Content-Type: multipart/mixed; boundary="===============0138832249940177494==" MIME-Version: 1.0 From: Peter Hurley To: lkp@lists.01.org Subject: Re: increased vmap_area_lock contentions on "n_tty: Move buffers into n_tty_data" Date: Thu, 26 Sep 2013 13:35:32 -0400 Message-ID: <52447064.5030902@hurleysoftware.com> In-Reply-To: <20130926150421.GA25545@kroah.com> List-Id: --===============0138832249940177494== Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: quoted-printable On 09/26/2013 11:04 AM, Greg KH wrote: > On Thu, Sep 26, 2013 at 07:31:47AM -0400, Peter Hurley wrote: >> On 09/26/2013 03:33 AM, Andrew Morton wrote: >>> On Tue, 17 Sep 2013 20:22:42 -0400 Peter Hurley wrote: >>> >>>> Looking over vmalloc.c, the critical section footprint of the vmap_are= a_lock >>>> could definitely be reduced (even nearly eliminated), but that's a pro= ject for >>>> another day :) >>> >>> 20bafb3d23d10 ("n_tty: Move buffers into n_tty_data") switched a >>> kmalloc (which is very fast) to a vmalloc (which is very slow) without >>> so much as mentioning it in the changelog. This should have been >>> picked up at review, btw. >>> >>> Revert that part of the patch and the problem will be solved. >>> >>> If we are really really worried that a ~9k kmalloc might fail or will >>> be slow, then implement a fallback to vmalloc() if kmalloc(GFP_NOWARN) >>> failed. This kinda sucks, but is practical, but really should only be >>> done if necessary - ie, if problems with using plain old kmalloc are >>> demonstrable. >>> >>> Or just revert all of 20bafb3d23d10 - it was supposed to be a small >>> performance improvement but turned out to be a significant performance >>> loss. Therefore zap. >> >> I have no particular objection to reverting the entire patch. > > How about just switching the call to vmalloc to kmalloc? Yes, it's a > larger size that is being allocated here, but we were allocating that > much memory anyway before, so it should be the same "speed", if not > faster than before (1 call to kmalloc instead of 3). The allocation itself isn't performance-critical. The speed difference between 1 kmalloc and 3 kmallocs here will be unmeasurable from any user-space test. And the only reason vmalloc has any measurable impact stems from the way reads of /proc/meminfo behave (to which there are a number of appropriate solutions). The issue with a single large kmalloc is that it may fail where 3 separate, page-or-less kmallocs would not have. Regards, Peter Hurley --===============0138832249940177494==--