Use XCB's new socket handoff mechanism rather than the old XCB Xlib lock.

Previously, Xlib/XCB used XCB's Xlib lock to prevent XCB from sending
requests between calls to Xlib's LockDisplay and UnlockDisplay macros.
Xlib/XCB then sent all of its requests using XCB's xcb_send_request, and
had to flush its requests when unlocking the display.

XCB 1.2 adds a new socket handoff mechanism, xcb_take_socket.  Replace
much of the existing Xlib/XCB implementation with the use of
xcb_take_socket to take ownership of the write side of the X connection
socket, and a return_socket callback which writes any outstanding requests
with xcb_writev.  This approach allows Xlib/XCB to use the same buffering
as traditional Xlib did.  In particular, programs which use Xlib/XCB and
never make XCB calls will never need to hand the socket back to XCB, and
vice versa.

This allows us to discard large quantities of synchronization code from
Xlib/XCB, together with the synchronization bugs present in that code.
Several test cases which previously failed now work perfectly, including
multi-threaded ico.  In addition, the infamous locking correctness
assertions, triggered when double-locking or when unlocking without a
previous lock, no longer exist, because Xlib/XCB no longer has any reason
to care more about application locking than traditional Xlib does.

Furthermore, the handoff approach provides great improvements to
performance.  Results from x11perf's XNoOp test, which represented the
worst case for the lock-based Xlib/XCB:

Traditional Xlib:       average 19100000/sec
Lock-based Xlib/XCB:    average  3350000/sec
Handoff-based Xlib/XCB: average 17400000/sec

Thus, for no-ops, the handoff mechanism provides more than a 4x speedup to
Xlib/XCB, bringing Xlib/XCB within 9% of traditional Xlib no-op
performance.  Of course, real-world workloads do not use no-op, so your
mileage may vary.  In particular, since no-ops represent the worst case,
we expect real workloads to more closely match the performance of
traditional Xlib.

While removing synchronization code, we changed _XReply to not drop any
locks when calling xcb_wait_for_reply; previously, we had to carefully
avoid a deadlock between the Display lock and the XCB Xlib lock. Holding
the locks reduces implementation complexity and should not impact
applications.

Commit by Jamey Sharp and Josh Triplett.
XCB's handoff mechanism inspired by Keith Packard.

6 files changed, 140 insertions, 346 deletions

diff --git a/configure.ac b/configure.ac
index 73bf00fb..9a254ad7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -53,6 +53,8 @@ no)
 	AC_DEFINE(USE_XCB, 0, [Use XCB for low-level protocol implementation])
 	;;
 *)
+	X11_REQUIRES="xcb >= 1.1.92"
+	X11_EXTRA_DEPS="xcb >= 1.1.92"
 	xdmauth="no" # XCB handles all auth
 	AC_DEFINE(USE_XCB, 1, [Use XCB for low-level protocol implementation])
 	;;
diff --git a/src/Makefile.am b/src/Makefile.am
index 564e03f3..3380f81c 100644
--- a/src/Makefile.am
+++ b/src/Makefile.am
@@ -342,7 +342,6 @@ EXTRA_DIST = \
 
 if XCB
 libX11_la_SOURCES += \
-                  xcb_lock.c \
                   xcb_disp.c \
                   xcb_io.c \
                   Xxcbint.h
diff --git a/src/OpenDis.c b/src/OpenDis.c
index d5089e81..aa2541fa 100644
--- a/src/OpenDis.c
+++ b/src/OpenDis.c
@@ -261,13 +261,6 @@ XOpenDisplay (
 		return(NULL);
 	}
 
-#if USE_XCB
-	if (!_XCBInitDisplayLock(dpy)) {
-	        OutOfMemory (dpy, setup);
-		return(NULL);
-	}
-#endif
-
 	if (!_XPollfdCacheInit(dpy)) {
 	        OutOfMemory (dpy, setup);
 		return(NULL);
@@ -292,6 +285,10 @@ XOpenDisplay (
          return(NULL);
     }
     dpy->bufmax = dpy->buffer + conn_buf_size;
+#if USE_XCB
+    dpy->xcb->real_bufmax = dpy->bufmax;
+    dpy->bufmax = dpy->buffer;
+#endif
 
