On 05/12/21 at 12:37pm, Mike Rapoport wrote: > On Wed, May 12, 2021 at 01:33:20PM +0800, Miles Chen wrote: > > On Tue, 2021-05-11 at 13:24 +0300, Mike Rapoport wrote: > > > On Tue, May 11, 2021 at 05:31:14PM +0800, Miles Chen wrote: > > > > In current implementation of node_data, if CONFIG_NEED_MULTIPLE_NODES=y, > > > > node_data is allocated by kzmalloc. If CONFIG_NEED_MULTIPLE_NODES=n, > > > > we use a global variable named "contig_page_data". > > > > > > > > If CONFIG_DEBUG_VIRTUAL is not enabled. __pa() can handle both kzalloc and > > > > symbol cases. But if CONFIG_DEBUG_VIRTUAL is set, we will have the > > > > "virt_to_phys used for non-linear address" warning when booting. > > > > > > Maybe we'll just allocate pgdat for CONFIG_NEED_MULTIPLE_NODES=n (which is > > > essentially !NUMA) case in, say, free_area_init()? > > > > > > thanks for your comment. > > > > I check the source tree and found that contig_page_data is used by > > crash_core.c as a symbol. I am not familiar with crash_core but I guess > > allocate pgdat may break this crash_core users. > > > > For example: some userspace scripts want to read the address of > > contig_page_data symbol from a corefile. > > > > kernel/crash_core.c:460: VMCOREINFO_SYMBOL(contig_page_data); > > > > #ifndef CONFIG_NEED_MULTIPLE_NODES > > VMCOREINFO_SYMBOL(mem_map); > > VMCOREINFO_SYMBOL(contig_page_data); > > #endif > > My understanding is that VMCOREINFO_SYMBOL() should correspond to actual > symbol. If there is no contig_page_data symbol, there is no need for > VMCOREINFO_SYMBOL() either. Yeah, it's exported for makedumpfile and crash utility to parse and get the memory layout of the corrupted kernel. If removing it, makedumpfile will get it from node_data[]. Looks like a good idea to unify code for numa|!numa on pglist_data instances. Add Kazu to CC since he maintain makedumpfile and Crash utilities. My concern is that that only happens on arm/arm64/riscv, does it mean the warning is not necessary, so can be removed? Or we need to check if CONFIG_DEBUG_VIRTUAL doesn't work well in this case. Thanks Baoquan
