From 7f0bc778c88ab2f565cc05d5d3d5ee4c8d1388a1 Mon Sep 17 00:00:00 2001 From: Jamey Sharp Date: Thu, 23 Feb 2006 22:41:59 -0800 Subject: [PATCH] Factor padding out of _xcb_out_write_block and into its callers, XCBSendRequest and write_setup. This requires dynamically allocating memory in XCBSendRequest, but this malloc/free pair turns out to cause a 30% speed hit for the 'x11perf -noop' test -- so for the moment I use alloca where available and fall back to malloc on other platforms. Later I think I'll change the contract of XCBSendRequest so the caller is responsible for memory allocation, because the caller ought to always be able to stack-allocate here. --- configure.ac | 1 + src/xcb_conn.c | 10 ++++- src/xcb_out.c | 114 +++++++++++++++++++++++++++++---------------------------- 3 files changed, 69 insertions(+), 56 deletions(-) diff --git a/configure.ac b/configure.ac index 5e4405c..a07a89e 100644 --- a/configure.ac +++ b/configure.ac @@ -45,6 +45,7 @@ AC_MSG_RESULT($XCBPROTO_XCBINCLUDEDIR) AC_SUBST(XCBPROTO_XCBINCLUDEDIR) AC_HEADER_STDC +AC_FUNC_ALLOCA AC_SEARCH_LIBS(gethostbyname, nsl) AC_SEARCH_LIBS(connect, socket) diff --git a/src/xcb_conn.c b/src/xcb_conn.c index 9a7399e..de55951 100644 --- a/src/xcb_conn.c +++ b/src/xcb_conn.c @@ -38,8 +38,9 @@ static int write_setup(XCBConnection *c, XCBAuthInfo *auth_info) { + static const char pad[3]; XCBConnSetupReq out; - struct iovec parts[3]; + struct iovec parts[6]; int count = 0; int endian = 0x01020304; int ret; @@ -57,14 +58,21 @@ static int write_setup(XCBConnection *c, XCBAuthInfo *auth_info) out.authorization_protocol_data_len = 0; parts[count].iov_len = sizeof(XCBConnSetupReq); parts[count++].iov_base = &out; + parts[count].iov_len = XCB_PAD(sizeof(XCBConnSetupReq)); + parts[count++].iov_base = (caddr_t) pad; if(auth_info) { parts[count].iov_len = out.authorization_protocol_name_len = auth_info->namelen; parts[count++].iov_base = auth_info->name; + parts[count].iov_len = XCB_PAD(out.authorization_protocol_name_len); + parts[count++].iov_base = (caddr_t) pad; parts[count].iov_len = out.authorization_protocol_data_len = auth_info->datalen; parts[count++].iov_base = auth_info->data; + parts[count].iov_len = XCB_PAD(out.authorization_protocol_data_len); + parts[count++].iov_base = (caddr_t) pad; } + assert(count <= sizeof(parts) / sizeof(*parts)); pthread_mutex_lock(&c->iolock); _xcb_out_write_block(c, parts, count); diff --git a/src/xcb_out.c b/src/xcb_out.c index 442fcf5..cecf7cb 100644 --- a/src/xcb_out.c +++ b/src/xcb_out.c @@ -31,6 +31,16 @@ #include #include +#ifndef __GNUC__ +# if HAVE_ALLOCA_H +# include +# else +# ifdef _AIX + #pragma alloca +# endif +# endif +#endif + #include "xcb.h" #include "xcbext.h" #include "xcbint.h" @@ -105,9 +115,11 @@ CARD32 XCBGetMaximumRequestLength(XCBConnection *c) int XCBSendRequest(XCBConnection *c, unsigned int *request, struct iovec *vector, const XCBProtocolRequest *req) { + static const char pad[3]; int ret; int i; - struct iovec prefix[2]; + struct iovec *padded; + int padlen = 0; CARD16 shortlen = 0; CARD32 longlen = 0; enum workarounds workaround = WORKAROUND_NONE; @@ -151,17 +163,37 @@ int XCBSendRequest(XCBConnection *c, unsigned int *request, struct iovec *vector longlen = 0; } + padded = +#ifdef HAVE_ALLOCA + alloca +#else + malloc +#endif + ((req->count * 2 + 3) * sizeof(struct iovec)); /* set the length field. */ ((CARD16 *) vector[0].iov_base)[1] = shortlen; if(!shortlen) { - prefix[0].iov_base = vector[0].iov_base; - prefix[0].iov_len = sizeof(CARD32); + padded[0].iov_base = vector[0].iov_base; + padded[0].iov_len = sizeof(CARD32); vector[0].iov_base = ((char *) vector[0].iov_base) + sizeof(CARD32); vector[0].iov_len -= sizeof(CARD32); ++longlen; - prefix[1].iov_base = &longlen; - prefix[1].iov_len = sizeof(CARD32); + padded[1].iov_base = &longlen; + padded[1].iov_len = sizeof(CARD32); + padlen = 2; + } + + for(i = 0; i < req->count; ++i) + { + if(!vector[i].iov_len) + continue; + padded[padlen].iov_base = vector[i].iov_base; + padded[padlen++].iov_len = vector[i].iov_len; + if(!XCB_PAD(vector[i].iov_len)) + continue; + padded[padlen].iov_base = (caddr_t) pad; + padded[padlen++].iov_len = XCB_PAD(vector[i].iov_len); } /* get a sequence number and arrange for delivery. */ @@ -169,6 +201,9 @@ int XCBSendRequest(XCBConnection *c, unsigned int *request, struct iovec *vector if(req->isvoid && !force_sequence_wrap(c)) { pthread_mutex_unlock(&c->iolock); +#ifndef HAVE_ALLOCA + free(padded); +#endif return -1; } @@ -177,13 +212,11 @@ int XCBSendRequest(XCBConnection *c, unsigned int *request, struct iovec *vector if(!req->isvoid) _xcb_in_expect_reply(c, *request, workaround); - if(!shortlen) - ret = _xcb_out_write_block(c, prefix, 2); - else - ret = 1; - if(ret > 0) - ret = _xcb_out_write_block(c, vector, req->count); + ret = _xcb_out_write_block(c, padded, padlen); pthread_mutex_unlock(&c->iolock); +#ifndef HAVE_ALLOCA + free(padded); +#endif return ret; } @@ -244,6 +277,7 @@ int _xcb_out_write(XCBConnection *c) for(i = 0; i < c->out.vec_len; ++i) if(c->out.vec[i].iov_len) return n; + c->out.vec = 0; c->out.vec_len = 0; } return n; @@ -251,55 +285,25 @@ int _xcb_out_write(XCBConnection *c) int _xcb_out_write_block(XCBConnection *c, struct iovec *vector, size_t count) { - static const char pad[3]; - int i; - int len = 0; - - for(i = 0; i < count; ++i) - len += XCB_CEIL(vector[i].iov_len); - - /* Is the queue about to overflow? */ - if(c->out.queue_len + len < sizeof(c->out.queue)) + while(count && c->out.queue_len + vector[0].iov_len < sizeof(c->out.queue)) { - /* No, this will fit. */ - for(i = 0; i < count; ++i) - { - memcpy(c->out.queue + c->out.queue_len, vector[i].iov_base, vector[i].iov_len); - if(vector[i].iov_len & 3) - memset(c->out.queue + c->out.queue_len + vector[i].iov_len, 0, XCB_PAD(vector[i].iov_len)); - c->out.queue_len += XCB_CEIL(vector[i].iov_len); - } - return len; + memcpy(c->out.queue + c->out.queue_len, vector[0].iov_base, vector[0].iov_len); + c->out.queue_len += vector[0].iov_len; + ++vector, --count; } + if(!count) + return 1; + + memmove(vector + 1, vector, count++ * sizeof(struct iovec)); + vector[0].iov_base = c->out.queue; + vector[0].iov_len = c->out.queue_len; + c->out.queue_len = 0; assert(!c->out.vec_len); assert(!c->out.vec); - c->out.vec = malloc(sizeof(struct iovec) * (1 + count * 2)); - if(!c->out.vec) - return -1; - if(c->out.queue_len) - { - c->out.vec[c->out.vec_len].iov_base = c->out.queue; - c->out.vec[c->out.vec_len++].iov_len = c->out.queue_len; - c->out.queue_len = 0; - } - for(i = 0; i < count; ++i) - { - if(!vector[i].iov_len) - continue; - c->out.vec[c->out.vec_len].iov_base = vector[i].iov_base; - c->out.vec[c->out.vec_len++].iov_len = vector[i].iov_len; - if(!XCB_PAD(vector[i].iov_len)) - continue; - c->out.vec[c->out.vec_len].iov_base = (void *) pad; - c->out.vec[c->out.vec_len++].iov_len = XCB_PAD(vector[i].iov_len); - } - if(!_xcb_out_flush(c)) - len = -1; - free(c->out.vec); - c->out.vec = 0; - - return len; + c->out.vec_len = count; + c->out.vec = vector; + return _xcb_out_flush(c); } int _xcb_out_flush(XCBConnection *c) -- 2.7.4