Le 13/09/2018 à 11:35, Sudeep Holla a écrit : > On Thu, Sep 13, 2018 at 10:39:10AM +0100, James Morse wrote: >> Hi Brice, >> >> On 13/09/18 06:51, Brice Goglin wrote: >>> Le 12/09/2018 à 11:49, Sudeep Holla a écrit : >>>>> Yes. Without this change, we hit the lscpu error in the commit message, >>>>> and get zero output about the system. We don't even get information >>>>> about the caches which are architecturally specified or how many cpus >>>>> are present. With this change, we get what we expect out of lscpu (and >>>>> also lstopo) including the cache(s) which are not architecturally >>>>> specified. >>>>> >>>> lscpu and lstopo are so broken. They just assume everything on CPU0. >>>> If you hotplug them out, you start seeing issues. So reading and file >>>> that doesn't exist and then bail out on other essential info though they >>>> are present, hmmm ... >>> Can you elaborate? >>> >>> I am not sure cpu0 is supposed to be offlineable on Linux. There's no >>> "online" file in /sys/devices/system/cpu/cpu0. That's why former lstopo >>> doesn't like CPU0 being hotplugged out. We are actually making that case >>> work for another non-standard corner case. But offlining "cpu0" this is >>> considered "normal", somebody must add that missing "online" sysfs >>> attribute for "cpu0" (change >>> https://elixir.bootlin.com/linux/latest/source/drivers/base/cpu.c#L375). >> On x86 you can't normally offline CPU0, its something to do with certain >> interrupts always being routed to CPU0, (oh, and hibernate). >> You should be able to enable this behaviour with 'cpu0_hotplug' on the kernel >> command line. >> >> (Kconfig's CONFIG_BOOTPARAM_HOTPLUG_CPU0 and CONFIG_DEBUG_HOTPLUG_CPU0 are also >> worth a look) >> >> On arm64 at least, cpu0 is just like the others, and can be offlined. >> > Thanks James, for providing all the details. > > To add to the issues I spotted with lscpu/lstopo around topology, it ignores > the updates to topology sibling masks when CPUs are hotplugged in and out. > > We have following in lscpu: > add_summary_n(tb, _("Core(s) per socket:"), > cores_per_socket ?: desc->ncores / desc->nsockets); > > Now when cores_per_socket = 1, (i.e when we don't have procfs entry), > if ncores = (ncores_max - few_cpus_hotplugged_out), core(s) per socket > will get computed as less than the actual number. > > IMO lscpu should be used only when all CPUs are online and it should have > a warning when all cores are not online. > >>> By the way, did anybody actually see an error with lstopo when there's >>> no "type" attribute for L3? I can't reproduce any issue, we just skip >>> that specific cache entirely, but everything else appears. If you guys >>> want to make that "no_cache" cache appear, I'll make it a Unified cache >>> unless you tell me what to show :) > IIUC, Jeffrey Hugo did see error as per his initial message: > " > This fixes the following lscpu issue where only the cache type sysfs file > is missing which results in no output providing a poor user experience in > the above system configuration. > lscpu: cannot open /sys/devices/system/cpu/cpu0/cache/index3/type: No such > file or directory > " > I don't know about lscpu (it's a different project), but lstopo shouldn't have any such problem. If you see an issue with lstopo, I'd be interesting in getting the tarball generated by hwloc-gather-topology (it dumps useful files from procfs and sysfs so that we may debug offline). Brice
