From: Stefan Metzmacher Date: Thu, 2 Apr 2009 08:36:03 +0000 (+0200) Subject: tsocket: optimize tdgram_bsd a lot X-Git-Tag: tdb-1.1.5~1031^2~118 X-Git-Url: http://git.samba.org/samba.git/?p=ira%2Fwip.git;a=commitdiff_plain;h=c59ee5a139421762adb6f3f4bbfc21723c2ce407;ds=sidebyside tsocket: optimize tdgram_bsd a lot The desire is to do as less syscalls during the tdgram_sendto_send/recv() and tdgram_recvfrom_send/recv() operations. 1. we first try the sendto()/recvfrom() syscall and only use a fd event if we got EAGAIN. 2. we cache the fd event and only change it's flags if really needed. For the highload case we do almost no epoll_ctl() and epoll_wait()/select() syscalls anymore. This speeds up the LDAP-BENCH-CLDAP test by more than 20%. (With a modified version of this test which let the server skip any ldb calls and just return success I'm getting about 8000 requests per second, while I'm getting just about 6000 requests per second without optimization) metze --- diff --git a/lib/tsocket/tsocket_bsd.c b/lib/tsocket/tsocket_bsd.c index 6c60ef2ebd1..db1fd38bdb0 100644 --- a/lib/tsocket/tsocket_bsd.c +++ b/lib/tsocket/tsocket_bsd.c @@ -1313,6 +1313,10 @@ static void tdgram_bsd_fde_handler(struct tevent_context *ev, return; } if (flags & TEVENT_FD_READ) { + if (!bsds->readable_handler) { + TEVENT_FD_NOT_READABLE(bsds->fde); + return; + } bsds->readable_handler(bsds->readable_private); return; } @@ -1328,19 +1332,25 @@ static int tdgram_bsd_set_readable_handler(struct tdgram_bsd *bsds, errno = EINVAL; return -1; } - + if (!bsds->readable_handler) { + return 0; + } bsds->readable_handler = NULL; bsds->readable_private = NULL; - TEVENT_FD_NOT_READABLE(bsds->fde); - if (bsds->fde && !bsds->writeable_handler) { - /* we don't need the fd event anymore */ - bsds->event_ptr = NULL; - TALLOC_FREE(bsds->fde); - } return 0; } + /* read and write must use the same tevent_context */ + if (bsds->event_ptr != ev) { + if (bsds->readable_handler || bsds->writeable_handler) { + errno = EINVAL; + return -1; + } + bsds->event_ptr = NULL; + TALLOC_FREE(bsds->fde); + } + if (bsds->fde == NULL) { bsds->fde = tevent_add_fd(ev, bsds, bsds->fd, TEVENT_FD_READ, @@ -1352,15 +1362,10 @@ static int tdgram_bsd_set_readable_handler(struct tdgram_bsd *bsds, /* cache the event context we're running on */ bsds->event_ptr = ev; + } else if (!bsds->readable_handler) { + TEVENT_FD_READABLE(bsds->fde); } - /* read and write must use the same tevent_context */ - if (bsds->event_ptr != ev) { - errno = EINVAL; - return -1; - } - - TEVENT_FD_READABLE(bsds->fde); bsds->readable_handler = handler; bsds->readable_private = private_data; @@ -1377,19 +1382,26 @@ static int tdgram_bsd_set_writeable_handler(struct tdgram_bsd *bsds, errno = EINVAL; return -1; } - + if (!bsds->writeable_handler) { + return 0; + } bsds->writeable_handler = NULL; bsds->writeable_private = NULL; TEVENT_FD_NOT_WRITEABLE(bsds->fde); - if (bsds->fde && !bsds->readable_handler) { - /* we don't need the fd event anymore */ - bsds->event_ptr = NULL; - TALLOC_FREE(bsds->fde); - } return 0; } + /* read and write must use the same tevent_context */ + if (bsds->event_ptr != ev) { + if (bsds->readable_handler || bsds->writeable_handler) { + errno = EINVAL; + return -1; + } + bsds->event_ptr = NULL; + TALLOC_FREE(bsds->fde); + } + if (bsds->fde == NULL) { bsds->fde = tevent_add_fd(ev, bsds, bsds->fd, TEVENT_FD_WRITE, @@ -1401,15 +1413,10 @@ static int tdgram_bsd_set_writeable_handler(struct tdgram_bsd *bsds, /* cache the event context we're running on */ bsds->event_ptr = ev; + } else if (!bsds->writeable_handler) { + TEVENT_FD_WRITEABLE(bsds->fde); } - /* read and write must use the same tevent_context */ - if (bsds->event_ptr != ev) { - errno = EINVAL; - return -1; - } - - TEVENT_FD_WRITEABLE(bsds->fde); bsds->writeable_handler = handler; bsds->writeable_private = private_data; @@ -1470,6 +1477,16 @@ static struct tevent_req *tdgram_bsd_recvfrom_send(TALLOC_CTX *mem_ctx, goto post; } + /* + * this is a fast path, not waiting for the + * socket to become explicit readable gains + * about 10%-20% performance in benchmark tests. + */ + tdgram_bsd_recvfrom_handler(req); + if (!tevent_req_is_in_progress(req)) { + goto post; + } + ret = tdgram_bsd_set_readable_handler(bsds, ev, tdgram_bsd_recvfrom_handler, req); @@ -1634,6 +1651,16 @@ static struct tevent_req *tdgram_bsd_sendto_send(TALLOC_CTX *mem_ctx, goto post; } + /* + * this is a fast path, not waiting for the + * socket to become explicit writeable gains + * about 10%-20% performance in benchmark tests. + */ + tdgram_bsd_sendto_handler(req); + if (!tevent_req_is_in_progress(req)) { + goto post; + } + ret = tdgram_bsd_set_writeable_handler(bsds, ev, tdgram_bsd_sendto_handler, req);