Microsoft Announces Linux-Powered "Azure Sphere" IoT Platform

[indepe]

How does one extinguish GPL licensed code

By taking over key customer groups, of course, and then slowly moving them to a “better” place. For example those looking fo “security innovations”.

The message: Linux better get its security story straight.

[RomuloP]

Uh, reasonably sure you'd want it on servers too, because if the *driver* has crashed anything running on those driver managed GPUs is likely corrupted as well. This is why you save your computational state as frequently as performance limits allow. It'll let you pick up where you were from before the crash, assuming something in your code isn't what's causing the crash. What you *don't* want is a driver crash bringing the whole system down which has been the traditional error state in Windows for most of its lifetime.

TDR (timeout detection and recovery) is a gambling, intended for the graphic stack (GUI), and it is really useful there, it's mostly based on timeouts as name says (and probably this is the best you can get at kernel mode), so if you have any CUDA intensive task that domine your GPU for enough seconds, TDR simply break everything to "recover" the graphic stack from a freezing, honestly, I can't even imagine the horrors it can generate in a vGPU dedicated host (no it did not save other state if not the graphic server... it is not cuda aware, neither virtualization aware, it is aware of the Windows graphic stack and mostly need modern DX, 9 up), it just works fine for the graphic stack because it is simply resilient, there is no much precision needed in GUI data as there is in a meteorologic simulator. But sure if you really think a GUI is a thing in a server... I can only think in one use for it, basic game servers (those that use the engine itself at server side to synchronize clients).

Anyway thanks God CUDA generally is not so propense to failure, depend on much less complicated components on the driver, leave the graphic stack out of it (and preferably leave a graphic stack out of the server...) and everything will run fine.

[stormcrow]

TDR (timeout detection and recovery) is a gambling, intended for the graphic stack (GUI), and it is really useful there, it's mostly based on timeouts as name says (and probably this is the best you can get at kernel mode), so if you have any CUDA intensive task that domine your GPU for enough seconds, TDR simply break everything to "recover" the graphic stack from a freezing, honestly, I can't even imagine the horrors it can generate in a vGPU dedicated host (no it did not save other state if not the graphic server... it is not cuda aware, neither virtualization aware, it is aware of the Windows graphic stack and mostly need modern DX, 9 up), it just works fine for the graphic stack because it is simply resilient, there is no much precision needed in GUI data as there is in a meteorologic simulator. But sure if you really think a GUI is a thing in a server... I can only think in one use for it, basic game servers (those that use the engine itself at server side to synchronize clients).

Anyway thanks God CUDA generally is not so propense to failure, depend on much less complicated components on the driver, leave the graphic stack out of it (and preferably leave a graphic stack out of the server...) and everything will run fine.

I was looking at it from a base hardware driver point of view rather than a graphics stack point of view, but I concede the point as far as a graphics stack dependent timer reset is concerned. I was considering it from a basic hardware watchdog reset - much like an auxiliary hardware module failing on a microkernel and simply resetting itself rather than "oh the graphics stack isn't responding!" due to preemption-- issues a premature reset and dumps my properly functioning OpenCL/CUDA program.

[RomuloP]

I was looking at it from a base hardware driver point of view rather than a graphics stack point of view, but I concede the point as far as a graphics stack dependent timer reset is concerned. I was considering it from a basic hardware watchdog reset - much like an auxiliary hardware module failing on a microkernel and simply resetting itself rather than "oh the graphics stack isn't responding!" due to preemption-- issues a premature reset and dumps my properly functioning OpenCL/CUDA program.

This is a common mistake about TDR, the Windows GPU drive still run exactly as any other modern system, at kernel mode, with some fancy reset that all graphic stack is aware of (and no it did not restore the last state so precisely all time, there exist all sort of BSOD failures that we can't figure where corruption really start and what is a sane state, or even if a failure is occurring, appart from a external timer, but how cares if you're at a similar image?) anything else you could (and is the right place to) implement recovering states at your own OpenCL/CUDA code (with a lot relevant performance impact I believe as you will need to rely on expensive synchronizations of cores to build a barrier concept, mostly why the driver or the graphic stack is not a good place for it).

[gnarlin]

How does one extinguish GPL licensed code

Tivoization is how.

[Drago]

Guys, doesn't Intel and AMD DRM drivers have GPU reset capability? Isn't it the same thing as this recovery in Windos?

[Delgarde]

are they required to release the source code then?

If it's GPL, yes, any kernel patches must be released – though with the usual caveat, that GPL requires only that the code be provided to users, not to the general public or to upstream. But since this is aimed at mass-market embedded hardware, there aren't any obvious loopholes about what constitutes a user...

Besides, remember that Microsoft is a lot friendlier towards Linux and GPL than they used to be – it actually makes up a large chunk of their profits these days, since something like 40% of the Azure cloud is running Linux (according to figures released last year). They've actually contributed a fair amount of upstream code in recent years...

[elvenbone]

But from a strict security point of view, all I can think of is that MS decided to use ACLs from the start, whereas Linux carried over the moronic "UGO" permissions model from Unix (and even today, its ACL support still sucks). It's obviously not a "Windows security innovation" though, ACLs have been used all the way back to MULTICS.

I get why ACLs are so important on servers, but MS' implementation on at least the desktop versions of Windows is a nightmare from end-user perspective.
Recently I set up a Windows 10 system for a relative, with the OS installed on a new SSD. The old HDD which previously hosted a Windows OS is now intended as data drive. However i couldn't figure out how to delete the 20GB Windows folder from there. I granted full read-write permissions to the relevant system-owned users and user-owned user accounts. Still no chance to delete that folder. Formatting the drive was not an option for practical reasons concerning user data.
Not to mention how long applying altered ACL permissions to many files can take.

[Yndoendo]

How does one extinguish GPL licensed code

Technically it could happen and does happen. By creating software that competes with GPL which is BSD or MIT or non-GPL license. Once people move to the alternative the original most likely will be no longer be used or maintained over time. Likely hood of doing to a largely used code base is close to zero. Even Mir (GPL) lives on when Wayland (MIT) was working to extinguish it, indirectly.

[Danielsan]

They can't extinguish the GPL but they have the resources to distort the way how the others perceive GNU and Linux, as it did for a while Ubuntu when overlapped the meaning of GNU/Linux distro. This is the power of the corporations that can invest money and time on marketing stuff. And it was exactly what has been doing Microsoft overlapping the meaning of personal computer with the Windows brand for a very long time.