This is the mega-ucontext commit. It replaces the task switching mechanism with a new one inspired by ucontext. It works under Linux, OS X and Windows, and is Valgrind clean on Linux and OS X (provided the runtime is built with gcc).

This commit also moves yield and join to the standard library, as requested in #42. Join is currently a no-op though.
This commit is contained in:
Eric Holk 2011-05-31 17:44:54 -07:00 committed by Graydon Hoare
parent 1595c9d767
commit d1857d30fc
30 changed files with 419 additions and 443 deletions

4
configure vendored
View File

@ -176,11 +176,11 @@ fi
step_msg "making directories"
for i in \
doc \
rt rt/isaac rt/bigint rt/sync rt/test \
rt rt/isaac rt/bigint rt/sync rt/test rt/arch/i386 \
rustllvm \
dl stage0 stage1 stage2 stage3 \
test/run-pass test/run-fail test/compile-fail \
test/bench/99-bottles test/bench/shootout
test/bench/99-bottles test/bench/shootout
do
make_dir $i
done

View File

@ -27,11 +27,12 @@ RUNTIME_CS := rt/sync/timer.cpp \
rt/memory_region.cpp \
rt/test/rust_test_harness.cpp \
rt/test/rust_test_runtime.cpp \
rt/test/rust_test_util.cpp
rt/test/rust_test_util.cpp \
rt/arch/i386/context.cpp \
RUNTIME_LL := rt/new_exit.ll rt/vec_append.ll
RUNTIME_LL := rt/vec_append.ll
RUNTIME_S := rt/activate_glue.s rt/yield_glue.s
RUNTIME_S := rt/arch/i386/_context.s
RUNTIME_HDR := rt/globals.h \
rt/rust.h \
@ -60,10 +61,12 @@ RUNTIME_HDR := rt/globals.h \
rt/memory.h \
rt/test/rust_test_harness.h \
rt/test/rust_test_runtime.h \
rt/test/rust_test_util.h
rt/test/rust_test_util.h \
rt/arch/i386/context.h \
RUNTIME_DEF := rt/rustrt$(CFG_DEF_SUFFIX)
RUNTIME_INCS := -I $(S)src/rt/isaac -I $(S)src/rt/uthash
RUNTIME_INCS := -I $(S)src/rt/isaac -I $(S)src/rt/uthash \
-I $(S)src/rt/arch/i386
RUNTIME_OBJS := $(RUNTIME_CS:.cpp=.o) $(RUNTIME_LL:.ll=.o) $(RUNTIME_S:.s=.o)
RUNTIME_LIBS := $(CFG_GCCISH_POST_LIB_FLAGS)

View File

@ -93,10 +93,6 @@ fn vec_append_glue_name() -> str {
ret "rust_vec_append_glue";
}
fn yield_glue_name() -> str {
ret "rust_yield_glue";
}
fn no_op_type_glue_name() -> str {
ret "rust_no_op_type_glue";
}

View File

@ -114,7 +114,7 @@ fn declare_upcalls(type_names tn, ModuleRef llmod) -> @upcalls {
T_ptr(T_tydesc(tn))),
new_task=d("new_task", [T_ptr(T_str())], T_taskptr(tn)),
start_task=d("start_task", [T_taskptr(tn),
T_int(), T_int()],
T_int(), T_int(), T_size_t()],
T_taskptr(tn)),
new_thread=d("new_thread", [T_ptr(T_i8())], T_taskptr(tn)),
start_thread=d("start_thread", [T_taskptr(tn), T_int(), T_int(),

View File

@ -85,8 +85,7 @@ state obj namegen(mutable int i) {
type derived_tydesc_info = rec(ValueRef lltydesc, bool escapes);
type glue_fns = rec(ValueRef yield_glue,
ValueRef no_op_type_glue,
type glue_fns = rec(ValueRef no_op_type_glue,
ValueRef vec_append_glue);
type tydesc_info = rec(ty::t ty,
@ -6454,13 +6453,12 @@ fn trans_spawn(&@block_ctxt cx,
auto wrapper = mk_spawn_wrapper(bcx, func, args_ty);
bcx = wrapper.bcx;
auto llfnptr_i = bcx.build.PointerCast(wrapper.val, T_int());
// TODO: this next line might be necessary...
//llfnptr_i = bcx.build.Load(llfnptr_i);
// And start the task
auto args_size = size_of(bcx, args_ty).val;
bcx.build.Call(bcx.fcx.lcx.ccx.upcalls.start_task,
[bcx.fcx.lltaskptr, new_task,
llfnptr_i, llargs_i]);
llfnptr_i, llargs_i, args_size]);
auto task_ty = node_ann_type(bcx.fcx.lcx.ccx, ann);
auto dropref = clean(bind drop_ty(_, new_task, task_ty));
@ -8545,8 +8543,7 @@ fn vec_p0(&@block_ctxt bcx, ValueRef v) -> ValueRef {
}
fn make_glues(ModuleRef llmod, &type_names tn) -> @glue_fns {
ret @rec(yield_glue = decl_glue(llmod, tn, abi::yield_glue_name()),
no_op_type_glue = decl_no_op_type_glue(llmod, tn),
ret @rec(no_op_type_glue = decl_no_op_type_glue(llmod, tn),
vec_append_glue = make_vec_append_glue(llmod, tn));
}