 	/* Set up the input event queue and input event queue parameters. */
 	dpy->head = dpy->tail = NULL;
@@ -652,6 +649,12 @@ XOpenDisplay (
 	UnlockDisplay(dpy);
 #endif /* !USE_XCB */
 
+#if USE_XCB
+	dpy->bigreq_size = xcb_get_maximum_request_length(dpy->xcb->connection);
+	if(dpy->bigreq_size <= dpy->max_request_size)
+		dpy->bigreq_size = 0;
+#endif /* USE_XCB */
+
 /*
  * Set up other stuff clients are always going to use.
  */
@@ -673,12 +676,6 @@ XOpenDisplay (
  */
 	(void) XSynchronize(dpy, _Xdebug);
 
-#if USE_XCB
-	dpy->bigreq_size = xcb_get_maximum_request_length(dpy->xcb->connection);
-	if(dpy->bigreq_size <= dpy->max_request_size)
-		dpy->bigreq_size = 0;
-#endif /* USE_XCB */
-
 /*
  * get availability of large requests, and
  * get the resource manager database off the root window.
@@ -883,10 +880,6 @@ void _XFreeDisplayStructure(dpy)
  	    Xfree (dpy->scratch_buffer);
 	FreeDisplayLock(dpy);
 
-#if USE_XCB
-	_XCBShutdownDisplayLock(dpy);
-#endif /* USE_XCB */
-
 	if (dpy->qfree) {
 	    register _XQEvent *qelt = dpy->qfree;
 
diff --git a/src/Xxcbint.h b/src/Xxcbint.h
index cc73749a..1b3137b9 100644
--- a/src/Xxcbint.h
+++ b/src/Xxcbint.h
@@ -5,34 +5,28 @@
 #define XXCBINT_H
 
 #include <assert.h>
+#include <stdint.h>
 #include <X11/Xlibint.h>
 #include <X11/Xlib-xcb.h>
-#include "locking.h" /* really just want X11/Xthreads.h but can't have it. */
 
-#define XCB_SEQUENCE_COMPARE(a,op,b)	((int) ((a) - (b)) op 0)
-#define assert_sequence_less(a,b) assert(XCB_SEQUENCE_COMPARE((a), <=, (b)))
+#define XLIB_SEQUENCE_COMPARE(a,op,b)	(((long) (a) - (long) (b)) op 0)
 
 typedef struct PendingRequest PendingRequest;
 struct PendingRequest {
 	PendingRequest *next;
-	xcondition_rec condition;
-	int waiters; /* Number of threads waiting; -1 if no wait needed */
-	unsigned int sequence;
+	unsigned long sequence;
 };
 
 typedef struct _X11XCBPrivate {
-	struct _XLockPtrs lock_fns;
 	xcb_connection_t *connection;
 	PendingRequest *pending_requests;
 	PendingRequest **pending_requests_tail;
 	xcb_generic_event_t *next_event;
-	const char *request_extra;
-	int request_extra_size;
-	char *partial_request;
-	int partial_request_offset;
+	char *real_bufmax;
 	char *reply_data;
 	int reply_length;
 	int reply_consumed;
+	uint64_t last_flushed;
 	enum XEventQueueOwner event_owner;
 	XID next_xid;
 } _X11XCBPrivate;
@@ -42,15 +36,4 @@ typedef struct _X11XCBPrivate {
 int _XConnectXCB(Display *dpy, _Xconst char *display, char **fullnamep, int *screenp);
 void _XFreeX11XCBStructure(Display *dpy);
 
