From mboxrd@z Thu Jan 1 00:00:00 1970 From: Frank Rowand Subject: Re: [PATCH v4 1/2] of: cache phandle nodes to reduce cost of of_find_node_by_phandle() Date: Wed, 28 Feb 2018 11:44:51 -0800 Message-ID: References: <1519844656-16443-1-git-send-email-frowand.list@gmail.com> <1519844656-16443-2-git-send-email-frowand.list@gmail.com> Mime-Version: 1.0 Content-Type: text/plain; charset=utf-8 Content-Transfer-Encoding: 7bit Return-path: In-Reply-To: Content-Language: en-US Sender: linux-kernel-owner@vger.kernel.org To: Andy Shevchenko Cc: Rob Herring , cpandya@codeaurora.org, devicetree , Linux Kernel Mailing List List-Id: devicetree@vger.kernel.org On 02/28/18 11:31, Andy Shevchenko wrote: > On Wed, Feb 28, 2018 at 9:04 PM, wrote: > >> Create a cache of the nodes that contain a phandle property. Use this >> cache to find the node for a given phandle value instead of scanning >> the devicetree to find the node. If the phandle value is not found >> in the cache, of_find_node_by_phandle() will fall back to the tree >> scan algorithm. >> >> The cache is initialized in of_core_init(). >> >> The cache is freed via a late_initcall_sync() if modules are not >> enabled. >> >> If the devicetree is created by the dtc compiler, with all phandle >> property values auto generated, then the size required by the cache >> could be 4 * (1 + number of phandles) bytes. This results in an O(1) >> node lookup cost for a given phandle value. Due to a concern that the >> phandle property values might not be consistent with what is generated >> by the dtc compiler, a mask has been added to the cache lookup algorithm. >> To maintain the O(1) node lookup cost, the size of the cache has been >> increased by rounding the number of entries up to the next power of >> two. >> >> The overhead of finding the devicetree node containing a given phandle >> value has been noted by several people in the recent past, in some cases >> with a patch to add a hashed index of devicetree nodes, based on the >> phandle value of the node. One concern with this approach is the extra >> space added to each node. This patch takes advantage of the phandle >> property values auto generated by the dtc compiler, which begin with >> one and monotonically increase by one, resulting in a range of 1..n >> for n phandle values. This implementation should also provide a good >> reduction of overhead for any range of phandle values that are mostly >> in a monotonic range. >> >> Performance measurements by Chintan Pandya >> of several implementations of patches that are similar to this one >> suggest an expected reduction of boot time by ~400ms for his test >> system. If the cache size was decreased to 64 entries, the boot >> time was reduced by ~340 ms. The measurements were on a 4.9.73 kernel >> for arch/arm64/boot/dts/qcom/sda670-mtp.dts, contains 2371 nodes and >> 814 phandle values. > > The question is why O(1) is so important? O(log(n)) wouldn't work? O(1) is not critical. It was just a nice side result. > Using radix_tree() I suppose allows to dynamically extend or shrink > the cache which would work with DT overlays. The memory usage of the phandle cache in this patch is fairly small. The memory overhead of a radix_tree() would not be justified.