Re: [PATCH 17/18] dax: Assign RAM regions to memory-hotplug by default

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



Gregory Price wrote:
> On Mon, Feb 06, 2023 at 02:15:56PM -0800, Dan Williams wrote:
> > Gregory Price wrote:
> > > On Sun, Feb 05, 2023 at 05:04:05PM -0800, Dan Williams wrote:
> > [..]
> > > 
> > > Stupid question: when defaulting to online, do these devices get placed
> > > into Zone Normal?  Is there a way for us, at a minimum, to online this
> > > as Zone Moveable in an effort to assist the "hope the kernel has not
> > > pinned a page" problem (and to try to keep kernel resources out of this
> > > zone in general).
> > > 
> > > If this is covered by a different patch or already set up this way,
> > > ignore me :]
> > 
> > Have a look in Documentation/admin-guide/mm/memory-hotplug.rst, the
> > 'daxctl recconfigure-device' man page and the the policy options of how
> > hot-added memory is brought online.
> > 
> > The routing can be anything from fully offline device-dax, to fully
> > online ZONE_NORMAL, or even a mix of device-dax subdivision,
> > ZONE_NORMAL, and ZONE_MOVABLE memblocks all within one CXL ram region.
> 
> Hm.
> 
> I'm just thinking, for early expander devices the assumption that the
> default behavior should be auto-online is good.  Just kinda working
> through this in my head for multi-headed devices and early pools, and I
> suppose the kernel paremeter covers that. Most of those will be used in
> specialty, vertically integrated systems, so we're good to go.
> 
> In the future when we get fully featured DCD's, I imagine we will have a
> memory region of size X, with N memory blocks, but only some of those
> blocks are online.  I suppose it could also be done such that DCD capacity
> add events drive the creation of new blocks under the region, as opposed
> to having the region pre-create the blocks.

In the case of DCD the proposal is "sparse" dax regions. So there will
be nothing to online by default until extents start arriving. I do
expect DCD systems will opt for memhp_default_state=offline so that
userspace policy can be applied.

> Either way I think the kernel parameter probably covers this case as
> well, I'm just wondering if at some point the default will yet again
> wish to be false, but i think that's a "CXL has become so successful all
> memory is now CXL DCDs and computers look radically different" scenario.

It's a distro policy. Fedora and RHEL, for example, pick different
defaults.




[Index of Archives]     [Linux ARM Kernel]     [Linux ARM]     [Linux Omap]     [Fedora ARM]     [IETF Annouce]     [Bugtraq]     [Linux OMAP]     [Linux MIPS]     [eCos]     [Asterisk Internet PBX]     [Linux API]

  Powered by Linux