Re: [PATCH] memblock: exclude NOMAP regions from kmemleak

[Catalin Marinas]

On Wed, Oct 20, 2021 at 10:38:23AM +0300, Mike Rapoport wrote:
> On Tue, Oct 19, 2021 at 09:33:11PM +0300, Mike Rapoport wrote:
> > On Tue, Oct 19, 2021 at 01:59:22PM -0400, Qian Cai wrote:
> > > [	0.000000][	T0] Booting Linux on physical CPU 0x0000000000 [0x503f0002]
> > > [	0.000000][	T0] Linux version 5.15.0-rc6-next-20211019+ (root@admin5) (gcc (Ubuntu 9.3.0-17ubuntu1~20.04) 9.3.0, GNU ld (GNU Binutils for Ubuntu) 2.34) #104 SMP Tue Oct 19 17:36:17 UTC 2021
> > > [	0.000000][	T0] earlycon: pl11 at MMIO32 0x0000000012600000 (options '')
> > > [	0.000000][	T0] printk: bootconsole [pl11] enabled
> > > [	0.000000][	T0] efi: Getting UEFI parameters from /chosen in DT:
> > > [	0.000000][	T0] efi:   System Table     	: 0x0000009ff7de0018
> > > [	0.000000][	T0] efi:   MemMap Address   	: 0x0000009fe6dae018
> > > [	0.000000][	T0] efi:   MemMap Size      	: 0x0000000000000600
> > > [	0.000000][	T0] efi:   MemMap Desc. Size	: 0x0000000000000030
> > > [	0.000000][	T0] efi:   MemMap Desc. Version : 0x0000000000000001
> > > [	0.000000][	T0] efi: EFI v2.70 by American Megatrends
> > > [	0.000000][	T0] efi: ACPI 2.0=0x9ff5b40000 SMBIOS 3.0=0x9ff686fd98 ESRT=0x9ff1d18298 MEMRESERVE=0x9fe6dacd98  
> > > [	0.000000][	T0] efi: Processing EFI memory map:
> > > [	0.000000][	T0] efi:   0x000090000000-0x000091ffffff [Conventional|   |  |  |  |  |  |  |  |  |   |WB|WT|WC|UC]
> > > [	0.000000][	T0] efi:   0x000092000000-0x0000928fffff [Runtime Data|RUN|  |  |  |  |  |  |  |  |   |WB|WT|WC|UC]
> > > [	0.000000][	T0] ------------[ cut here ]------------
> > > [	0.000000][	T0] kernel BUG at mm/kmemleak.c:1140!
> > > [	0.000000][	T0] Internal error: Oops - BUG: 0 [#1] SMP
> > > 
> > > I did not quite figure out where this BUG() was triggered and I did not
> > 
> > This is from here:
> > arch/arm64/include/asm/memory.h:
> > 
> > #define PHYS_OFFSET         ({ VM_BUG_ON(memstart_addr & 1); memstart_addr; })
> > 
> > kmemleak_free_part_phys() does __va() which uses PHYS_OFFSET and all this
> > happens before memstart_addr is set.
> > 
> > I'll try to see how this can be untangled...
>  
> This late in the cycle I can only think of reverting kmemleak wavier from
> memblock_mark_nomap() and putting it in
> early_init_dt_alloc_reserved_memory_arch() being the only user setting
> MEMBLOCK_NOMAP to an allocated chunk rather than marking NOMAP "unusable"
> memory reported by firmware.

It makes sense, there aren't many places or nomap is called.

I think arch_reserve_mem_area() called from acpi_table_upgrade() also
follows a memblock allocation. But I'd call kmemleak in
acpi_table_upgrade() directly rather than in the arch back-end.

Regarding which callback, I think kmemleak_ignore_phys() is better
suited here since kmemleak still keeps track of the object but won't
scan it.

-- 
Catalin