#
✕ nick:kettenis
#asahi-dev · 2022-04-13 10:27
kettenisI was going to say that you're not allowed to work on new stuff before nvme has been upstreamed ;)
#asahi-dev · 2022-04-13 08:56
ketteniskevans91: yes that needs more work to support suspend/resume and runtime power management
#asahi-dev · 2022-04-11 06:59
ketteniss/nu/no/
#asahi-dev · 2022-04-11 06:59
ketteniskevans91: there are nu PCI bus master capable devices on bus 0
#asahi-dev · 2022-04-09 17:59
kettenisso I've left it in place for now
#asahi-dev · 2022-04-09 17:59
ketteniskevans91: it will be needed again when I add support for running interrupts on other cores
#asahi-dev · 2022-04-09 08:44
kettenisthanks for spotting that
#asahi-dev · 2022-04-09 08:44
ketteniskevans91: LOL, yes, that can't be right (it seems to work just fine though)
#asahi-dev · 2022-04-07 14:08
kettenisthe "compatible" property should be the definitve answer
#asahi-dev · 2022-04-07 13:56
kettenisfor OpenBSD a single one on the /soc node would be enough, but I'm not sure that works for Linux
#asahi-dev · 2022-04-07 13:55
kettenisstill, I think we need to sprinkle some dma-coherent properties into the device trees
#asahi-dev · 2022-04-07 13:54
kettenissven: On OpenBSD I couldn't really detect any performance benefits of skipping the cache flushes, but that may just be because the CPUs are too fast
#asahi-dev · 2022-04-07 13:15
kettenisbut nvme doesn't use the DART for example
#asahi-dev · 2022-04-07 13:14
kettenisthat's effectively what I do for OpenBSD (force the flag that tells the core code DMA is cache coherent in the DART driver)
#asahi-dev · 2022-04-07 13:10
kettenisby allocating those as non-cached in the first place?
#asahi-dev · 2022-04-07 13:08
kettenisnote that our device trees don't have dma-coherent attributes, which I think that Linux does all the appropriate cache flushing and invalidation
#asahi-dev · 2022-04-07 13:07
kettenisdoesn't make sense to me, but that doesn't carry much weight ;)
#asahi-dev · 2022-04-07 13:06
kettenissven: you mean the cache would still snoop DART transactions, but DART transaction would no longer consult the cache?
#asahi-dev · 2022-04-04 19:55
ketteniss/oqn/own/
#asahi-dev · 2022-04-04 19:55
kettenisah wait, I'm misinterpreting my oqn code
#asahi-dev · 2022-04-04 18:57
kettenissven: by writing the appropriate bit to AIC_SW_CLR?
#asahi-dev · 2022-04-02 14:54
kettenisright
#asahi-dev · 2022-04-02 14:53
kettenisI don't understand the C++ memory ordering semantics ;)
#asahi-dev · 2022-04-02 14:53
kettenisi.e. you may end up with old memory contents in X2, X1 and new memory contents in X0
#asahi-dev · 2022-04-02 14:53
kettenisso actually what I think this means is that if [X3] and [X0] are overlapping memory addresses the values in X2, X1 and X0 are not necessarily consistent after the two instructions have executed
#asahi-dev · 2022-04-02 14:49
kettenisnote that C3.2.4 talks about address dependencies
#asahi-dev · 2022-04-02 14:49
ketteniswell, I mean, true that my interpretation is probably wrong
#asahi-dev · 2022-04-02 14:48
kettenistrue
#asahi-dev · 2022-04-02 14:33
kettenisso if there is a 2nd thread that writes to [X3] you may end up in the situation where LDNP loads from the old address at [X3] but X0 will hold the new address
#asahi-dev · 2022-04-02 14:29
kettenisah, I think I understand now
#asahi-dev · 2022-04-02 14:23
kettenisbut the page you quoted does indeed seem to suggest that it actually affects the observer executing the code
#asahi-dev · 2022-04-02 14:22
kettenisanyway, C3.2.4 talks about *other* observers withing the sharability domain
#asahi-dev · 2022-04-02 14:17
kettenislooking at G.b here
#asahi-dev · 2022-04-02 14:15
kettenishmm, wait
#asahi-dev · 2022-04-02 14:13
kettenisand C6.2.129 makes it obvious it does indeed do the load
#asahi-dev · 2022-04-02 14:09
kettenisindeed
#asahi-dev · 2022-04-02 14:08
kettenishmm, maybe I'm reading this wrong
#asahi-dev · 2022-04-02 14:03
ketteniswhich is ok I suppose since it isn't an actual load; just a hint
#asahi-dev · 2022-04-02 14:03
kettenisbut yes, it seems like LDNP doesn't "interlock"
#asahi-dev · 2022-04-02 14:02
kettenisYichaoYu: The ARM ARM is a bit more explicit about this
#asahi-dev · 2022-03-29 19:20
kettenisjannau: pushed that m1 ultra support patch to the u-boot asahi branch
#asahi-dev · 2022-03-29 17:58
kettenisstill useful to continue the discussion with ardb I suppose
#asahi-dev · 2022-03-29 17:58
kettenisheh
#asahi-dev · 2022-03-29 17:57
kettenisso does this mean you have the efi interface for this in u-boot implemented?
#asahi-dev · 2022-03-29 17:53
kettenisj`ey: I believe you'll need to implement CPU_SUSPEND and add the idle state stuff from Documentation/devicetree/bindings/arm/idle-states.yaml in the device tree
#asahi-dev · 2022-03-29 17:48
kettenisthose appear to balance fairly well when not running on battery
#asahi-dev · 2022-03-29 17:47
kettenisOpenBSD looks at both PDTR and PSTR which are interpreted as input power and system power
#asahi-dev · 2022-03-28 15:55
kettenishaving additional I/O ranges for devices that aren't present shouldn't hurt
#asahi-dev · 2022-03-28 15:55
kettenisbut maybe simply extending the m1 max map to be the m1 ultra map would work as well
#asahi-dev · 2022-03-28 15:53
kettenisjannau: I think my preference is to simply copy the m1 max map and add additional regions to it

plus ›