linux-kernel - Re: [PATCH RFC ticketlock] Auto-queued ticketlock

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives
Hash Suite: Windows password security audit tool. GUI, reports in PDF.
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <1371067399.1746.47.camel@buesod1.americas.hpqcorp.net>
Date:	Wed, 12 Jun 2013 13:03:19 -0700
From:	Davidlohr Bueso <davidlohr.bueso@...com>
To:	Linus Torvalds <torvalds@...ux-foundation.org>
Cc:	Al Viro <viro@...iv.linux.org.uk>,
	Steven Rostedt <rostedt@...dmis.org>,
	Paul McKenney <paulmck@...ux.vnet.ibm.com>,
	Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
	Ingo Molnar <mingo@...e.hu>,
	赖江山 <laijs@...fujitsu.com>,
	Dipankar Sarma <dipankar@...ibm.com>,
	Andrew Morton <akpm@...ux-foundation.org>,
	Mathieu Desnoyers <mathieu.desnoyers@...icios.com>,
	Josh Triplett <josh@...htriplett.org>, niv@...ibm.com,
	Thomas Gleixner <tglx@...utronix.de>,
	Peter Zijlstra <peterz@...radead.org>,
	Valdis Kletnieks <Valdis.Kletnieks@...edu>,
	David Howells <dhowells@...hat.com>,
	Eric Dumazet <edumazet@...gle.com>,
	Darren Hart <darren@...art.com>,
	Frédéric Weisbecker <fweisbec@...il.com>,
	Silas Boyd-Wickizer <sbw@....edu>,
	Waiman Long <Waiman.Long@...com>
Subject: Re: [PATCH RFC ticketlock] Auto-queued ticketlock

On Wed, 2013-06-12 at 11:15 -0700, Linus Torvalds wrote:
> On Wed, Jun 12, 2013 at 10:50 AM, Davidlohr Bueso
> <davidlohr.bueso@...com> wrote:
> >
> > * short: is the big winner for this patch, +69% throughput improvement
> > with 100-2000 users. This makes a lot of sense since the workload spends
> > a ridiculous amount of time trying to acquire the d_lock:
> >
> >   84.86%    1569902            reaim  [kernel.kallsyms]            [k] _raw_spin_lock
> >                       |
> >                       --- _raw_spin_lock
> >                          |
> >                          |--49.96%-- dget_parent
> >                          |          __fsnotify_parent
> >                          |--49.71%-- dput
> 
> Ugh. Do you have any idea what the heck that thing actually does?

Waiman's dcache patchet were actually an attempt to address these exact
issues: http://lkml.org/lkml/2013/5/22/716

According to him:

"the short workload calls security functions like getpwnam(),
getpwuid(), getgrgid() a couple of times. These functions open
the /etc/passwd or /etc/group files, read their content and close the
files. 
It is the intensive open/read/close sequence from multiple threads that 
is causing 80%+ contention in the d_lock on a system with large number 
of cores. The MIT's MOSBench paper also outlined dentry reference 
counting as a scalability roadblock for its Apache and Exim tests."

> 
> Normally, we shouldn't see lots of dget contention, since the dcache
> these days does everything but the last path component locklessly.
> 
> But there's a few exceptions, like symlinks (act as "last component"
> in the middle). And obviously, if some crazy threaded program opens
> the *same* file concurrently over and over again, then that "last
> component" will hammer on the dentry lock of that particular path. But
> that "open the same file concurrently" seems totally unrealistic -
> although maybe that's what AIM does..
> 
> Anybody know the AIM subtests?
> 
> Also, we *may* actually be able to optimize this by making
> dentry->d_count atomic, which will allow us to often do dget_parent
> and put() without taking the dcache lock at all. That's what it used
> to be, but the RCU patches actually made it be protected by the
> d_lock. It made sense at the time, as a step in the sequence, and many
> of the dentry d_count accesses are under the lock, but now that the
> remaining hot-paths are dget_parent and dput and many of the dentry
> d_count increments are gone from the hot-paths, we might want to
> re-visit that decision.  It could go either way.

I did a quick attempt at this (patch attached). For the short workload,
we now have:

   76.90%     928688            reaim  [kernel.kallsyms]        [k] _raw_spin_lock                            
                      |
                      --- _raw_spin_lock
                         |          
                         |--99.69%-- dget_parent
                         |          __fsnotify_parent
                         |          |          
                         |          |--20.23%-- fsnotify_access
                         |          |          vfs_read
                         |          |--20.13%-- __fput
                         |          |          ____fput
                         |          |          task_work_run
                         |          |--20.07%-- security_file_permission
                         |          |          rw_verify_area
                         |          |          vfs_read
                         |          |--19.97%-- do_sys_open
                         |          |          SyS_open
                         |           --19.60%-- security_file_open
                         |                     do_dentry_open

Still 76%!!! Throughput wise we do have a very nice boost when compared
to the vanilla kernel:

10-100 users: +47%
100-1000 users: +76%
1000-2000 users: +76%

Thanks,
Davidlohr


View attachment "atomic_dcount.patch" of type "text/x-patch" (15385 bytes)