From mboxrd@z Thu Jan 1 00:00:00 1970 X-Msuck: nntp://news.gmane.org/gmane.linux.lib.musl.general/13383 Path: news.gmane.org!.POSTED!not-for-mail From: Rich Felker Newsgroups: gmane.linux.lib.musl.general Subject: Re: =?utf-8?B?562U5aSN?= =?utf-8?Q?=3A?= [musl] musl: about malloc 'expand heap' issue Date: Tue, 30 Oct 2018 23:44:36 -0400 Message-ID: <20181031034436.GV5150@brightrain.aerifal.cx> References: <20181030150039.GT5150@brightrain.aerifal.cx> Reply-To: musl@lists.openwall.com NNTP-Posting-Host: blaine.gmane.org Mime-Version: 1.0 Content-Type: text/plain; charset=utf-8 Content-Transfer-Encoding: 8bit X-Trace: blaine.gmane.org 1540957497 11570 195.159.176.226 (31 Oct 2018 03:44:57 GMT) X-Complaints-To: usenet@blaine.gmane.org NNTP-Posting-Date: Wed, 31 Oct 2018 03:44:57 +0000 (UTC) User-Agent: Mutt/1.5.21 (2010-09-15) Cc: "musl@lists.openwall.com" , "Jianing (OS-LAB)" , "Huangqiang (H)" , leijitang , wanghaozhan To: "zhangwentao (M)" Original-X-From: musl-return-13399-gllmg-musl=m.gmane.org@lists.openwall.com Wed Oct 31 04:44:52 2018 Return-path: Envelope-to: gllmg-musl@m.gmane.org Original-Received: from mother.openwall.net ([195.42.179.200]) by blaine.gmane.org with smtp (Exim 4.84_2) (envelope-from ) id 1gHhR6-0002uN-Ek for gllmg-musl@m.gmane.org; Wed, 31 Oct 2018 04:44:52 +0100 Original-Received: (qmail 32045 invoked by uid 550); 31 Oct 2018 03:47:01 -0000 Mailing-List: contact musl-help@lists.openwall.com; run by ezmlm Precedence: bulk List-Post: List-Help: List-Unsubscribe: List-Subscribe: List-ID: Original-Received: (qmail 32022 invoked from network); 31 Oct 2018 03:47:00 -0000 Content-Disposition: inline In-Reply-To: Original-Sender: Rich Felker Xref: news.gmane.org gmane.linux.lib.musl.general:13383 Archived-At: On Wed, Oct 31, 2018 at 01:19:11AM +0000, zhangwentao (M) wrote: > Hi > > Now we are using a global lock to solve the issue. And as you said the performance maybe cost too much. > And if you have solution about this and the fix is not that expensive, that's great. > If you finish it, would you send the patch to me ? If I get a solution that's acceptable for upstream, it will be committed in git master and I'll follow up on this thread to mention it. Unfortunately, looking again at where the spurious empty-bin situation happens (as a result of alloc_rev/_fwd during free), I don't see an easy fix that's not costly. The best we can do might be something like this: In malloc, only use the per-bin lock to obtain a chunk if the exact-sized bin is non-empty. If it's empty, take a global lock (free_lock, but its semantics might need to be adjusted somewhat) to ensure a consistent view of what larger free chunks are available for splitting (which might end up revealing that a chunk of the exact desired bin was actually available). Ideally this will not have much impact on malloc performance under conditions where lots of free chunks are available (heavy load). Rich > -----邮件原件----- > 发件人: Rich Felker [mailto:dalias@aerifal.cx] 代表 Rich Felker > 发送时间: 2018年10月30日 23:01 > 收件人: musl@lists.openwall.com > 主题: Re: [musl] musl: about malloc 'expand heap' issue > > On Tue, Oct 30, 2018 at 11:11:07AM +0000, zhangwentao (M) wrote: > > > > Hi all, > > I am using musl in my project and I found an issue about the malloc function in musl: > > > > Issue Description: > > * When in muti-threads environment, malloc/free are called in high concurrency. > > > > Malloc: > > Will find 'struct bin' from bitmap(without lock), and allocate memory from the bin (with lock). > > > > Free: > > Will merge the chunk together if the free memory is 'connected' to the existing chunk. > > > > ? It will remove the old chunk first then combine the chunk to a larger one.. > > > > ? After merge operation done, insert the chunk to the bin list. > > > > ? Each of the chunk operation is locked while merging, but the whole steps aren't within a lock. > > > > So here is the issue: > > > > 1. There is only one chunk in largest bin list, and Free is on process, just remove the largest bins chunk from bin, the bitmap(mal.binmap) on that bit will be zero. > > > > 2. A malloc comes, the bitmap is zero, and goes to expand heap. (Actually there is enough memories in process) > > > > 3. Free operation goes on, and put the merged big chunk to bins. > > > > But in operation 2, the process has expand heap. > > > > If we have a loop on step 1-3, the process will expand heap frequently. > > So it will cost more Virtual Memory (of course, physical memory would be freed by calling '__madvise' if the chunk is big enough) > > > > In my environment , we do not have that much virtual memory. I think stop expand heap would a better choice. > > > > Do you have plan to fix it ?? > > This is a known issue, and intended to be fixed in the complete > redesign of malloc. Fixing it right in the current design seems to > impose significant performance costs that I thought were equivalent > to, or worse than, just using one global lock. However if it's causing > major problems I may be able to make a quick fix that's not too > expensive -- I'll take a look again today or tomorrow. > > Rich