auto merge of #9834 : alexcrichton/rust/morestack, r=brson

This commit re-introduces the functionality of __morestack in a way that it was
not originally anticipated. Rust does not currently have segmented stacks,
rather just large stack segments. We do not detect when these stack segments are
overrun currently, but this commit leverages __morestack in order to check this.

This commit purges a lot of the old __morestack and stack limit C++
functionality, migrating the necessary chunks to rust. The stack limit is now
entirely maintained in rust, and the "main logic bits" of __morestack are now
also implemented in rust as well.

I put my best effort into validating that this currently builds and runs successfully on osx and linux 32/64 bit, but I was unable to get this working on windows. We never did have unwinding through __morestack frames, and although I tried poking at it for a bit, I was unable to understand why we don't get unwinding right now.

A focus of this commit is to implement as much of the logic in rust as possible. This involved some liberal usage of `no_split_stack` in various locations, along with some use of the `asm!` macro (scary). I modified a bit of C++ to stop calling `record_sp_limit` because this is no longer defined in C++, rather in rust.

Another consequence of this commit is that `thread_local_storage::{get, set}` must both be flagged with `#[rust_stack]`. I've briefly looked at the implementations on osx/linux/windows to ensure that they're pretty small stacks, and I'm pretty sure that they're definitely less than 20K stacks, so we probably don't have a lot to worry about.

Other things worthy of note:
* The default stack size is now 4MB instead of 2MB. This is so that when we request 2MB to call a C function you don't immediately overflow because you have consumed any stack at all.
* `asm!` is actually pretty cool, maybe we could actually define context switching with it?
* I wanted to add links to the internet about all this jazz of storing information in TLS, but I was only able to find a link for the windows implementation. Otherwise my suggestion is just "disassemble on that arch and see what happens"
* I put my best effort forward on arm/mips to tweak __morestack correctly, we have no ability to test this so an extra set of eyes would be useful on these spots.
* This is all really tricky stuff, so I tried to put as many comments as I thought were necessary, but if anything is still unclear (or I completely forgot to take something into account), I'm willing to write more!

1 files changed, 253 insertions, 30 deletions

diff --git a/src/libstd/rt/context.rs b/src/libstd/rt/context.rs
index 222f9a44b17..7f7545ca230 100644
--- a/src/libstd/rt/context.rs
+++ b/src/libstd/rt/context.rs
@@ -11,9 +11,12 @@
 use option::*;
 use super::stack::StackSegment;
 use libc::c_void;
+use uint;
 use cast::{transmute, transmute_mut_unsafe,
            transmute_region, transmute_mut_region};
 
+pub static RED_ZONE: uint = 20 * 1024;
+
 // FIXME #7761: Registers is boxed so that it is 16-byte aligned, for storing
 // SSE regs.  It would be marginally better not to do this. In C++ we
 // use an attribute on a struct.
@@ -24,14 +27,17 @@ pub struct Context {
     /// The context entry point, saved here for later destruction
     start: Option<~~fn()>,
     /// Hold the registers while the task or scheduler is suspended
-    regs: ~Registers
+    regs: ~Registers,
+    /// Lower bound and upper bound for the stack
+    stack_bounds: Option<(uint, uint)>,
 }
 
 impl Context {
     pub fn empty() -> Context {
         Context {
             start: None,
-            regs: new_regs()
+            regs: new_regs(),
+            stack_bounds: None,
         }
     }
 
@@ -47,7 +53,6 @@ impl Context {
 
         let fp: *c_void = task_start_wrapper as *c_void;
         let argp: *c_void = unsafe { transmute::<&~fn(), *c_void>(&*start) };
-        let stack_base: *uint = stack.start();
         let sp: *uint = stack.end();
         let sp: *mut uint = unsafe { transmute_mut_unsafe(sp) };
         // Save and then immediately load the current context,
@@ -57,11 +62,23 @@ impl Context {
             swap_registers(transmute_mut_region(&mut *regs), transmute_region(&*regs));
         };
 
