From mboxrd@z Thu Jan 1 00:00:00 1970 From: ANNIE LI Subject: Re: [PATCH 2/4] xen/netback: Split one page pool into two(tx/rx) page pool. Date: Fri, 16 Nov 2012 11:10:01 +0800 Message-ID: <50A5AE89.3090500@oracle.com> References: <1352962987-541-1-git-send-email-annie.li@oracle.com> <1352963089-599-1-git-send-email-annie.li@oracle.com> <1352970913.3499.47.camel@zakaz.uk.xensource.com> Mime-Version: 1.0 Content-Type: text/plain; charset=UTF-8; format=flowed Content-Transfer-Encoding: 7bit Cc: "xen-devel@lists.xensource.com" , "netdev@vger.kernel.org" , "konrad.wilk@oracle.com" To: Ian Campbell Return-path: Received: from aserp1040.oracle.com ([141.146.126.69]:24149 "EHLO aserp1040.oracle.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751084Ab2KPDKN (ORCPT ); Thu, 15 Nov 2012 22:10:13 -0500 In-Reply-To: <1352970913.3499.47.camel@zakaz.uk.xensource.com> Sender: netdev-owner@vger.kernel.org List-ID: On 2012-11-15 17:15, Ian Campbell wrote: > On Thu, 2012-11-15 at 07:04 +0000, Annie Li wrote: >> For tx path, this implementation simplifies the work of searching out >> grant page from page pool based on grant reference. > It's still a linear search though, and it doesn't look much simpler to > me: > for (i = 0; i< count; i++) { > if (tx_pool) > vif = netbk->gnttab_tx_vif[i]; > else > vif = netbk->gnttab_rx_vif[i]; > > pers_entry = vif->persistent_gnt; > gnt_count =&vif->persistent_gntcnt; > gnt_total = MAXIMUM_OUTSTANDING_BLOCK_REQS; > becomes: > for (i = 0; i< count; i++) { > if (tx_pool) { > vif = netbk->gnttab_tx_vif[i]; > gnt_count =&vif->persistent_tx_gntcnt; > gnt_total = XEN_NETIF_TX_RING_SIZE; > pers_entry = vif->persistent_tx_gnt; > } else { > vif = netbk->gnttab_rx_vif[i]; > gnt_count =&vif->persistent_rx_gntcnt; > gnt_total = 2*XEN_NETIF_RX_RING_SIZE; > pers_entry = vif->persistent_rx_gnt; > } Yes, the code is not simpler. If we make netback per-VIF based, then these code will disappear. The simplifying here means for tx path, the max search index is XEN_NETIF_TX_RING_SIZE(256 here), and this change can save some time when searching out grant page for specific grant reference. > >> @@ -111,8 +109,16 @@ struct xenvif { >> >> wait_queue_head_t waiting_to_free; >> >> - struct persistent_entry *persistent_gnt[MAXIMUM_OUTSTANDING_BLOCK_REQS]; >> - unsigned int persistent_gntcnt; >> + struct persistent_entry *persistent_tx_gnt[XEN_NETIF_TX_RING_SIZE]; >> + >> + /* >> + * 2*XEN_NETIF_RX_RING_SIZE is for the case of each head/fragment page > Shouldn't that been incorporated into MAXIMUM_OUTSTANDING_BLOCK_REQS > (sic) too? Yes, the total value is same as MAXIMUM_OUTSTANDING_BLOCK_REQS. But here 2*XEN_NETIF_RX_RING_SIZE means it is only used by rx path, and it is used just like other elements in netback structure, such as grant_copy_op, meta, etc. >> + * using 2 copy operations. >> + */ >> + struct persistent_entry *persistent_rx_gnt[2*XEN_NETIF_RX_RING_SIZE]; > What is the per-vif memory overhead after this change? Per-vif memory overhead is following, for tx path, it is about XEN_NETIF_RX_RING_SIZE*PAGE_SIZE (256 PAGE_SIZE here) for rx path, it is about 2*XEN_NETIF_RX_RING_SIZE*PAGE_SIZE (512 PAGE_SIZE here) I can add some comment here. Thanks Annie > > Ian. >