Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1751815AbdF1Ufo (ORCPT ); Wed, 28 Jun 2017 16:35:44 -0400 Received: from relay1.mentorg.com ([192.94.38.131]:53807 "EHLO relay1.mentorg.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751501AbdF1Ufg (ORCPT ); Wed, 28 Jun 2017 16:35:36 -0400 From: Jim Baxter To: , , , Oliver Neukum CC: , =?UTF-8?q?Bj=C3=B8rn=20Mork?= , David Laight Subject: [PATCH V2 0/1] Reduce cdc_ncm memory use when kernel memory low Date: Wed, 28 Jun 2017 21:35:28 +0100 Message-ID: <1498682129-9129-1-git-send-email-jim_baxter@mentor.com> X-Mailer: git-send-email 1.9.1 MIME-Version: 1.0 Content-Type: text/plain X-Originating-IP: [137.202.0.87] X-ClientProxiedBy: svr-ies-mbx-01.mgc.mentorg.com (139.181.222.1) To SVR-IES-MBX-04.mgc.mentorg.com (139.181.222.4) Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 3013 Lines: 92 Problem ------- We are using an ARM embedded platform and require 16KiB NTB's to allow for fast data transfer. Unfortunately we have found that there are times after running the kernel for a while and transferring a lot of data over the CDC-NCM connection that it can become harder to find 16KiB pages of memory for allocation. This results in a disconnection of the NCM Gadget attached to the host platform. We are running with reduced buffers to not cross over into the 32KiB page boundary by setting the buffer sizes to: tx_max=16000 rx_max=16000 Analysis -------- We identified through investigation that the lack of 16KiB pages would be short lived as the kernel would compact the buddy list soon after the failure which results in pages being available within seconds. Solution -------- In order to avoid disconnections I implemented a patch that will attempt to use a 2048 Byte minimum size NTB if the allocation of the maximum size NTB fails. This allows the connection to limp along until the memory has been recovered which was usually between 1 and 4 NTB's on our heavy traffic system. The algorithm will wait for an increasing number of small allocations each time we have a failure to not burden a system short on memory. --- V1: Sent to linux-usb for review. V2: Added code to increase amount of time spent making small allocations to reduce the burden on the system. This is the diff between Version 1 and 2 of the patches. -- File: drivers/net/usb/cdc_ncm.c 41c51,60 < @@ -1055,10 +1055,10 @@ static struct usb_cdc_ncm_ndp16 *cdc_ncm_ndp(struct cdc_ncm_ctx *ctx, struct sk_ --- > @@ -89,6 +89,8 @@ struct cdc_ncm_stats { > CDC_NCM_SIMPLE_STAT(rx_ntbs), > }; > > +#define CDC_NCM_LOW_MEM_MAX_CNT 10 > + > static int cdc_ncm_get_sset_count(struct net_device __always_unused *netdev, int sset) > { > switch (sset) { > @@ -1055,10 +1057,10 @@ static struct usb_cdc_ncm_ndp16 *cdc_ncm_ndp(struct cdc_ncm_ctx *ctx, struct sk_ 59,60c78,91 < + ctx->tx_curr_size = ctx->tx_max; < + skb_out = alloc_skb(ctx->tx_curr_size, GFP_ATOMIC); --- > + if (ctx->tx_low_mem_val == 0) { > + ctx->tx_curr_size = ctx->tx_max; > + skb_out = alloc_skb(ctx->tx_curr_size, GFP_ATOMIC); > + /* If the memory allocation fails we will wait longer > + * each time before attempting another full size > + * allocation again to not overload the system > + * further. > + */ > + if (skb_out == NULL) { > + ctx->tx_low_mem_max_cnt = min(ctx->tx_low_mem_max_cnt + 1, > + (unsigned)CDC_NCM_LOW_MEM_MAX_CNT); > + ctx->tx_low_mem_val = ctx->tx_low_mem_max_cnt; > + } > + } 84a116 > + ctx->tx_low_mem_val--; -- File: include/linux/usb/cdc_ncm.h 130a163,164 > + u32 tx_low_mem_max_cnt; > + u32 tx_low_mem_val; Jim Baxter (1): net: cdc_ncm: Reduce memory use when kernel memory low drivers/net/usb/cdc_ncm.c | 54 +++++++++++++++++++++++++++++++++++---------- include/linux/usb/cdc_ncm.h | 3 +++ 2 files changed, 45 insertions(+), 12 deletions(-) -- 1.9.1