-/* xcb_lock.c */
-
-int _XCBInitDisplayLock(Display *dpy);
-void _XCBShutdownDisplayLock(Display *dpy);
-
-/* _XGetXCBBuffer and _XPutXCBBuffer calls must be paired and must not
- * be nested. */
-
-void _XGetXCBBuffer(Display *dpy);
-void _XPutXCBBuffer(Display *dpy);
-
 #endif /* XXCBINT_H */
diff --git a/src/xcb_io.c b/src/xcb_io.c
index dcbe5a83..830ddb9f 100644
--- a/src/xcb_io.c
+++ b/src/xcb_io.c
@@ -5,12 +5,50 @@
 #include "locking.h"
 #include "Xxcbint.h"
 #include <xcb/xcbext.h>
-#include <xcb/xcbxlib.h>
 
 #include <assert.h>
+#include <inttypes.h>
+#include <stdint.h>
 #include <stdlib.h>
 #include <string.h>
 
+static void return_socket(void *closure)
+{
+	Display *dpy = closure;
+	LockDisplay(dpy);
+	_XSend(dpy, 0, 0);
+	dpy->bufmax = dpy->buffer;
+	UnlockDisplay(dpy);
+}
+
+static void require_socket(Display *dpy)
+{
+	if(dpy->bufmax == dpy->buffer)
+	{
+		uint64_t sent;
+		int flags = 0;
+		/* if we don't own the event queue, we have to ask XCB
+		 * to set our errors aside for us. */
+		if(dpy->xcb->event_owner != XlibOwnsEventQueue)
+			flags = XCB_REQUEST_CHECKED;
+		if(!xcb_take_socket(dpy->xcb->connection, return_socket, dpy,
+		                    flags, &sent))
+			_XIOError(dpy);
+		/* Xlib uses unsigned long for sequence numbers.  XCB
+		 * uses 64-bit internally, but currently exposes an
+		 * unsigned int API.  If these differ, Xlib cannot track
+		 * the full 64-bit sequence number if 32-bit wrap
+		 * happens while Xlib does not own the socket.  A
+		 * complete fix would be to make XCB's public API use
+		 * 64-bit sequence numbers. */
+		assert(!(sizeof(unsigned long) > sizeof(unsigned int)
+		         && dpy->xcb->event_owner == XlibOwnsEventQueue
+		         && (sent - dpy->last_request_read >= (UINT64_C(1) << 32))));
+		dpy->xcb->last_flushed = dpy->request = sent;
+		dpy->bufmax = dpy->xcb->real_bufmax;
+	}
+}
+
 /* Call internal connection callbacks for any fds that are currently
  * ready to read. This function will not block unless one of the
  * callbacks blocks.
@@ -68,15 +106,6 @@ static void check_internal_connections(Display *dpy)
 		}
 }
 
-static void condition_wait(Display *dpy, xcondition_t cv)
-{
-	_XPutXCBBuffer(dpy);
-	xcb_xlib_unlock(dpy->xcb->connection);
-	ConditionWait(dpy, cv);
-	xcb_xlib_lock(dpy->xcb->connection);
-	_XGetXCBBuffer(dpy);
-}
-
 static void call_handlers(Display *dpy, xcb_generic_reply_t *buf)
 {
 	_XAsyncHandler *async, *next;
@@ -105,7 +134,17 @@ static xcb_generic_event_t * wait_or_poll_for_event(Display *dpy, int wait)
 	return event;
 }
 
-static void process_responses(Display *dpy, int wait_for_first_event, xcb_generic_error_t **current_error, unsigned int current_request)
+/* Widen a 32-bit sequence number into a native-word-size (unsigned long)
+ * sequence number.  Treating the comparison as a 1 and shifting it avoids a
+ * conditional branch, and shifting by 16 twice avoids a compiler warning when
+ * sizeof(unsigned long) == 4. */
+static void widen(unsigned long *wide, unsigned int narrow)
+{
+	unsigned long new = (*wide & ~0xFFFFFFFFUL) | narrow;
+	*wide = new + ((unsigned long) (new < *wide) << 16 << 16);
+}
+
+static void process_responses(Display *dpy, int wait_for_first_event, xcb_generic_error_t **current_error, unsigned long current_request)
 {
 	void *reply;
 	xcb_generic_event_t *event = dpy->xcb->next_event;
@@ -114,13 +153,18 @@ static void process_responses(Display *dpy, int wait_for_first_event, xcb_generi
 	if(!event && dpy->xcb->event_owner == XlibOwnsEventQueue)
 		event = wait_or_poll_for_event(dpy, wait_for_first_event);
 
