Ability to adopt file descriptors

[ruby_posix_mq.git] / ext / posix_mq / posix_mq.c

#define _XOPEN_SOURCE 600
#ifdef HAVE_SYS_SELECT_H
#  include <sys/select.h>
#endif
#ifdef HAVE_SIGNAL_H
#  include <signal.h>
#endif
#ifdef HAVE_PTHREAD_H
#  include <pthread.h>
#endif
#include <ruby.h>

#ifndef NUM2TIMET
#  define NUM2TIMET NUM2INT
#endif

#include <time.h>
#include <mqueue.h>
#include <fcntl.h>
#include <sys/stat.h>
#include <errno.h>
#include <assert.h>
#include <unistd.h>
#include <float.h>
#include <math.h>

#if defined(__linux__)
#  define MQD_TO_FD(mqd) (int)(mqd)
#  define FD_TO_MQD(fd) (mqd_t)(fd)
#elif defined(HAVE___MQ_OSHANDLE) /* FreeBSD */
#  define MQD_TO_FD(mqd) __mq_oshandle(mqd)
#else
#  define MQ_IO_MARK(mq) ((void)(0))
#  define MQ_IO_SET(mq,val) ((void)(0))
#  define MQ_IO_CLOSE(mq) ((int)(0))
#  define MQ_IO_NIL_P(mq) ((int)(1))
#endif

struct posix_mq {
        mqd_t des;
        struct mq_attr attr;
        VALUE name;
        VALUE thread;
#ifdef MQD_TO_FD
        VALUE io;
#endif
};

#ifdef MQD_TO_FD
#  define MQ_IO_MARK(mq) rb_gc_mark((mq)->io)
#  define MQ_IO_SET(mq,val) do { (mq)->io = (val); } while (0)
#  define MQ_IO_NIL_P(mq) NIL_P((mq)->io)
static int MQ_IO_CLOSE(struct posix_mq *mq)
{
        if (NIL_P(mq->io))
                return 0;

        /* not safe during GC */
        rb_io_close(mq->io);
        mq->io = Qnil;

        return 1;
}
#endif

# define PMQ_WANTARRAY (1<<0)
# define PMQ_TRY       (1<<1)

static VALUE cAttr;
static ID id_new, id_kill, id_fileno, id_divmod;
static ID id_flags, id_maxmsg, id_msgsize, id_curmsgs;
static VALUE sym_r, sym_w, sym_rw;
static const mqd_t MQD_INVALID = (mqd_t)-1;

/* Ruby 1.8.6+ macros (for compatibility with Ruby 1.9) */
#ifndef RSTRING_PTR
#  define RSTRING_PTR(s) (RSTRING(s)->ptr)
#endif
#ifndef RSTRING_LEN
#  define RSTRING_LEN(s) (RSTRING(s)->len)
#endif
#ifndef RFLOAT_VALUE
#  define RFLOAT_VALUE(f) (RFLOAT(f)->value)
#endif

#ifndef HAVE_RB_STR_SET_LEN
/* this is taken from Ruby 1.8.7, 1.8.6 may not have it */
#  ifdef RUBINIUS
#    error upgrade Rubinius, rb_str_set_len should be available
#  endif
static void rb_18_str_set_len(VALUE str, long len)
{
        RSTRING(str)->len = len;
        RSTRING(str)->ptr[len] = '\0';
}
#define rb_str_set_len rb_18_str_set_len
#endif /* !defined(HAVE_RB_STR_SET_LEN) */

#if defined(HAVE_RB_THREAD_CALL_WITHOUT_GVL) && defined(HAVE_RUBY_THREAD_H)
/* Ruby 2.0+ */
#  include <ruby/thread.h>
#  define WITHOUT_GVL(fn,a,ubf,b) \
        rb_thread_call_without_gvl((fn),(a),(ubf),(b))
#elif defined(HAVE_RB_THREAD_BLOCKING_REGION)
typedef VALUE (*my_blocking_fn_t)(void*);
#  define WITHOUT_GVL(fn,a,ubf,b) \
        rb_thread_blocking_region((my_blocking_fn_t)(fn),(a),(ubf),(b))

#else /* Ruby 1.8 */
/* partial emulation of the 1.9 rb_thread_blocking_region under 1.8 */
#  include <rubysig.h>
#  define RUBY_UBF_IO ((rb_unblock_function_t *)-1)
typedef void rb_unblock_function_t(void *);
typedef void * rb_blocking_function_t(void *);
static void * WITHOUT_GVL(rb_blocking_function_t *func, void *data1,
                        rb_unblock_function_t *ubf, void *data2)
{
        void *rv;

        assert(RUBY_UBF_IO == ubf && "RUBY_UBF_IO required for emulation");

