#
✕ nick:maz
#asahi · 2021-11-09 10:47
mazindeed. I started reading the USB4 spec, and it's not fun.
#asahi · 2021-11-09 10:47
mazsomehow, the PMU was more fun.
#asahi · 2021-11-09 10:46
mazI need to find a moment to tackle the PCIe hot-plug stuff.
#asahi · 2021-11-05 15:55
mazPMU driver updated with a mapping table of which event can count where. Probably not complete, but that's what I could find with a bit of experimentation.
#asahi · 2021-11-05 14:05
maza whole bunch are restricted to 5,6,7.
#asahi · 2021-11-05 14:04
mazevent 0x8c is really odd. it can count everywhere if placed at least in PMCR7...
#asahi · 2021-11-05 13:53
mazalmost halfway through (6.5 sec per event)
#asahi · 2021-11-05 13:53
mazbgb: I'm generating the event-map myself. after all, it is just a matter of iterating through the event space with all counters configured the same way, and see what sticks.
#asahi · 2021-11-05 13:11
mazj_ey: the theme is *songs* by FZ, not albums: http://lib.ru/SONGS/frankzappa/listofsongs.txt
#asahi · 2021-11-05 12:46
mazso not all counters are created equal.. what a pain.
#asahi · 2021-11-05 12:45
mazbgb: yup, 0x8c in counter #7 does the trick: https://paste.debian.net/1218324/
#asahi · 2021-11-05 11:27
mazbgb: I'll give it a shot later, debugging something else ATM.
#asahi · 2021-11-05 11:04
mazthis is obviously copyrighted material.
#asahi · 2021-11-05 11:04
mazI would really refrain from doing so.
#asahi · 2021-11-05 11:02
mazbgb: that'd involve booting macos, and I'm not doing that.
#asahi · 2021-11-05 10:58
mazis there a public repo with this file? dougall's python script doesn't seem to do anything...
#asahi · 2021-11-05 10:50
maz0x8c (aka INST_ALL) never counts here.
#asahi · 2021-11-05 10:45
mazbgb: I only use a mini, so no kbd.
#asahi · 2021-11-05 10:42
mazbgb: not that I can think of. but I'm not using much apart from PCIe.
#asahi · 2021-11-05 10:35
mazI'll give this a try, that looks like what I need!
#asahi · 2021-11-05 08:44
maz_jannau_: ah, I see. yes, that seems to be a good approach.
#asahi · 2021-11-05 08:41
maz'and if it discloses'*
#asahi · 2021-11-05 08:40
maz_jannau_: if you have the equivalent of perf in macos and it disclose the raw event numbers, that'd be great!
#asahi · 2021-11-05 08:34
mazI'm also pretty sure that counter #1 (instruction count) is completely synthetic. there is no other event that counts anything close.
#asahi · 2021-11-05 08:31
maz_jannau_: the most recent stuff is Hurricane, which predates the M1 cores by at least 4 years. In general, events are reorganised pretty often to fit the uarch, so it is quite understandable that the event numbering is vastly different.
#asahi · 2021-11-05 08:22
maz_jannau_: the events are for much older CPUs. they all have A32/T32 events, which makes no sense on a 64bit only CPU.
#asahi · 2021-11-05 08:14
mazshenki: there is still plenty of work to be done on the PMU. only *two* events are know. it you have time to write test cases and work out some of the most obvious events (cache miss, for example), that'd be useful.
#asahi · 2021-10-28 08:46
mazI refuse to run a hypervisor. end of. :D
#asahi · 2021-10-28 08:41
mazfor bare-metal work, nothing replaces the serial console, unfortunately.
#asahi · 2021-10-26 10:45
mazagraf: in KVM, I consider that a S2 exit on d-side CMO for non-memory always succeeds. since there is no mapping, there is nothing to cache.
#asahi · 2021-10-25 18:58
mazas far as the core is concerned, this could be CMO over avian carrier.
#asahi · 2021-10-25 18:57
mazinterconnect fabric such as AMBA ® AXI are, strictly, outside the scope of the architecture."
#asahi · 2021-10-25 18:57
mazmechanisms by which different memory type and Cacheability attributes are presented on an interface to an
#asahi · 2021-10-25 18:57
maz"This means the
#asahi · 2021-10-25 18:57
maznobody says anything about AXI. the only mention of AXI in the architecture is as follows:
#asahi · 2021-10-25 18:53
mazkettenis: who cares? :D
#asahi · 2021-10-25 18:53
mazagraf: it is supposed to do what it says on the tin: a cache clean and invalidate to the PoC. the end-point is allowed to barf, but the instruction is still valid.
#asahi · 2021-10-25 17:53
mazkettenis: agraf: Page D4-2652 in version G.b of the ARMv8 ARM explicitly calls out CMOs on device memory as allowed.
#asahi · 2021-10-25 17:24
mazmarcan: sure, I'll do that either later today or tomorrow morning, depending on how annoyed I'm in this call...
#asahi · 2021-10-25 15:00
mazj_ey: at least it is a noun followed by a verb (or the other way around).
#asahi · 2021-10-18 08:26
mazI'm not indifferent to non-invasive surgery, but given that I already have the tooling... :D
#asahi · 2021-10-18 08:18
mazj_ey: I'll survive that.
#asahi · 2021-10-18 08:17
mazj_ey: booting into macos involves finding a keyboard and a screen... I think I like sven's approach better.
#asahi · 2021-10-18 08:15
mazs/at//
#asahi · 2021-10-18 08:15
mazsound? and I was trying to prevent the mini from making that sound at on reset... ;-)
#asahi · 2021-10-18 08:07
mazGlanzmann: no worries
#asahi · 2021-10-18 08:03
mazGlanzmann: no, I have no such thing. I don't touch the corellium stuff at all.
#asahi · 2021-10-18 07:46
mazGlanzmann: try *what* out? my tree is at git://git.kernel.org/pub/scm/linux/kernel/git/maz/arm-platforms.git , but there is no single M1 branch. rather a collection of series.
#asahi · 2021-10-17 12:27
mazeta: compared to other European countries, the UK trains are pretty crap. even France does better, even with the strikes! ;-)
#asahi · 2021-10-17 12:23
mazs/lay/pay/

plus ›