+	require_socket(dpy);
+
 	while(1)
 	{
 		PendingRequest *req = dpy->xcb->pending_requests;
-		assert(!(req && current_request && !XCB_SEQUENCE_COMPARE(req->sequence, <=, current_request)));
-		if(event && (!req || XCB_SEQUENCE_COMPARE(event->full_sequence, <=, req->sequence)))
+		unsigned long event_sequence = dpy->last_request_read;
+		if(event)
+			widen(&event_sequence, event->full_sequence);
+		assert(!(req && current_request && !XLIB_SEQUENCE_COMPARE(req->sequence, <=, current_request)));
+		if(event && (!req || XLIB_SEQUENCE_COMPARE(event_sequence, <=, req->sequence)))
 		{
-			dpy->last_request_read = event->full_sequence;
+			dpy->last_request_read = event_sequence;
 			if(event->response_type != X_Error)
 			{
 				/* GenericEvents may be > 32 bytes. In this
@@ -139,7 +183,7 @@ static void process_responses(Display *dpy, int wait_for_first_event, xcb_generi
 				_XEnq(dpy, (xEvent *) event);
 				wait_for_first_event = 0;
 			}
-			else if(current_error && event->full_sequence == current_request)
+			else if(current_error && event_sequence == current_request)
 			{
 				/* This can only occur when called from
 				 * _XReply, which doesn't need a new event. */
@@ -152,18 +196,9 @@ static void process_responses(Display *dpy, int wait_for_first_event, xcb_generi
 			free(event);
 			event = wait_or_poll_for_event(dpy, wait_for_first_event);
 		}
-		else if(req && req->waiters != -1)
+		else if(req && req->sequence == current_request)
 		{
-			if(req->sequence == current_request)
-				break;
-			if(!current_request && !wait_for_first_event)
-				break;
-			dpy->xcb->next_event = event;
-			req->waiters++;
-			assert(req->waiters > 0);
-			condition_wait(dpy, &req->condition);
-			--req->waiters;
-			event = dpy->xcb->next_event;
+			break;
 		}
 		else if(req && xcb_poll_for_reply(dpy->xcb->connection, req->sequence, &reply, &error))
 		{
@@ -192,7 +227,7 @@ static void process_responses(Display *dpy, int wait_for_first_event, xcb_generi
 	if(xcb_connection_has_error(c))
 		_XIOError(dpy);
 
-	assert_sequence_less(dpy->last_request_read, dpy->request);
+	assert(XLIB_SEQUENCE_COMPARE(dpy->last_request_read, <=, dpy->request));
 	assert(!wait_for_first_event);
 }
 
@@ -234,32 +269,61 @@ void _XReadEvents(Display *dpy)
  */
 void _XSend(Display *dpy, const char *data, long size)
 {
+	static const xReq dummy_request;
+	static char const pad[3];
+	struct iovec vec[3];
+	uint64_t requests;
+	_XExtension *ext;
 	xcb_connection_t *c = dpy->xcb->connection;
 	if(dpy->flags & XlibDisplayIOError)
 		return;
 
