Content-Type: multipart/mixed; boundary="===============3937851257418758275==" 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 13:22:08 -0400 Message-ID: <52446D40.9040504@hurleysoftware.com> In-Reply-To: <20130926153208.GN18242@two.firstfloor.org> List-Id: --===============3937851257418758275== Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: quoted-printable On 09/26/2013 11:32 AM, Andi Kleen wrote: > On Thu, Sep 26, 2013 at 07:52:23AM -0400, Peter Hurley wrote: >> On 09/25/2013 11:20 PM, Andi Kleen wrote: >>> Lin Ming writes: >>>> >>>> Would you like below patch? >>> >>> The loop body keeps rather complex state. It could easily >>> get confused by parallel RCU changes. >>> >>> So if the list changes in parallel you may suddenly >>> report very bogus values, as the va_start - prev_end >>> computation may be bogus. >>> >>> Perhaps it's ok (may report bogus gaps), but it seems a bit risky. >> >> I don't understand how the computed gap would be bogus; there >> _was_ a list state in which that particular gap existed. The fact > > It could change any time as you don't have an atomic view > of vm_end / vm_start. It is valid to change the fields > with the lock held. va_start and va_end are constant for the lifetime of their vmap_area (if it's accessible by traversing the vmap_area_list), so it is not possible for an rcu-based list traversal to see different values of these individual fields than the spin-locked version. In addition, for the rcu-based traversal to have arrived at any given vmap_area requires that the previous vmap_area was its adjacent lower range at the instant in time when the list cursor was advanced; again, this is no different than if the spin-locked version had happened to begin at that same instant. Regards, Peter Hurley --===============3937851257418758275==--