Details

Reviewers

Group Reviewers

Commits

rG6fddab804a17: amd64: Reload CPU ext features after resume or cr4 changes
rG050f5a8405c6: amd64: Reload CPU ext features after resume or cr4 changes

Diff Detail

Repository

rG FreeBSD src repository

Lint

Lint Not Applicable

Unit

Tests Not Applicable

Event Timeline

dchagin created this revision.Jun 22 2022, 11:42 PM

Herald added a subscriber: imp. · View Herald TranscriptJun 22 2022, 11:42 PM

dchagin requested review of this revision.Jun 22 2022, 11:42 PM

Harbormaster completed remote builds in B46069: Diff 107264.Jun 22 2022, 11:42 PM

dchagin retitled this revision from linprocfs: Save some cpu features to avoid overwriting by the kernel to linprocfs: Save ext cpu feature3 as it is zeroed by the kernel.Jun 22 2022, 11:50 PM

dchagin added reviewers: Linux Emulation, kib.

dchagin changed the repository for this revision from rS FreeBSD src repository - subversion to rG FreeBSD src repository.

read and local feature2 to reflect cr4 changes

Harbormaster completed remote builds in B46092: Diff 107290.Jun 23 2022, 10:59 AM

If these are indeed the problems, they should be fixed in the kernel proper, not worked around in linprocfs.

sys/compat/linprocfs/linprocfs.c
149 ↗	(On Diff #107290)	If you are about OSPKE, then I believe that the right fix is to re-read feature2 after setting the bit in CR4.
150 ↗	(On Diff #107290)	Can you explain more, where feature3 is zeroed?

In D35555#807163, @kib wrote:

If these are indeed the problems, they should be fixed in the kernel proper, not worked around in linprocfs.

sure, I wrote a reply to the 986c4ca3 a few days ago about this

sys/compat/linprocfs/linprocfs.c
149 ↗	(On Diff #107290)	agree, I commented on it above
150 ↗	(On Diff #107290)	https://cgit.freebsd.org/src/tree/sys/amd64/acpica/acpi_wakeup.c#n219 seems on wakeup we should reload ext features before initializecpu(), at first look

So why not write a proper patch instead?

In D35555#807315, @kib wrote:

So why not write a proper patch instead?

Target fixation, ok, I just got home, I'm on it.

Replace linprocfs workaround by kernel fix

Harbormaster completed remote builds in B46131: Diff 107430.Jun 27 2022, 11:48 AM

dchagin mentioned this in D35556: linprocfs: Decode more CPU flags in cpuinfo.Jun 27 2022, 11:48 AM

dchagin retitled this revision from linprocfs: Save ext cpu feature3 as it is zeroed by the kernel to amd64: Reload cpu ext features after resume or cr4 changes.Jun 27 2022, 11:49 AM

kib added inline comments.Jun 27 2022, 5:16 PM

sys/amd64/amd64/initcpu.c
296	I suggest only do this once, see the IS_BSP() check above. Otherwise you introduce the unpleasant race where all APs are writing to the variables. I believe that resume occurs on BSP, please recheck.