-	assert(!dpy->xcb->request_extra);
-	dpy->xcb->request_extra = data;
-	dpy->xcb->request_extra_size = size;
+	if(dpy->bufptr == dpy->buffer && !size)
+		return;
+
+	/* iff we asked XCB to set aside errors, we must pick those up
+	 * eventually. iff there are async handlers, we may have just
+	 * issued requests that will generate replies. in either case,
+	 * we need to remember to check later. */
+	if(dpy->xcb->event_owner != XlibOwnsEventQueue || dpy->async_handlers)
+	{
+		unsigned int sequence;
+		for(sequence = dpy->xcb->last_flushed; sequence < dpy->request; ++sequence)
+		{
+			PendingRequest *req = malloc(sizeof(PendingRequest));
+			assert(req);
+			req->next = 0;
+			req->sequence = sequence;
+			*dpy->xcb->pending_requests_tail = req;
+			dpy->xcb->pending_requests_tail = &req->next;
+		}
+	}
+	requests = dpy->request - dpy->xcb->last_flushed;
+	dpy->xcb->last_flushed = dpy->request;
 
-	/* give dpy->buffer to XCB */
-	_XPutXCBBuffer(dpy);
+	vec[0].iov_base = dpy->buffer;
+	vec[0].iov_len = dpy->bufptr - dpy->buffer;
+	vec[1].iov_base = (caddr_t) data;
+	vec[1].iov_len = size;
+	vec[2].iov_base = (caddr_t) pad;
+	vec[2].iov_len = -size & 3;
 
-	if(xcb_flush(c) <= 0)
-		_XIOError(dpy);
+	for(ext = dpy->flushes; ext; ext = ext->next_flush)
+	{
+		int i;
+		for(i = 0; i < 3; ++i)
+			if(vec[i].iov_len)
+				ext->before_flush(dpy, &ext->codes, vec[i].iov_base, vec[i].iov_len);
+	}
 
-	/* get a new dpy->buffer */
-	_XGetXCBBuffer(dpy);
+	if(xcb_writev(c, vec, 3, requests) < 0)
+		_XIOError(dpy);
+	dpy->bufptr = dpy->buffer;
+	dpy->last_req = (char *) &dummy_request;
 
 	check_internal_connections(dpy);
 
-	/* A straight port of XlibInt.c would call _XSetSeqSyncFunction
-	 * here. However that does no good: unlike traditional Xlib,
-	 * Xlib/XCB almost never calls _XFlush because _XPutXCBBuffer
-	 * automatically pushes requests down into XCB, so Xlib's buffer
-	 * is empty most of the time. Since setting a synchandler has no
-	 * effect until after UnlockDisplay returns, we may as well do
-	 * the check in _XUnlockDisplay. */
+	_XSetSeqSyncFunction(dpy);
 }
 
 /*
@@ -268,6 +332,7 @@ void _XSend(Display *dpy, const char *data, long size)
  */
 void _XFlush(Display *dpy)
 {
+	require_socket(dpy);
 	_XSend(dpy, 0, 0);
 
 	_XEventsQueued(dpy, QueuedAfterReading);
@@ -308,10 +373,18 @@ XID _XAllocID(Display *dpy)
 void _XAllocIDs(Display *dpy, XID *ids, int count)
 {
 	int i;
-	_XPutXCBBuffer(dpy);
+#ifdef XTHREADS
+	if (dpy->lock)
+		(*dpy->lock->user_lock_display)(dpy);
+	UnlockDisplay(dpy);
+#endif
 	for (i = 0; i < count; i++)
 		ids[i] = xcb_generate_id(dpy->xcb->connection);
-	_XGetXCBBuffer(dpy);
+#ifdef XTHREADS
+	LockDisplay(dpy);
+	if (dpy->lock)
+		(*dpy->lock->user_unlock_display)(dpy);
+#endif
 }
 