        TRAP_BEG;
        rv = func(data1);
        TRAP_END;

        return rv;
}
#endif /* ! HAVE_RB_THREAD_BLOCKING_REGION */

/* used to pass arguments to mq_open inside blocking region */
struct open_args {
        mqd_t des;
        int argc;
        const char *name;
        int oflags;
        mode_t mode;
        struct mq_attr attr;
};

/* used to pass arguments to mq_send/mq_receive inside blocking region */
struct rw_args {
        mqd_t des;
        union {
                ssize_t received;
                int retval;
        };
        char *msg_ptr;
        size_t msg_len;
        unsigned msg_prio;
        struct timespec *timeout;
};

#ifndef HAVE_MQ_TIMEDSEND
static mqd_t
not_timedsend(mqd_t mqdes, const char *msg_ptr,
              size_t msg_len, unsigned msg_prio,
              const struct timespec *abs_timeout)
{
        rb_bug("mq_timedsend workaround failed");
        return (mqd_t)-1;
}
#  define mq_timedsend not_timedsend
#endif
#ifndef HAVE_MQ_TIMEDRECEIVE
static ssize_t
not_timedreceive(mqd_t mqdes, char *msg_ptr,
                 size_t msg_len, unsigned *msg_prio,
                 const struct timespec *abs_timeout)
{
        rb_bug("mq_timedreceive workaround failed");
        return (mqd_t)-1;
}
#  define mq_timedreceive not_timedreceive
#endif

#if defined(HAVE_MQ_TIMEDRECEIVE) && defined(HAVE_MQ_TIMEDSEND)
static void num2timespec(struct timespec *ts, VALUE t)
{
        switch (TYPE(t)) {
        case T_FIXNUM:
        case T_BIGNUM:
                ts->tv_sec = NUM2TIMET(t);
                ts->tv_nsec = 0;
                break;
        case T_FLOAT: {
                double f, d;
                double val = RFLOAT_VALUE(t);

                d = modf(val, &f);
                if (d >= 0) {
                        ts->tv_nsec = (long)(d * 1e9 + 0.5);
                } else {
                        ts->tv_nsec = (long)(-d * 1e9 + 0.5);
                        if (ts->tv_nsec > 0) {
                                ts->tv_nsec = 1000000000 - ts->tv_nsec;
                                f -= 1;
                        }
                }
                ts->tv_sec = (time_t)f;
                if (f != ts->tv_sec)
                        rb_raise(rb_eRangeError, "%f out of range", val);
                }
                break;
        default: {
                VALUE f;
                VALUE ary = rb_funcall(t, id_divmod, 1, INT2FIX(1));

                Check_Type(ary, T_ARRAY);

                ts->tv_sec = NUM2TIMET(rb_ary_entry(ary, 0));
                f = rb_ary_entry(ary, 1);
                f = rb_funcall(f, '*', 1, INT2FIX(1000000000));
                ts->tv_nsec = NUM2LONG(f);
                }
        }
}
#else
static void num2timespec(struct timespec *ts, VALUE t)
{
        rb_raise(rb_eNotImpError,
                 "mq_timedsend and/or mq_timedreceive missing");
}
#endif

static struct timespec *convert_timeout(struct timespec *dest, VALUE t)
{
        struct timespec ts, now;

        if (NIL_P(t))
                return NULL;

        num2timespec(&ts, t);
        clock_gettime(CLOCK_REALTIME, &now);
        dest->tv_sec = now.tv_sec + ts.tv_sec;
        dest->tv_nsec = now.tv_nsec + ts.tv_nsec;

        if (dest->tv_nsec > 1000000000) {
                dest->tv_nsec -= 1000000000;
                ++dest->tv_sec;
        }

        return dest;
}

/* (may) run without GVL */
static void * xopen(void *ptr)
{
        struct open_args *x = ptr;

        switch (x->argc) {
        case 2: x->des = mq_open(x->name, x->oflags); break;
        case 3: x->des = mq_open(x->name, x->oflags, x->mode, NULL); break;
        case 4: x->des = mq_open(x->name, x->oflags, x->mode, &x->attr); break;
        default: x->des = MQD_INVALID;
        }

        return NULL;
}

/* runs without GVL */
static void *xsend(void *ptr)
{
        struct rw_args *x = ptr;

        x->retval = x->timeout ?
                mq_timedsend(x->des, x->msg_ptr, x->msg_len,
                             x->msg_prio, x->timeout) :
                mq_send(x->des, x->msg_ptr, x->msg_len, x->msg_prio);

        return NULL;
}

/* runs without GVL */
static void * xrecv(void *ptr)
{
        struct rw_args *x = ptr;

        x->received = x->timeout ?
                mq_timedreceive(x->des, x->msg_ptr, x->msg_len,
                                &x->msg_prio, x->timeout) :
                mq_receive(x->des, x->msg_ptr, x->msg_len, &x->msg_prio);

        return NULL;
}

/* called by GC */
static void mark(void *ptr)
{
        struct posix_mq *mq = ptr;

        rb_gc_mark(mq->name);
        rb_gc_mark(mq->thread);
        MQ_IO_MARK(mq);
}

/* called by GC */
static void _free(void *ptr)
{
        struct posix_mq *mq = ptr;

        if (mq->des != MQD_INVALID && MQ_IO_NIL_P(mq)) {
                /* we ignore errors when gc-ing */
                mq_close(mq->des);
                errno = 0;
        }
        xfree(ptr);
}

