interrupts masked, we can't use the old mechanism of delivering
signals immediately when they arrive. Signals must instead be
delivered when the asyncs run *after* the end of the critical
section in scm_internal_select. But this also means after context
switch so that the signal will be delivered to a different thread.
To avoid this, I have changed the protocol of
coop_wait_for_runnable_thread and friends so that they are allowed
to return the original thread. So, if a signal arrives during
scm_internal_select, we won't any longer be forced do a context
switch, but can remain in the same thread and deliver the signal
to it.
* iselect.c: Small fixes.
interrupts masked, we can't use the old mechanism of delivering
signals immediately when they arrive. Signals must instead be
delivered when the asyncs run *after* the end of the critical
section in scm_internal_select. But this also means after context
switch so that the signal will be delivered to a different thread.
To avoid this, I have changed the protocol of
coop_wait_for_runnable_thread and friends so that they are allowed
to return the original thread. So, if a signal arrives during
scm_internal_select, we won't any longer be forced do a context
switch, but can remain in the same thread and deliver the signal
to it.
* coop.c (coop_mutex_init, coop_mutex_lock, coop_mutex_unlock,
coop_condition_variable_init, coop_condition_variable_wait,
coop_condition_variable_signal): Changed return type from `void'
to `int'. This is to adhere closer to the pthreads interface.
This, in turn, is part of an attempt to provide C versions of the
mutex and condition variable primitives which can be part of a
frontend to COOP or pthreads.
* coop.c (coop_mutex_destroy, coop_condition_variable_wait_mutex,
coop_condition_variable_destroy): New functions.
scheme_handler_bootstrip, scheme_launch_thread, c_launch_data,
c_body_bootstrip, c_handler_bootstrip, c_launch_thread): Add an
extra layer of functions around the body and handler of a thread.
This extra layer makes sure that the handler is called in the
dynamic context of the surround (= empty dynwind list), but under
the *dynamic root* of the body. We can not use the dynamic root
of the surround since that root belongs to another thread => stack
is not handled correctly. It may seem ugly to use this extra
layer, but the extra cost in terms of execution time is really
negligible compared to the total amount of time required to create
a thread, and it would reduce maintainability to duplicate the
crucial and complicated steps performed by cwdr.
* coop-threads.c (scm_wait_condition_variable): Use
coop_condition_variable_wait_mutex.
(SCM_ALLOW_INTS): Added thread switching code before interrupts
get re-enabled. The important effect of this is that interrupts
are blocked during thread switching so that thread data structures
don't risk getting messed up by an unfortunate signal.
(SCM_REDEFER_INTS, SCM_REALLOW_INTS): It turned out that gcc-2.8.0
seems to do more aggressive optimization which actually move
instructions around in these macros in a fatal way. Therefore:
Introduce Anthony's SCM_FENCE macro! (And I who thought he was
just superstitious...)
(SCM_TICK): Maybe do a context switch and take care of asyncs.
This macro should be used instead of SCM_ASYNC_TICK since the
latter doesn't do context switches any more.
coop_condition_variable_init, coop_condition_variable_wait,
coop_condition_variable_signal): Changed return type from `void'
to `int'. This is to adhere closer to the pthreads interface.
This, in turn, is part of an attempt to provide C versions of the
mutex and condition variable primitives which can be part of a
frontend to COOP or pthreads.
* coop.c (coop_mutex_destroy, coop_condition_variable_wait_mutex,
coop_condition_variable_destroy): New functions.
* coop-threads.c (scm_wait_condition_variable): Use
coop_condition_variable_wait_mutex.
* coop-threads.h, coop-defs.h (coop_q_t, coop_m, coop_c):
Definitions moved to coop-defs.h.
* coop-defs.h (scm_mutex_init, scm_mutex_lock, scm_mutex_unlock,
scm_mutex_destroy, scm_cond_init, scm_cond_wait, scm_cond_signal,
scm_cond_destroy): New C interface to mutecis and cond vars.
ioext.c (%read-line): Return a cons of the input line and its terminator.
fports.c, genio.c, genio.h, ports.c, ports.h: Change fgets methods
to return the string length as well as its contents.
(fill_select_type): Return the highest file descriptor.
(scm_select): Tell select about the highest file descriptor. On
some systems the SELECT_SET_SIZE can be as much as 128 bytes.
Therefore the extra overhead for calculating the maximum fd seems
to be more than compensated. Is this correct? In any case,
scm_internal_select will be much faster with this info.
(scm_select, fill_select_type, set_element): Don't accept any kind
of object in the file descriptor list or vector.
* configure.in: AC_CHECK_FUNCS: add "system".
* simpos.c (scm_system): always define: use sysmissing if not
available. Check for HAVE_SYSTEM instead of _Windows (does
Windows lack system or does it have an unusable one?).
Check for error conditions -1 and 127. Use SCM_DEFER_INTS.
Let the argument be optional: if not supplied, call system(NULL).
* ports.c (scm_close_port): relax the type check from OPPORTP to
PORTP; closing a closed port is allowed.
newline. When fgets returns a string whose length is `size-1', it
is ambiguous whether a whole line was retrieved, so we must check
explicitly whether a line terminator is present.
(scm_init_backtrace): Initialize `the-last-stack' to a fluid.
(scm_backtrace): `the-last-stack' is now a fluid.
* init.c (scm_boot_guile_1): Moved call to scm_init_backtrace
after scm_init_fluids.
coop_wait_for_runnable_thread): Disable interrupts so that no
async is executed before a potential error_revive.
(scm_internal_select): Disable interrupts during the parts of the
code which manipulate the sleep queue and the file descriptors.
* coop.c (coop_qinit): Initialize fields used by
scm_internal_select.
(coop_qget, coop_qget, coop_tmp_queue): Made global.
(coop_next_runnable_thread): If GUILE_ISELECT enabled, use
replacement in iselect.c.
(coop_mutex_lock, coop_condition_variable_wait, coop_abort,
coop_join): If GUILE_ISELECT enabled, use
coop_wait_for_runnable_thread instead of
coop_next_runnable_thread.
(usleep, sleep): New replacements for system functions if
GUILE_ISELECT is enabled.
* coop-threads.h: Declare coop_wait_for_runnable_thread.
* coop-defs.h (coop_t): Added fields used by scm_internal_select.
* filesys.c: Added #include "iselect.h". Moved FD-macros to
iselect.h. Implement Scheme level `select' using
scm_internal_select. (See NEWS.)
* genio.c (scm_getc): Block with scm_internal_select. (See NEWS.)
* init.c: Call scm_init_iselect.
* fports.c (scm_fgets): Use malloc/free rather than scm_must_malloc
and scm_must_free, since ultimately the string returned will be copied
by scm_makfrom0str anyway. Also, read any characters that may have
been pushed onto the port with scm_ungetc.
* ports.c (scm_generic_fgets): Same as for scm_fgets.
* ioext.c (scm_read_line): Free string after Guilifying it.
* ports.h (scm_ptobfuns): fgets method returns a char *, not a char.
