CINXE.COM

LKML: Thomas Gleixner: [patch V3 14/18] posix-timers: Avoid false cacheline sharing

<?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: Thomas Gleixner: [patch V3 14/18] posix-timers: Avoid false cacheline sharing</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 Thomas Gleixner" href="/groupie.php?aid=" /><!--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/2025"> [2025]</a> 聽 <a class="nb" href="/lkml/2025/3"> [Mar]</a> 聽 <a class="nb" href="/lkml/2025/3/8"> [8]</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/2025/3/8/441" onclick="this.href='/lkml/headers'+'/2025/3/8/441';">[headers]</a>聽 <a href="/lkml/bounce/2025/3/8/441">[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/2025/3/8/426">First message in thread</a></li><li><a href="/lkml/2025/3/8/426">Thomas Gleixner</a><ul><li><a href="/lkml/2025/3/8/427">Thomas Gleixner</a><ul><li><a href="/lkml/2025/3/8/616">Frederic Weisbecker</a></li><li><a href="/lkml/2025/3/13/634">"tip-bot2 for Thomas Gleixner"</a></li></ul></li><li><a href="/lkml/2025/3/8/428">Thomas Gleixner</a><ul><li><a href="/lkml/2025/3/11/914">Frederic Weisbecker</a><ul><li><a href="/lkml/2025/3/11/977">Thomas Gleixner</a></li></ul></li><li><a href="/lkml/2025/3/13/633">"tip-bot2 for Eric Dumazet"</a></li></ul></li><li><a href="/lkml/2025/3/8/430">Thomas Gleixner</a><ul><li><a href="/lkml/2025/3/13/631">"tip-bot2 for Eric Dumazet"</a></li></ul></li><li><a href="/lkml/2025/3/8/431">Thomas Gleixner</a><ul><li><a href="/lkml/2025/3/13/630">"tip-bot2 for Thomas Gleixner"</a></li></ul></li><li><a href="/lkml/2025/3/8/432">Thomas Gleixner</a><ul><li><a href="/lkml/2025/3/13/632">"tip-bot2 for Thomas Gleixner"</a></li></ul></li><li><a href="/lkml/2025/3/8/433">Thomas Gleixner</a><ul><li><a href="/lkml/2025/3/13/629">"tip-bot2 for Thomas Gleixner"</a></li></ul></li><li><a href="/lkml/2025/3/8/434">Thomas Gleixner</a><ul><li><a href="/lkml/2025/3/13/627">"tip-bot2 for Thomas Gleixner"</a></li></ul></li><li><a href="/lkml/2025/3/8/435">Thomas Gleixner</a><ul><li><a href="/lkml/2025/3/13/626">"tip-bot2 for Thomas Gleixner"</a></li></ul></li><li><a href="/lkml/2025/3/8/436">Thomas Gleixner</a><ul><li><a href="/lkml/2025/3/10/13">Frederic Weisbecker</a><ul><li><a href="/lkml/2025/3/10/148">Thomas Gleixner</a></li></ul></li><li><a href="/lkml/2025/3/10/256">Thomas Gleixner</a><ul><li><a href="/lkml/2025/3/13/628">"tip-bot2 for Thomas Gleixner"</a></li></ul></li></ul></li><li><a href="/lkml/2025/3/8/437">Thomas Gleixner</a><ul><li><a href="/lkml/2025/3/10/634">Frederic Weisbecker</a><ul><li><a href="/lkml/2025/3/10/1329">Thomas Gleixner</a><ul><li><a href="/lkml/2025/3/10/1664">Frederic Weisbecker</a></li></ul></li></ul></li><li><a href="/lkml/2025/3/13/625">"tip-bot2 for Peter Zijlstra"</a></li></ul></li><li><a href="/lkml/2025/3/8/438">Thomas Gleixner</a><ul><li><a href="/lkml/2025/3/10/1716">Frederic Weisbecker</a></li><li><a href="/lkml/2025/3/11/927">Frederic Weisbecker</a></li><li><a href="/lkml/2025/3/13/623">"tip-bot2 for Eric Dumazet"</a></li></ul></li><li><a href="/lkml/2025/3/8/439">Thomas Gleixner</a><ul><li><a href="/lkml/2025/3/11/932">Frederic Weisbecker</a></li><li><a href="/lkml/2025/3/13/624">"tip-bot2 for Thomas Gleixner"</a></li></ul></li><li><a href="/lkml/2025/3/8/440">Thomas Gleixner</a><ul><li><a href="/lkml/2025/3/13/622">"tip-bot2 for Thomas Gleixner"</a></li></ul></li><li class="origin"><a href="/lkml/2025/3/11/940">Thomas Gleixner</a><ul><li><a href="/lkml/2025/3/11/940">Frederic Weisbecker</a></li><li><a href="/lkml/2025/3/13/621">"tip-bot2 for Thomas Gleixner"</a></li><li><a href="/lkml/2025/3/13/1624">David Laight</a></li><li><a href="/lkml/2025/3/17/208">"Nysal Jan K.A."