Enable Thread Local Storage (TLS)

Add arch_prctl and pass necessary ELF info to libc via aux vector

src/libos/src/process/arch_prctl.rs

@@ -0,0 +1,43 @@
+use super::*;
+
+#[allow(non_camel_case_types)]
+#[derive(Debug)]
+pub enum ArchPrctlCode {
+    ARCH_SET_GS = 0x1001,
+    ARCH_SET_FS = 0x1002,
+    ARCH_GET_FS = 0x1003,
+    ARCH_GET_GS = 0x1004,
+}
+
+impl ArchPrctlCode {
+    pub fn from_u32(bits: u32) -> Result<ArchPrctlCode, Error> {
+        match bits {
+            0x1001 => Ok(ArchPrctlCode::ARCH_SET_GS),
+            0x1002 => Ok(ArchPrctlCode::ARCH_SET_FS),
+            0x1003 => Ok(ArchPrctlCode::ARCH_GET_FS),
+            0x1004 => Ok(ArchPrctlCode::ARCH_GET_GS),
+            _ => errno!(EINVAL, "Unknown code for arch_prctl"),
+        }
+    }
+}
+
+pub fn do_arch_prctl(code: ArchPrctlCode, addr: *mut usize) -> Result<(), Error> {
+    info!("do_arch_prctl: code: {:?}, addr: {:#o}", code, addr as usize);
+    match code {
+       ArchPrctlCode::ARCH_SET_FS => {
+            let current_ref = get_current();
+            let mut current = current_ref.lock().unwrap();
+            let task = &mut current.task;
+            task.user_fsbase_addr = addr as usize;
+       },
+       ArchPrctlCode::ARCH_GET_FS => {
+            let current_ref = get_current();
+            let current = current_ref.lock().unwrap();
+            let task = &current.task;
+            unsafe { *addr = task.user_fsbase_addr; }
+       },
+       ArchPrctlCode::ARCH_SET_GS | ArchPrctlCode::ARCH_GET_GS
+           => return errno!(EINVAL, "GS cannot be accessed from the user space"),
+    }
+    Ok(())
+}

src/libos/src/process/mod.rs

@@ -8,6 +8,7 @@ pub use self::spawn::{do_spawn, FileAction};
 pub use self::wait::{WaitQueue, Waiter};
 pub use self::thread::{do_clone, CloneFlags, ThreadGroup};
 pub use self::futex::{FutexOp, FutexFlags, futex_op_and_flags_from_u32, futex_wake, futex_wait};
+pub use self::arch_prctl::{ArchPrctlCode, do_arch_prctl};
 
 #[allow(non_camel_case_types)]
 pub type pid_t = u32;
@@ -65,6 +66,7 @@ mod task;
 mod wait;
 mod thread;
 mod futex;
+mod arch_prctl;
 
 use self::task::Task;
 use super::*;

src/libos/src/process/spawn/elf_helper.rs

@@ -6,6 +6,22 @@ use xmas_elf::symbol_table::Entry;
 use xmas_elf::symbol_table::{DynEntry64, Entry64};
 use xmas_elf::{program, sections, ElfFile, P64};
 
+#[derive(Clone, Default, Copy, Debug)]
+pub struct ProgramHeaderInfo {
+    pub addr: usize,
+    pub entry_size: usize,
+    pub entry_num: usize,
+}
+
+pub fn get_program_header_info(elf_file: &ElfFile) -> Result<ProgramHeaderInfo, Error> {
+    let elf_header = &elf_file.header.pt2;
+    Ok(ProgramHeaderInfo {
+        addr: elf_header.ph_offset() as usize,
+        entry_size: elf_header.ph_entry_size() as usize,
+        entry_num: elf_header.ph_count() as usize,
+    })
+}
+
 pub fn print_program_headers(elf_file: &ElfFile) -> Result<(), Error> {
     println!("Program headers:");
     let ph_iter = elf_file.program_iter();
@@ -87,18 +103,8 @@ pub fn get_code_program_header<'b, 'a: 'b>(
 }
 
 pub fn get_start_address<'b, 'a: 'b>(elf_file: &'b ElfFile<'a>) -> Result<usize, Error> {
-    let sym_entries = get_sym_entries(elf_file)?;
-
-    for sym_entry in sym_entries {
-        let sym_str = sym_entry
-            .get_name(elf_file)
-            .map_err(|e| Error::new(Errno::ENOEXEC, "Failed to get the name of a symbol"))?;
-        if sym_str == "_start" {
-            return Ok(sym_entry.value() as usize);
-        }
-    }
-
-    Err((Errno::ENOEXEC, "Failed to get the _start symbol").into())
+    let elf_header = &elf_file.header.pt2;
+    Ok(elf_header.entry_point() as usize)
 }
 