/* automatically called at creation (before initialize) */
static VALUE alloc(VALUE klass)
{
        struct posix_mq *mq;
        VALUE rv = Data_Make_Struct(klass, struct posix_mq, mark, _free, mq);

        mq->des = MQD_INVALID;
        mq->attr.mq_flags = 0;
        mq->attr.mq_maxmsg = 0;
        mq->attr.mq_msgsize = -1;
        mq->attr.mq_curmsgs = 0;
        mq->name = Qnil;
        mq->thread = Qnil;
        MQ_IO_SET(mq, Qnil);

        return rv;
}

/* unwraps the posix_mq struct from self */
static struct posix_mq *get(VALUE self, int need_valid)
{
        struct posix_mq *mq;

        Data_Get_Struct(self, struct posix_mq, mq);

        if (need_valid && mq->des == MQD_INVALID)
                rb_raise(rb_eIOError, "closed queue descriptor");

        return mq;
}

static void check_struct_type(VALUE astruct)
{
        if (CLASS_OF(astruct) == cAttr)
                return;
        astruct = rb_inspect(astruct);
        rb_raise(rb_eTypeError, "not a POSIX_MQ::Attr: %s",
                 StringValuePtr(astruct));
}

static void rstruct2mqattr(struct mq_attr *attr, VALUE astruct, int all)
{
        VALUE tmp;

        check_struct_type(astruct);
        attr->mq_flags = NUM2LONG(rb_funcall(astruct, id_flags, 0));

        tmp = rb_funcall(astruct, id_maxmsg, 0);
        if (all || !NIL_P(tmp))
                attr->mq_maxmsg = NUM2LONG(tmp);

        tmp = rb_funcall(astruct, id_msgsize, 0);
        if (all || !NIL_P(tmp))
                attr->mq_msgsize = NUM2LONG(tmp);

        tmp = rb_funcall(astruct, id_curmsgs, 0);
        if (!NIL_P(tmp))
                attr->mq_curmsgs = NUM2LONG(tmp);
}

#ifdef FD_TO_MQD
/*
 * call-seq:
 *      POSIX_MQ.for_fd(socket) => mq
 *
 * Adopts a socket as a POSIX message queue. Argument will be
 * checked to ensure it is a POSIX message queue socket.
 *
 * This is useful for adopting systemd sockets passed via the
 * ListenMessageQueue directive.
 * Returns a +POSIX_MQ+ instance.  This method is only available
 * under Linux and FreeBSD and is not intended to be portable.
 *
 */
static VALUE for_fd(VALUE klass, VALUE socket)
{
        VALUE mqv = alloc(klass);
        struct posix_mq *mq = get(mqv, 0);

        mq->name = Qnil;
        mq->des = FD_TO_MQD(NUM2INT(socket));

        if (mq_getattr(mq->des, &mq->attr) < 0)
                rb_sys_fail("provided file descriptor is not a POSIX MQ");

        return mqv;
}
#endif /* FD_TO_MQD */

/*
 * call-seq:
 *      POSIX_MQ.new(name [, flags [, mode [, mq_attr]])        => mq
 *
 * Opens a POSIX message queue given by +name+.  +name+ should start
 * with a slash ("/") for portable applications.
 *
 * If a Symbol is given in place of integer +flags+, then:
 *
 * * +:r+ is equivalent to IO::RDONLY
 * * +:w+ is equivalent to IO::CREAT|IO::WRONLY
 * * +:rw+ is equivalent to IO::CREAT|IO::RDWR
 *
 * +mode+ is an integer and only used when IO::CREAT is used.
 * +mq_attr+ is a POSIX_MQ::Attr and only used if IO::CREAT is used.
 * If +mq_attr+ is not specified when creating a queue, then the
 * system defaults will be used.
 *
 * See the manpage for mq_open(3) for more details on this function.
 */
static VALUE init(int argc, VALUE *argv, VALUE self)
{
        struct posix_mq *mq = get(self, 0);
        struct open_args x;
        VALUE name, oflags, mode, attr;

        rb_scan_args(argc, argv, "13", &name, &oflags, &mode, &attr);

