On 2019/9/3 15:11, Peter Zijlstra wrote: > On Tue, Sep 03, 2019 at 02:19:04PM +0800, Yunsheng Lin wrote: >> On 2019/9/2 20:56, Peter Zijlstra wrote: >>> On Mon, Sep 02, 2019 at 08:25:24PM +0800, Yunsheng Lin wrote: >>>> On 2019/9/2 15:25, Peter Zijlstra wrote: >>>>> On Mon, Sep 02, 2019 at 01:46:51PM +0800, Yunsheng Lin wrote: >>>>>> On 2019/9/1 0:12, Peter Zijlstra wrote: >>>>> >>>>>>> 1) because even it is not set, the device really does belong to a node. >>>>>>> It is impossible a device will have magic uniform access to memory when >>>>>>> CPUs cannot. >>>>>> >>>>>> So it means dev_to_node() will return either NUMA_NO_NODE or a >>>>>> valid node id? >>>>> >>>>> NUMA_NO_NODE := -1, which is not a valid node number. It is also, like I >>>>> said, not a valid device location on a NUMA system. >>>>> >>>>> Just because ACPI/BIOS is shit, doesn't mean the device doesn't have a >>>>> node association. It just means we don't know and might have to guess. >>>> >>>> How do we guess the device's location when ACPI/BIOS does not set it? >>> >>> See device_add(), it looks to the device's parent and on NO_NODE, puts >>> it there. >>> >>> Lacking any hints, just stick it to node0 and print a FW_BUG or >>> something. >>> >>>> It seems dev_to_node() does not do anything about that and leave the >>>> job to the caller or whatever function that get called with its return >>>> value, such as cpumask_of_node(). >>> >>> Well, dev_to_node() doesn't do anything; nor should it. It are the >>> callers of set_dev_node() that should be taking care. >>> >>> Also note how device_add() sets the device node to the parent device's >>> node on NUMA_NO_NODE. Arguably we should change it to complain when it >>> finds NUMA_NO_NODE and !parent. >> >> Is it possible that the node id set by device_add() become invalid >> if the node is offlined, then dev_to_node() may return a invalid >> node id. > > In that case I would expect the device to go away too. Once the memory > controller goes away, the PCI bus connected to it cannot continue to > function. Ok. To summarize the discussion in order to for me to understand it correctly: 1) Make sure device_add() set to default node0 to a device if ACPI/BIOS does not set the node id and it has not no parent device. 2) Use '(unsigned)node_id >= nr_node_ids' to fix the CONFIG_DEBUG_PER_CPU_MAPS version of cpumask_of_node() for x86 and arm64, x86 just has a fix from you now, a patch for arm64 is also needed. 3) Maybe fix some other the sign bug for node id checking through the kernel using the '(unsigned)node_id >= nr_node_ids'. Please see if I understand it correctly or miss something. Maybe I can begin by sending a patch about item one to see if everyone is ok with the idea? > >> From the comment in select_fallback_rq(), it seems that a node can >> be offlined, not sure if node offline process has taken cared of that? >> >> /* >> * If the node that the CPU is on has been offlined, cpu_to_node() >> * will return -1. There is no CPU on the node, and we should >> * select the CPU on the other node. >> */ > > Ugh, so I disagree with that notion. cpu_to_node() mapping should be > fixed, you simply cannot change it after boot, too much stuff relies on > it. > > Setting cpu_to_node to -1 on node offline is just wrong. But alas, it > seems this is already so.
