CINXE.COM

LKML: Esben Nielsen: Re: Recursion bug in -rt

<?xml version="1.0" encoding="UTF-8" standalone="yes"?> <!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Transitional//EN" "http://www.w3.org/TR/xhtml1/DTD/xhtml1-transitional.dtd"><html xmlns="http://www.w3.org/1999/xhtml"><head><meta http-equiv="Content-Type" content="text/html; charset=UTF-8" /><title>LKML: Esben Nielsen: Re: Recursion bug in -rt</title><link href="/css/message.css" rel="stylesheet" type="text/css" /><link href="/css/wrap.css" rel="alternate stylesheet" type="text/css" title="wrap" /><link href="/css/nowrap.css" rel="stylesheet" type="text/css" title="nowrap" /><link href="/favicon.ico" rel="shortcut icon" /><script src="/js/simple-calendar.js" type="text/javascript"></script><script src="/js/styleswitcher.js" type="text/javascript"></script><link rel="alternate" type="application/rss+xml" title="lkml.org : last 100 messages" href="/rss.php" /><link rel="alternate" type="application/rss+xml" title="lkml.org : last messages by Esben Nielsen" href="/groupie.php?aid=9881" /><!--Matomo--><script> var _paq = window._paq = window._paq || []; /* tracker methods like "setCustomDimension" should be called before "trackPageView" */ _paq.push(["setDoNotTrack", true]); _paq.push(["disableCookies"]); _paq.push(['trackPageView']); _paq.push(['enableLinkTracking']); (function() { var u="//m.lkml.org/"; _paq.push(['setTrackerUrl', u+'matomo.php']); _paq.push(['setSiteId', '1']); var d=document, g=d.createElement('script'), s=d.getElementsByTagName('script')[0]; g.async=true; g.src=u+'matomo.js'; s.parentNode.insertBefore(g,s); })(); </script><!--End Matomo Code--></head><body onload="es.jasper.simpleCalendar.init();" itemscope="itemscope" itemtype="http://schema.org/BlogPosting"><table border="0" cellpadding="0" cellspacing="0"><tr><td width="180" align="center"><a href="/"><img style="border:0;width:135px;height:32px" src="/images/toprowlk.gif" alt="lkml.org" /></a></td><td width="32">聽</td><td class="nb"><div><a class="nb" href="/lkml"> [lkml]</a> 聽 <a class="nb" href="/lkml/2005"> [2005]</a> 聽 <a class="nb" href="/lkml/2005/12"> [Dec]</a> 聽 <a class="nb" href="/lkml/2005/12/20"> [20]</a> 聽 <a class="nb" href="/lkml/last100"> [last100]</a> 聽 <a href="/rss.php"><img src="/images/rss-or.gif" border="0" alt="RSS Feed" /></a></div><div>Views: <a href="#" class="nowrap" onclick="setActiveStyleSheet('wrap');return false;">[wrap]</a><a href="#" class="wrap" onclick="setActiveStyleSheet('nowrap');return false;">[no wrap]</a> 聽 <a class="nb" href="/lkml/mheaders/2005/12/20/290" onclick="this.href='/lkml/headers'+'/2005/12/20/290';">[headers]</a>聽 <a href="/lkml/bounce/2005/12/20/290">[forward]</a>聽 </div></td><td width="32">聽</td></tr><tr><td valign="top"><div class="es-jasper-simpleCalendar" baseurl="/lkml/"></div><div class="threadlist">Messages in this thread</div><ul class="threadlist"><li class="root"><a href="/lkml/2005/12/14/298">First message in thread</a></li><li><a href="/lkml/2005/12/20/281">Esben Nielsen</a><ul><li><a href="/lkml/2005/12/20/285">Steven Rostedt</a><ul><li class="origin"><a href="/lkml/2005/12/21/336">Esben Nielsen</a><ul><li><a href="/lkml/2005/12/21/336">david singleton</a></li></ul></li></ul></li></ul></li></ul></td><td width="32" rowspan="2" class="c" valign="top"><img src="/images/icornerl.gif" width="32" height="32" alt="/" /></td><td class="c" rowspan="2" valign="top" style="padding-top: 1em"><table><tr><td><table><tr><td class="lp">Date</td><td class="rp" itemprop="datePublished">Wed, 21 Dec 2005 00:55:56 +0100 (MET)</td></tr><tr><td class="lp">From</td><td class="rp" itemprop="author">Esben Nielsen &lt;&gt;</td></tr><tr><td class="lp">Subject</td><td class="rp" itemprop="name">Re: Recursion bug in -rt</td></tr></table></td><td></td></tr></table><pre itemprop="articleBody">On Tue, 20 Dec 2005, Steven Rostedt wrote:<br /><br />&gt; <br />&gt; On Tue, 20 Dec 2005, Esben Nielsen wrote:<br />&gt; <br />&gt; &gt; &gt;<br />&gt; &gt;<br />&gt; &gt; The same lock taken twice is just a special case of deadlocking. It would<br />&gt; &gt; be very hard to check for the general case in the futex code without<br />&gt; &gt; "fixing" the rt_mutex. Not that the rt_mutex code is broken - it just<br />&gt; &gt; doesn't handle deadlocks very well as it wasn't supposed to. But as the<br />&gt; &gt; futex indirectly exposes the rt_mutex to userspace it becomes a problem.