-        initialize_call_frame(&mut *regs, fp, argp, sp, stack_base);
+        initialize_call_frame(&mut *regs, fp, argp, sp);
 
+        // Scheduler tasks don't have a stack in the "we allocated it" sense,
+        // but rather they run on pthreads stacks. We have complete control over
+        // them in terms of the code running on them (and hopefully they don't
+        // overflow). Additionally, their coroutine stacks are listed as being
+        // zero-length, so that's how we detect what's what here.
+        let stack_base: *uint = stack.start();
+        let bounds = if sp as uint == stack_base as uint {
+            None
+        } else {
+            Some((stack_base as uint, sp as uint))
+        };
         return Context {
             start: Some(start),
-            regs: regs
+            regs: regs,
+            stack_bounds: bounds,
         }
     }
 
@@ -79,8 +96,25 @@ impl Context {
         let in_regs: &Registers = match in_context {
             &Context { regs: ~ref r, _ } => r
         };
-        rtdebug!("doing raw swap");
-        unsafe { swap_registers(out_regs, in_regs) };
+
+        rtdebug!("noting the stack limit and doing raw swap");
+
+        unsafe {
+            // Right before we switch to the new context, set the new context's
+            // stack limit in the OS-specified TLS slot. This also  means that
+            // we cannot call any more rust functions after record_stack_bounds
+            // returns because they would all likely fail due to the limit being
+            // invalid for the current task. Lucky for us `swap_registers` is a
+            // C function so we don't have to worry about that!
+            match in_context.stack_bounds {
+                Some((lo, hi)) => record_stack_bounds(lo, hi),
+                // If we're going back to one of the original contexts or
+                // something that's possibly not a "normal task", then reset
+                // the stack limit to 0 to make morestack never fail
+                None => record_stack_bounds(0, uint::max_value),
+            }
+            swap_registers(out_regs, in_regs)
+        }
     }
 }
 
@@ -89,6 +123,29 @@ extern {
     fn swap_registers(out_regs: *mut Registers, in_regs: *Registers);
 }
 
+// Register contexts used in various architectures
+//
+// These structures all represent a context of one task throughout its
+// execution. Each struct is a representation of the architecture's register
+// set. When swapping between tasks, these register sets are used to save off
+// the current registers into one struct, and load them all from another.
+//
+// Note that this is only used for context switching, which means that some of
+// the registers may go unused. For example, for architectures with
+// callee/caller saved registers, the context will only reflect the callee-saved
+// registers. This is because the caller saved registers are already stored
+// elsewhere on the stack (if it was necessary anyway).
+//
+// Additionally, there may be fields on various architectures which are unused
+// entirely because they only reflect what is theoretically possible for a
+// "complete register set" to show, but user-space cannot alter these registers.
+// An example of this would be the segment selectors for x86.
+//
+// These structures/functions are roughly in-sync with the source files inside
+// of src/rt/arch/$arch. The only currently used function from those folders is
+// the `swap_registers` function, but that's only because for now segmented
+// stacks are disabled.
+
 #[cfg(target_arch = "x86")]
 struct Registers {
     eax: u32, ebx: u32, ecx: u32, edx: u32,
@@ -109,7 +166,7 @@ fn new_regs() -> ~Registers {
 
 #[cfg(target_arch = "x86")]
 fn initialize_call_frame(regs: &mut Registers, fptr: *c_void, arg: *c_void,
-                         sp: *mut uint, _stack_base: *uint) {
+                         sp: *mut uint) {
 
     let sp = align_down(sp);
     let sp = mut_offset(sp, -4);
@@ -125,6 +182,8 @@ fn initialize_call_frame(regs: &mut Registers, fptr: *c_void, arg: *c_void,
     regs.ebp = 0;
 }
 
+// windows requires saving more registers (both general and XMM), so the windows
+// register context must be larger.
 #[cfg(windows, target_arch = "x86_64")]
 type Registers = [uint, ..34];
 #[cfg(not(windows), target_arch = "x86_64")]
@@ -137,29 +196,14 @@ fn new_regs() -> ~Registers { ~([0, .. 22]) }
 
