From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Message-ID: <1455674441.2925.204.camel@hpe.com> Subject: Re: [PATCH 3/3] ACPI: Change NFIT driver to set PMEM type to iomem entry From: Toshi Kani Date: Tue, 16 Feb 2016 19:00:41 -0700 In-Reply-To: References: <1454439311-23690-1-git-send-email-toshi.kani@hpe.com> <1454439311-23690-4-git-send-email-toshi.kani@hpe.com> <1455316203.2925.132.camel@hpe.com> Content-Type: text/plain; charset="UTF-8" Mime-Version: 1.0 Content-Transfer-Encoding: 8bit Sender: linux-kernel-owner@vger.kernel.org To: Dan Williams Cc: Ingo Molnar , Borislav Petkov , "Rafael J. Wysocki" , Andrew Morton , "linux-nvdimm@lists.01.org" , Linux ACPI , "linux-kernel@vger.kernel.org" List-ID: On Fri, 2016-02-12 at 15:32 -0800, Dan Williams wrote: > On Fri, Feb 12, 2016 at 2:30 PM, Toshi Kani wrote: > > On Fri, 2016-02-12 at 11:41 -0800, Dan Williams wrote: > > > On Tue, Feb 2, 2016 at 10:55 AM, Toshi Kani wro  : > > > Hmm, if we set the type on driver load, should we clear the type on > > > driver unload? > > > > I think this type update should stay for the life-cycle of this iomem > > entry itself since this range is PMEM even after the driver is > > unloaded.  This is an extension of the boot-time iomem table > > initialization from e820/EFI, which allows ACPI to set a correct > > type.  This is independent from driver's resource allocations. > > > > > Actually it might be more straightforward to specify a type at > > > request_region() time.  That way it gets released at > > > release_region().  We're already setting a resource name at > > > request_region time, adding a type annotation at the time seems > > > appropriate. > > > > I first considered simply setting "namespaceX.X" as PMEM.  However, > > region_intersects() and its friends only check the top-level entries, > > not their children, of the iomem table.  And I think a child should > > have the same type as the parent as I fixed it in patch 1/3. > > Did we investigate updating region_intersects() to check children? > When a child sub-divides a region with different types it may be the > wrong answer to check the parent.  Is there a problem with moving > checking to the child? Here are three options I can think of. 1) Set pmem type to "reserved" (This patch-set)  - Add a new iomem_set_desc(), which sets a given type to a top-level entry.  Change the ACPI NFIT driver to call it to set pmem type to "reserved" entry.  - region_intersects() finds a pmem entry by checking top-level entries (no change). 2) Change region_intersects() to check children's type  - Add a new request_region_ext(), which is an extension to request_region() to allow specifying a type of resource.  It puts a new child entry under "reserved".  Change the pmem driver to call this func.  - Change region_intersects() to check children's type for finding this child pmem entry. 3) Pmem driver to call insert_resource()  - Change the pmem driver to call insert_resource(), which puts a new pmem entry as the parent of "reserved".  - region_intersects() finds a pmem entry by checking top-level entries (no change).  - Add a new release_resource_self(), which releases a given entry and keeps its children if any.  Change the pmem driver to call it for release. This patch-set implements 1).  The pmem type is set to "reserved" for its life-cycle.  This option is simplest. For 2), the changes to region_intersects() may be too complex for maintenance.  Here are a few examples when region_intersects() is called with addr [1-10] where iomem has entry P and its children. Case A: P is fully covered by children C1 & C2.  region_intersects() ignores P's type, but checks C1 and C2's.      P [1-10] + C1 [1-5]            + C2 [6-10] Case B: C2 is fully covered by C3, but P is not.  region_intersects() ignores C2's type, but checks P, C1, C3's.   P [1-10] + C1 [1-2]            + C2 [6-10] + C3 [6-10] I think region_intersects() will need to construct a flat table from the tree while making recursive calls to walk thru all children. 3) is similar to 2), but avoids the changes to region_intersects() since insert_resource() inserts a new entry as the parent to "reserved".  However, a new interface is necessary to put "reserved" back to top-level when releasing the added entry. My recommendation is go with either 1) or 3).  What do you think? Thanks, -Toshi