On Wed, Sep 13, 2023 at 9:23 AM WANG Xuerui <[email protected]> wrote:
>
> On 9/13/23 08:49, Huacai Chen wrote:
> > On Wed, Sep 13, 2023 at 12:08 AM WANG Xuerui <[email protected]> wrote:
> >> On 9/11/23 17:28, Huacai Chen wrote:
> >>> After commit 61167ad5fecdea ("mm: pass nid to reserve_bootmem_region()")
> >>> we get a panic if DEFERRED_STRUCT_PAGE_INIT is enabled:
> >>>
> >>> [snip]
> >>>
> >>> The reason is early memblock_reserve() in memblock_init() set node id
> >> Why is it that only "early" but not "late" memblock_reserve() matters? I
> >> failed to see the reason because the arch-specific memblock_init() isn't
> >> even in the backtrace, which means that *neither* is the culprit.
> > Late memblock_reserve() operates on subregions of memblock.memory
> > regions. These reserved regions will be set to the correct node at the
> > first iteration of memmap_init_reserved_pages().
> Thanks for the clarification. According to the code behavior (and the
> comment I left on the reordering change below) I'm now sure the intended
> meaning is "calling memblock_reserve() after memblock_set_node() is
> effectively leaving those regions with nid=MAX_NUMNODES" (or something
> like that, pointing out that the memblock_set_node() call actually had
> no effect in this case). "Early" and "late" in the context of init code
> can be especially confusing IMO :-)
The "early call" specifically means these lines:
/* Reserve the first 2MB */
memblock_reserve(PHYS_OFFSET, 0x200000);
/* Reserve the kernel text/data/bss */
memblock_reserve(__pa_symbol(&_text),
__pa_symbol(&_end) - __pa_symbol(&_text));
These two regions can be out of initial memory maps. Other later
memblock_reserve() regions must be in initial memory maps, so will get
a correct node id.
Huacai
> >
> > Huacai
> >
> >>> to MAX_NUMNODES, which causes NODE_DATA(nid) be a NULL dereference in
> >> "making NODE_DATA(nid) a NULL ..."
> >>> reserve_bootmem_region() -> init_reserved_page(). So set all reserved
> >>> memblocks on Node#0 at initialization to avoid this panic.
> >>>
> >>> Reported-by: WANG Xuerui <[email protected]>
> >>> Signed-off-by: Huacai Chen <[email protected]>
> >>> ---
> >>> arch/loongarch/kernel/mem.c | 4 +++-
> >>> 1 file changed, 3 insertions(+), 1 deletion(-)
> >>>
> >>> diff --git a/arch/loongarch/kernel/mem.c b/arch/loongarch/kernel/mem.c
> >>> index 4a4107a6a965..aed901c57fb4 100644
> >>> --- a/arch/loongarch/kernel/mem.c
> >>> +++ b/arch/loongarch/kernel/mem.c
> >>> @@ -50,7 +50,6 @@ void __init memblock_init(void)
> >>> }
> >>>
> >>> memblock_set_current_limit(PFN_PHYS(max_low_pfn));
> >>> - memblock_set_node(0, PHYS_ADDR_MAX, &memblock.memory, 0);
> >>>
> >>> /* Reserve the first 2MB */
> >>> memblock_reserve(PHYS_OFFSET, 0x200000);
> >>> @@ -58,4 +57,7 @@ void __init memblock_init(void)
> >>> /* Reserve the kernel text/data/bss */
> >>> memblock_reserve(__pa_symbol(&_text),
> >>> __pa_symbol(&_end) - __pa_symbol(&_text));
> >>> +
> >>> + memblock_set_node(0, PHYS_ADDR_MAX, &memblock.memory, 0);
> >>> + memblock_set_node(0, PHYS_ADDR_MAX, &memblock.reserved, 0);
> >> So the reordering is for being able to override the newly added
> >> memblocks' nids to 0, and additionally doing the same for
> >> memblock.reserved is the actual fix. Looks okay.
> >>> }
> >> And I've tested the patch on the 2-way 3C5000L server, and it now
> >> correctly boots with deferred struct page init enabled. Thanks for
> >> providing such a quick fix!
> >>
> >> Tested-by: WANG Xuerui <[email protected]>
> >> Reviewed-by: WANG Xuerui <[email protected]> # with nits addressed
> >>
> >> --
> >> WANG "xen0n" Xuerui
> >>
> >> Linux/LoongArch mailing list: https://lore.kernel.org/loongarch/
> >>
> >>
> --
> WANG "xen0n" Xuerui
>
> Linux/LoongArch mailing list: https://lore.kernel.org/loongarch/
>