</a></li></ul></li><li><a href="/lkml/2025/3/8/442">Thomas Gleixner</a><ul><li><a href="/lkml/2025/3/11/1110">Frederic Weisbecker</a></li><li><a href="/lkml/2025/3/13/620">"tip-bot2 for Thomas Gleixner"</a></li></ul></li><li><a href="/lkml/2025/3/8/443">Thomas Gleixner</a><ul><li><a href="/lkml/2025/3/9/14">Cyrill Gorcunov</a></li><li><a href="/lkml/2025/3/11/1103">Frederic Weisbecker</a></li><li><a href="/lkml/2025/3/13/619">"tip-bot2 for Thomas Gleixner"</a></li></ul></li><li><a href="/lkml/2025/3/8/444">Thomas Gleixner</a><ul><li><a href="/lkml/2025/3/9/13">Cyrill Gorcunov</a></li><li><a href="/lkml/2025/3/11/1588">Frederic Weisbecker</a><ul><li><a href="/lkml/2025/3/11/1615">Thomas Gleixner</a><ul><li><a href="/lkml/2025/3/11/1618">Thomas Gleixner</a></li></ul></li><li><a href="/lkml/2025/3/12/649">Cyrill Gorcunov</a></li></ul></li></ul></li><li><a href="/lkml/2025/3/8/445">Thomas Gleixner</a><ul><li><a href="/lkml/2025/3/10/254">Thomas Gleixner</a><ul><li><a href="/lkml/2025/3/11/1599">Frederic Weisbecker</a></li><li><a href="/lkml/2025/3/13/617">"tip-bot2 for Thomas Gleixner"</a></li></ul></li></ul></li></ul></li></ul><div class="threadlist">Patch in this message</div><ul class="threadlist"><li><a href="/lkml/diff/2025/3/8/441/1">Get diff 1</a></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">From</td><td class="rp" itemprop="author">Thomas Gleixner &lt;&gt;</td></tr><tr><td class="lp">Subject</td><td class="rp" itemprop="name">[patch V3 14/18] posix-timers: Avoid false cacheline sharing</td></tr><tr><td class="lp">Date</td><td class="rp" itemprop="datePublished">Sat, 8 Mar 2025 17:48:42 +0100 (CET)</td></tr></table></td><td></td></tr></table><pre itemprop="articleBody">struct k_itimer has the hlist_node, which is used for lookup in the hash<br />bucket, and the timer lock in the same cache line.<br /><br />That's obviously bad, if one CPU fiddles with a timer and the other is<br />walking the hash bucket on which that timer is queued.<br /><br />Avoid this by restructuring struct k_itimer, so that the read mostly (only<br />modified during setup and teardown) fields are in the first cache line and<br />the lock and the rest of the fields which get written to are in cacheline<br />2-N.<br /><br />Reduces cacheline contention in a test case of 64 processes creating and<br />accessing 20000 timers each by almost 30% according to perf.<br /><br />Signed-off-by: Thomas Gleixner &lt;tglx&#64;linutronix.de&gt;<br /><br />---<br />V2: New patch<br />---<br /> include/linux/posix-timers.h | 21 ++++++++++++---------<br /> kernel/time/posix-timers.c | 4 ++--<br /> 2 files changed, 14 insertions(+), 11 deletions(-)<br /><br />--- a/include/linux/posix-timers.h<br />+++ b/include/linux/posix-timers.h<br />&#64;&#64; -177,23 +177,26 &#64;&#64; static inline void posix_cputimers_init_<br /> * &#64;rcu: RCU head for freeing the timer.<br /> */<br /> struct k_itimer {<br />- struct hlist_node list;<br />- struct hlist_node ignored_list;<br />+ /* 1st cacheline contains read-mostly fields */<br /> struct hlist_node t_hash;<br />- spinlock_t it_lock;<br />- const struct k_clock *kclock;<br />- clockid_t it_clock;<br />+ struct hlist_node list;<br /> timer_t it_id;<br />+ clockid_t it_clock;<br />+ int it_sigev_notify;<br />+ enum pid_type it_pid_type;<br />+ struct signal_struct *it_signal;<br />+ const struct k_clock *kclock;<br />+<br />+ /* 2nd cacheline and above contain fields which are modified regularly */<br />+ spinlock_t it_lock;<br /> int it_status;<br /> bool it_sig_periodic;<br /> s64 it_overrun;<br /> s64 it_overrun_last;<br /> unsigned int it_signal_seq;<br /> unsigned int it_sigqueue_seq;<br />- int it_sigev_notify;<br />- enum pid_type it_pid_type;<br /> ktime_t it_interval;<br />- struct signal_struct *it_signal;<br />+ struct hlist_node ignored_list;<br /> union {<br /> struct pid *it_pid;<br /> struct task_struct *it_process;<br />&#64;&#64; -210,7 +213,7 &#64;&#64; struct k_itimer {<br /> } alarm;<br /> } it;<br /> struct rcu_head rcu;<br />-};<br />+} ____cacheline_aligned_in_smp;<br /> <br /> void run_posix_cpu_timers(void);<br /> void posix_cpu_timers_exit(struct task_struct *task);<br />--- a/kernel/time/posix-timers.c<br />+++ b/kernel/time/posix-timers.c<br />&#64;&#64; -260,8 +260,8 &#64;&#64; static int posix_get_hrtimer_res(clockid<br /> <br /> static __init int init_posix_timers(void)<br /> {<br />- posix_timers_cache = kmem_cache_create("posix_timers_cache", sizeof(struct k_itimer), 0,<br />- SLAB_ACCOUNT, NULL);<br />+ posix_timers_cache = kmem_cache_create("posix_timers_cache", sizeof(struct k_itimer),<br />+ __alignof__(struct k_itimer), SLAB_ACCOUNT, NULL);<br /> return 0;<br /> }<br /> __initcall(init_posix_timers);<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: 2025-03-08 17:51 聽聽 [W:1.339 / U:0.262 seconds]<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