 #[cfg(target_arch = "x86_64")]
 fn initialize_call_frame(regs: &mut Registers, fptr: *c_void, arg: *c_void,
-                         sp: *mut uint, stack_base: *uint) {
+                         sp: *mut uint) {
 
-    // Redefinitions from regs.h
+    // Redefinitions from rt/arch/x86_64/regs.h
     static RUSTRT_ARG0: uint = 3;
     static RUSTRT_RSP: uint = 1;
     static RUSTRT_IP: uint = 8;
     static RUSTRT_RBP: uint = 2;
 
-    #[cfg(windows)]
-    fn initialize_tib(regs: &mut Registers, sp: *mut uint, stack_base: *uint) {
-        // Redefinitions from regs.h
-        static RUSTRT_ST1: uint = 11; // stack bottom
-        static RUSTRT_ST2: uint = 12; // stack top
-        regs[RUSTRT_ST1] = sp as uint;
-        regs[RUSTRT_ST2] = stack_base as uint;
-    }
-    #[cfg(not(windows))]
-    fn initialize_tib(_: &mut Registers, _: *mut uint, _: *uint) {
-    }
-
-    // Win64 manages stack range at TIB: %gs:0x08 (top) and %gs:0x10 (bottom)
-    initialize_tib(regs, sp, stack_base);
-
     let sp = align_down(sp);
     let sp = mut_offset(sp, -1);
 
@@ -167,9 +211,9 @@ fn initialize_call_frame(regs: &mut Registers, fptr: *c_void, arg: *c_void,
     unsafe { *sp = 0; }
 
     rtdebug!("creating call frame");
-    rtdebug!("fptr {}", fptr as uint);
-    rtdebug!("arg {}", arg as uint);
-    rtdebug!("sp {}", sp as uint);
+    rtdebug!("fptr {}", fptr);
+    rtdebug!("arg {}", arg);
+    rtdebug!("sp {}", sp);
 
     regs[RUSTRT_ARG0] = arg as uint;
     regs[RUSTRT_RSP] = sp as uint;
@@ -187,7 +231,7 @@ fn new_regs() -> ~Registers { ~([0, .. 32]) }
 
 #[cfg(target_arch = "arm")]
 fn initialize_call_frame(regs: &mut Registers, fptr: *c_void, arg: *c_void,
-                         sp: *mut uint, _stack_base: *uint) {
+                         sp: *mut uint) {
     let sp = align_down(sp);
     // sp of arm eabi is 8-byte aligned
     let sp = mut_offset(sp, -2);
@@ -208,7 +252,7 @@ fn new_regs() -> ~Registers { ~([0, .. 32]) }
 