<br />&gt; &gt;<br />&gt; &gt; The only _hack_ I can see is to force all robust futex calls to go through<br />&gt; &gt; one global lock to prevent the futex deadlocks becomming rt_mutex<br />&gt; &gt; deadlocks which again can turn into spin-lock deadlocks.<br />&gt; &gt;<br />&gt; &gt; I instead argue for altering the premisses for the rt_mutex such<br />&gt; &gt; they can handle deadlocks without turning them into spin-lock deadlocks<br />&gt; &gt; blocking the whole system. Then a futex deadlock will become a rt_mutex<br />&gt; &gt; deadlock which can be handled.<br />&gt; &gt;<br />&gt; <br />&gt; For the type of deadlock you are talking about is the following:<br />&gt; <br />&gt; P1 -- grabs futex A (no system call)<br />&gt; P2 -- grabs futex B (no system call)<br />&gt; <br />&gt; P1 -- tries to grab futex B (system call to block and boost P2)<br />&gt; But holds no other kernel rt_mutex!<br />&gt; P2 -- tries to grab futex A (system call to block and boost P1)<br />&gt; spinning deadlock here,<br />&gt; <br />&gt; So, before P2 blocks on P1, can there be a circular check t see if this is<br />&gt; a deadlock. You don't need to worry about other kernel rt_mutexes, you<br />&gt; only need to worry about blocked process.<br />&gt; <br />&gt; Is this feasible?<br />Ofcourse it is - but it the exact same kind of traversal you do in the<br />rt_mutex part to resolve the PI boosting. Thus by making the futex code do<br />this by it's own you essentially just move the complexity in there and<br />make the total more complex. Notice, that the futex code can't rely on<br />user-space flag - it can't be trusted for this (opening a local DOS<br />attack). So it has to rely on the rt_mutex structure to do this - and<br />therefore also the locks in there to protect against simultanious<br />unlocking.<br /><br />Esben<br /><br />&gt; <br />&gt; -- Steve<br />&gt; <br /><br />-<br />To unsubscribe from this list: send the line "unsubscribe linux-kernel" in<br />the body of a message to majordomo&#64;vger.kernel.org<br />More majordomo info at <a href="http://vger.kernel.org/majordomo-info.html">http://vger.kernel.org/majordomo-info.html</a><br />Please read the FAQ at <a href="http://www.tux.org/lkml/">http://www.tux.org/lkml/</a><br /><br /></pre></td><td width="32" rowspan="2" class="c" valign="top"><img src="/images/icornerr.gif" width="32" height="32" alt="\" /></td></tr><tr><td align="right" valign="bottom"> 聽 </td></tr><tr><td align="right" valign="bottom">聽</td><td class="c" valign="bottom" style="padding-bottom: 0px"><img src="/images/bcornerl.gif" width="32" height="32" alt="\" /></td><td class="c">聽</td><td class="c" valign="bottom" style="padding-bottom: 0px"><img src="/images/bcornerr.gif" width="32" height="32" alt="/" /></td></tr><tr><td align="right" valign="top" colspan="2"> 聽 </td><td class="lm">Last update: 2005-12-21 00:58 聽聽 [from the cache]<br />漏2003-2020 <a href="http://blog.jasper.es/"><span itemprop="editor">Jasper Spaans</span></a>|hosted at <a href="https://www.digitalocean.com/?refcode=9a8e99d24cf9">Digital Ocean</a> and my Meterkast|<a href="http://blog.jasper.es/categories.html#lkml-ref">Read the blog</a></td><td>聽</td></tr></table><script language="javascript" src="/js/styleswitcher.js" type="text/javascript"></script></body></html>

Pages: 1 2 3 4 5 6 7 8 9 10