#
✕ nick:marcan
#asahi · 2021-05-26 09:31
marcanyup, makes sense
#asahi · 2021-05-26 09:24
marcanah, right :)
#asahi · 2021-05-26 09:22
marcanbut of course this all makes sense, I just wanted to double check
#asahi · 2021-05-26 09:22
marcanmaz: sure, though normally those UNDEFs are sent straight to EL1, which is different from these which trap to EL2 and have to be reinjected, so I wasn't sure what the behavior was :)
#asahi · 2021-05-26 09:21
marcanMisthios: I guess your webclient is missing the Hypervisor Progress Glyphs unicode page
#asahi · 2021-05-26 09:20
marcanso no problem then :)
#asahi · 2021-05-26 09:20
marcanit would if you terminated the VM instead or something
#asahi · 2021-05-26 09:19
marcanmaz: it isn't if you inject UNDEF
#asahi · 2021-05-26 06:30
marcanyeah, but APTs like that have better exploits anyway, as I said in the faq... :)
#asahi · 2021-05-26 06:14
marcanand certainly not in scenarios sensitive enough where this kind of covert channel is an issue
#asahi · 2021-05-26 06:14
marcanI would not deliberately run fully untrusted, presumed-malicious code in containers as policy
#asahi · 2021-05-26 06:14
marcanI consider containers to be a security technology for, a priori, non-malicious software
#asahi · 2021-05-26 06:13
marcanthis is veering on the malware analysis side of things, and you are quite frankly insane if you are doing that with containers; that needs a VM
#asahi · 2021-05-26 06:13
marcanexactly, the use cases for "I want to run two containers but talking to each other breaks my security model" are very few
#asahi · 2021-05-26 05:47
marcane.g. I bet someone can come up with something funky with cache timing attacks on vDSO pages, or stuff like that
#asahi · 2021-05-26 05:46
marcanbut there are probably many ways of doing that anyway, with other CPU side channels and such (just this one is a particularly trivial one)
#asahi · 2021-05-26 05:46
marcanOS-based containers (e.g. docker on linux) yes, unfortunately
#asahi · 2021-05-26 03:38
marcanseems we were spared by choosing OFTC instead, but I'm still going to just purge the asahi channels
#asahi · 2021-05-26 03:37
marcanlooks like Freenode just purged all channels with the word "libera" in the topic
#asahi · 2021-05-26 03:35
marcanor perhaps just re-inject the exception as an unknown into the guest (which would then handle it as a userspace fault)
#asahi · 2021-05-26 03:34
marcan(which is what Hypervisor.framework does)
#asahi · 2021-05-26 03:34
marcanif it just kills the VM, then this thing turns into a DoS vuln in KVM, so it'd have to handle it and ignore reads/writes instead
#asahi · 2021-05-26 03:34
marcanmaz: I was just thinking, what does KVM do with impdef traps? I know it does set TIDCP
#asahi · 2021-05-26 03:02
marcan:p
#asahi · 2021-05-26 03:02
marcanthere, the silly thing is out: https://twitter.com/marcan42/status/1397387066044911618
#asahi · 2021-05-25 19:32
marcan��%
#asahi · 2021-05-25 17:17
marcan*trap
#asahi · 2021-05-25 17:17
marcan*tra@
#asahi · 2021-05-25 17:17
marcanwell, now that I can probably actually grap gxf registers I guess I'll find out
#asahi · 2021-05-25 17:16
marcan(maybe the PPL is only half there? Apple says it's not like iPhones; clearly there's GXF in use but how much exactly?)
#asahi · 2021-05-25 17:12
marcan... but the PAN thing plus PPL apparently being not called into that often and running fast conspired to confuse me
#asahi · 2021-05-25 17:12
marcanand because I wasn't seeing the GL2 exceptions
#asahi · 2021-05-25 17:12
marcansven had already worked most of it out, but I'd been *assuming* GL1 somehow trapped to EL2 because, well, it was getting way too far for that not to be the case!
#asahi · 2021-05-25 17:11
marcaneret goes back to GL1 or EL1 as required
#asahi · 2021-05-25 17:11
marcanand that just makes the whole thing work, and agnostic to whether it's in GL2 or EL2
#asahi · 2021-05-25 17:11
marcanso right now I just point the GL2 VBAR at the same vectors as the EL2 VBAR, and have wrappers to read/write the context registers that pick the right one based on the current mode
#asahi · 2021-05-25 17:10
marcanthey basically operate in parallel, roughly identically other than the separate context registers for GLx
#asahi · 2021-05-25 17:10
marcanit seems it's always EL1 -> EL2 and GL1 -> GL2
#asahi · 2021-05-25 17:10
marcanyes!
#asahi · 2021-05-25 17:10
marcanbut of course changing configuration settings would do things like make some operations slower, or more likely to trap while in GL1
#asahi · 2021-05-25 17:09
marcanbut I guess the guest spends little enough time in GL1 until that point in boot that most of the time it didn't crash
#asahi · 2021-05-25 17:09
marcanthe HV timer fires at 1000Hz
#asahi · 2021-05-25 17:09
marcanit's the timer firing *while* it's in GL1 that makes it go into GL2
#asahi · 2021-05-25 17:04
marcanit's never gotten past sec cpu bringup (possibly because for whatever reason that puts it in gl1 long enough to reliably die; large page table changes?) but it's randomly died way earlier, sometimes consistently depending on settings
#asahi · 2021-05-25 17:03
marcanj_ey: like to clarify, xnu is multitasking at this point, so non-det breadcrumbs make some sense, but what *didn't* make sense is crashing way earlier/way later depending on random configuration changes
#asahi · 2021-05-25 15:30
marcanI bet trapping GXF works fine now
#asahi · 2021-05-25 15:30
marcanbecause it would probably genter then trap on something, and that would hang
#asahi · 2021-05-25 15:30
marcanoh, and this also explains why trapping the GXF registers breaks
#asahi · 2021-05-25 15:30
marcanlike trapping more/fewer registers etc
#asahi · 2021-05-25 15:30
marcanj_ey: no, just the general "it crashes in a different place depending on completely random changes" thing

plus ›