 #[cfg(target_arch = "mips")]
 fn initialize_call_frame(regs: &mut Registers, fptr: *c_void, arg: *c_void,
-                         sp: *mut uint, _stack_base: *uint) {
+                         sp: *mut uint) {
     let sp = align_down(sp);
     // sp of mips o32 is 8-byte aligned
     let sp = mut_offset(sp, -2);
@@ -236,3 +280,182 @@ pub fn mut_offset<T>(ptr: *mut T, count: int) -> *mut T {
     use mem::size_of;
     (ptr as int + count * (size_of::<T>() as int)) as *mut T
 }
+
+#[inline(always)]
+pub unsafe fn record_stack_bounds(stack_lo: uint, stack_hi: uint) {
+    // When the old runtime had segmented stacks, it used a calculation that was
+    // "limit + RED_ZONE + FUDGE". The red zone was for things like dynamic
+    // symbol resolution, llvm function calls, etc. In theory this red zone
+    // value is 0, but it matters far less when we have gigantic stacks because
+    // we don't need to be so exact about our stack budget. The "fudge factor"
+    // was because LLVM doesn't emit a stack check for functions < 256 bytes in
+    // size. Again though, we have giant stacks, so we round all these
+    // calculations up to the nice round number of 20k.
+    record_sp_limit(stack_lo + RED_ZONE);
+
+    return target_record_stack_bounds(stack_lo, stack_hi);
+
+    #[cfg(not(windows))] #[cfg(not(target_arch = "x86_64"))] #[inline(always)]
+    unsafe fn target_record_stack_bounds(_stack_lo: uint, _stack_hi: uint) {}
+    #[cfg(windows, target_arch = "x86_64")] #[inline(always)]
+    unsafe fn target_record_stack_bounds(stack_lo: uint, stack_hi: uint) {
+        // Windows compiles C functions which may check the stack bounds. This
+        // means that if we want to perform valid FFI on windows, then we need
+        // to ensure that the stack bounds are what they truly are for this
+        // task. More info can be found at:
+        //   https://github.com/mozilla/rust/issues/3445#issuecomment-26114839
+        //
+        // stack range is at TIB: %gs:0x08 (top) and %gs:0x10 (bottom)
+        asm!("mov $0, %gs:0x08" :: "r"(stack_lo) :: "volatile");
+        asm!("mov $0, %gs:0x10" :: "r"(stack_hi) :: "volatile");
+    }
+}
+
+/// Records the current limit of the stack as specified by `end`.
+///
+/// This is stored in an OS-dependent location, likely inside of the thread
+/// local storage. The location that the limit is stored is a pre-ordained
+/// location because it's where LLVM has emitted code to check.
+///
+/// Note that this cannot be called under normal circumstances. This function is
+/// changing the stack limit, so upon returning any further function calls will
+/// possibly be triggering the morestack logic if you're not careful.
+///
+/// Also note that this and all of the inside functions are all flagged as
+/// "inline(always)" because they're messing around with the stack limits.  This
+/// would be unfortunate for the functions themselves to trigger a morestack
+/// invocation (if they were an actual function call).
+#[inline(always)]
+pub unsafe fn record_sp_limit(limit: uint) {
+    return target_record_sp_limit(limit);
+
+    // x86-64
+    #[cfg(target_arch = "x86_64", target_os = "macos")] #[inline(always)]
+    unsafe fn target_record_sp_limit(limit: uint) {
+        asm!("movq $$0x60+90*8, %rsi
+              movq $0, %gs:(%rsi)" :: "r"(limit) : "rsi" : "volatile")
+    }
+    #[cfg(target_arch = "x86_64", target_os = "linux")] #[inline(always)]
+    unsafe fn target_record_sp_limit(limit: uint) {
+        asm!("movq $0, %fs:112" :: "r"(limit) :: "volatile")
+    }
+    #[cfg(target_arch = "x86_64", target_os = "win32")] #[inline(always)]
+    unsafe fn target_record_sp_limit(limit: uint) {
+        // see: http://en.wikipedia.org/wiki/Win32_Thread_Information_Block
+        // store this inside of the "arbitrary data slot", but double the size
+        // because this is 64 bit instead of 32 bit
+        asm!("movq $0, %gs:0x28" :: "r"(limit) :: "volatile")
+    }
+    #[cfg(target_arch = "x86_64", target_os = "freebsd")] #[inline(always)]
+    unsafe fn target_record_sp_limit(limit: uint) {
+        asm!("movq $0, %fs:24" :: "r"(limit) :: "volatile")
+    }
+
+    // x86
+    #[cfg(target_arch = "x86", target_os = "macos")] #[inline(always)]
