[ced29b]: doc / internals / specials.texinfo Maximize Restore History

Download this file

specials.texinfo    78 lines (63 with data), 3.5 kB

@node Specials
@comment  node-name,  next,  previous,  up
@chapter Specials

@menu
* Overview::
* Binding and unbinding::
@end menu

@node Overview
@section Overview

Unithread SBCL uses a shallow binding scheme: the current value of a
symbol is stored directly in its value slot. Accessing specials is
pretty fast but it's still a lot slower than accessing lexicals.

With multithreading it's slightly more complicated. The symbol's value
slot contains the global value and each symbol has a @code{TLS-INDEX}
slot that - when it's first bound - is set to a unique index of the
thread local area reserved for this purpose. The tls index is
initially zero and at index zero in the tls @code{NO-TLS-VALUE-MARKER}
resides. @code{NO-TLS-VALUE-MARKER} is different from
@code{UNBOUND-MARKER} to allow @code{PROGV} to bind a special to no
value locally in a thread.

@node Binding and unbinding
@section Binding and unbinding

Binding goes like this: the binding stack pointer (bsp) is bumped, old
value and symbol are stored at bsp - 1, new value is stored in symbol's
value slot or the tls. On multithreaded builds, @code{TLS-INDEX} is
stored on the binding stack in place of the symbol.

Unbinding: the symbol's value is restored from bsp - 1, value and
symbol at bsp - 1 are set to zero, and finally bsp is decremented.

The @code{UNBIND-TO-HERE} VOP assists in unwinding the stack. It
iterates over the bindings on the binding stack until it reaches the
prescribed point. For each binding with a non-zero symbol it does an
@code{UNBIND}.

How can a binding's symbol be zero? @code{BIND} is not pseudo atomic
(for performance reasons) and it can be interrupted by a signal. If
the signal hits after the bsp is incremented but before the values on
the stack are set the symbol is zero because a thread starts with a
zeroed tls plus @code{UNBIND} and @code{UNBIND-TO-HERE} both zero the
binding being unbound.

Zeroing the binding's symbol would not be enough as the binding's
value can be moved or garbage collected and if the above interrupt
initiates gc (or be @code{SIG_STOP_FOR_GC}) it will be greeted by a
garbage pointer.

Furthermore, @code{BIND} must always write the value to the binding
stack first and the symbol second because the symbol being non-zero
means validity to @code{UNBIND-TO-HERE}. For similar reasons
@code{UNBIND} also zeroes the symbol first. But if it is interrupted
by a signal that does an async unwind then @code{UNBIND-TO-HERE} can
be triggered when the symbol is zeroed but the value is not. In this
case @code{UNBIND-TO-HERE} must zero out the value to avoid leaving
garbage around that may wreck the ship on the next @code{BIND}.

In other words, the invariant is that the binding stack above bsp only
contains zeros. This makes @code{BIND} safe in face of gc triggered at
any point during its execution.

On platforms with the @code{UNWIND-TO-FRAME-AND-CALL-VOP} feature, it's
possible to restart frames in the debugger, unwinding the binding stack.
To know how much to unwind, @code{BIND-SENTINEL} in the beginning of a
function puts the current frame pointer on the binding stack with
@code{UNBOUND-MARKER-WIDETAG} instead of the symbol/tls-index.
@code{UNBIND-SENTINEL} removes it before returning. The debugger then
search for @code{UNBOUND-MARKER-WIDETAG} with the value being equal to
the desired frame, and calls @code{UNBIND-TO-HERE}. Consequently,
@code{UNBIND-TO-HERE} treats @code{UNBOUND-MARKER-WIDETAG} the same way
as zeros.