Unmap the shared memory
+    if (munmap(shm_buf, SHM_SIZE) < 0) {
+        THROW_ERROR("munmap failed");
+    }
+
+    // Unlink the shared memory object
+    shm_unlink(SHM_OBJ);
+    return 0;
+}
+
+int consumer_process() {
+    // Shared memory file descriptor
+    int shm_fd;
+    // Shared memory buffer
+    void *shm_buf;
+
+    // Create the shared memory object
+    shm_fd = shm_open(SHM_OBJ, O_CREAT | O_RDWR, 0666);
+    if (shm_fd < 0) {
+        THROW_ERROR("shm_open failed");
+    }
+
+    // Configure the size of the shared memory object
+    if (ftruncate(shm_fd, SHM_SIZE) < 0) {
+        THROW_ERROR("ftruncate error");
+    }
+
+    // Memory map the shared memory object
+    shm_buf = mmap(NULL, SHM_SIZE, PROT_READ | PROT_WRITE, MAP_SHARED, shm_fd, 0);
+    if (shm_buf == MAP_FAILED) {
+        THROW_ERROR("mmap(MAP_SHARED) failed");
+    }
+
+    while (1) {
+        if (strncmp(shm_buf, MSG0, MSG_SIZE) != 0) {
+            sleep(1);
+            continue;
+        }
+        printf("[Consumer] receive %s\n", MSG0);
+        strncpy(shm_buf, MSG1, MSG_SIZE);
+        printf("[Consumer] send %s\n", MSG1);
+        while (1) {
+            if (strncmp(shm_buf, MSG2, MSG_SIZE) != 0) {
+                sleep(1);
+                continue;
+            }
+            printf("[Consumer] receive %s\n", MSG2);
+            strncpy(shm_buf, MSG3, MSG_SIZE);
+            printf("[Consumer] send %s\n", MSG3);
+            break;
+        }
+        break;
+    }
+
+    // Unmap the shared memory
+    if (munmap(shm_buf, SHM_SIZE) < 0) {
+        THROW_ERROR("munmap failed");
+    }
+
+    // Unlink the shared memory object
+    shm_unlink(SHM_OBJ);
+    return 0;
+}
+
+int test_posix_shm() {
+    return producer_process();
+}
+
+// ============================================================================
+// Test suite main
+// ============================================================================
+
+static test_case_t test_cases[] = {
+    TEST_CASE(test_posix_shm),
+};
+
+int main(int argc, const char *argv[]) {
+    if (argc == 1) {
+        // Producer process
+        return test_suite_run(test_cases, ARRAY_SIZE(test_cases));
+    } else {
+        // Consumer process
+        return consumer_process();
+    }
+}