Mark Knecht posted on Tue, 03 Sep 2013 13:25:15 -0700 as excerpted:
> I suspect I can disassociate the CUDA stuff if I have to. I don't have
> to tell X11 that the 465 is in the machine, and I have 3 PCI Express
> slots so I could potentially have 2 VGAs and 1 CUDA. (I think...)
There's actually some brand new developments in that regard -- hit the
Linux news in the last week, and scheduled for 3.12.
So what's the big deal? They're splitting the formerly single-device KMS/
DRM into two separate devices, one of which will be render/compute-only,
and thus require lower privs -- classic Unix user/group file permissions,
not the root or CAP_SYS_ADMIN that the current device requires, and a
second that will be mode-setting/display-controller only, accessible via
clone from the current DRM-control node (which is currently unused).
Because no new resources are created -- they're simply cloned from the
existing control node, required privs here can be reduced as well. The
practical effect of the latter will be another step toward allowing user-
priv-only X (and wayland).
But the former is bigger news in srovider. For just $5 per month, you can
get yourself a Cloud server with 512 MB of RAM, 20 GB super-fast SSD,
free snapshots, plus backups for a minimal fee. All via a simple
graphical interface.
And by signing up with this referral link, you can help support this
website.
If you are reading this, your ad could also be occupying this space.
Contact us to make it happen.
2 Comments
ome ways, as it means any app running as a user/group with suitable file
permissions will be able to open compute nodes, no root or CAP_SYS_ADMIN
required. That will in turn dramatically open up the possibilities for
"ordinary application" use of compute-nodes, thus allowing pretty much
any app to use those resources, instead of forcing the severe privilege
restrictions currently needed to avoid huge security issues, currently.
So why is all this happening now? It turns out that while on x86/amd64,
it has historically been the same hardware device handling both rendering
and mode-setting, on arm, which is where the big money is now since
that's what all the mobile devices seem to run on, these functions are
often split into two different pieces of hardware. The existing unified
setup could be made to work, but it wasn't a particularly good fit, while
the new setup is a natural fit.
Meanwhile, while they could have solved their own little hardware problem
in their corner of the computing world in just their own drivers, the
bigger solution was to split the functionality up in general, thus
allowing solutions to a bunch of thorny problems with the existing
unified hardware as well. =:^) In addition to the security and compute
resource availability issues mentioned above that this solves, it also
makes solving problems such as the newer embedded/dedicated gpu pairing
on laptops, with the display only attached to the embedded, thus making
the dedicated card that's supposed to activate for games and stuff
useless, under the current unified setup. Similarly, it'll now be
possible to run different user X sessions on different displays driven by
the same GPU hardware -- multi-seat-on- single-gpu-multi-display, which
currently isn't possible, because only one app can be DRM-master on a gpu
at once, and that app must be run by one user, it can't be shared.
So the separate render/modeset hardware on arm is driving better and far
more flexible solutions for (normally) unified render/modeset hardware on
x86/amd64, too! =:^)
Original blog announcement:
https://dvdhrm.wordpress.com/2013/09/01/splitting-drm-and-kms-device- nodes/
Initial LWN coverage. (Not much there besides the single paragraph
pointer to the above at present, but kernel devs often comment there, and
LWN will very likely have a feature article covering it in some depth in
one of its coming weekly editions, too.)
http://lwn.net/Articles/565464 Phoronix (pre-blog announcement coverage):
http://www.phoronix.com/scan.php?page=news_item&px=MTQ0OTI (earlier...)
http://www.phoronix.com/scan.php?page=news_item&px=MTQ0MzQ (earlier still, July...)
http://www.phoronix.com/scan.php?page=news_item&px=MTQwNTA --
Duncan - List replies preferred. No HTML msgs.
"Every nonfree program has a lord, a master --
and if you use the program, he is your master." Richard Stallman