On Fri, Aug 19, 2016 at 03:44:23PM +0530, Pratyush Anand wrote: > On 19/08/2016:04:19:31 PM, AKASHI Takahiro wrote: > > Pratyush, Geoff > > > > # If I had Seattle, I could debug easily :) > > > > On Wed, Aug 10, 2016 at 11:26:48PM +0530, Pratyush Anand wrote: > > > Hi Geoff and Takahiro, > > > > > > I am having some issues with kexec+kdump while working with Seattle platform. On > > > top level, kernel crashes in copy_oldmem_page(), because it gets wrong offset > > > for log_buf during vmcore-dmesg save. > > > > > > Here is the detail: > > > > > > (1) From /proc/iomem, these are the "System RAM" Components: > > > > > > 8000000000-8001e7ffff : System RAM > > > 8001e80000-83ff17ffff : System RAM > > > 8002080000-8002b3ffff : Kernel code > > > 8002c40000-800348ffff : Kernel data > > > 807fe00000-80ffdfffff : Crash kernel > > > 83ff180000-83ff1cffff : System RAM > > > 83ff1d0000-83ff21ffff : System RAM > > > 83ff220000-83ffe4ffff : System RAM > > > 83ffe50000-83ffffffff : System RAM > > > > > > (2) From kexec-tools debug print I see following: > > > elf_arm64_load: e_entry: fffffc0008080000 -> 0000008000080000 > > > elf_arm64_load: p_vaddr: fffffc0008080000 -> 0000008000080000 > > > elf_arm64_load: header_offset: 0000000000000000 > > > elf_arm64_load: text_offset: 0000000000080000 > > > elf_arm64_load: image_size: 0000000001410000 > > > elf_arm64_load: phys_offset: 0000008000000000 > > > elf_arm64_load: page_offset: fffffc0008000000 > > > > > > I understand that "Kernel Code start physical address" 0x8002080000 should map > > > to e_entry vaddr which is 0xfffffc0008080000. However, kexec-tools debug print > > > shows that e_entry vaddr maps to PA 8000080000 which seems wrong. > > > > Obviously, virt_to_phys() is wrong. > > As you know, we have to calculate virt_to_phys() in a different way > > depending on whether the vaddr is greater than PAGE_OFFSET (linear mapping) > > or not (kernel image mapping). > > See the kernel's include/asm/memory.h. > > > > Geoff, please update this function. > > > > > (3) further page_offset (or vp_offset in your new code) is calculated > > > as:arm64_mem.page_offset = ehdr.e_entry - arm64_mem.text_offset; > > > > > > Current calcualtion of page_offset leads to wrong configuration of VA of alls > > > PT_LOAD (see below). Ultimately, this is also leading to kernel crash during > > > vmcore-dmesg and vmcore save operations, because we pass an offset to pread() > > > system call which maps to wrong physical address. > > > > > > Elf header: p_type = 1, p_offset = 0x8000000000 p_paddr = 0x8000000000 > > > p_vaddr = 0xfffffc0008000000 p_filesz = 0x1e80000 p_memsz = 0x1e80000 > > > [0xfffffc0008000000 should be mapping to 0x8002000000 and not 0x8000000000] > > > > I think that your comment here is wrong. > > This program header indicates the first memory region in the 1st kernel, > > that is, > > 8000000000-8001e7ffff : System RAM > > > > Is this region really part of "System RAM?" > > Can you show me the "Virtual kernel memory layout" in dmesg? > > [ 0.000000] Virtual kernel memory layout: > [ 0.000000] modules : 0xfffffc0000000000 - 0xfffffc0008000000 ( 128 MB) > [ 0.000000] vmalloc : 0xfffffc0008000000 - 0xfffffdff5fff0000 ( 2045 GB) > [ 0.000000] .text : 0xfffffc0008080000 - 0xfffffc00087d0000 ( 7488 KB) > [ 0.000000] .rodata : 0xfffffc00087d0000 - 0xfffffc0008b40000 ( 3520 KB) > [ 0.000000] .init : 0xfffffc0008b40000 - 0xfffffc0008c40000 ( 1024 KB) > [ 0.000000] .data : 0xfffffc0008c40000 - 0xfffffc0008d93e00 ( 1360 KB) > [ 0.000000] .bss : 0xfffffc0008d93e00 - 0xfffffc0009438248 ( 6802 KB) > [ 0.000000] fixed : 0xfffffdff7e7d0000 - 0xfffffdff7ec00000 ( 4288 KB) > [ 0.000000] PCI I/O : 0xfffffdff7ee00000 - 0xfffffdff7fe00000 ( 16 MB) > [ 0.000000] vmemmap : 0xfffffdff80000000 - 0xfffffe0000000000 ( 2 GB maximum) > [ 0.000000] 0xfffffdff80000000 - 0xfffffdff81000000 ( 16 MB actual) > [ 0.000000] memory : 0xfffffe0000000000 - 0xfffffe0400000000 ( 16384 MB) > > > > > > Elf header: p_type = 1, p_offset = 0x8001e80000 p_paddr = 0x8001e80000 > > > p_vaddr = 0xfffffc0009e80000 p_filesz = 0x7df80000 p_memsz = 0x7df80000 > > > Elf header: p_type = 1, p_offset = 0x80ffe00000 p_paddr = 0x80ffe00000 > > > p_vaddr = 0xfffffc0107e00000 p_filesz = 0x2ff380000 p_memsz = 0x2ff380000 > > > Elf header: p_type = 1, p_offset = 0x83ff180000 p_paddr = 0x83ff180000 > > > p_vaddr = 0xfffffc0407180000 p_filesz = 0x50000 p_memsz = 0x50000 > > > Elf header: p_type = 1, p_offset = 0x83ff1d0000 p_paddr = 0x83ff1d0000 > > > p_vaddr = 0xfffffc04071d0000 