 static void _XFreeReplyData(Display *dpy, Bool force)
@@ -325,16 +398,9 @@ static void _XFreeReplyData(Display *dpy, Bool force)
 static PendingRequest * insert_pending_request(Display *dpy)
 {
 	PendingRequest **cur = &dpy->xcb->pending_requests;
-	while(*cur && XCB_SEQUENCE_COMPARE((*cur)->sequence, <, dpy->request))
+	while(*cur && XLIB_SEQUENCE_COMPARE((*cur)->sequence, <, dpy->request))
 		cur = &((*cur)->next);
-	if(*cur && (*cur)->sequence == dpy->request)
-	{
-		/* Replacing an existing PendingRequest should only happen once,
-		   when calling _XReply, and the replaced PendingRequest must
-		   not have a condition set. */
-		assert((*cur)->waiters == -1);
-	}
-	else
+	if(!*cur || (*cur)->sequence != dpy->request)
 	{
 		PendingRequest *node = malloc(sizeof(PendingRequest));
 		assert(node);
@@ -344,8 +410,6 @@ static PendingRequest * insert_pending_request(Display *dpy)
 			dpy->xcb->pending_requests_tail = &(node->next);
 		*cur = node;
 	}
-	(*cur)->waiters = 0;
-	xcondition_init(&((*cur)->condition));
 	return *cur;
 }
 
@@ -367,27 +431,15 @@ Status _XReply(Display *dpy, xReply *rep, int extra, Bool discard)
 	if(dpy->flags & XlibDisplayIOError)
 		return 0;
 
-	/* Internals of UnlockDisplay done by hand here, so that we can
-	   insert_pending_request *after* we _XPutXCBBuffer, but before we
-	   unlock the display. */
-	_XPutXCBBuffer(dpy);
+	_XSend(dpy, 0, 0);
 	current = insert_pending_request(dpy);
-	if(!dpy->lock || dpy->lock->locking_level == 0)
-		xcb_xlib_unlock(dpy->xcb->connection);
-	if(dpy->xcb->lock_fns.unlock_display)
-		dpy->xcb->lock_fns.unlock_display(dpy);
+	/* FIXME: drop the Display lock while waiting?
+	 * Complicates process_responses. */
 	reply = xcb_wait_for_reply(c, current->sequence, &error);
-	LockDisplay(dpy);
 
 	check_internal_connections(dpy);
 	process_responses(dpy, 0, &error, current->sequence);
 