 pub fn get_sym_entries<'b, 'a: 'b>(elf_file: &'b ElfFile<'a>) -> Result<&'a [Entry64], Error> {

src/libos/src/process/spawn/init_stack.rs

@@ -58,7 +58,6 @@ pub fn do_init(
     let envp_cloned = clone_cstrings_on_stack(&stack_buf, envp)?;
     let argv_cloned = clone_cstrings_on_stack(&stack_buf, argv)?;
     dump_auxtbl_on_stack(&stack_buf, auxtbl)?;
-    dump_auxtbl_on_stack(&stack_buf, auxtbl)?;
     dump_cstrptrs_on_stack(&stack_buf, &envp_cloned);
     dump_cstrptrs_on_stack(&stack_buf, &argv_cloned);
     stack_buf.put(argv.len() as u64);

src/libos/src/process/spawn/init_vm.rs

@@ -25,10 +25,10 @@ pub fn do_init(elf_file: &ElfFile, elf_buf: &[u8]) -> Result<ProcessVM, Error> {
 
     // Calculate the "real" addresses
     let process_base_addr = process_vm.get_base_addr();
-    let code_start = process_base_addr + code_start;
-    let code_end = process_base_addr + code_end;
-    let data_start = process_base_addr + data_start;
-    let data_end = process_base_addr + data_end;
+    let code_start = code_start + process_base_addr;
+    let code_end = code_end + process_base_addr;
+    let data_start = data_start + process_base_addr;
+    let data_end = data_end + process_base_addr;
     code_seg.set_runtime_info(process_base_addr, code_start, code_end);
     data_seg.set_runtime_info(process_base_addr, data_start, data_end);
 
@@ -54,8 +54,8 @@ fn reloc_symbols(process_base_addr: usize, elf_file: &ElfFile) -> Result<(), Err
              rela_entry.get_addend());
         */
 
-        /* reloc type == R_X86_64_RELATIVE */
         match rela_entry.get_type() {
+            // reloc type == R_X86_64_RELATIVE
             8 if rela_entry.get_symbol_table_index() == 0 => {
                 let rela_addr = process_base_addr + rela_entry.get_offset() as usize;
                 let rela_val = process_base_addr + rela_entry.get_addend() as usize;

src/libos/src/process/spawn/mod.rs

@@ -54,16 +54,18 @@ pub fn do_spawn<P: AsRef<Path>>(
     let (new_pid, new_process_ref) = {
         let cwd = parent_ref.lock().unwrap().get_cwd().to_owned();
         let vm = init_vm::do_init(&elf_file, &elf_buf[..])?;
+        let base_addr = vm.get_base_addr();
+        let program_entry = {
+            let program_entry = base_addr + elf_helper::get_start_address(&elf_file)?;
+            if !vm.get_code_vma().contains_obj(program_entry, 16) {
+                return Err(Error::new(Errno::EINVAL, "Invalid program entry"));
+            }
+            program_entry
+        };
+        let auxtbl = init_auxtbl(base_addr, program_entry, &elf_file)?;
         let task = {
-            let program_entry = {
-                let program_entry = vm.get_base_addr() + elf_helper::get_start_address(&elf_file)?;
-                if !vm.get_code_vma().contains_obj(program_entry, 16) {
-                    return Err(Error::new(Errno::EINVAL, "Invalid program entry"));
-                }
-                program_entry
-            };
             let stack_top = vm.get_stack_top();
-            init_task(program_entry, stack_top, argv, envp)?
+            init_task(program_entry, stack_top, argv, envp, &auxtbl)?
         };
         let vm_ref = Arc::new(SgxMutex::new(vm));
         let files_ref = {
@@ -119,8 +121,9 @@ fn init_task(
     stack_top: usize,
     argv: &[CString],
     envp: &[CString],
+    auxtbl: &AuxTable,
 ) -> Result<Task, Error> {
-    let user_stack = init_stack(stack_top, argv, envp)?;
+    let user_stack = init_stack::do_init(stack_top, 4096, argv, envp, auxtbl)?;
     Ok(Task {
         user_stack_addr: user_stack,
         user_entry_addr: user_entry,
@@ -128,7 +131,7 @@ fn init_task(
     })
 }
 