View File

@ -18,7 +18,7 @@ mod str;
mod io;
mod sys;
mod _task;
mod task;
// Utility modules.
@ -33,7 +33,7 @@ auth os_fs = unsafe;
auth run = unsafe;
auth str = unsafe;
auth vec = unsafe;
auth _task = unsafe;
auth task = unsafe;
auth dbg = unsafe;

View File

@ -1,5 +1,6 @@
native "rust" mod rustrt {
fn task_sleep(uint time_in_us);
fn task_yield();
}
/**
@ -11,6 +12,15 @@ fn sleep(uint time_in_us) {
ret rustrt::task_sleep(time_in_us);
}
fn yield() {
ret rustrt::task_yield();
}
fn join(task t) {
// TODO: figure out how to pass tasks to the runtime and call the builtin
// join.
}
// Local Variables:
// mode: rust;
// fill-column: 78;

View File

@ -1,89 +0,0 @@
/*
* This is a bit of glue-code.
*
* - save regs on C stack
* - save sp to task.runtime_sp (runtime_sp is thus always aligned)
* - load saved task sp (switch stack)
* - restore saved task regs
* - return to saved task pc
*
* Our incoming stack looks like this:
*
* *esp+4 = [arg1 ] = task ptr
* *esp = [retpc ]
*/
.globl new_rust_activate_glue
.balign 4
new_rust_activate_glue:
movl 4(%esp), %ecx # ecx = rust_task
pushl %ebp
pushl %edi
pushl %esi
pushl %ebx
movl %esp, 12(%ecx)
movl 16(%ecx), %esp
/*
* There are two paths we can arrive at this code from:
*
*
* 1. We are activating a task for the first time. When we switch
* into the task stack and 'ret' to its first instruction, we'll
* start doing whatever the first instruction says. Probably
* saving registers and starting to establish a frame. Harmless
* stuff, doesn't look at task->rust_sp again except when it
* clobbers it during a later native call.
*
*
* 2. We are resuming a task that was descheduled by the yield glue
* below. When we switch into the task stack and 'ret', we'll be
* ret'ing to a very particular instruction:
*
* "esp <- task->rust_sp"
*
* this is the first instruction we 'ret' to after this glue,
* because it is the first instruction following *any* native
* call, and the task we are activating was descheduled
* mid-native-call.
*
* Unfortunately for us, we have already restored esp from
* task->rust_sp and are about to eat the 5 words off the top of
* it.
*
*
* | ... | <-- where esp will be once we restore + ret, below,
* | retpc | and where we'd *like* task->rust_sp to wind up.
* | ebp |
* | edi |
* | esi |
* | ebx | <-- current task->rust_sp == current esp
*
*
* This is a problem. If we return to "esp <- task->rust_sp" it
* will push esp back down by 5 words. This manifests as a rust
* stack that grows by 5 words on each yield/reactivate. Not
* good.
*
* So what we do here is just adjust task->rust_sp up 5 words as
* well, to mirror the movement in esp we're about to
* perform. That way the "esp <- task->rust_sp" we 'ret' to below
* will be a no-op. Esp won't move, and the task's stack won't
* grow.
*/
addl $20, 16(%ecx)
/*
* In most cases, the function we're returning to (activating)
* will have saved any caller-saves before it yielded via native call,
* so no work to do here. With one exception: when we're initially
* activating, the task needs to be in the fastcall 2nd parameter
* expected by the rust main function. That's edx.
*/
mov %ecx, %edx
popl %ebx
popl %esi
popl %edi
popl %ebp
ret

