Date: Mon, 29 Nov 2010 21:04:00 +0000 (UTC) From: Juli Mallett <jmallett@FreeBSD.org> To: src-committers@freebsd.org, svn-src-all@freebsd.org, svn-src-head@freebsd.org Subject: svn commit: r216064 - head/sys/mips/cavium/octe Message-ID: <201011292104.oATL40J3086207@svn.freebsd.org>
next in thread | raw e-mail | index | archive | help
Author: jmallett Date: Mon Nov 29 21:04:00 2010 New Revision: 216064 URL: http://svn.freebsd.org/changeset/base/216064 Log: Don't free the work queue entry that we're using to hold the scatter-gather list on exit from the transmit path. The scatter-gather list itself can be asynchronously DMAed to the transmit hardware, and we could actually lock up the transmitter if any of a number of races around this were lost. Instead, let the PKO free the scatter-gather list when it is done with it, and use the "i" bit in each segment of the packet to avoid having them go into the FPA. This fixes an unrecoverable transmit stall under transmit load. MFC after: 3 days Modified: head/sys/mips/cavium/octe/ethernet-tx.c Modified: head/sys/mips/cavium/octe/ethernet-tx.c ============================================================================== --- head/sys/mips/cavium/octe/ethernet-tx.c Mon Nov 29 20:43:06 2010 (r216063) +++ head/sys/mips/cavium/octe/ethernet-tx.c Mon Nov 29 21:04:00 2010 (r216064) @@ -143,6 +143,7 @@ int cvm_oct_xmit(struct mbuf *m, struct /* Build the PKO command */ pko_command.u64 = 0; pko_command.s.segs = 1; + pko_command.s.dontfree = 1; /* Do not put this buffer into the FPA. */ work = NULL; } else { @@ -164,6 +165,7 @@ int cvm_oct_xmit(struct mbuf *m, struct /* Build the PKO buffer pointer */ hw_buffer.u64 = 0; + hw_buffer.s.i = 1; /* Do not put this buffer into the FPA. */ hw_buffer.s.addr = cvmx_ptr_to_phys(n->m_data); hw_buffer.s.pool = 0; hw_buffer.s.size = n->m_len; @@ -182,12 +184,11 @@ int cvm_oct_xmit(struct mbuf *m, struct pko_command.u64 = 0; pko_command.s.segs = segs; pko_command.s.gather = 1; + pko_command.s.dontfree = 0; /* Put the WQE above back into the FPA. */ } /* Finish building the PKO command */ pko_command.s.n2 = 1; /* Don't pollute L2 with the outgoing packet */ - pko_command.s.dontfree = 1; - pko_command.s.reg0 = priv->fau+qos*4; pko_command.s.reg0 = priv->fau+qos*4; pko_command.s.total_bytes = m->m_pkthdr.len; pko_command.s.size0 = CVMX_FAU_OP_SIZE_32; @@ -200,6 +201,11 @@ int cvm_oct_xmit(struct mbuf *m, struct pko_command.s.ipoffp1 = ETHER_HDR_LEN + 1; } + /* + * XXX + * Could use a different free queue (and different FAU address) per + * core instead of per QoS, to reduce contention here. + */ IF_LOCK(&priv->tx_free_queue[qos]); if (USE_ASYNC_IOBDMA) { /* Get the number of mbufs in use by the hardware */ @@ -242,8 +248,6 @@ int cvm_oct_xmit(struct mbuf *m, struct /* Pass it to any BPF listeners. */ ETHER_BPF_MTAP(ifp, m); } - if (work != NULL) - cvmx_fpa_free(work, CVMX_FPA_WQE_POOL, DONT_WRITEBACK(1)); /* Free mbufs not in use by the hardware */ if (_IF_QLEN(&priv->tx_free_queue[qos]) > in_use) {
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?201011292104.oATL40J3086207>