p_filesz = 0x50000 p_memsz = 0x50000 > > > Elf header: p_type = 1, p_offset = 0x83ff220000 p_paddr = 0x83ff220000 > > > p_vaddr = 0xfffffc0407220000 p_filesz = 0xc30000 p_memsz = 0xc30000 > > > Elf header: p_type = 1, p_offset = 0x83ffe50000 p_paddr = 0x83ffe50000 > > > p_vaddr = 0xfffffc0407e50000 p_filesz = 0x1b0000 p_memsz = 0x1b0000 > > > > > > May be following should be better. > > > arm64_mem.page_offset = ehdr.e_entry - "kernel Code Start PA" + phys_offset. > > > > Probably not. > > page_offset is the virtual address of phys_offset, right? > > ehdr.e_entry is the virtual address of "kernel Code Start PA", right? > > If yes, then why should n't above be correct for all linear regions. What I was thinking of is that we'd better fake arm64_mem.text_offset instead of calculating a page_offset because text_offset is also used in arm64_load_other_segments(). Here, if the first segment doesn't not contain kernel text, image_base will be incorrect. (It still works, though, now that the kernel can be loaded in any address.) > Other option could be what kernel is doing. But I am not sure, how can > kexec-tools get kimage_voffset. > > > > > > (4) Further more, vmcore must have first PT_LOAD segment as kernel text area. > > > > Which part of the code do you think depends on this assumption? > > It seems that there was no issue while analyzing with crash utility, even when > first PT_LOAD segment is not for kernel text area. > > However, from the comment in kexec/crashdump-elf.c: FUNC(), it seems that kernel > text is expected as first segment. > > 218 /* We already prepared the header for kernel text. Map > 219 * rest of the memory segments to kernel linearly mapped > 220 * memory region. > 221 */ Well, I guess, the point is not that the first segment has kernel text, but that there is one segment, in any order, which contains actual virtual address range of kernel text/data if the kernel is "relocatable," that is, KASLR-enabled. In this case, the vmcore file may refer to randomized addresses of symbols which don't match ones in vmlinux. So we will not be able to examine the core with, say, gdb. Crash utility, on the other hand, is *KASLR-aware*. I will try to support "kern_size" on kexec-tools for arm64. Thanks, -Takahiro AKASHI > ~Pratyush > > > > > Thanks, > > -Takahiro AKASHI > > > > > In this platform we have first "System RAM" area as 8000000000-8001e7ffff which > > > is not matching to "Kernel code" area. Therefore, we should provide support of > > > "kern_size" so that first PT_LOAD is kernel text area. > > > > > > ~Pratyush > > > On 09/08/2016:11:00:25 AM, AKASHI Takahiro wrote: > > > > My kernel patches of kdump suport on arm64 are currently under reviews [1]. > > > > > > > > This patchset is synced with them (v24) and provides necessary changes for > > > > kexec-tools. It should be applied on top of Geoff's kexec-tools patches > > > > v3[2] along with a bugfix[3]. > > > > > > > > [1] http://lists.infradead.org/pipermail/linux-arm-kernel/2016-August/447597.html > > > > [2] http://lists.infradead.org/pipermail/kexec/2016-August/016768.html > > > > [3] http://lists.infradead.org/pipermail/kexec/2016-July/016664.html > > > > > > > > Changes for v2: > > > > - Trim a temoprary buffer in setup_2nd_dtb() > > > > - Add patch#6("kexec: generalize and rename get_kernel_stext_sym()") > > > > - Update patch#7 from Pratyush > > > > (re-worked by akashi) > > > > > > > > AKASHI Takahiro (5): > > > > arm64: kdump: identify memory regions > > > > arm64: kdump: add elf core header segment > > > > arm64: kdump: set up kernel image segment > > > > arm64: kdump: set up other segments > > > > arm64: kdump: add DT properties to crash dump kernel's dtb > > > > > > > > Pratyush Anand (2): > > > > kexec: generalize and rename get_kernel_stext_sym() > > > > arm64: kdump: Add support for binary image files > > > > > > > > kexec/arch/arm/crashdump-arm.c | 40 +------ > > > > kexec/arch/arm64/Makefile | 2 + > > > > kexec/arch/arm64/crashdump-arm64.c | 188 +++++++++++++++++++++++++++++++- > > > > kexec/arch/arm64/crashdump-arm64.h | 18 ++- > > > > kexec/arch/arm64/include/arch/options.h | 8 +- > > > > kexec/arch/arm64/kexec-arm64.c | 91 ++++++++++++++-- > > > > kexec/arch/arm64/kexec-elf-arm64.c | 23 +++- > > > > kexec/arch/arm64/kexec-image-arm64.c | 60 +++++++++- > > > > kexec/arch/i386/crashdump-x86.c | 32 +----- > > > > kexec/crashdump.c | 37 +++++++ > > > > kexec/crashdump.h | 1 + > > > > 11 files changed, 407 insertions(+), 93 deletions(-) > > > > > > > > -- > > > > 2.9.0