        switch (TYPE(oflags)) {
        case T_NIL:
                x.oflags = O_RDONLY;
                break;
        case T_SYMBOL:
                if (oflags == sym_r)
                        x.oflags = O_RDONLY;
                else if (oflags == sym_w)
                        x.oflags = O_CREAT|O_WRONLY;
                else if (oflags == sym_rw)
                        x.oflags = O_CREAT|O_RDWR;
                else {
                        oflags = rb_inspect(oflags);
                        rb_raise(rb_eArgError,
                                 "symbol must be :r, :w, or :rw: %s",
                                 StringValuePtr(oflags));
                }
                break;
        case T_BIGNUM:
        case T_FIXNUM:
                x.oflags = NUM2INT(oflags);
                break;
        default:
                rb_raise(rb_eArgError, "flags must be an int, :r, :w, or :wr");
        }

        x.name = StringValueCStr(name);
        x.argc = 2;

        switch (TYPE(mode)) {
        case T_FIXNUM:
                x.argc = 3;
                x.mode = NUM2UINT(mode);
                break;
        case T_NIL:
                if (x.oflags & O_CREAT) {
                        x.argc = 3;
                        x.mode = 0666;
                }
                break;
        default:
                rb_raise(rb_eArgError, "mode not an integer");
        }

        switch (TYPE(attr)) {
        case T_STRUCT:
                x.argc = 4;
                rstruct2mqattr(&x.attr, attr, 1);

                /* principle of least surprise */
                if (x.attr.mq_flags & O_NONBLOCK)
                        x.oflags |= O_NONBLOCK;
                break;
        case T_NIL:
                break;
        default:
                check_struct_type(attr);
        }

        (void)xopen(&x);
        mq->des = x.des;
        if (mq->des == MQD_INVALID) {
                switch (errno) {
                case ENOMEM:
                case EMFILE:
                case ENFILE:
                case ENOSPC:
                        rb_gc();
                        (void)xopen(&x);
                        mq->des = x.des;
                }
                if (mq->des == MQD_INVALID)
                        rb_sys_fail("mq_open");
        }

        mq->name = rb_str_dup(name);
        if (x.oflags & O_NONBLOCK)
                mq->attr.mq_flags = O_NONBLOCK;

        return self;
}

/*
 * call-seq:
 *      POSIX_MQ.unlink(name) =>        1
 *
 * Unlinks the message queue given by +name+.  The queue will be destroyed
 * when the last process with the queue open closes its queue descriptors.
 */
static VALUE s_unlink(VALUE self, VALUE name)
{
        int rv = mq_unlink(StringValueCStr(name));

        if (rv == -1)
                rb_sys_fail("mq_unlink");

        return INT2NUM(1);
}

/*
 * call-seq:
 *      mq.unlink =>    mq
 *
 * Unlinks the message queue to prevent other processes from accessing it.
 * All existing queue descriptors to this queue including those opened by
 * other processes are unaffected.  The queue will only be destroyed
 * when the last process with open descriptors to this queue closes
 * the descriptors.
 */
static VALUE _unlink(VALUE self)
{
        struct posix_mq *mq = get(self, 0);
        int rv;

        if (NIL_P(mq->name)) {
                rb_raise(rb_eArgError, "can not unlink an adopted socket");
        }

        assert(TYPE(mq->name) == T_STRING && "mq->name is not a string");

        rv = mq_unlink(RSTRING_PTR(mq->name));
        if (rv == -1)
                rb_sys_fail("mq_unlink");

        return self;
}

static void setup_send_buffer(struct rw_args *x, VALUE buffer)
{
        buffer = rb_obj_as_string(buffer);
        x->msg_ptr = RSTRING_PTR(buffer);
        x->msg_len = (size_t)RSTRING_LEN(buffer);
}

static VALUE _send(int sflags, int argc, VALUE *argv, VALUE self);

/*
 * call-seq:
 *      mq.send(string [,priority[, timeout]])  => true
 *
 * Inserts the given +string+ into the message queue with an optional,
 * unsigned integer +priority+.  If the optional +timeout+ is specified,
 * then Errno::ETIMEDOUT will be raised if the operation cannot complete
 * before +timeout+ seconds has elapsed.  Without +timeout+, this method
 * may block until the queue is writable.
 *
 * On some older systems, the +timeout+ argument is not currently
 * supported and may raise NotImplementedError if +timeout+ is used.
 */
static VALUE my_send(int argc, VALUE *argv, VALUE self)
{
        return _send(0, argc, argv, self);
}

static VALUE _send(int sflags, int argc, VALUE *argv, VALUE self)
{
        struct posix_mq *mq = get(self, 1);
        struct rw_args x;
        VALUE buffer, prio, timeout;
        struct timespec expire;

        rb_scan_args(argc, argv, "12", &buffer, &prio, &timeout);

        setup_send_buffer(&x, buffer);
        x.des = mq->des;
        x.timeout = convert_timeout(&expire, timeout);
        x.msg_prio = NIL_P(prio) ? 0 : NUM2UINT(prio);

