Message-ID: <1364338776.15745.27.camel@misato.fc.hp.com>
Subject: Re: [PATCH v2] ACPI: Add sysfs links from memory device to memblocks
From: Toshi Kani <toshi.kani@hp.com>
To: "Rafael J. Wysocki" <rjw@sisk.pl>
Cc: linux-acpi@vger.kernel.org, linux-kernel@vger.kernel.org,
        isimatu.yasuaki@jp.fujitsu.com, vasilis.liaskovitis@profitbricks.com
Date: Tue, 26 Mar 2013 16:59:36 -0600
In-Reply-To: <3205964.WeZJN7sRdd@vostro.rjw.lan>
References: <1361826130-31062-1-git-send-email-toshi.kani@hp.com>
	 <2207657.9htT4cP6Jv@vostro.rjw.lan>
	 <1364312526.11659.134.camel@misato.fc.hp.com>
	 <3205964.WeZJN7sRdd@vostro.rjw.lan>
Content-Type: text/plain; charset="UTF-8"
Mime-Version: 1.0
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 4539
Lines: 94

On Tue, 2013-03-26 at 23:39 +0100, Rafael J. Wysocki wrote:
> On Tuesday, March 26, 2013 09:42:06 AM Toshi Kani wrote:
> > On Tue, 2013-03-26 at 14:04 +0100, Rafael J. Wysocki wrote:
> > > On Monday, February 25, 2013 02:02:10 PM Toshi Kani wrote:
> > > > In order to eject a memory device object represented as "PNP0C80:%d"
> > > > in sysfs, its associated memblocks (system/memory/memory%d) need to
> > > > be off-lined.  However, there is no user friendly way to correlate
> > > > between a memory device object and its memblocks in sysfs.
> > > > 
> > > > This patch creates sysfs links to memblocks under a memory device
> > > > object so that a user can easily checks and manipulates its memblocks
> > > > in sysfs.
> > > > 
> > > > For example, when PNP0C80:05 is associated with memory8 and memory9,
> > > > the following two links are created under PNP0C80:05.  This allows
> > > > a user to access memory8/9 directly from PNP0C80:05.
> > > > 
> > > >   # ll /sys/devices/LNXSYSTM:00/LNXSYBUS:00/PNP0C80:05
> > > >   lrwxrwxrwx. memory8 -> ../../../system/memory/memory8
> > > >   lrwxrwxrwx. memory9 -> ../../../system/memory/memory9
> > > > 
> > > > Signed-off-by: Toshi Kani <toshi.kani@hp.com>
> > > 
> > > Here I have some doubts.
> > > 
> > > This adds a very specific interface for user space that we're going to need to
> > > maintain going forward if the user space starts to use it.  However, it kind of
> > > duplicates the existing "physical_node" interface that we have for "regular"
> > > devices.
> > > 
> > > So, if possible, I'd like the memory subsystem to utilize the existing
> > > interface instead of creating an entirely new one.  Namely, why don't we create
> > > a struct device-based object for each memory block and associated those new
> > > "devices" with the PNP0C80 ACPI object through the functions in glue.c?
> > > Then, we could add an "offline/online" interface to those "devices" too.
> > 
> > This patch simply adds symbolic links to system/memory/memoryN, which
> > the memory subsystem already provides for the online/offline interface
> > of memory blocks.  So, it does not introduce a new interface, but guides
> > users (and user tools) to know which memory blocks need to be off-lined
> > in order to hot-delete any particular memory device PNP0C80:X.  A cpu
> > device LNXCPU:X also has a similar symbolic link "sysdev" that links to
> > system/cpu/cpuN.  I could not use the same "sysdev" for PNP0C80:X since
> > it typically associates with multiple memory blocks.
> > 
> > I thought about using glue.c to create symbolic links between memoryN
> > and PNP0C80:X.  However, it has an ordering issue.  During boot-time,
> > memoryN gets created before PNP0C80:X.  But during hot-add, PNP0C80:X
> > gets created before memoryN.
> 
> Quite frankly, this sounds like a bug to me.  Namely, what is memoryN really
> good for without PNP0C80:X?  If it is not good for anything in that case,
> it should never be created befor PNP0C80:X.

memoryN works without PNP0C80:X and does not depend on ACPI.  A memoryN
represents a memblk, each of which is merely a 128MB (in case of x86) of
memory chunk sliced from the entire memory ranges.  This is why it is
hard to associate between memoryN and PNP0C80:X without these symbolic
links (otherwise, you will have to calculate from memory address.)

The memory subsystem also obtains the memory ranges from EFI or e820
during boot, and ACPI is not necessary to construct memoryN at boot.
Since EFI / e820 only provides the boot-time configuration, ACPI is used
to update the memory ranges during hot-add/delete.

> > This patch calls
> > acpi_setup_mem_blk_links() in a point that solves this ordering issue
> > since this point guarantees that both memoryN and PNP0C80X are created
> > for both boot-time and hot-add.
> 
> I would prefer the ordering of creation to be the same in both cases.
> Otherwise it really looks like we need to work around a problem that we're
> creating for ourselves.
> 
> How exactly are memoryN created during boot?

memoryN is created in memory_dev_init().  This is even before
do_initcalls().  I'd also prefer the same ordering, but I felt that
changing this ordering would be rather challenging.

do_basic_setup()
  driver_init()
    memory_dev_init()

Thanks,
-Toshi


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/