From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Return-Path: Date: Tue, 19 Mar 2019 00:38:13 +0300 From: Cyrill Gorcunov Subject: Re: [tarantool-patches] Re: [PATCH] lib/core/fiber: Initialize stack_watermark where appropriate Message-ID: <20190318213813.GH2469@uranus.lan> References: <20190318172352.14128-1-gorcunov@gmail.com> <20190318175246.22utl4hoc3jnxuk2@esperanza> <20190318185529.GA11507@chai> <20190318190446.GD2469@uranus.lan> <20190318191008.GA13557@chai> <20190318191746.GE2469@uranus.lan> <20190318211919.GB13557@chai> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <20190318211919.GB13557@chai> To: Konstantin Osipov Cc: tarantool-patches@freelists.org, vdavydov.dev@gmail.com List-ID: On Tue, Mar 19, 2019 at 12:19:19AM +0300, Konstantin Osipov wrote: > > > > > > Well, then let's increase the number of elements to 32 or 64, or, > > > better yet, 128, to make the situation better. > > > > > > Why did you guys choose 8? > > > > To cover low 1024 bytes. Maximal path is about 4K limit > > so I think in average ~1/4 of page should catch paths > > on stack. > > Okay, well, that at least a plausible cause, I guess I'd be nice to > see it in the comment for the poison array. However, you could use > 4 integers and a larger poison offset?-)) Which decreases probability, but yes :) Kostya, do you think 4 writes vs 8 writes would give us that much benefit? Technically it will be a couple of lines patch but... in the worst case where we have all marks present but stack actually consumes more pages, (and instead of shrinking it we wrongly "think" that we don't need to call madvise), will lead to very bad consequences -- rss will bloat as hell (just like Volodya pointed me about a stack-eager task which jumps over all fibers in a system).