#asahi-dev · 2026-08-15 14:39
yukabut I guess this is not a problem since the kernel does not blanket map everything?#asahi-dev · 2026-08-15 14:38
yukathey are only unmapped inside of m1n1, and not once the kernel re-inits the mmu?#asahi-dev · 2026-08-15 14:38
yukathe ones read from "plane_base+i*tz->stride+tz->{start,end}", and then unmapped#asahi-dev · 2026-08-15 14:36
yukamcc carveouts#asahi-dev · 2026-08-15 14:30
svener, wait, you mean mcc carveouts and not fdt carveouts i guess?#asahi-dev · 2026-08-15 14:30
svenand doesn't linux reserve the carveouts in its own memory map and never hands out that memory to anyhting?#asahi-dev · 2026-08-15 14:30
svennothing, it just has to be enabled#asahi-dev · 2026-08-15 14:24
yukaand if so, what about mcc even matters to linux other than it being enabled?#asahi-dev · 2026-08-15 14:22
yukado the carveouts even matter once linux takes over?#asahi-dev · 2026-08-15 14:08
yukado we also want an mcc_disable_cache and ensure the cache is disabled at some point?#asahi-dev · 2026-08-15 14:05
yukaI guess we still need mcc_unmap_carveouts() then even if these are enabled at boot?#asahi-dev · 2026-08-15 14:04
yukathey start out with 0x2800 = 0x8000000c, 0x2804 = 0xc000c00, and if I write 0 or some other <0xc value 'X' to the last nibble of the enable reg, it is echoed back as 0xX000X00 in the status reg. So I guess that means CACHE_WAYS = 12.#asahi-dev · 2026-08-15 14:00
yukat8140 & t6040 are different from this: both have a moved CACHE_ENABLE/CACHE_STATUS (0x2800/0x2804 instead of 0x1c00/0x1c04) and are already enabled on startup#asahi-dev · 2026-08-15 13:58
yukajust a different reg_offset, and the mcc_count can't be estimated using the remaining regs since there are trailing regs which are not mccs#asahi-dev · 2026-08-15 13:57
yukasame cache-ways#asahi-dev · 2026-08-15 13:56
yukasame tz_regs#asahi-dev · 2026-08-15 13:56
yukat8132 mcc is very similar to t6030#asahi-dev · 2026-08-15 10:45
jannaulooks like a factor 10 error is mine. tracing counts 220k thunderbolt_net:tbnet_rx_ip_frame events per second and not 25k I calculated with#asahi-dev · 2026-08-15 10:32
jannau~20 us per map or unmap page (~1 gigabyte per second and 4k payload per package) feels at least a order of magnitude too slow#asahi-dev · 2026-08-15 10:07
svenah well, that sounds like the issue then#asahi-dev · 2026-08-15 10:07
jannauerr, I missed that it unmaps and reallocates buffers in tbnet_poll (the receive path)#asahi-dev · 2026-08-15 10:07
sveni also wonder if we even need to do tlb maintenance when adding mappings to dart#asahi-dev · 2026-08-15 09:53
svenhm.. weird that the speeds are so different then#asahi-dev · 2026-08-15 09:53
jannauit allocs a fixed number of buffers for rx and tx#asahi-dev · 2026-08-15 09:51
svenlet's see if i can reproduce that usb3 no longer works after a tbt connection was made#asahi-dev · 2026-08-15 09:51
svenso now devices don't wedge the port forever anymore#asahi-dev · 2026-08-15 09:51
sventhe new thunderbolt WARN was surprisingly easy to fix i think \o/#asahi-dev · 2026-08-15 09:40
svensounds like one path re-uses buffers while the other constantly maps/unmaps them#asahi-dev · 2026-08-15 09:40
sveni bet it's very different#asahi-dev · 2026-08-15 09:40
svenif i were to be nerdsniped into fixing this i'd look into how it does dma buffer alloc/free in the RX vs. the TX path :D#asahi-dev · 2026-08-15 09:40
jannauyes#asahi-dev · 2026-08-15 09:40
sventhis is usb4net, right?#asahi-dev · 2026-08-15 09:39
svenit essentially batches unmaps and then does them all at once after some timeout or when the queue is filled#asahi-dev · 2026-08-15 09:39
svenyeah, -FQ only affects unmap#asahi-dev · 2026-08-15 09:38
jannauthe perf profile shifted towards dart_map_pages https://paste.centos.org/view/9576827b#asahi-dev · 2026-08-15 09:37
sventhere must be some difference the way dma buffers are handled in the RX and TX path#asahi-dev · 2026-08-15 09:36
svenokay, so it's iommu crap#asahi-dev · 2026-08-15 09:36
jannaufurther increase to 10.5 Gbits/sec#asahi-dev · 2026-08-15 09:33
svencan you try echo DMA-FQ > /sys/kernel/iommu_groups/$GROUP/type for the iommu group (should be in dmesg during acio bringup i think)? i'm not sure that's actually gonna do anything with the current driver though#asahi-dev · 2026-08-15 09:32
jannauincreased throughput from 7.2 to 8.8 Gbits/sec by forcing the irq to performance core and switching cpufreq gov to performance#asahi-dev · 2026-08-15 09:32
svencould be weird interaction between usb4net and dart#asahi-dev · 2026-08-15 09:26
jannaunot sure that perf samples correctly (I see less samples than I would expect) but https://paste.centos.org/view/cc49f0f1#asahi-dev · 2026-08-15 09:22
jannauwe can on t8103#asahi-dev · 2026-08-15 09:21
sveni don't think we can control that, can we?#asahi-dev · 2026-08-15 09:21
jannauwhy are all AIC interrupts served on core 0 (t8103)#asahi-dev · 2026-08-15 09:20
svencan you see what it's doing in that softirq context?#asahi-dev · 2026-08-15 09:18
jannaupossibly with https://github.com/AsahiLinux/m1n1/pull/563#asahi-dev · 2026-08-15 09:17
chaos_princesscan we turn off the e-cores and still have linux boot?#asahi-dev · 2026-08-15 09:17
jannausmells like irq though. on the receiving side a single core is at ~90% in soft irq state. for some reason the first (efficiency) core#asahi-dev · 2026-08-15 09:10
svenoh, true