retry:
        WITHOUT_GVL(xsend, &x, RUBY_UBF_IO, 0);
        if (x.retval == -1) {
                if (errno == EINTR)
                        goto retry;
                if (errno == EAGAIN && (sflags & PMQ_TRY))
                        return Qfalse;
                rb_sys_fail("mq_send");
        }

        return Qtrue;
}

/*
 * call-seq:
 *      mq << string    => mq
 *
 * Inserts the given +string+ into the message queue with a
 * default priority of 0 and no timeout.
 *
 * Returns itself so its calls may be chained.  This use is only
 * recommended only for users who expect blocking behavior from
 * the queue.
 */
static VALUE send0(VALUE self, VALUE buffer)
{
        struct posix_mq *mq = get(self, 1);
        struct rw_args x;

        setup_send_buffer(&x, buffer);
        x.des = mq->des;
        x.timeout = NULL;
        x.msg_prio = 0;

retry:
        WITHOUT_GVL(xsend, &x, RUBY_UBF_IO, 0);
        if (x.retval == -1) {
                if (errno == EINTR)
                        goto retry;
                rb_sys_fail("mq_send");
        }

        return self;
}

#ifdef MQD_TO_FD
/*
 * call-seq:
 *      mq.to_io        => IO
 *
 * Returns an IO.select-able +IO+ object.  This method is only available
 * under Linux and FreeBSD and is not intended to be portable.
 */
static VALUE to_io(VALUE self)
{
        struct posix_mq *mq = get(self, 1);
        int fd = MQD_TO_FD(mq->des);

        if (NIL_P(mq->io))
                mq->io = rb_funcall(rb_cIO, id_new, 1, INT2NUM(fd));

        return mq->io;
}
#endif

static VALUE _receive(int rflags, int argc, VALUE *argv, VALUE self);

/*
 * call-seq:
 *      mq.receive([buffer, [timeout]])         => [ message, priority ]
 *
 * Takes the highest priority message off the queue and returns
 * an array containing the message as a String and the Integer
 * priority of the message.
 *
 * If the optional +buffer+ is present, then it must be a String
 * which will receive the data.
 *
 * If the optional +timeout+ is present, then it may be a Float
 * or Integer specifying the timeout in seconds.  Errno::ETIMEDOUT
 * will be raised if +timeout+ has elapsed and there are no messages
 * in the queue.
 *
 * On some older systems, the +timeout+ argument is not currently
 * supported and may raise NotImplementedError if +timeout+ is used.
 */
static VALUE receive(int argc, VALUE *argv, VALUE self)
{
        return _receive(PMQ_WANTARRAY, argc, argv, self);
}

/*
 * call-seq:
 *      mq.shift([buffer, [timeout]])           => message
 *
 * Takes the highest priority message off the queue and returns
 * the message as a String.
 *
 * If the optional +buffer+ is present, then it must be a String
 * which will receive the data.
 *
 * If the optional +timeout+ is present, then it may be a Float
 * or Integer specifying the timeout in seconds.  Errno::ETIMEDOUT
 * will be raised if +timeout+ has elapsed and there are no messages
 * in the queue.
 *
 * On some older systems, the +timeout+ argument is not currently
 * supported and may raise NotImplementedError if +timeout+ is used.
 */
static VALUE shift(int argc, VALUE *argv, VALUE self)
{
        return _receive(0, argc, argv, self);
}

static VALUE _receive(int rflags, int argc, VALUE *argv, VALUE self)
{
        struct posix_mq *mq = get(self, 1);
        struct rw_args x;
        VALUE buffer, timeout;
        struct timespec expire;

        if (mq->attr.mq_msgsize < 0) {
                if (mq_getattr(mq->des, &mq->attr) < 0)
                        rb_sys_fail("mq_getattr");
        }

        rb_scan_args(argc, argv, "02", &buffer, &timeout);
        x.timeout = convert_timeout(&expire, timeout);

        if (NIL_P(buffer)) {
                buffer = rb_str_new(0, mq->attr.mq_msgsize);
        } else {
                StringValue(buffer);
                rb_str_modify(buffer);
                rb_str_resize(buffer, mq->attr.mq_msgsize);
        }
        OBJ_TAINT(buffer);
        x.msg_ptr = RSTRING_PTR(buffer);
        x.msg_len = (size_t)mq->attr.mq_msgsize;
        x.des = mq->des;

retry:
        WITHOUT_GVL(xrecv, &x, RUBY_UBF_IO, 0);
        if (x.received < 0) {
                if (errno == EINTR)
                        goto retry;
                if (errno == EAGAIN && (rflags & PMQ_TRY))
                        return Qnil;
                rb_sys_fail("mq_receive");
        }

        rb_str_set_len(buffer, x.received);

        if (rflags & PMQ_WANTARRAY)
                return rb_ary_new3(2, buffer, UINT2NUM(x.msg_prio));
        return buffer;
}

