LWN: Comments on "Moving physical pages from user space"

Moving physical pages from user space

kaesaecracker — Mon, 02 Oct 2023 08:40:11 +0000

I think it might make it better, because you could add rules for which process can influence which process and because it would not tell you the physical addresses

Moving physical pages from user space

Aissen — Thu, 21 Sep 2023 16:46:26 +0000

I know about CXL, but I fail to see why one would want "profiling mechanisms" to live outside the kernel and move physical pages around. What am I missing ?

Moving physical pages from user space

corbet — Wed, 20 Sep 2023 07:43:27 +0000

FWIW, there is a V2 patch set out there now, without feedback so far.

Moving physical pages from user space

willy — Tue, 19 Sep 2023 21:54:00 +0000

For whatever reason, this patch set never made it to the linux-mm mailing list, which may be why it has had so little feedback.

It's an ugly design that solves the wrong problem in the wrong way.

Moving physical pages from user space

nim-nim — Tue, 19 Sep 2023 08:59:09 +0000

Given that the original motivation is

> For a process executing on a given node, memory attached to that same node will be faster than memory on other nodes, so the placement of memory matters.

But the proposed implementation is trying to

> move pages between memory types without the need for an awareness of which processes are using those pages.

I seriously doubt success will be achieved.

If user-space detects a fast tier is under utilised but does not want to analyse process by process which one would benefit from a relocation, surely an API prompting the kernel to fill the fast tier, looking itself for the processes that would benefit most would be more appropriate ?

Moving physical pages from user space

florianfainelli — Tue, 19 Sep 2023 02:45:16 +0000

Would a system call taking a PID work a bit better, security wise, while still meeting the initial intent?

Moving physical pages from user space

xecycle — Tue, 19 Sep 2023 00:47:26 +0000

Seconded. Obtain a pseudo fd and map it, copy over the old page, let the user resolve thread races, then remap over the old address.

Moving physical pages from user space

gerdesj — Mon, 18 Sep 2023 23:04:14 +0000

"It seems BPF"

I'm just missing "Rust" for bingo. We are at the writing numbers on the table stage now ...

Moving physical pages from user space

ibukanov — Mon, 18 Sep 2023 20:31:00 +0000

It seems BPF should be more suitable for this kind of management decision than a user space application.

Alternatively, if the app knows so much about physical memory, can kernel just disable virtual memory or at least make it map 1-to-1to physical memory for an app running exclusively on some Numa node?

Moving physical pages from user space

caliloo — Mon, 18 Sep 2023 18:42:45 +0000

I wonder what the thinking is when it comes to rights management around such a feature… there also seems to little tangible information about a use case. I’m very curious…

Moving physical pages from user space

alonz — Mon, 18 Sep 2023 18:14:36 +0000

Another potential security issue is that moving pages to a slower memory tier can help attackers using timing side channels.

Moving physical pages from user space

quotemstr — Mon, 18 Sep 2023 15:53:12 +0000

Why can't the process wanting to migrate the page just mmap it, even use the virtual memory APIs? How does it know what the right PFN even is? This API seems bad.

Moving physical pages from user space

NightMonkey — Mon, 18 Sep 2023 15:33:44 +0000

I am confident that there would be many eyes on all patches resulting from this conversation, and my ignorance is vast, but this would seem to open the "attack surface" of the kernel wider. Wouldn't you have to worry more about unauthorized attempts to access memory allocated to other processes? Or devices driver mapped memory?

And what about subtle interactions between the mechanisms used for virtual mappings and physical mapping by the kernel? If these subsystems are not "aware" of each other, could new bugs be introduced that would be hard to test for?

I'm not saying I believe one path or the other are better, but the security angle seems interesting here.

Cheers!