#
#asahi-dev · 2026-10-25 22:32
JamesCalligeros[m]sofus and a-ramses both have DT MRs that supposedly work
#asahi-dev · 2026-10-25 22:32
chaos_princessbut max is problematic
#asahi-dev · 2026-10-25 22:32
chaos_princesssomeone got one of those two working (don't remember which)
#asahi-dev · 2026-10-25 22:32
jannauis it only m3 max? dcp m3 and m3 pro works? I mostly ignored m3* dcp for now
#asahi-dev · 2026-10-25 22:31
svenwould be nice if this all works without hacking exclave support into all this
#asahi-dev · 2026-10-25 22:31
svenalright, it's crashing some some sysreg inside XNU now that I don't patch out yet. enough progress for today
#asahi-dev · 2026-10-25 22:30
mmediouni[m]Apple and Android-land most likely going to stick with pMTE
#asahi-dev · 2026-10-25 22:29
JamesCalligeros[m]oh even better....
#asahi-dev · 2026-10-25 22:29
mmediouni[m]MS asked for FEAT_VMTE and x86's memory tagging extension is also virtual memory-based
#asahi-dev · 2026-10-25 22:29
chaos_princessand yes, things go wrong quite early, even before they start chatting over afk/epic
#asahi-dev · 2026-10-25 22:29
JamesCalligeros[m]i hope we dont have to feed it a bunch of property chunks before starting things now, that would suck big time
#asahi-dev · 2026-10-25 22:28
mmediouni[m]lower priority item, I expect it to come at some point
#asahi-dev · 2026-10-25 22:28
JamesCalligeros[m]oh yeah i remember you mentioning this a little while ago now
#asahi-dev · 2026-10-25 22:28
svendo you know why they don't expose MTE in vms?
#asahi-dev · 2026-10-25 22:28
svenmmediouni[m]: ah, makes sense!
#asahi-dev · 2026-10-25 22:27
chaos_princessalso starting endpoints in macos order did not help (i think)
#asahi-dev · 2026-10-25 22:27
mmediouni[m]sven: virtual MTE (on virtual memory instead of physical, so that you can have different tags on the same underlying physical page depending on the view) is an ARMv10 feature - FEAT_VMTE
#asahi-dev · 2026-10-25 22:27
JamesCalligeros[m]the tracer in main should be good enough to see whats going on now
#asahi-dev · 2026-10-25 22:27
JamesCalligeros[m]thats annoying
#asahi-dev · 2026-10-25 22:27
chaos_princesscan confirm the above
#asahi-dev · 2026-10-25 22:26
yukaJamesCalligeros[m] Yes on M3 Max
#asahi-dev · 2026-10-25 22:26
svenyuuup, creating a SPTMDebug memory region and just making it 16 KiB (which it actually is supposed to be I think) + 2 GiB seems to work :D
#asahi-dev · 2026-10-25 22:26
JamesCalligeros[m]yuka: is this still 14.8.3 on m3*?
#asahi-dev · 2026-10-25 22:25
mmediouni[m]tags are stored in a tag memory region that's accessible only to SPTM but with a twist. If the pages that reference that tag page are not using MTE, then the underlying MTE tag page is usable as regular memory
#asahi-dev · 2026-10-25 22:24
mmediouni[m]for macs
#asahi-dev · 2026-10-25 22:24
mmediouni[m]chaos_princess: no MTE exposed through virtualization right now
#asahi-dev · 2026-10-25 22:22
yuka> - [ap_link.c:308]shared memory dva: 0x1000003c000 0xffffffff004f0000
#asahi-dev · 2026-10-25 22:22
yukabut yeah this log line never appears:
#asahi-dev · 2026-10-25 22:21
yukawill continue from here on tuesday probably
#asahi-dev · 2026-10-25 22:20
yuka(reproducer is currently still linux)
#asahi-dev · 2026-10-25 22:20
yukaI sadly only managed to set up the reproducer for dcp and get the traces for comparison. I see a few differences in the endpoints started, in invalidations etc. But probably need to get the m1n1 experiment into either state in order to isolate the cause.
#asahi-dev · 2026-10-25 22:18
svenif it doesn't I'll just go back to patching out that panic and steal memory from the beginning of RAM again
#asahi-dev · 2026-10-25 22:18
svenlet's see if that works
#asahi-dev · 2026-10-25 22:18
svenbut I think I may have found a way to steal some memory (just a little 2 GiB hole) from SPTM without it complaining :D
#asahi-dev · 2026-10-25 22:18
svenyeah, it's tagging (real) physical addresses
#asahi-dev · 2026-10-25 21:51
svenah.. I think there's no need for a separate EL1 storage because the allocation tags are for _physical_ addresses
#asahi-dev · 2026-10-25 21:38
chaos_princesscool
#asahi-dev · 2026-10-25 21:38
svenapparently "tag ram" stores the allocation tags and that seems to be IMPDEF? I dunno
#asahi-dev · 2026-10-25 21:37
svenno idea
#asahi-dev · 2026-10-25 21:35
chaos_princesssince that can probably be remapped
#asahi-dev · 2026-10-25 21:35
chaos_princesswhere does el1 mte store tags?
#asahi-dev · 2026-10-25 21:34
svenbut I'd assume mte can just be enabled at el1 no matter what the state at el2 is
#asahi-dev · 2026-10-25 21:34
svenno idea
#asahi-dev · 2026-10-25 21:34
svenyeah, I think that might work. [fake mte tag region][sprr shadow pagetables][real mte tag region] and then just map fake to real via stage 2 and pretend RAM ends after the fake region
#asahi-dev · 2026-10-25 21:34
chaos_princesslike, qemu, or utm, or w/e
#asahi-dev · 2026-10-25 21:33
chaos_princessdoes macos expose mte to vms?
#asahi-dev · 2026-10-25 21:33
svenI don't know if SPTM tries to write to that region.. though I guess I could remap those writes via stage 2
#asahi-dev · 2026-10-25 21:32
chaos_princessand virtualize said reg
#asahi-dev · 2026-10-25 21:32
chaos_princesswhat about having shadow page tables right before it in physical memory, but hide them from sptm
#asahi-dev · 2026-10-25 21:31
svenwriting to that IMPDEF register to move that region results in an serror ofc :(

plus ›