Re: [PATCH] dax: adding fsync/msync support for device DAX

From: Jeff Moyer <jmoyer@redhat.com>
To: Dan Williams <dan.j.williams@intel.com>
Cc: Christoph Hellwig <hch@infradead.org>,
	linux-nvdimm <linux-nvdimm@lists.01.org>
Subject: Re: [PATCH] dax: adding fsync/msync support for device DAX
Date: Wed, 11 Apr 2018 13:27:53 -0400	[thread overview]
Message-ID: <x498t9tps9y.fsf@segfault.boston.devel.redhat.com> (raw)
In-Reply-To: <CAPcyv4guech76UWc7xo+SAzKOv8dT83HTUoUPVUCx5Kjuxy5Vg@mail.gmail.com> (Dan Williams's message of "Wed, 11 Apr 2018 09:27:22 -0700")

Dan Williams <dan.j.williams@intel.com> writes:

> On Wed, Apr 11, 2018 at 9:06 AM, Jeff Moyer <jmoyer@redhat.com> wrote:
>> Christoph Hellwig <hch@infradead.org> writes:
>>
>>> On Fri, Apr 06, 2018 at 03:41:39PM -0700, Dan Williams wrote:
>>>> Yes, but the trust interface definition is what is missing, especially
>>>> when we consider memmap=ss!nn and qemu-kvm. For example do we turn off
>>>> DAX and/or MAP_SYNC on all platforms that don't provide a positive "I
>>>> have ADR" indication (ACPI 6.2 Section 5.2.25.9 NFIT Platform
>>>> Capabilities Structure)?
>>>
>>> Sounds like a default.
>>
>> Which do you turn off?  DAX or MAP_SYNC (or both)?
>
> Only MAP_SYNC I would say, because then userspace can't opt out of
> fsync/msync and it gives us a chance to "Do The Right Thing (TM)" with
> respect to WPQ flush.

That works for me.

>> Systems that support ACPI versions earlier than 6.2 could provide
>> flush hint addresses.  In that case, we could assume no ADR, and turn
>> off MAP_SYNC, but still allow DAX.  Note that I'm not aware of any
>> hardware that actually falls into this category.
>>
>> Platforms prior to 6.2 that don't support flush hints are currently
>> assumed to implement ADR.  This proposal would change that default.
>>
>>>> Require opt-in when the user has trust in the hardware config that
>>>> the kernel can't verify?
>>>
>>> And that sounds like a good config nob ontop of that default.
>>
>> Well, we could also make a white-list for known good platforms.  I
>> assume HPE's systems would be on that list.  Otherwise we'd have to cook
>> up udev rules, I guess?
>
> udev rules sound good. We can distribute them in the ndctl package. I
> think I would handle this by making
> /sys/bus/nd/devices/regionX/persistence_domain writable to opt-in to a
> user specified persistence_domain. If the persistence_domain attribute
> file is not writable then you know you're on a kernel that blindly
> trusts all pmem descriptions as MAP_SYNC capable.
>
>> Dan, is this something you're working on now?
>
> No, it's behind finalizing dax-dma-vs-truncate and memcpy_mcsafe for
> user copies in my queue...

OK, I'll take a stab at it.

Thanks,
Jeff
_______________________________________________
Linux-nvdimm mailing list
Linux-nvdimm@lists.01.org
https://lists.01.org/mailman/listinfo/linux-nvdimm