View File

@ -0,0 +1,86 @@
.text
/*
Callee save registers:
ebp, ebx, esi, edi
Caller save registers:
eax, ecx, edx
*/
/*
Saves a set of registers. This is used by our implementation of
getcontext.
The registers_t variable is in (%esp)
*/
.globl get_registers
get_registers:
movl 4(%esp), %eax
movl %eax, 0(%eax)
movl %ebx, 4(%eax)
movl %ecx, 8(%eax)
movl %edx, 12(%eax)
movl %ebp, 16(%eax)
movl %esi, 20(%eax)
movl %edi, 24(%eax)
movl %esp, 28(%eax)
movw %cs, 32(%eax)
movw %ds, 34(%eax)
movw %ss, 36(%eax)
movw %es, 38(%eax)
movw %fs, 40(%eax)
movw %gs, 42(%eax)
// save the flags
pushf
popl %ecx
movl %ecx, 44(%eax)
// save the return address as the instruction pointer
movl 0(%esp), %ecx
movl %ecx, 48(%eax)
// return 0
xor %eax, %eax
ret
.globl set_registers
set_registers:
movl 4(%esp), %eax
movl 4(%eax), %ebx
// save ecx for later...
movl 12(%eax), %edx
movl 16(%eax), %ebp
movl 20(%eax), %esi
movl 24(%eax), %edi
movl 28(%eax), %esp
// We can't actually change this...
//movl 32(%eax), %cs
movw 34(%eax), %ds
movw 36(%eax), %ss
movw 38(%eax), %es
movw 40(%eax), %fs
movw 42(%eax), %gs
// restore the flags
movl 44(%eax), %ecx
push %ecx
popf
// get ready to return back to the old eip
// We could write this directly to 0(%esp), but Valgrind on OS X
// complains.
pop %ecx
mov 48(%eax), %ecx
push %ecx
//movl %ecx, 0(%esp)
// okay, now we can restore ecx.
movl 8(%eax), %ecx
// return 1 to the saved eip
movl $1, %eax
ret

View File

@ -0,0 +1,81 @@
#include "context.h"
#include "../../rust.h"
#include <stdio.h>
#include <stdlib.h>
extern "C" uint32_t CDECL get_registers(registers_t *regs)
asm ("get_registers");
extern "C" uint32_t CDECL set_registers(registers_t *regs)
asm ("set_registers");
context::context()
: next(NULL)
{
get_registers(&regs);
}
void context::set()
{
//printf("Activating %p...\n", this);
set_registers(&regs);
}
void context::swap(context &out)
{
//printf("Swapping to %p and saving in %p\n", this, &out);
uint32_t r = get_registers(&out.regs);
//printf("get_registers = %d, sp = 0x%x\n", r, out.regs.esp);
if(!r) {
set();
}
//printf("Resumed %p...\n", &out);
}
void context::call(void *f, void *arg, void *stack) {
// set up the trampoline frame
uint32_t *sp = (uint32_t *)stack;
*--sp = (uint32_t)this;
*--sp = (uint32_t)arg;
*--sp = 0xdeadbeef; //(uint32_t)ctx_trampoline1;
*--sp = 0xdeadbeef;
regs.esp = (uint32_t)sp;
regs.eip = (uint32_t)f;
}
#if 0
// This is some useful code to check how the registers struct got
// layed out in memory.
int main() {
registers_t regs;
printf("Register offsets\n");
#define REG(r) \
printf(" %6s: +%ld\n", #r, (intptr_t)&regs.r - (intptr_t)&regs);
REG(eax);
REG(ebx);
REG(ecx);
REG(edx);
REG(ebp);
REG(esi);
REG(edi);
REG(esp);
REG(cs);
REG(ds);
REG(ss);
REG(es);
REG(fs);
REG(gs);
REG(eflags);
REG(eip);
return 0;
}
#endif

View File

@ -0,0 +1,35 @@
// -*- mode: c++ -*-
#ifndef CONTEXT_H
#define CONTEXT_H
#include <inttypes.h>
struct registers_t {
// general purpose registers
uint32_t eax, ebx, ecx, edx, ebp, esi, edi, esp;
// segment registers
uint16_t cs, ds, ss, es, fs, gs;
uint32_t eflags;
uint32_t eip;
};
class context {
registers_t regs;
public:
context();
context *next;
void set();
void swap(context &out);
void call(void *f, void *arg, void *sp);
};
#endif

View File

@ -1,9 +1,12 @@
#ifndef GLOBALS_H
#define GLOBALS_H
#ifndef RUST_INTERNAL_H
// these are defined in two files and GCC complains
#define __STDC_LIMIT_MACROS 1
#define __STDC_CONSTANT_MACROS 1
#define __STDC_FORMAT_MACROS 1
#endif
#include <stdlib.h>
#include <stdint.h>
@ -32,4 +35,16 @@ extern "C" {
#error "Platform not supported."
#endif
#define CHECKED(call) \
{ \
int res = (call); \
if(0 != res) { \
fprintf(stderr, \
#call " failed in %s at line %d, result = %d " \
"(%s) \n", \
__FILE__, __LINE__, res, strerror(res)); \
abort(); \
} \
}
#endif /* GLOBALS_H */