-	if(current->waiters)
-	{ /* The ConditionBroadcast macro contains an if; braces needed here. */
-		ConditionBroadcast(dpy, &current->condition);
-	}
-	--current->waiters;
-
 	if(error)
 	{
 		_XExtension *ext;
diff --git a/src/xcb_lock.c b/src/xcb_lock.c
deleted file mode 100644
index 71b23835..00000000
--- a/src/xcb_lock.c
+++ /dev/null
@@ -1,235 +0,0 @@
-/* Copyright (C) 2003-2006 Jamey Sharp, Josh Triplett
- * This file is licensed under the MIT license. See the file COPYING. */
-
-#ifdef HAVE_CONFIG_H
-#include <config.h>
-#endif
-
-#include "Xlibint.h"
-#include "locking.h"
-#include "Xxcbint.h"
-#include <xcb/xcbext.h>
-#include <xcb/xcbxlib.h>
-
-#include <pthread.h>
-
-static void _XCBLockDisplay(Display *dpy)
-{
-    if(dpy->xcb->lock_fns.lock_display)
-	dpy->xcb->lock_fns.lock_display(dpy);
-    if(!dpy->lock || dpy->lock->locking_level == 0)
-	xcb_xlib_lock(dpy->xcb->connection);
-    if(!(dpy->flags & XlibDisplayIOError))
-	_XGetXCBBuffer(dpy);
-}
-
-/* XXX: If you change this function, update _XReply's copy of its guts! */
-static void _XCBUnlockDisplay(Display *dpy)
-{
-    if(!(dpy->flags & XlibDisplayIOError))
-    {
-	_XPutXCBBuffer(dpy);
-	assert(dpy->xcb->partial_request == 0);
-	assert(xcb_get_request_sent(dpy->xcb->connection) == dpy->request);
-
-	/* Traditional Xlib does this in _XSend; see the Xlib/XCB version
-	 * of that function for why we do it here instead. */
-	_XSetSeqSyncFunction(dpy);
-    }
-
-    if(!dpy->lock || dpy->lock->locking_level == 0)
-	xcb_xlib_unlock(dpy->xcb->connection);
-    if(dpy->xcb->lock_fns.unlock_display)
-	dpy->xcb->lock_fns.unlock_display(dpy);
-}
-
-int _XCBInitDisplayLock(Display *dpy)
-{
-    if(!dpy->lock_fns && !(dpy->lock_fns = Xcalloc(1, sizeof(*dpy->lock_fns))))
-	return 0;
-    dpy->xcb->lock_fns.lock_display = dpy->lock_fns->lock_display;
-    dpy->lock_fns->lock_display = _XCBLockDisplay;
-    dpy->xcb->lock_fns.unlock_display = dpy->lock_fns->unlock_display;
-    dpy->lock_fns->unlock_display = _XCBUnlockDisplay;
-    return 1;
-}
-
-void _XCBShutdownDisplayLock(Display *dpy)
-{
-    if(dpy->lock_fns) {
-	Xfree((char *)dpy->lock_fns);
-	dpy->lock_fns = NULL;
-    }
-}
-
-void _XGetXCBBuffer(Display *dpy)
-{
-    static const xReq dummy_request;
-    unsigned int xcb_req = xcb_get_request_sent(dpy->xcb->connection);
-    if(xcb_connection_has_error(dpy->xcb->connection))
-	_XIOError(dpy);
-
-    /* if Xlib has a partial request pending then XCB doesn't know about
-     * the current request yet */
-    if(dpy->xcb->partial_request)
-	++xcb_req;
-
-    assert(XCB_SEQUENCE_COMPARE(xcb_req, >=, dpy->request));
-    dpy->request = xcb_req;
-
-    dpy->last_req = (char *) &dummy_request;
-}
-
-static size_t request_length(struct iovec *vec)
-{
-    /* we have at least part of a request. dig out the length field.
-     * note that length fields are always in vec[0]: Xlib doesn't split
-     * fixed-length request parts. */
-    size_t len;
-    assert(vec[0].iov_len >= 4);
-    len = ((uint16_t *) vec[0].iov_base)[1];
-    if(len == 0)
-    {
-	/* it's a bigrequest. dig out the *real* length field. */
-	assert(vec[0].iov_len >= 8);
-	len = ((uint32_t *) vec[0].iov_base)[1];
-    }
-    return len << 2;
-}
-
-static inline int issue_complete_request(Display *dpy, int veclen, struct iovec *vec)
-{
-    xcb_protocol_request_t xcb_req = { 0 };
-    unsigned int sequence;
-    int flags = XCB_REQUEST_RAW;
-    int i;
-    size_t len;
-
-    /* skip empty iovecs. if no iovecs remain, we're done. */
-    assert(veclen >= 0);
-    while(veclen > 0 && vec[0].iov_len == 0)
-	--veclen, ++vec;
-    if(!veclen)
-	return 0;
-
-    len = request_length(vec);
-
-    /* do we have enough data for a complete request? how many iovec
-     * elements does it span? */
-    for(i = 0; i < veclen; ++i)
-    {