/*
 * call-seq:
 *      mq.attr =>       mq_attr
 *
 * Returns a POSIX_MQ::Attr struct containing the attributes
 * of the message queue.  See the mq_getattr(3) manpage for
 * more details.
 */
static VALUE getattr(VALUE self)
{
        struct posix_mq *mq = get(self, 1);

        if (mq_getattr(mq->des, &mq->attr) < 0)
                rb_sys_fail("mq_getattr");

        return rb_funcall(cAttr, id_new, 4,
                          LONG2NUM(mq->attr.mq_flags),
                          LONG2NUM(mq->attr.mq_maxmsg),
                          LONG2NUM(mq->attr.mq_msgsize),
                          LONG2NUM(mq->attr.mq_curmsgs));
}

/*
 * call-seq:
 *      mq.attr = POSIX_MQ::Attr(IO::NONBLOCK)          => mq_attr
 *
 * Only the IO::NONBLOCK flag may be set or unset (zero) in this manner.
 * See the mq_setattr(3) manpage for more details.
 *
 * Consider using the POSIX_MQ#nonblock= method as it is easier and
 * more natural to use.
 */
static VALUE setattr(VALUE self, VALUE astruct)
{
        struct posix_mq *mq = get(self, 1);
        struct mq_attr newattr;

        rstruct2mqattr(&newattr, astruct, 0);

        if (mq_setattr(mq->des, &newattr, NULL) < 0)
                rb_sys_fail("mq_setattr");

        return astruct;
}

/*
 * call-seq:
 *      mq.close        => nil
 *
 * Closes the underlying message queue descriptor.
 * If this descriptor had a registered notification request, the request
 * will be removed so another descriptor or process may register a
 * notification request.  Message queue descriptors are automatically
 * closed by garbage collection.
 */
static VALUE _close(VALUE self)
{
        struct posix_mq *mq = get(self, 1);

        if (! MQ_IO_CLOSE(mq)) {
                if (mq_close(mq->des) == -1)
                        rb_sys_fail("mq_close");
        }
        mq->des = MQD_INVALID;

        return Qnil;
}

/*
 *  call-seq:
 *      mq.closed?      => true or false
 *
 *  Returns +true+ if the message queue descriptor is closed and therefore
 *  unusable, otherwise +false+
 */
static VALUE closed(VALUE self)
{
        struct posix_mq *mq = get(self, 0);

        return mq->des == MQD_INVALID ? Qtrue : Qfalse;
}

/*
 * call-seq:
 *      mq.name         => string
 *
 * Returns the string name of message queue associated with +mq+
 */
static VALUE name(VALUE self)
{
        struct posix_mq *mq = get(self, 0);

        if (NIL_P(mq->name)) {
                /*
                 * We could use readlink(2) on /proc/self/fd/N, but lots of
                 * care required.
                 * http://stackoverflow.com/questions/1188757/
                 */
                rb_raise(rb_eArgError, "can not get name of an adopted socket");
        }

        return rb_str_dup(mq->name);
}

static int lookup_sig(VALUE sig)
{
        static VALUE list;
        const char *ptr;
        long len;

        sig = rb_obj_as_string(sig);
        len = RSTRING_LEN(sig);
        ptr = RSTRING_PTR(sig);

        if (len > 3 && !memcmp("SIG", ptr, 3))
                sig = rb_str_new(ptr + 3, len - 3);

        if (!list) {
                VALUE mSignal = rb_const_get(rb_cObject, rb_intern("Signal"));

                list = rb_funcall(mSignal, rb_intern("list"), 0, 0);
                rb_global_variable(&list);
        }

        sig = rb_hash_aref(list, sig);
        if (NIL_P(sig))
                rb_raise(rb_eArgError, "invalid signal: %s\n", ptr);

        return NUM2INT(sig);
}

/*
 * TODO: Under Linux, we could just use netlink directly
 * the same way glibc does...
 */
/* we spawn a thread just to write ONE byte into an fd (usually a pipe) */
static void thread_notify_fd(union sigval sv)
{
        int fd = sv.sival_int;

        while ((write(fd, "", 1) < 0) && (errno == EINTR || errno == EAGAIN));
}

static void my_mq_notify(mqd_t des, struct sigevent *not)
{
        int rv = mq_notify(des, not);

        if (rv == -1) {
                if (errno == ENOMEM) {
                        rb_gc();
                        rv = mq_notify(des, not);
                }
                if (rv == -1)
                        rb_sys_fail("mq_notify");
        }
}

static void lower_stack_size(pthread_attr_t *attr)
{
/* some OSes have ridiculously small stack sizes */
#ifdef PTHREAD_STACK_MIN
        size_t stack_size = PTHREAD_STACK_MIN;
        size_t min_size = 4096;

        if (stack_size < min_size)
                stack_size = min_size;
        pthread_attr_setstacksize(attr, stack_size);
#endif
}