View File

@ -1,43 +0,0 @@
declare fastcc i32 @"\01rust_native_rust_local_copy"(i32, i32)
module asm "\09.globl rust_native_rust_local_copy"
module asm "\09.balign 4"
module asm "rust_native_rust_local_copy:"
module asm "\09.cfi_startproc"
module asm "\09pushl %ebp"
module asm "\09.cfi_def_cfa_offset 8"
module asm "\09.cfi_offset %ebp, -8"
module asm "\09pushl %edi"
module asm "\09.cfi_def_cfa_offset 12"
module asm "\09pushl %esi"
module asm "\09.cfi_def_cfa_offset 16"
module asm "\09pushl %ebx"
module asm "\09.cfi_def_cfa_offset 20"
module asm "\09movl %esp, %ebp # ebp = rust_sp"
module asm "\09.cfi_def_cfa_register %ebp"
module asm "\09movl %esp, 16(%edx)"
module asm "\09movl 12(%edx), %esp"
module asm "\09subl $4, %esp # esp -= args"
module asm "\09andl $~0xf, %esp # align esp down"
module asm "\09movl %edx, (%esp)"
module asm "\09movl %edx, %edi # save task from edx to edi"
module asm "\09call *%ecx # call *%ecx"
module asm "\09movl %edi, %edx # restore edi-saved task to edx"
module asm "\09movl 16(%edx), %esp"
module asm "\09popl %ebx"
module asm "\09popl %esi"
module asm "\09popl %edi"
module asm "\09popl %ebp"
module asm "\09ret"
module asm "\09.cfi_endproc"
declare i32 @upcall_exit(i32)
define void @rust_new_exit_task_glue(i32, i32, i32, i32, i32) {
entry:
%5 = inttoptr i32 %0 to void (i32, i32, i32, i32)*
tail call fastcc void %5(i32 %1, i32 %2, i32 %3, i32 %4)
%6 = tail call fastcc i32 @"\01rust_native_rust_local_copy"(i32 ptrtoint (i32 (i32)* @upcall_exit to i32), i32 %2)
ret void
}

View File

@ -93,9 +93,13 @@ rust_start(uintptr_t main_fn, int argc, char **argv, void* crate_map) {
DLOG(dom, dom, "startup: arg[%d] = '%s'", i, args->argv[i]);
}
/*
uintptr_t main_args[4] = {0, 0, 0, (uintptr_t)args->args};
dom->root_task->start(main_fn,
(uintptr_t)&main_args, sizeof(main_args));
*/
dom->root_task->start(main_fn,
(uintptr_t)args->args, sizeof(args->args));
int ret = dom->start_main_loop();
delete args;

View File

@ -10,11 +10,14 @@
// 'cdecl' ABI only means anything on i386
#ifdef __WIN32__
#define CDECL __cdecl
#define FASTCALL __fastcall
#else
#define CDECL __attribute__((cdecl))
#define FASTCALL __attribute__((fastcall))
#endif
#else
#define CDECL
#define FASTCALL
#endif
/*

View File

@ -354,6 +354,16 @@ task_sleep(rust_task *task, size_t time_in_us) {
upcall_sleep(task, time_in_us);
}
extern "C" CDECL void
task_yield(rust_task *task) {
task->yield(1);
}
extern "C" CDECL void
task_join(rust_task *task, rust_task *join_task) {
// TODO
}
/* Debug builtins for std.dbg. */
static void

View File

@ -1,6 +1,7 @@
#include <stdarg.h>
#include "rust_internal.h"
#include "globals.h"
rust_dom::rust_dom(rust_kernel *kernel,
rust_message_queue *message_queue, rust_srv *srv,
@ -44,13 +45,17 @@ rust_dom::~rust_dom() {
#endif
}
extern "C" void new_rust_activate_glue(rust_task *)
asm("new_rust_activate_glue");
void
rust_dom::activate(rust_task *task) {
curr_task = task;
new_rust_activate_glue(task);
context ctx;
task->ctx.next = &ctx;
DLOG(this, task, "descheduling...");
task->ctx.swap(ctx);
DLOG(this, task, "task has returned");
curr_task = NULL;
}
@ -308,10 +313,14 @@ rust_dom::start_main_loop() {
scheduled_task->state->name,
scheduled_task->rust_sp);
/*
// These invariants are no longer valid, as rust_sp is not
// updated.
I(this, scheduled_task->rust_sp >=
(uintptr_t) &scheduled_task->stk->data[0]);
I(this, scheduled_task->rust_sp < scheduled_task->stk->limit);
*/
reap_dead_tasks();
}

View File

@ -1,9 +1,12 @@
#ifndef RUST_INTERNAL_H
#define RUST_INTERNAL_H
#ifndef GLOBALS_H
// these are defined in two files, and GCC complains.
#define __STDC_LIMIT_MACROS 1
#define __STDC_CONSTANT_MACROS 1
#define __STDC_FORMAT_MACROS 1
#endif
#define ERROR 0
@ -203,8 +206,6 @@ struct rust_timer {
#include "rust_util.h"
typedef void CDECL (*activate_glue_ty)(rust_task *);
struct type_desc {
// First part of type_desc is known to compiler.
// first_param = &descs[1] if dynamic, null if static.

