Documentation for kdump - the kexec-based crash dumping solution ================================================================ DESIGN ====== Kdump uses kexec to reboot to a second kernel whenever a dump needs to be taken. This second kernel is booted with very little memory. The first kernel reserves the section of memory that the second kernel uses. This ensures that on-going DMA from the first kernel does not corrupt the second kernel. All the necessary information about Core image is encoded in ELF format and stored in reserved area of memory before crash. Physical address of start of ELF header is passed to new kernel through command line parameter elfcorehdr=. On i386, the first 640 KB of physical memory is needed to boot, irrespective of where the kernel loads. Hence, this region is backed up by kexec just before rebooting into the new kernel. In the second kernel, "old memory" can be accessed in two ways. - The first one is through a /dev/oldmem device interface. A capture utility can read the device file and write out the memory in raw format. This is raw dump of memory and analysis/capture tool should be intelligent enough to determine where to look for the right information. ELF headers (elfcorehdr=) can become handy here. - The second interface is through /proc/vmcore. This exports the dump as an ELF format file which can be written out using any file copy command (cp, scp, etc). Further, gdb can be used to perform limited debugging on the dump file. This method ensures methods ensure that there is correct ordering of the dump pages (corresponding to the first 640 KB that has been relocated). SETUP ===== 1) Download the upstream kexec-tools userspace package from http://www.xmission.com/~ebiederm/files/kexec/kexec-tools-1.101.tar.gz. Apply the latest consolidated kdump patch on top of kexec-tools-1.101 from http://lse.sourceforge.net/kdump/. This arrangment has been made till all the userspace patches supporting kdump are integrated with upstream kexec-tools userspace. 2) Download and build the appropriate (2.6.13-rc1 onwards) vanilla kernels. Two kernels need to be built in order to get this feature working. Following are the steps to properly configure the two kernels specific to kexec and kdump features: A) First kernel or regular kernel: ---------------------------------- a) Enable "kexec system call" feature (in Processor type and features). CONFIG_KEXEC=y b) Enable "sysfs file system support" (in Pseudo filesystems). CONFIG_SYSFS=y c) make d) Boot into first kernel with the command line parameter "crashkernel=Y@X". Use appropriate values for X and Y. Y denotes how much memory to reserve for the second kernel, and X denotes at what physical address the reserved memory section starts. For example: "crashkernel=64M@16M". B) Second kernel or dump capture kernel: --------------------------------------- a) For i386 architecture enable Highmem support CONFIG_HIGHMEM=y b) Enable "kernel crash dumps" feature (under "Processor type and features") CONFIG_CRASH_DUMP=y c) Make sure a suitable value for "Physical address where the kernel is loaded" (under "Processor type and features"). By default this value is 0x1000000 (16MB) and it should be same as X (See option d above), e.g., 16 MB or 0x1000000. CONFIG_PHYSICAL_START=0x1000000 d) Enable "/proc/vmcore support" (Optional, under "Pseudo filesystems"). CONFIG_PROC_VMCORE=y 3) After booting to regular kernel or first kernel, load the second kernel using the following command: kexec -p <second-kernel> --args-linux --elf32-core-headers --append="root=<root-dev> init 1 irqpoll maxcpus=1" Notes: ====== i) <second-kernel> has to be a vmlinux image ie uncompressed elf image. bzImage will not work, as of now. ii) --args-linux has to be speicfied as if kexec it loading an elf image, it needs to know that the arguments supplied are of linux type. iii) By default ELF headers are stored in ELF64 format to support systems with more than 4GB memory. Option --elf32-core-headers forces generation of ELF32 headers. The reason for this option being, as of now gdb can not open vmcore file with ELF64 headers on a 32 bit systems. So ELF32 headers can be used if one has non-PAE systems and hence memory less than 4GB. iv) Specify "irqpoll" as command line parameter. This reduces driver initialization failures in second kernel due to shared interrupts. v) <root-dev> needs to be specified in a format corresponding to the root device name in the output of mount command. vi) If you have built the drivers required to mount root file system as modules in <second-kernel>, then, specify --initrd=<initrd-for-second-kernel>. vii) Specify maxcpus=1 as, if during first kernel run, if panic happens on non-boot cpus, second kernel doesn't seem to be boot up all the cpus. The other option is to always built the second kernel without SMP support ie CONFIG_SMP=n 4) After successfully loading the second kernel as above, if a panic occurs system reboots into the second kernel. A module can be written to force the panic or "ALT-SysRq-c" can be used initiate a crash dump for testing purposes. 5) Once the second kernel has booted, write out the dump file using cp /proc/vmcore <dump-file> Dump memory can also be accessed as a /dev/oldmem device for a linear/raw view. To create the device, type: mknod /dev/oldmem c 1 12 Use "dd" with suitable options for count, bs and skip to access specific portions of the dump. Entire memory: dd if=/dev/oldmem of=oldmem.001 ANALYSIS ======== Limited analysis can be done using gdb on the dump file copied out of /proc/vmcore. Use vmlinux built with -g and run gdb vmlinux <dump-file> Stack trace for the task on processor 0, register display, memory display work fine. Note: gdb cannot analyse core files generated in ELF64 format for i386. Latest "crash" (crash-4.0-2.18) as available on Dave Anderson's site http://people.redhat.com/~anderson/ works well with kdump format. TODO ==== 1) Provide a kernel pages filtering mechanism so that core file size is not insane on systems having huge memory banks. 2) Relocatable kernel can help in maintaining multiple kernels for crashdump and same kernel as the first kernel can be used to capture the dump. CONTACT ======= Vivek Goyal (vgoyal@in.ibm.com) Maneesh Soni (maneesh@in.ibm.com)