[Crash-utility] ARM support for CONFIG_SPARSEMEM:(was Re:DDimage)

[Dave Anderson]

----- Original Message -----
> Dave, Mika,
> memory.h I am using is probably the same as this,
> https://www.codeaurora.org/gitweb/quic/la/?p=kernel/msm.git;a=blob;f=arch/arm/mach-msm/include/mach/memory.h;h=afe6051b9826d788849aafa52b55b3760b6cf1ec;hb=debd216b762a87da0faf0eb4f991ed10c7a052cc
> 
> 19 /* physical offset of RAM */
> 20 #define PHYS_OFFSET UL(CONFIG_PHYS_OFFSET) <-- 0x00200000
> 21
> 22 #define MAX_PHYSMEM_BITS 32
> 23 #define SECTION_SIZE_BITS 28
> 24
> 25 /* Certain configurations of MSM7x30 have multiple memory banks.
> 26 * One or more of these banks can contain holes in the memory map as well.
> 27 * These macros define appropriate conversion routines between the physical
> 28 * and virtual address domains for supporting these configurations using
> 29 * SPARSEMEM and a 3G/1G VM split.
> 30 */
> 31
> 32 #if defined(CONFIG_ARCH_MSM7X30)
> 33
> 34 #define EBI0_PHYS_OFFSET PHYS_OFFSET
> 35 #define EBI0_PAGE_OFFSET PAGE_OFFSET
> 36 #define EBI0_SIZE 0x10000000
> 37
> 38 #define EBI1_PHYS_OFFSET 0x40000000
> 39 #define EBI1_PAGE_OFFSET (EBI0_PAGE_OFFSET + EBI0_SIZE)
> 40
> 41 #if (defined(CONFIG_SPARSEMEM) && defined(CONFIG_VMSPLIT_3G))
> 42
> 43 #define __phys_to_virt(phys) \
> 44 ((phys) >= EBI1_PHYS_OFFSET ? \
> 45 (phys) - EBI1_PHYS_OFFSET + EBI1_PAGE_OFFSET : \
> 46 (phys) - EBI0_PHYS_OFFSET + EBI0_PAGE_OFFSET)
> 47
> 48 #define __virt_to_phys(virt) \
> 49 ((virt) >= EBI1_PAGE_OFFSET ? \
> 50 (virt) - EBI1_PAGE_OFFSET + EBI1_PHYS_OFFSET : \
> 51 (virt) - EBI0_PAGE_OFFSET + EBI0_PHYS_OFFSET)
> 52
> 53 #endif
> 54
> 55 #endif
> 
> 
> So my previous description of actual memory mapping was not correct.
> The correct description is,
> Virtual 0xc0000000-0xcfdfffff -> Physical 0x00200000-0x0fffffff
> and
> Virtual 0xd0000000-0xdfffffff -> Physical 0x40000000-0x4fffffff
> Tomorrow I will check if vmcore with the following header is
> recognized by crash.
> Type Offset VirtAddr PhysAddr FileSiz MemSiz Flg Align
> NOTE 0x000094 0x00000000 0x00000000 0x00000 0x00000 0
> LOAD 0x000094 0xc0000000 0x00200000 0x0fe00000 0x0fe00000 RWE
> LOAD 0x0fe00094 0xc0000000 0x40000000 0x10000000 0x10000000 RWE

Shouldn't the second PT_LOAD segment have a p_vaddr of 0xc0000000?

Although, in the case of ARM, I believe that the p_vaddr field
is not used by the crash utility, as it is only interested in the
p_paddr and p_memsz/p_filesz fields.  But I presume you meant 
0xd0000000 for the second one.

> 
> And I guess VTOP/PTOV needs modification in accordance with
> __phys_to_virt and __virt_to_phys.

Right...

Ultimately it will be advisable to extract the ARM VTOP() and PTOV()
macros into machine-dependent functions, i.e., similar to the X86_64
and IA64 architectures.  And in those functions, intelligence will
have to be applied to determine how to handle the various ARM types.

> 
> For your information,
> The vmcore file with the following header is recognized by crash and
> many commands works fine,
> Type Offset VirtAddr PhysAddr FileSiz MemSiz Flg Align
> NOTE 0x000074 0x00000000 0x00000000 0x00000 0x00000 0
> LOAD 0x000074 0xc0000000 0x00000000 0x20000000 0x20000000 RWE
> if the patches for unwind_arm.c, arm.c and defs.h posted in this ML
> thread applied and readmem error_handle for FILL_PTBL is change to
> RETURN_ON_ERROR.
> Without the last modification above crash exits when readmem fails at
> FILL_PTBL before reaching the first prompt.

But I'm thinking that if the VTOP/PTOV issue is resolved, then you won't
see that readmem() error in FILL_PTBL(), because the address that was
failing was generated by PTOV():

        /*
         * pte_offset_map(pmd, vaddr)
         */
        page_table = (ulong *)PTOV(pmd_page_addr(pmd_pte)) + PTE_OFFSET(vaddr);

        FILL_PTBL(PAGEBASE(page_table), KVADDR, PAGESIZE());

Dave