View File

@ -187,8 +187,11 @@ rust_kernel::~rust_kernel() {
KLOG("freeing handles");
free_handles(_task_handles);
KLOG("..task handles freed");
free_handles(_port_handles);
KLOG("..port handles freed");
free_handles(_dom_handles);
KLOG("..dom handles freed");
KLOG("freeing queues");
@ -214,6 +217,7 @@ rust_kernel::free_handles(hash_map<T*, rust_handle<T>* > &map) {
T* key;
rust_handle<T> *value;
while (map.pop(&key, &value)) {
KLOG("...freeing " PTR, value);
delete value;
}
}

View File

@ -1,3 +1,4 @@
// -*- c++ -*-
#ifndef RUST_KERNEL_H
#define RUST_KERNEL_H

View File

@ -8,6 +8,8 @@
#include <execinfo.h>
#endif
#include "globals.h"
// Stacks
// FIXME (issue #151): This should be 0x300; the change here is for
@ -50,30 +52,6 @@ del_stk(rust_dom *dom, stk_seg *stk)
size_t const n_callee_saves = 4;
size_t const callee_save_fp = 0;
static uintptr_t
align_down(uintptr_t sp)
{
// There is no platform we care about that needs more than a
// 16-byte alignment.
return sp & ~(16 - 1);
}
static uintptr_t*
align_down(uintptr_t* sp)
{
return (uintptr_t*) align_down((uintptr_t)sp);
}
static void
make_aligned_room_for_bytes(uintptr_t*& sp, size_t n)
{
uintptr_t tmp = (uintptr_t) sp;
tmp = align_down(tmp - n) + n;
sp = (uintptr_t*) tmp;
}
rust_task::rust_task(rust_dom *dom, rust_task_list *state,
rust_task *spawner, const char *name) :
maybe_proxy<rust_task>(this),
@ -94,6 +72,7 @@ rust_task::rust_task(rust_dom *dom, rust_task_list *state,
handle(NULL)
{
LOGPTR(dom, "new task", (uintptr_t)this);
DLOG(dom, task, "sizeof(task) = %d (0x%x)", sizeof *this, sizeof *this);
if (spawner == NULL) {
ref_count = 0;
@ -135,6 +114,42 @@ rust_task::~rust_task()
extern "C" void rust_new_exit_task_glue();
struct spawn_args {
rust_task *task;
uintptr_t a3;
uintptr_t a4;
void (*FASTCALL f)(int *, rust_task *,
uintptr_t, uintptr_t);
};
// TODO: rewrite this in LLVM assembly so we can be sure the calling
// conventions will match.
extern "C" CDECL
void task_start_wrapper(spawn_args *a)
{
rust_task *task = a->task;
int rval = 42;
// This is used by the context switching code. LLVM generates fastcall
// functions, but ucontext needs cdecl functions. This massages the
// calling conventions into the right form.
a->f(&rval, task, a->a3, a->a4);
LOG(task, task, "task exited with value %d", rval);
// TODO: the old exit glue does some magical argument copying stuff. This
// is probably still needed.
// This is duplicated from upcall_exit, which is probably dead code by
// now.
LOG(task, task, "task ref_count: %d", task->ref_count);
A(task->dom, task->ref_count >= 0,
"Task ref_count should not be negative on exit!");
task->die();
task->notify_tasks_waiting_to_join();
task->yield(1);
}
void
rust_task::start(uintptr_t spawnee_fn,
uintptr_t args,
@ -142,54 +157,23 @@ rust_task::start(uintptr_t spawnee_fn,
{
LOGPTR(dom, "from spawnee", spawnee_fn);
// Set sp to last uintptr_t-sized cell of segment
rust_sp -= sizeof(uintptr_t);
I(dom, stk->data != NULL);
// Begin synthesizing the exit_task_glue frame. We will return to
// exit_task_glue and it is responsible for calling the user code
// and passing the value returned by the user to the system
// exit routine.
uintptr_t *spp = (uintptr_t *)rust_sp;
char *sp = (char *)stk->limit;
uintptr_t dummy_ret = (uintptr_t) spp--;
sp -= sizeof(spawn_args);
uintptr_t args_size = callsz - 3*sizeof(uintptr_t);
uintptr_t frame_size = args_size + 4*sizeof(uintptr_t);
spawn_args *a = (spawn_args *)sp;
a->task = this;
a->a3 = 0xca11ab1e;
a->a4 = args;
void **f = (void **)&a->f;
*f = (void *)spawnee_fn;
// NB: Darwin needs "16-byte aligned" stacks *at the point of the call
// instruction in the caller*. This means that the address at which the
// word before retpc is pushed must always be 16-byte aligned.
//
// see: "Mac OS X ABI Function Call Guide"
make_aligned_room_for_bytes(spp, frame_size - sizeof(uintptr_t));
// Copy args from spawner to spawnee.
uintptr_t *src = (uintptr_t *)args;
src += 1; // spawn-call output slot
src += 1; // spawn-call task slot
src += 1; // spawn-call closure-or-obj slot
*spp-- = (uintptr_t) *src; // vec
*spp-- = (uintptr_t) 0x0; // closure-or-obj
*spp-- = (uintptr_t) this; // task
*spp-- = (uintptr_t) dummy_ret; // output address
I(dom, spp == align_down(spp));
*spp-- = (uintptr_t) (uintptr_t) spawnee_fn;
*spp-- = (uintptr_t) 0x0; // retp
*spp-- = (uintptr_t) rust_new_exit_task_glue;
for (size_t j = 0; j < n_callee_saves; ++j) {
*spp-- = (uintptr_t)NULL;
}
// Back up one, we overshot where sp should be.
rust_sp = (uintptr_t) (spp+1);
ctx.call((void *)task_start_wrapper, a, sp);
yield_timer.reset(0);
