#asahi · 2021-08-23 12:52
mazmarcan: has anyone documented the ESR values reported for an IMPDEF SError?#asahi · 2021-08-20 11:39
mazcgv[m]: you can't really directly associate what is reported by CLIDR_EL1 to actual cache levels. all you know is that it is ordered before something that would report '2'. And yes, I would fully expect the DSB to be where the cost can be measured.#asahi · 2021-08-20 11:37
mazfair enough. KVM uses it extensively, as it makes everything much simpler when dynamically mapping/unmapping guest pages. for the m1n1 case, it probably doesn't have any advantage.#asahi · 2021-08-20 11:33
mazyou may also see different behaviours if you run in a guest where the hypervisor has set FWB.#asahi · 2021-08-20 11:32
mazcgv[m]: right. that's totally expected, and inline with the architecture. you should observe timing differences with cleans to the PoC though, unless that one also reports 0...#asahi · 2021-08-20 11:30
mazcgv[m]: what does CLIDR_EL1.LoUIS report?#asahi · 2021-08-20 11:29
mazactually, s/SWB/S2FWB/#asahi · 2021-08-20 11:28
mazs/SWB/FWB/#asahi · 2021-08-20 11:28
mazcgv[m]: it has FEAT_SWB, so it is likely that clean to the PoU are implemented as NOPs (the i-side can snoop the d-side).#asahi · 2021-08-20 09:26
mazchadmed: IMPDEF stuff is all fine. things that actively break the architecture, less so. somehow, Apple is allowed to do that, while it is frowned upon for most other architecture licensees. #asahi · 2021-08-20 09:11
mazit does run Linux, and you have one in the office.#asahi · 2021-08-20 09:11
mazand VMs are, by definition, systems without EL3.#asahi · 2021-08-20 09:10
mazthat's assuming that EL3 exists. there are at least another example of a system without EL3.#asahi · 2021-08-20 09:05
mazthe E2H bit is another story altogether.#asahi · 2021-08-20 09:05
maztheir usage of FIQs is actually legal. just pretty uncommon.#asahi · 2021-08-19 14:10
mazalyssa: high quality stuff, I'm sure!#asahi · 2021-07-26 08:02
mazkettenis: gah. life got in the way and my brain is pretty similar to that of a goldfish these days. please respin the series, and I'll get on it (I need to revive my own PCIe series).#asahi · 2021-06-10 17:10
mazsorear: the ARM architecture isn't virtualisable. not in the Popek/Goldberg sense anyway. even with NV2, it is all a massive hack that only allows a small portion of the architecture to be approximated.#asahi · 2021-06-10 16:18
mazsorear: contentious? why?#asahi · 2021-06-10 15:38
mazyup, and it will always say EL0.#asahi · 2021-06-10 15:37
maz(and even by doing that, CurrentEL would still return the wrong thing).#asahi · 2021-06-10 15:35
mazby*#asahi · 2021-06-10 15:35
mazyou could emulate NV my running everything at EL0. good luck with that.#asahi · 2021-06-04 07:37
mazkettenis: yup, I'll try and get to it.#asahi · 2021-06-03 07:39
mazsven: that's pretty awesome, thanks for pulling everything together. It'd be good to work on a series that enable that upstream.#asahi · 2021-06-03 07:31
mazkettenis: ^^#asahi · 2021-06-03 07:31
mazsven: ah, I see you caught my stupid MSI address bug! thanks for that! :D#asahi · 2021-06-02 07:47
mazsven: sure, no worries. just let me know when you have something I can resync on.#asahi · 2021-06-01 07:20
mazsven: do you have a branch somewhere with the whole dart+pcie?#asahi · 2021-05-28 09:47
mazand the GICv3 virtualisation that allowed the guest to remove all interrupts from the host...#asahi · 2021-05-28 09:46
mazwell, they alsi ignored the requirement for ASIDs at EL2 with VHE...#asahi · 2021-05-28 09:46
mazso did I, and then thrown the box again! :D#asahi · 2021-05-28 09:45
mazpinskia: they completely fucked the v8.1 stuff in all TX1 implementations.#asahi · 2021-05-28 09:02
mazpinskia: yeah, this was added at the last minute (right *after* the arch was frozen) at the request of particular vendor, which explains the very bizarre encoding in ID_AA64_MMFR0_EL1.#asahi · 2021-05-28 08:04
mazEmantor: dremel?#asahi · 2021-05-28 08:03
mazsure. it just doubles the number of kernels to ship, and the size of the validation space.#asahi · 2021-05-28 08:02
maz(or a M1-specific distro).#asahi · 2021-05-28 08:02
mazwhich means no distro kernels for M1.#asahi · 2021-05-28 08:01
mazthen 16kB it has to be. we can't have different granularities between MMU and IOMMU.#asahi · 2021-05-28 07:59
maz(alpha style).#asahi · 2021-05-28 07:59
mazwe *may* be able to do something in the DMA layer to just map all memory in the PCI DARTs and treat it as a bus offset.#asahi · 2021-05-28 07:58
mazyeah, this is a total bummer.#asahi · 2021-05-28 07:58
mazwhich obviously would be 16kB for M1 thanks to the DART limitation...#asahi · 2021-05-28 07:57
mazmarcan: no, that's way to hard. you'd have to assume 64kB for all page alignments, turn the VA range into something dynamic... you'd be better off having 3 kernels, compiled for each page size, and get the bootloader to pick the one you want.#asahi · 2021-05-26 09:31
mazmarcan: yup, UNDEF is the default handler.#asahi · 2021-05-26 09:25
maz(without NV)#asahi · 2021-05-26 09:25
mazmarcan: that's what it looks like: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/arch/arm64/kvm/hyp/exception.c#asahi · 2021-05-26 09:23
mazmarcan: thankfully, from the guest PoV, it cannot make the difference. we just emulate that part of the architecture (and the NV support requires that at all times anyway).#asahi · 2021-05-26 09:21
mazmarcan: you execute an UNDEF, you take an exception. that's what the architecture says.#asahi · 2021-05-26 09:20
mazmarcan: no. the guest has done something that was never advertised as supported.