/* :nodoc: */
static VALUE setnotify_exec(VALUE self, VALUE io, VALUE thr)
{
        int fd = NUM2INT(rb_funcall(io, id_fileno, 0, 0));
        struct posix_mq *mq = get(self, 1);
        struct sigevent not;
        pthread_attr_t attr;

        errno = pthread_attr_init(&attr);
        if (errno) rb_sys_fail("pthread_attr_init");

        errno = pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED);
        if (errno) rb_sys_fail("pthread_attr_setdetachstate");

        lower_stack_size(&attr);
        not.sigev_notify = SIGEV_THREAD;
        not.sigev_notify_function = thread_notify_fd;
        not.sigev_notify_attributes = &attr;
        not.sigev_value.sival_int = fd;

        if (!NIL_P(mq->thread))
                rb_funcall(mq->thread, id_kill, 0, 0);
        mq->thread = thr;

        my_mq_notify(mq->des, &not);

        return thr;
}

/* :nodoc: */
static VALUE notify_cleanup(VALUE self)
{
        struct posix_mq *mq = get(self, 1);

        if (!NIL_P(mq->thread)) {
                rb_funcall(mq->thread, id_kill, 0, 0);
                mq->thread = Qnil;
        }
        return Qnil;
}

/*
 * call-seq:
 *      mq.notify = signal      => signal
 *
 * Registers the notification request to deliver a given +signal+
 * to the current process when message is received.
 * If +signal+ is +nil+, it will unregister and disable the notification
 * request to allow other processes to register a request.
 * If +signal+ is +false+, it will register a no-op notification request
 * which will prevent other processes from registering a notification.
 * If +signal+ is an +IO+ object, it will spawn a thread upon the
 * arrival of the next message and write one "\\0" byte to the file
 * descriptor belonging to that IO object.
 * Only one process may have a notification request for a queue
 * at a time, Errno::EBUSY will be raised if there is already
 * a notification request registration for the queue.
 *
 * Notifications are only fired once and processes must reregister
 * for subsequent notifications.
 *
 * For readers of the mq_notify(3) manpage, passing +false+
 * is equivalent to SIGEV_NONE, and passing +nil+ is equivalent
 * of passing a NULL notification pointer to mq_notify(3).
 */
static VALUE setnotify(VALUE self, VALUE arg)
{
        struct posix_mq *mq = get(self, 1);
        struct sigevent not;
        struct sigevent * notification = &not;
        VALUE rv = arg;

        notify_cleanup(self);
        not.sigev_notify = SIGEV_SIGNAL;

        switch (TYPE(arg)) {
        case T_FALSE:
                not.sigev_notify = SIGEV_NONE;
                break;
        case T_NIL:
                notification = NULL;
                break;
        case T_FIXNUM:
                not.sigev_signo = NUM2INT(arg);
                break;
        case T_SYMBOL:
        case T_STRING:
                not.sigev_signo = lookup_sig(arg);
                rv = INT2NUM(not.sigev_signo);
                break;
        default:
                rb_raise(rb_eArgError, "must be a signal or nil");
        }

        my_mq_notify(mq->des, notification);

        return rv;
}

/*
 * call-seq:
 *      mq.nonblock?    => true or false
 *
 * Returns the current non-blocking state of the message queue descriptor.
 */
static VALUE nonblock_p(VALUE self)
{
        struct posix_mq *mq = get(self, 1);

        if (mq_getattr(mq->des, &mq->attr) < 0)
                rb_sys_fail("mq_getattr");
        return mq->attr.mq_flags & O_NONBLOCK ? Qtrue : Qfalse;
}

/*
 * call-seq:
 *      mq.nonblock = boolean   => boolean
 *
 * Enables or disables non-blocking operation for the message queue
 * descriptor.  Errno::EAGAIN will be raised in situations where
 * the queue would block.  This is not compatible with +timeout+
 * arguments to POSIX_MQ#send and POSIX_MQ#receive.
 */
static VALUE setnonblock(VALUE self, VALUE nb)
{
        struct mq_attr newattr;
        struct posix_mq *mq = get(self, 1);

        if (nb == Qtrue)
                newattr.mq_flags = O_NONBLOCK;
        else if (nb == Qfalse)
                newattr.mq_flags = 0;
        else
                rb_raise(rb_eArgError, "must be true or false");

        if (mq_setattr(mq->des, &newattr, &mq->attr) < 0)
                rb_sys_fail("mq_setattr");

        mq->attr.mq_flags = newattr.mq_flags;

        return nb;
}

/*
 * call-seq:
 *      mq.trysend(string [,priority[, timeout]]) => +true+ or +false+
 *
 * Exactly like POSIX_MQ#send, except it returns +false+ instead of raising
 * Errno::EAGAIN when non-blocking operation is desired and returns +true+
 * on success instead of +nil+.
 *
 * This does not guarantee non-blocking behavior, the message queue must
 * be made non-blocking before calling this method.
 */
static VALUE trysend(int argc, VALUE *argv, VALUE self)
{
        return _send(PMQ_TRY, argc, argv, self);
}