transition(&dom->newborn_tasks, &dom->running_tasks);
}
@ -201,112 +185,6 @@ rust_task::grow(size_t n_frame_bytes)
// the presence of non-word-aligned pointers.
abort();
#if 0
stk_seg *old_stk = this->stk;
uintptr_t old_top = (uintptr_t) old_stk->limit;
uintptr_t old_bottom = (uintptr_t) &old_stk->data[0];
uintptr_t rust_sp_disp = old_top - this->rust_sp;
size_t ssz = old_top - old_bottom;
DLOG(dom, task, "upcall_grow_task(%" PRIdPTR
"), old size %" PRIdPTR " bytes (old lim: 0x%" PRIxPTR ")",
n_frame_bytes, ssz, old_top);
ssz *= 2;
if (ssz < n_frame_bytes)
ssz = n_frame_bytes;
ssz = next_power_of_two(ssz);
DLOG(dom, task, "upcall_grow_task growing stk 0x%"
PRIxPTR " to %d bytes", old_stk, ssz);
stk_seg *nstk = new_stk(dom, ssz);
uintptr_t new_top = (uintptr_t) &nstk->data[ssz];
size_t n_copy = old_top - old_bottom;
DLOG(dom, task,
"copying %d bytes of stack from [0x%" PRIxPTR ", 0x%" PRIxPTR "]"
" to [0x%" PRIxPTR ", 0x%" PRIxPTR "]",
n_copy,
old_bottom, old_bottom + n_copy,
new_top - n_copy, new_top);
VALGRIND_MAKE_MEM_DEFINED((void*)old_bottom, n_copy);
memcpy((void*)(new_top - n_copy), (void*)old_bottom, n_copy);
nstk->limit = new_top;
this->stk = nstk;
this->rust_sp = new_top - rust_sp_disp;
DLOG(dom, task, "processing relocations");
// FIXME (issue #32): this is the most ridiculously crude
// relocation scheme ever. Try actually, you know, writing out
// reloc descriptors?
size_t n_relocs = 0;
for (uintptr_t* p = (uintptr_t*)(new_top - n_copy);
p < (uintptr_t*)new_top; ++p) {
if (old_bottom <= *p && *p < old_top) {
//DLOG(dom, mem, "relocating pointer 0x%" PRIxPTR
// " by %d bytes", *p, (new_top - old_top));
n_relocs++;
*p += (new_top - old_top);
}
}
DLOG(dom, task, "processed %d relocations", n_relocs);
del_stk(dom, old_stk);
LOGPTR(dom, "grown stk limit", new_top);
#endif
}
void
push_onto_thread_stack(uintptr_t &sp, uintptr_t value)
{
asm("xchgl %0, %%esp\n"
"push %2\n"
"xchgl %0, %%esp\n"
: "=r" (sp)
: "0" (sp), "r" (value)
: "eax");
}
void
rust_task::run_after_return(size_t nargs, uintptr_t glue)
{
// This is only safe to call if we're the currently-running task.
check_active();
uintptr_t sp = runtime_sp;
// The compiler reserves nargs + 1 word for oldsp on the stack and
// then aligns it.
sp = align_down(sp - nargs * sizeof(uintptr_t));
uintptr_t *retpc = ((uintptr_t *) sp) - 1;
DLOG(dom, task,
"run_after_return: overwriting retpc=0x%" PRIxPTR
" @ runtime_sp=0x%" PRIxPTR
" with glue=0x%" PRIxPTR,
*retpc, sp, glue);
// Move the current return address (which points into rust code)
// onto the rust stack and pretend we just called into the glue.
push_onto_thread_stack(rust_sp, *retpc);
*retpc = glue;
}
void
rust_task::run_on_resume(uintptr_t glue)
{
// This is only safe to call if we're suspended.
check_suspended();
// Inject glue as resume address in the suspended frame.
uintptr_t* rsp = (uintptr_t*) rust_sp;
rsp += n_callee_saves;
DLOG(dom, task,
"run_on_resume: overwriting retpc=0x%" PRIxPTR
" @ rust_sp=0x%" PRIxPTR
" with glue=0x%" PRIxPTR,
*rsp, rsp, glue);
*rsp = glue;
}
void
@ -314,20 +192,17 @@ rust_task::yield(size_t nargs) {
yield(nargs, 0);
}
extern "C" void new_rust_yield_glue(void) asm("new_rust_yield_glue");
void
rust_task::yield(size_t nargs, size_t time_in_us) {
LOG(this, task, "task %s @0x%" PRIxPTR " yielding for %d us",
name, this, time_in_us);
yield_timer.reset(time_in_us);
run_after_return(nargs, (uintptr_t) new_rust_yield_glue);
}
static inline uintptr_t
get_callee_save_fp(uintptr_t *top_of_callee_saves)
{
return top_of_callee_saves[n_callee_saves - (callee_save_fp + 1)];
// TODO: what is nargs for, and is it safe to ignore?
yield_timer.reset(time_in_us);
// Return to the scheduler.
ctx.next->swap(ctx);
}
void
@ -410,20 +285,6 @@ rust_task::notify_tasks_waiting_to_join() {
}
}
uintptr_t
rust_task::get_fp() {
// sp in any suspended task points to the last callee-saved reg on
// the task stack.
return get_callee_save_fp((uintptr_t*)rust_sp);
}
uintptr_t
rust_task::get_previous_fp(uintptr_t fp) {
// FIXME: terribly X86-specific.
// *fp == previous_fp.
return *((uintptr_t*)fp);
}
frame_glue_fns*
rust_task::get_frame_glue_fns(uintptr_t fp) {
fp -= sizeof(uintptr_t);
@ -548,10 +409,10 @@ rust_task::free(void *p, bool is_gc)
void
rust_task::transition(rust_task_list *src, rust_task_list *dst) {
I(dom, state == src);
DLOG(dom, task,
"task %s " PTR " state change '%s' -> '%s'",
name, (uintptr_t)this, src->name, dst->name);
"task %s " PTR " state change '%s' -> '%s' while in '%s'",
name, (uintptr_t)this, src->name, dst->name, state->name);
I(dom, state == src);
src->remove(this);
dst->append(this);
state = dst;

