From owner-svn-src-head@FreeBSD.ORG Thu Dec 3 23:57:06 2009 Return-Path: Delivered-To: svn-src-head@freebsd.org Received: from mx1.freebsd.org (mx1.freebsd.org [IPv6:2001:4f8:fff6::34]) by hub.freebsd.org (Postfix) with ESMTP id CE0D010656A9; Thu, 3 Dec 2009 23:57:06 +0000 (UTC) (envelope-from yongari@FreeBSD.org) Received: from svn.freebsd.org (svn.freebsd.org [IPv6:2001:4f8:fff6::2c]) by mx1.freebsd.org (Postfix) with ESMTP id BC27A8FC21; Thu, 3 Dec 2009 23:57:06 +0000 (UTC) Received: from svn.freebsd.org (localhost [127.0.0.1]) by svn.freebsd.org (8.14.3/8.14.3) with ESMTP id nB3Nv6I0029774; Thu, 3 Dec 2009 23:57:06 GMT (envelope-from yongari@svn.freebsd.org) Received: (from yongari@localhost) by svn.freebsd.org (8.14.3/8.14.3/Submit) id nB3Nv6dx029772; Thu, 3 Dec 2009 23:57:06 GMT (envelope-from yongari@svn.freebsd.org) Message-Id: <200912032357.nB3Nv6dx029772@svn.freebsd.org> From: Pyun YongHyeon Date: Thu, 3 Dec 2009 23:57:06 +0000 (UTC) To: src-committers@freebsd.org, svn-src-all@freebsd.org, svn-src-head@freebsd.org X-SVN-Group: head MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Cc: Subject: svn commit: r200088 - head/sys/dev/bge X-BeenThere: svn-src-head@freebsd.org X-Mailman-Version: 2.1.5 Precedence: list List-Id: SVN commit messages for the src tree for head/-current List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , X-List-Received-Date: Thu, 03 Dec 2009 23:57:07 -0000 Author: yongari Date: Thu Dec 3 23:57:06 2009 New Revision: 200088 URL: http://svn.freebsd.org/changeset/base/200088 Log: Add workaround to overcome hardware limitation which allows only a single outstanding DMA read operation. Most controllers targeted to client with PCIe bus interface(e.g. BCM5761) may have this limitation. All controllers for servers does not have this limitation. Collapsing mbuf chains to reduce number of memory reads before transmitting was most effective way to workaround this. I got about 940Mbps from 850Mbps with mbuf collapsing on BCM5761. However it takes a lot of CPU cycles to collapse mbuf chains so add tunable to control the number of allowed TX buffers before collapsing. The default value is 0 which effectively disables the forced collapsing. For most cases 2 would yield best performance(about 930Mbps) without much sacrificing CPU cycles. Note the collapsing is only activated when the controller is on PCIe bus and the frame does not need TSO operation. TSO does not seem to suffer from the hardware limitation because the payload size is much bigger than normal IP datagram. Thanks to davidch@ who told me the limitation of client controllers and actually gave possible workarounds to mitigate the limitation. Reviewed by: davidch, marius Modified: head/sys/dev/bge/if_bge.c Modified: head/sys/dev/bge/if_bge.c ============================================================================== --- head/sys/dev/bge/if_bge.c Thu Dec 3 23:24:12 2009 (r200087) +++ head/sys/dev/bge/if_bge.c Thu Dec 3 23:57:06 2009 (r200088) @@ -483,12 +483,29 @@ DRIVER_MODULE(bge, pci, bge_driver, bge_ DRIVER_MODULE(miibus, bge, miibus_driver, miibus_devclass, 0, 0); static int bge_allow_asf = 1; +/* + * A common design characteristic for many Broadcom client controllers + * is that they only support a single outstanding DMA read operation + * on the PCIe bus. This means that it will take twice as long to fetch + * a TX frame that is split into header and payload buffers as it does + * to fetch a single, contiguous TX frame (2 reads vs. 1 read). For + * these controllers, coalescing buffers to reduce the number of memory + * reads is effective way to get maximum performance(about 940Mbps). + * Without collapsing TX buffers the maximum TCP bulk transfer + * performance is about 850Mbps. However forcing coalescing mbufs + * consumes a lot of CPU cycles, so leave it off by default. + */ +static int bge_forced_collapse = 0; TUNABLE_INT("hw.bge.allow_asf", &bge_allow_asf); +TUNABLE_INT("hw.bge.forced_collapse", &bge_forced_collapse); SYSCTL_NODE(_hw, OID_AUTO, bge, CTLFLAG_RD, 0, "BGE driver parameters"); SYSCTL_INT(_hw_bge, OID_AUTO, allow_asf, CTLFLAG_RD, &bge_allow_asf, 0, "Allow ASF mode if available"); +SYSCTL_INT(_hw_bge, OID_AUTO, forced_collapse, CTLFLAG_RD, &bge_forced_collapse, + 0, "Number of fragmented TX buffers of a frame allowed before " + "forced collapsing"); #define SPARC64_BLADE_1500_MODEL "SUNW,Sun-Blade-1500" #define SPARC64_BLADE_1500_PATH_BGE "/pci@1f,700000/network@2" @@ -3915,6 +3932,26 @@ bge_encap(struct bge_softc *sc, struct m csum_flags |= BGE_TXBDFLAG_IP_FRAG; } + if ((m->m_pkthdr.csum_flags & CSUM_TSO) == 0 && + bge_forced_collapse > 0 && (sc->bge_flags & BGE_FLAG_PCIE) != 0 && + m->m_next != NULL) { + /* + * Forcedly collapse mbuf chains to overcome hardware + * limitation which only support a single outstanding + * DMA read operation. + */ + if (bge_forced_collapse == 1) + m = m_defrag(m, M_DONTWAIT); + else + m = m_collapse(m, M_DONTWAIT, bge_forced_collapse); + if (m == NULL) { + m_freem(*m_head); + *m_head = NULL; + return (ENOBUFS); + } + *m_head = m; + } + map = sc->bge_cdata.bge_tx_dmamap[idx]; error = bus_dmamap_load_mbuf_sg(sc->bge_cdata.bge_tx_mtag, map, m, segs, &nsegs, BUS_DMA_NOWAIT);