+    unsafe fn target_record_sp_limit(limit: uint) {
+        asm!("movl $$0x48+90*4, %eax
+              movl $0, %gs:(%eax)" :: "r"(limit) : "eax" : "volatile")
+    }
+    #[cfg(target_arch = "x86", target_os = "linux")]
+    #[cfg(target_arch = "x86", target_os = "freebsd")] #[inline(always)]
+    unsafe fn target_record_sp_limit(limit: uint) {
+        asm!("movl $0, %gs:48" :: "r"(limit) :: "volatile")
+    }
+    #[cfg(target_arch = "x86", target_os = "win32")] #[inline(always)]
+    unsafe fn target_record_sp_limit(limit: uint) {
+        // see: http://en.wikipedia.org/wiki/Win32_Thread_Information_Block
+        // store this inside of the "arbitrary data slot"
+        asm!("movl $0, %fs:0x14" :: "r"(limit) :: "volatile")
+    }
+
+    // mips, arm - Some brave soul can port these to inline asm, but it's over
+    //             my head personally
+    #[cfg(target_arch = "mips")]
+    #[cfg(target_arch = "arm")] #[inline(always)]
+    unsafe fn target_record_sp_limit(limit: uint) {
+        return record_sp_limit(limit as *c_void);
+        extern {
+            #[rust_stack]
+            fn record_sp_limit(limit: *c_void);
+        }
+    }
+}
+
+/// The counterpart of the function above, this function will fetch the current
+/// stack limit stored in TLS.
+///
+/// Note that all of these functions are meant to be exact counterparts of their
+/// brethren above, except that the operands are reversed.
+///
+/// As with the setter, this function does not have a __morestack header and can
+/// therefore be called in a "we're out of stack" situation.
+#[inline(always)]
+// NOTE: after the next snapshot, can remove the initialization before inline
+//       assembly due to an improvement in how it's handled, then this specific
+//       allow directive should get removed.
+#[allow(dead_assignment)]
+pub unsafe fn get_sp_limit() -> uint {
+    return target_get_sp_limit();
+
+    // x86-64
+    #[cfg(target_arch = "x86_64", target_os = "macos")] #[inline(always)]
+    unsafe fn target_get_sp_limit() -> uint {
+        let mut limit: uint = 0;
+        asm!("movq $$0x60+90*8, %rsi
+              movq %gs:(%rsi), $0" : "=r"(limit) :: "rsi" : "volatile");
+        return limit;
+    }
+    #[cfg(target_arch = "x86_64", target_os = "linux")] #[inline(always)]
+    unsafe fn target_get_sp_limit() -> uint {
+        let mut limit: uint = 0;
+        asm!("movq %fs:112, $0" : "=r"(limit) ::: "volatile");
+        return limit;
+    }
+    #[cfg(target_arch = "x86_64", target_os = "win32")] #[inline(always)]
+    unsafe fn target_get_sp_limit() -> uint {
+        let mut limit: uint = 0;
+        asm!("movq %gs:0x28, $0" : "=r"(limit) ::: "volatile");
+        return limit;
+    }
+    #[cfg(target_arch = "x86_64", target_os = "freebsd")] #[inline(always)]
+    unsafe fn target_get_sp_limit() -> uint {
+        let mut limit: uint = 0;
+        asm!("movq %fs:24, $0" : "=r"(limit) ::: "volatile");
+        return limit;
+    }
+
+    // x86
+    #[cfg(target_arch = "x86", target_os = "macos")] #[inline(always)]
+    unsafe fn target_get_sp_limit() -> uint {
+        let mut limit: uint = 0;
+        asm!("movl $$0x48+90*4, %eax
+              movl %gs:(%eax), $0" : "=r"(limit) :: "eax" : "volatile");
+        return limit;
+    }
+    #[cfg(target_arch = "x86", target_os = "linux")]
+    #[cfg(target_arch = "x86", target_os = "freebsd")] #[inline(always)]
+    unsafe fn target_get_sp_limit() -> uint {
+        let mut limit: uint = 0;
+        asm!("movl %gs:48, $0" : "=r"(limit) ::: "volatile");
+        return limit;
+    }
+    #[cfg(target_arch = "x86", target_os = "win32")] #[inline(always)]
+    unsafe fn target_get_sp_limit() -> uint {
+        let mut limit: uint = 0;
+        asm!("movl %fs:0x14, $0" : "=r"(limit) ::: "volatile");
+        return limit;
+    }
+
+    // mips, arm - Some brave soul can port these to inline asm, but it's over
+    //             my head personally
+    #[cfg(target_arch = "mips")]
+    #[cfg(target_arch = "arm")] #[inline(always)]
+    unsafe fn target_get_sp_limit() -> uint {
+        return get_sp_limit() as uint;
+        extern {
+            #[rust_stack]
+            fn get_sp_limit() -> *c_void;
+        }
+    }
+}