View File

@ -7,6 +7,8 @@
#include "util/array_list.h"
#include "context.h"
struct
rust_task : public maybe_proxy<rust_task>,
public dom_owned<rust_task>
@ -47,6 +49,8 @@ rust_task : public maybe_proxy<rust_task>,
rust_handle<rust_task> *handle;
context ctx;
// Only a pointer to 'name' is kept, so it must live as long as this task.
rust_task(rust_dom *dom,
rust_task_list *state,
@ -83,14 +87,6 @@ rust_task : public maybe_proxy<rust_task>,
// Print a backtrace, if the "bt" logging option is on.
void backtrace();
// Swap in some glue code to run when we have returned to the
// task's context (assuming we're the active task).
void run_after_return(size_t nargs, uintptr_t glue);
// Swap in some glue code to run when we're next activated
// (assuming we're the suspended task).
void run_on_resume(uintptr_t glue);
// Save callee-saved registers and return to the main loop.
void yield(size_t nargs);
@ -114,8 +110,6 @@ rust_task : public maybe_proxy<rust_task>,
rust_handle<rust_task> * get_handle();
uintptr_t get_fp();
uintptr_t get_previous_fp(uintptr_t fp);
frame_glue_fns *get_frame_glue_fns(uintptr_t fp);
rust_crate_cache * get_crate_cache();
};

View File

@ -462,11 +462,22 @@ upcall_new_task(rust_task *spawner, rust_vec *name) {
return task;
}
// TODO: This is copied from rust_task.cpp. Both copies should be moved to a
// common location.
static uintptr_t
align_down(uintptr_t sp)
{
// There is no platform we care about that needs more than a
// 16-byte alignment.
return sp & ~(16 - 1);
}
extern "C" CDECL rust_task *
upcall_start_task(rust_task *spawner,
rust_task *task,
uintptr_t spawnee_fn,
uintptr_t args) {
uintptr_t args,
size_t args_sz) {
LOG_UPCALL_ENTRY(spawner);
rust_dom *dom = spawner->dom;
@ -478,7 +489,14 @@ upcall_start_task(rust_task *spawner,
// we used to be generating this tuple in rustc, but it's easier to do it
// here.
uintptr_t start_args[] = {0, 0, 0, args};
//
// The args tuple is stack-allocated. We need to move it over to the new
// stack.
task->rust_sp -= args_sz;
memcpy((void*)task->rust_sp, (void*)args, args_sz);
uintptr_t start_args[] = {0, 0, 0, task->rust_sp};
task->rust_sp = align_down(task->rust_sp);
task->start(spawnee_fn, (uintptr_t)start_args, sizeof(start_args));
return task;

View File

@ -36,6 +36,8 @@ str_push_byte
str_slice
str_vec
task_sleep
task_yield
task_join
unsafe_vec_to_mut
unsupervise
upcall_clone_chan