-	size_t oldlen = len;
-	len -= vec[i].iov_len;
-	/* if len is now 0 or has wrapped, we have enough data. */
-	if((len - 1) > oldlen)
-	    break;
-    }
-    if(i == veclen)
-	return 0;
-
-    /* we have enough data to issue one complete request. the remaining
-     * code can't fail. */
-
-    /* len says how far we overshot our data needs. (it's "negative" if
-     * we actually overshot, or 0 if we're right on.) */
-    vec[i].iov_len += len;
-    xcb_req.count = i + 1;
-    xcb_req.opcode = ((uint8_t *) vec[0].iov_base)[0];
-
-    /* if we don't own the event queue, we have to ask XCB to set our
-     * errors aside for us. */
-    if(dpy->xcb->event_owner != XlibOwnsEventQueue)
-	flags |= XCB_REQUEST_CHECKED;
-
-    /* XCB will always skip request 0; account for that in the Xlib count */
-    if (xcb_get_request_sent(dpy->xcb->connection) == 0xffffffff)
-	dpy->request++;
-    /* send the accumulated request. */
-    sequence = xcb_send_request(dpy->xcb->connection, flags, vec, &xcb_req);
-    if(!sequence)
-	_XIOError(dpy);
-
-    /* update the iovecs to refer only to data not yet sent. */
-    vec[i].iov_len = -len;
-
-    /* iff we asked XCB to set aside errors, we must pick those up
-     * eventually. iff there are async handlers, we may have just
-     * issued requests that will generate replies. in either case,
-     * we need to remember to check later. */
-    if(flags & XCB_REQUEST_CHECKED || dpy->async_handlers)
-    {
-	PendingRequest *req = malloc(sizeof(PendingRequest));
-	assert(req);
-	req->next = 0;
-	req->waiters = -1;
-	req->sequence = sequence;
-	*dpy->xcb->pending_requests_tail = req;
-	dpy->xcb->pending_requests_tail = &req->next;
-    }
-    return 1;
-}
-
-void _XPutXCBBuffer(Display *dpy)
-{
-    static char const pad[3];
-    const int padsize = -dpy->xcb->request_extra_size & 3;
-    xcb_connection_t *c = dpy->xcb->connection;
-    _XExtension *ext;
-    struct iovec iov[6];
-
-    assert_sequence_less(dpy->last_request_read, dpy->request);
-    assert_sequence_less(xcb_get_request_sent(c), dpy->request);
-
-    for(ext = dpy->flushes; ext; ext = ext->next_flush)
-    {
-	ext->before_flush(dpy, &ext->codes, dpy->buffer, dpy->bufptr - dpy->buffer);
-	if(dpy->xcb->request_extra)
-	{
-	    ext->before_flush(dpy, &ext->codes, dpy->xcb->request_extra, dpy->xcb->request_extra_size);
-	    if(padsize)
-		ext->before_flush(dpy, &ext->codes, pad, padsize);
-	}
-    }
-
-    iov[2].iov_base = dpy->xcb->partial_request;
-    iov[2].iov_len = dpy->xcb->partial_request_offset;
-    iov[3].iov_base = dpy->buffer;
-    iov[3].iov_len = dpy->bufptr - dpy->buffer;
-    iov[4].iov_base = (caddr_t) dpy->xcb->request_extra;
-    iov[4].iov_len = dpy->xcb->request_extra_size;
-    iov[5].iov_base = (caddr_t) pad;
-    iov[5].iov_len = padsize;
-
-    while(issue_complete_request(dpy, 4, iov + 2))
-	/* empty */;
-
-    /* first discard any completed partial_request. */
-    if(iov[2].iov_len == 0 && dpy->xcb->partial_request)
-    {
-	free(dpy->xcb->partial_request);
-	dpy->xcb->partial_request = 0;
-	dpy->xcb->partial_request_offset = 0;
-    }
-
-    /* is there anything to copy into partial_request? */
-    if(iov[3].iov_len != 0 || iov[4].iov_len != 0 || iov[5].iov_len != 0)
-    {
-	int i;
-	if(!dpy->xcb->partial_request)
-	{
-	    size_t len = request_length(iov + 3);
-	    assert(!dpy->xcb->partial_request_offset);
-	    dpy->xcb->partial_request = malloc(len);
-	    assert(dpy->xcb->partial_request);
-	}
-	for(i = 3; i < sizeof(iov) / sizeof(*iov); ++i)
-	{
-	    memcpy(dpy->xcb->partial_request + dpy->xcb->partial_request_offset, iov[i].iov_base, iov[i].iov_len);
-	    dpy->xcb->partial_request_offset += iov[i].iov_len;
-	}
-    }
-
-    dpy->xcb->request_extra = 0;
-    dpy->xcb->request_extra_size = 0;
-    dpy->bufptr = dpy->buffer;
-}