[PATCH v3 2/2] of: reserved_mem: allocate and map the reserved_mem array early

[Marek Szyprowski]

Hi Geert,

On 03.11.2026 19:04, Geert Uytterhoeven wrote:
> On Thu, 15 Oct 2026 at 10:29, Marek Szyprowski <m.szyprowski at samsung.com> wrote:
>> Get rid of the static, limited-size reserved_mem array and replace it
>> with array allocated by memblock_alloc_raw() and accessed directly or
>> through a temporary early_memremap() mapping (depending on the
>> architecture support). Such mapping is needed for some architectures
>> (like ARM64), where linear map is not yet available during early boot
>> scan. The array is allocated from the bottom of RAM to avoid placing it
>> in highmem area, because some architectures (e.g. Xtensa, C-SKY, MIPS32)
>> set the memblock limit to the end of lowmem only after the early scan.
>> The early mapping is not used on PPC64, where the scan is done before
>> early_ioremap_setup(), but the memory allocated by memblock is already
>> accessible at that point.
>>
>> Having a single, writeable array with all reserved regions removes the
>> need to perform two step initialization introduced by commit
>> 8a6e02d0c00e ("of: reserved_mem: Restructure how the reserved memory
>> regions are processed"), so all regions can be processed during the
>> early scan again.
>>
>> The reserved_mem array is allocated only after all static (defined by
>> the 'reg' property) regions are marked as reserved or no-map, so it
>> cannot be placed in any of them. Errors of the early reservations are
>> not recorded per region, only the fact that some of them failed is
>> stored and a message asking to fix the device-tree is printed. The
>> region specific init functions for the static regions are then called
>> in a separate pass, once the array is available. Then all dynamically
>> allocated reserved regions are processed.
>>
>> Suggested-by: Mike Rapoport <rppt at kernel.org>
>> Link: https://lore.kernel.org/all/asc6brs_g0ky0wOM@kernel.org/
>> Signed-off-by: Marek Szyprowski <m.szyprowski at samsung.com>
> Thanks for your patch, which is now commit aa05faf815836c47 ("of:
> reserved_mem: allocate and map the reserved_mem array early")
> in dt-rh/for-next.
>
> This causes an early crash on BeagleV Starlight Beta:
>
>  [...]
>  earlycon: uart0 MMIO32:0x0000000012440000 (options '115200n8')
>  printk: legacy bootconsole [uart0] enabled
>  printk: debug: skip boot console de-registration.
>  efi: UEFI not found.
> -OF: reserved mem: 0x00000000fb000000..0x00000000fcffffff (32768 KiB)
> map non-reusable framebuffer at fb000000
> -OF: reserved mem: 0x00000000f9000000..0x00000000f9ffffff (16384 KiB)
> map non-reusable framebuffer at f9000000
> -OF: reserved mem: 0x00000000d0000000..0x00000000f7ffffff (655360 KiB)
> map non-reusable framebuffer at d0000000
> -OF: reserved mem: 0x00000000c9000000..0x00000000ccffffff (65536 KiB)
> map non-reusable framebuffer at c9000000
> -OF: reserved mem: 0x00000000fa000000..0x00000000faffffff (16384 KiB)
> nomap non-reusable dma-reserved at fa000000
> -Reserved memory: created DMA memory pool at 0x000000107a000000, size 16384 KiB
> -OF: reserved mem: initialized node linux,dma at 107a000000, compatible
> id shared-dma-pool
> -OF: reserved mem: 0x000000107a000000..0x000000107affffff (16384 KiB)
> nomap non-reusable linux,dma at 107a000000
> -SBI HSM extension detected
> -[...]
> +Oops - store (or AMO) access fault [#1]
> +Modules linked in:
> +CPU: 0 UID: 0 PID: 0 Comm: swapper Not tainted
> 7.4.0-rc1-starlight-00002-gaa05faf81583 #411 PREEMPT
> +Hardware name: BeagleV Starlight Beta (DT)
> +epc : __memset+0x68/0xfe
> + ra : fdt_scan_reserved_mem+0x3a4/0x80a
> +epc : ffffffff8087fd88 ra : ffffffff80a313c8 sp : ffffffff81403c20
> + gp : ffffffff814dc6c0 tp : ffffffff8140cf80 t0 : ffffffc4fea38ff0
> + t1 : 0000000080200000 t2 : ffffffc4fec03008 s0 : ffffffff81403d10
> + s1 : ffffffffffffffff a0 : ffffffc4fea39000 a1 : 0000000000000000
> + a2 : 00000000000000f0 a3 : ffffffc4fea390f0 a4 : 0000000000000008
> + a5 : ffffffff8087fd88 a6 : 00000000000000f0 a7 : 0000000000001000
> + s2 : ffffffc4fec00000 s3 : 00000000000000f0 s4 : 0000000000000000
> + s5 : 0000000000002de0 s6 : 0000000000000006 s7 : fffffffffefefefe
> + s8 : ffffffc4fec02fd4 s9 : ffffffff8148e318 s10: fffffffffefefeff
> + s11: ffffffc4fea39000 t3 : ffffffff8150df00 t4 : ffffffff8150df00
> + t5 : ffffffff8150df00 t6 : ffffffff8150df78 ssp : 0000000000000000
> +status: 0000000200000100 badaddr: ffffffc4fea39000 cause: 0000000000000007
> +[<ffffffff8087fd88>] __memset+0x68/0xfe
> +[<ffffffff80a2f63a>] early_init_fdt_scan_reserved_mem+0xa4/0x110
> +[<ffffffff80a06c8e>] paging_init+0x17c/0x46a
> +[<ffffffff80a046dc>] setup_arch+0xdc/0x584
> +[<ffffffff80a009b8>] start_kernel+0x98/0x940
> +Code: 0000 8793 00e7 8305 97ba 8782 b023 00b2 b423 00b2 (b823) 00b2
> +---[ end trace 0000000000000000 ]---
> +Kernel panic - not syncing: Fatal exception in interrupt
> +---[ end Kernel panic - not syncing: Fatal exception in interrupt ]---
>
> Reverting the commit fixes the issue.

Does moving early_init_fdt_scan_reserved_mem() to the end of setup_bootmem()
in arch/riscv/mm/init.c change anything?

Could You provide a full boot log from the working kernel and /proc/iomem dump?


Best regards
-- 
Marek Szyprowski, PhD
Samsung R&D Institute Poland