-fn init_stack(stack_top: usize, argv: &[CString], envp: &[CString]) -> Result<usize, Error> {
+fn init_auxtbl(base_addr: usize, program_entry: usize, elf_file: &ElfFile) -> Result<AuxTable, Error> {
     let mut auxtbl = AuxTable::new();
     auxtbl.set_val(AuxKey::AT_PAGESZ, 4096)?;
     auxtbl.set_val(AuxKey::AT_UID, 0)?;
@@ -137,7 +140,16 @@ fn init_stack(stack_top: usize, argv: &[CString], envp: &[CString]) -> Result<us
     auxtbl.set_val(AuxKey::AT_EGID, 0)?;
     auxtbl.set_val(AuxKey::AT_SECURE, 0)?;
 
-    init_stack::do_init(stack_top, 4096, argv, envp, &auxtbl)
+    let ph = elf_helper::get_program_header_info(elf_file)?;
+    auxtbl.set_val(AuxKey::AT_PHDR, (base_addr + ph.addr) as u64)?;
+    auxtbl.set_val(AuxKey::AT_PHENT, ph.entry_size as u64)?;
+    auxtbl.set_val(AuxKey::AT_PHNUM, ph.entry_num as u64)?;
+
+    auxtbl.set_val(AuxKey::AT_ENTRY, program_entry as u64)?;
+    // TODO: init AT_EXECFN
+    // auxtbl.set_val(AuxKey::AT_EXECFN, "program_name")?;
+
+    Ok(auxtbl)
 }
 
 fn parent_adopts_new_child(parent_ref: &ProcessRef, child_ref: &ProcessRef) {

src/libos/src/syscall/mod.rs

@@ -122,6 +122,8 @@ pub extern "C" fn dispatch_syscall(
 
         SYS_GETTIMEOFDAY => do_gettimeofday(arg0 as *mut timeval_t),
 
+        SYS_ARCH_PRCTL => do_arch_prctl(arg0 as u32, arg1 as *mut usize),
+
         _ => do_unknown(num),
     };
     debug!("syscall return: {:?}", ret);
@@ -653,3 +655,10 @@ fn do_unlink(path: *const i8) -> Result<isize, Error> {
     fs::do_unlink(&path)?;
     Ok(0)
 }
+
+fn do_arch_prctl(code: u32, addr: *mut usize) -> Result<isize, Error> {
+    let code = process::ArchPrctlCode::from_u32(code)?;
+    check_mut_ptr(addr)?;
+    process::do_arch_prctl(code, addr).map(|_| 0)
+}
+

src/libos/src/syscall/syscall_entry_x86-64.S

@@ -19,8 +19,10 @@ __occlum_syscall:
     bndcl %rsp, %bnd0
     bndcu %rsp, %bnd0
 
-    // Save the user stack
+    // Save the callee-saved registers
     pushq %rbp
+    pushq %r12
+    // Save the user stack
     movq %rsp, %rbp
 
     // Get current task
@@ -39,11 +41,13 @@ __occlum_syscall:
     // addq 0x08, %rsp
 
     // Use user fsbase
-    movq TASK_KERNEL_FSBASE_ADDR(%r12), %r11
+    movq TASK_USER_FSBASE_ADDR(%r12), %r11
     wrfsbase %r11
 
-    // Restore the user stack
+    // Switch to the user stack
     movq %rbp, %rsp
+    // Restore callee-saved registers
+    popq %r12
     popq %rbp
 
     // Check return target is a valid instruction (i.e., a cfi_label)

test/Makefile

@@ -4,7 +4,7 @@ PROJECT_DIR := $(realpath $(CUR_DIR)/../)
 # Dependencies: need to be compiled but not to run by any Makefile target
 TEST_DEPS := dev_null
 # Tests: need to be compiled and run by test-% target
-TESTS := empty argv hello_world malloc file getpid spawn pipe time truncate readdir mkdir link clone
+TESTS := empty argv hello_world malloc file getpid spawn pipe time truncate readdir mkdir link clone tls
 # Benchmarks: need to be compiled and run by bench-% target
 BENCHES := spawn_and_exit_latency pipe_throughput
 

test/tls/Makefile

@@ -0,0 +1,5 @@
+include ../test_common.mk
+
+EXTRA_C_FLAGS :=
+EXTRA_LINK_FLAGS :=
+BIN_ARGS :=

test/tls/main.c

@@ -0,0 +1,11 @@
+volatile int g_int = 0;
+static void use_int(int* a) {
+    g_int += *a;
+}
+
+__thread int tls_g_int = 0;
+
+int main(int argc, const char* argv[]) {
+    use_int(&tls_g_int);
+    return g_int;
+}