View File

@ -22,8 +22,8 @@ lock_and_signal::lock_and_signal() {
#else
lock_and_signal::lock_and_signal() {
pthread_cond_init(&_cond, NULL);
pthread_mutex_init(&_mutex, NULL);
CHECKED(pthread_cond_init(&_cond, NULL));
CHECKED(pthread_mutex_init(&_mutex, NULL));
}
#endif
@ -31,8 +31,8 @@ lock_and_signal::~lock_and_signal() {
#if defined(__WIN32__)
CloseHandle(_event);
#else
pthread_cond_destroy(&_cond);
pthread_mutex_destroy(&_mutex);
CHECKED(pthread_cond_destroy(&_cond));
CHECKED(pthread_mutex_destroy(&_mutex));
#endif
}
@ -40,7 +40,7 @@ void lock_and_signal::lock() {
#if defined(__WIN32__)
EnterCriticalSection(&_cs);
#else
pthread_mutex_lock(&_mutex);
CHECKED(pthread_mutex_lock(&_mutex));
#endif
}
@ -48,7 +48,7 @@ void lock_and_signal::unlock() {
#if defined(__WIN32__)
LeaveCriticalSection(&_cs);
#else
pthread_mutex_unlock(&_mutex);
CHECKED(pthread_mutex_unlock(&_mutex));
#endif
}
@ -66,14 +66,14 @@ void lock_and_signal::timed_wait(size_t timeout_in_ns) {
EnterCriticalSection(&_cs);
#else
if (timeout_in_ns == 0) {
pthread_cond_wait(&_cond, &_mutex);
CHECKED(pthread_cond_wait(&_cond, &_mutex));
} else {
timeval time_val;
gettimeofday(&time_val, NULL);
timespec time_spec;
time_spec.tv_sec = time_val.tv_sec + 0;
time_spec.tv_nsec = time_val.tv_usec * 1000 + timeout_in_ns;
pthread_cond_timedwait(&_cond, &_mutex, &time_spec);
CHECKED(pthread_cond_timedwait(&_cond, &_mutex, &time_spec));
}
#endif
}
@ -85,7 +85,7 @@ void lock_and_signal::signal() {
#if defined(__WIN32__)
SetEvent(_event);
#else
pthread_cond_signal(&_cond);
CHECKED(pthread_cond_signal(&_cond));
#endif
}
@ -96,7 +96,7 @@ void lock_and_signal::signal_all() {
#if defined(__WIN32__)
SetEvent(_event);
#else
pthread_cond_broadcast(&_cond);
CHECKED(pthread_cond_broadcast(&_cond));
#endif
}

View File

@ -1,42 +0,0 @@
/* More glue code, this time the 'bottom half' of yielding.
*
* We arrived here because an native call decided to deschedule the
* running task. So the native call's return address got patched to the
* first instruction of this glue code.
*
* When the native call does 'ret' it will come here, and its esp will be
* pointing to the last argument pushed on the C stack before making
* the native call: the 0th argument to the native call, which is always
* the task ptr performing the native call. That's where we take over.
*
* Our goal is to complete the descheduling
*
* - Switch over to the task stack temporarily.
*
* - Save the task's callee-saves onto the task stack.
* (the task is now 'descheduled', safe to set aside)
*
* - Switch *back* to the C stack.
*
* - Restore the C-stack callee-saves.
*
* - Return to the caller on the C stack that activated the task.
*
*/
.globl new_rust_yield_glue
.balign 4
new_rust_yield_glue:
movl 0(%esp), %ecx # ecx = rust_task
movl 16(%ecx), %esp
pushl %ebp
pushl %edi
pushl %esi
pushl %ebx
movl %esp, 16(%ecx)
movl 12(%ecx), %esp
popl %ebx
popl %esi
popl %edi
popl %ebp
ret

View File

@ -8,12 +8,14 @@ fn main() {
spawn child(chan(p));
let int y;
p |> y;
log "received";
log y;
assert (y == 10);
log_err "received";
log_err y;
//assert (y == 10);
}
fn child(chan[int] c) {
log_err "sending";
c <| 10;
log_err "value sent"
}

View File

@ -1,23 +1,24 @@
// xfail-stage0
// xfail-stage1
// xfail-stage2
// -*- rust -*-
use std;
import std::task::*;
fn main() {
auto other = spawn child();
log "1";
yield;
log "2";
yield;
log "3";
join other;
log_err "1";
yield();
log_err "2";
yield();
log_err "3";
join(other);
}
fn child() {
log "4";
yield;
log "5";
yield;
log "6";
log_err "4";
yield();
log_err "5";
yield();
log_err "6";
}

View File

@ -0,0 +1,17 @@
// xfail-stage0
// -*- rust -*-
use std;
import std::task::*;
fn main() {
auto other = spawn child();
log_err "1";
yield();
join(other);
}
fn child() {
log_err "2";
}

View File

@ -1,13 +1,13 @@
// xfail-stage0
// xfail-stage1
// xfail-stage2
// -*- rust -*-
use std;
fn main() {
let int i = 0;
while (i < 100) {
i = i + 1;
log i;
yield;
log_err i;
std::task::yield();
}
}