/*
 * call-seq:
 *      mq.tryshift([buffer [, timeout]])       => message or nil
 *
 * Exactly like POSIX_MQ#shift, except it returns +nil+ instead of raising
 * Errno::EAGAIN when non-blocking operation is desired.
 *
 * This does not guarantee non-blocking behavior, the message queue must
 * be made non-blocking before calling this method.
 */
static VALUE tryshift(int argc, VALUE *argv, VALUE self)
{
        return _receive(PMQ_TRY, argc, argv, self);
}

/*
 * call-seq:
 *      mq.tryreceive([buffer [, timeout]])     => [ message, priority ] or nil
 *
 * Exactly like POSIX_MQ#receive, except it returns +nil+ instead of raising
 * Errno::EAGAIN when non-blocking operation is desired.
 *
 * This does not guarantee non-blocking behavior, the message queue must
 * be made non-blocking before calling this method.
 */
static VALUE tryreceive(int argc, VALUE *argv, VALUE self)
{
        return _receive(PMQ_WANTARRAY|PMQ_TRY, argc, argv, self);
}

void Init_posix_mq_ext(void)
{
        VALUE cPOSIX_MQ = rb_define_class("POSIX_MQ", rb_cObject);
        rb_define_alloc_func(cPOSIX_MQ, alloc);
        cAttr = rb_const_get(cPOSIX_MQ, rb_intern("Attr"));

        /*
         * The maximum number of open message descriptors supported
         * by the system.  This may be -1, in which case it is dynamically
         * set at runtime.  Consult your operating system documentation
         * for system-specific information about this.
         */
        rb_define_const(cPOSIX_MQ, "OPEN_MAX",
                        LONG2NUM(sysconf(_SC_MQ_OPEN_MAX)));

        /*
         * The maximum priority that may be specified for POSIX_MQ#send
         * On POSIX-compliant systems, this is at least 31, but some
         * systems allow higher limits.
         * The minimum priority is always zero.
         */
        rb_define_const(cPOSIX_MQ, "PRIO_MAX",
                        LONG2NUM(sysconf(_SC_MQ_PRIO_MAX)));

        rb_define_singleton_method(cPOSIX_MQ, "unlink", s_unlink, 1);

        rb_define_private_method(cPOSIX_MQ, "initialize", init, -1);
        rb_define_method(cPOSIX_MQ, "send", my_send, -1);
        rb_define_method(cPOSIX_MQ, "<<", send0, 1);
        rb_define_method(cPOSIX_MQ, "trysend", trysend, -1);
        rb_define_method(cPOSIX_MQ, "receive", receive, -1);
        rb_define_method(cPOSIX_MQ, "tryreceive", tryreceive, -1);
        rb_define_method(cPOSIX_MQ, "shift", shift, -1);
        rb_define_method(cPOSIX_MQ, "tryshift", tryshift, -1);
        rb_define_method(cPOSIX_MQ, "attr", getattr, 0);
        rb_define_method(cPOSIX_MQ, "attr=", setattr, 1);
        rb_define_method(cPOSIX_MQ, "close", _close, 0);
        rb_define_method(cPOSIX_MQ, "closed?", closed, 0);
        rb_define_method(cPOSIX_MQ, "unlink", _unlink, 0);
        rb_define_method(cPOSIX_MQ, "name", name, 0);
        rb_define_method(cPOSIX_MQ, "notify=", setnotify, 1);
        rb_define_method(cPOSIX_MQ, "nonblock=", setnonblock, 1);
        rb_define_private_method(cPOSIX_MQ, "notify_exec", setnotify_exec, 2);
        rb_define_private_method(cPOSIX_MQ, "notify_cleanup", notify_cleanup, 0);
        rb_define_method(cPOSIX_MQ, "nonblock?", nonblock_p, 0);
#ifdef MQD_TO_FD
        rb_define_method(cPOSIX_MQ, "to_io", to_io, 0);
#endif

#ifdef FD_TO_MQD
        rb_define_module_function(cPOSIX_MQ, "for_fd", for_fd, 1);
#endif

        id_new = rb_intern("new");
        id_kill = rb_intern("kill");
        id_fileno = rb_intern("fileno");
        id_divmod = rb_intern("divmod");
        id_flags = rb_intern("flags");
        id_maxmsg = rb_intern("maxmsg");
        id_msgsize = rb_intern("msgsize");
        id_curmsgs = rb_intern("curmsgs");
        sym_r = ID2SYM(rb_intern("r"));
        sym_w = ID2SYM(rb_intern("w"));
        sym_rw = ID2SYM(rb_intern("rw"));
}