Hi Mike,

On Wed, Jun 19, 2019 at 4:18 PM Mike Rapoport <[email protected]> wrote:
> On Wed, Jun 19, 2019 at 09:39:40AM +0200, Geert Uytterhoeven wrote:
> > On Wed, Jun 19, 2019 at 9:06 AM Geert Uytterhoeven <[email protected]> 
> > wrote:
> > > On Tue, Jun 18, 2019 at 8:10 AM Mike Rapoport <[email protected]> wrote:
> > > > On Mon, Jun 17, 2019 at 10:00:32PM +0200, Geert Uytterhoeven wrote:
> > > > > On Thu, May 16, 2019 at 8:03 AM Mike Rapoport <[email protected]> 
> > > > > wrote:
> > > > > > These pacthes replace DISCONTIGMEM with SPARSEMEM on m68k for 
> > > > > > systems with
> > > > > > !SINGLE_MEMORY_CHUNK set.
> > > > > >
> > > > > > With SPARSEMEM there is a single node for the entire physical 
> > > > > > memory and to
> > > > > > cope with holes in the physical address space it is divided to 
> > > > > > sections of
> > > > > > up to 16M.
> > > > > >
> > > > > > Each section has it's own memory map which size depends on actual 
> > > > > > populated
> > > > > > memory.
> > > > > >
> > > > > > The section size of 16M was chosen pretty much arbitrarily as I 
> > > > > > couldn't
> > > > > > find specs for systems with e.g. Zorro memory extensions.
> > >
> > > > > Unfortunately it crashes on my Amiga, cfr. the logs below.
> > > > >
> > > > > Then I realized the "section size of 16M". My Amiga has a single block
> > > > > of 12 MiB of FastRAM at 0x07400000, which is not aligned to 16 MiB.
> > > > > (Yes, base address of motherboard RAM is 0x08000000 - ramsize ;-)
> > > > >
> > > > > I've tried:
> > > > >
> > > > > -#define MAX_PHYSMEM_BITS       32
> > > > > -#define SECTION_SIZE_BITS      24
> > > > > +#define MAX_PHYSMEM_BITS       30
> > > > > +#define SECTION_SIZE_BITS      22
> > > > >
> > > > > but that doesn't seem to make a difference.
> > > > >
> > > > > Do you have a clue? Thanks!
> > > >
> > > > Not really, at least yet.
> > > > Can you please send the entire log with
> > > >
> > > > "mminit_loglevel=4 memblock=debug debug"
> > > >
> > > > in the command line?
> > >
> > > Attached, with debug_boot_weak_hash added, which reveals it's a real
> > > NULL (=0x0) pointer dereference.
> > >
> > > Looking at the disassembly, it happens in clear_page().
> > > Call chain:
> > >
> > >     get_page_from_freelist()
> > >         prep_new_page()
> > >             clear_highpage()
> > >                 void *kaddr = kmap_atomic(page);
> > >                     clear_page(kaddr);
> > >
> > > get_page_from_freelist() verifies page is non-zero before calling
> > > prep_new_page(), so it must be the kmap_atomic() that returns NULL.
> > >
> > > kmap_atomic() basically does page_address(page).
> > > As m68k defines WANT_PAGE_VIRTUAL, that evaluates to page->virtual,
> > > which I assume to be NULL.
> > > Is there a call to set_page_address() missing in the sparsemem code?
> > >
> > > Questions:
> > >   1. Why does it work on Atari/ARAnyM?
> > >   2. Why does it work on SPARC64, which also uses WANT_PAGE_VIRTUAL
> > >      and has SPARSEMEM support? (arc uses WANT_PAGE_VIRTUAL, but no
> > >      SPARSEMEM)
> >
> > Oh, I have CONFIG_SINGLE_MEMORY_CHUNK=y, so SPARSEMEM is
> > _not_ used.
>
> Well, that means I've managed to break FLATMEM :)

Indeed ^-(

> My guess would be that pfn_to_page()/page_to_pfn() are wrong with memory
> starting at !0 address.
>
> Can you please try the patch below to see if this helps with your machine?
> If yes, I'll think of a proper way to fix the pfn <-> page conversion.
>
> diff --git a/arch/m68k/include/asm/page_mm.h b/arch/m68k/include/asm/page_mm.h
> index 2fa176b1e583..25e300413e0f 100644
> --- a/arch/m68k/include/asm/page_mm.h
> +++ b/arch/m68k/include/asm/page_mm.h
> @@ -167,6 +167,7 @@ static inline __attribute_const__ int 
> __virt_to_node_shift(void)
>         ((__p) - pgdat->node_mem_map) + pgdat->node_start_pfn;          \
>  })
>  #else
> +#define ARCH_PFN_OFFSET (0x07400000 >> PAGE_SHIFT)
>  #include <asm-generic/memory_model.h>
>  #endif

Thanks, that hack did fix CONFIG_SINGLE_MEMORY_CHUNK=y.

Back to sparsemem...

With CONFIG_SINGLE_MEMORY_CHUNK=n, and CONFIG_SPARSEMEM=y,
it also fails.  Diff between working single memory chunk and failing
sparsemem:

   -On node 0 totalpages: 3072
   -  DMA zone: 27 pages used for memmap
   +On node 0 totalpages: 32768
   +  DMA zone: 288 pages used for memmap
      DMA zone: 0 pages reserved
   -  DMA zone: 3072 pages, LIFO batch:0
   +  DMA zone: 32768 pages, LIFO batch:7
    initrd: 00b61166 - 00c00000
    pcpu-alloc: s0 r0 d32768 u32768 alloc=1*32768
    pcpu-alloc: [0] 0
   -Built 1 zonelists, mobility grouping off.  Total pages: 3045
   +Built 1 zonelists, mobility grouping on.  Total pages: 32480
    Kernel command line: console=ttyS0 debug=mem root=/dev/ram
   -Dentry cache hash table entries: 2048 (order: 1, 8192 bytes)
   -Inode-cache hash table entries: 1024 (order: 0, 4096 bytes)
   +Dentry cache hash table entries: 16384 (order: 4, 65536 bytes)
   +Inode-cache hash table entries: 8192 (order: 3, 32768 bytes)
    Sorting __ex_table...
   -Memory: 7796K/12288K available (2555K kernel code, 259K rwdata,
700K rodata, 136K init, 153K bss, 4492K reserved, 0K cma-reserved)
   +Memory: 7816K/131072K available (2556K kernel code, 261K rwdata,
700K rodata, 136K init, 157K bss, 123256K reserved, 0K cma-reserved)

Oops, looks like it thinks there's memory from 0x00000000-0x08000000,
instead of 0x07400000-0x08000000.

And of course it crashes later:

    Trying to unpack rootfs image as initramfs...
    rootfs image is not initramfs (no cpio magic); looks like an initrd
    Unable to handle kernel access at virtual address (ptrval)

Full log with more debugging attached.

Thanks for your comments!

Gr{oetje,eeting}s,

                        Geert

--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- [email protected]

In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
                                -- Linus Torvalds

Attachment: dmesg-5.2.0-rc5-amiga-nomod-01404-g7e958d83f2ef4cd1.1
Description: Binary data

Reply via email to