On 2014/5/29 12:39, AKASHI Takahiro wrote: > Wang, Will > > I'm now working on kdump support for arm64 on top of Geoff's kexec patch. > > On 05/20/2014 12:22 PM, Wang Nan wrote: >> On 2014/5/20 0:09, Will Deacon wrote: >>> On Mon, May 19, 2014 at 02:54:03AM +0100, Wang Nan wrote: >>>> When SPARSEMEM and CRASH_DUMP both selected, simple pfn_valid prevents >>>> the second kernel ioremap first kernel's memory if the address falls >>>> into second kernel section. This limitation requires the second kernel >>>> occupies a full section, and elfcorehdr must resides in another section. >>>> >>>> This patch makes crash dump kernel use strict pfn_valid, removes such >>>> limitation. >>>> >>>> For example: >>>> >>>> For a platform with SECTION_SIZE_BITS == 28 (256MiB) and >>>> crashkernel=128M at 0x28000000 in kernel cmdline, the second >>>> kernel is loaded at 0x28000000. Kexec puts elfcorehdr at >>>> 0x2ff00000, and passes 'elfcorehdr=0x2ff00000 mem=130048K' to >>>> second kernel. When second kernel start, it tries to use >>>> ioremap to retrive its elfcorehrd. In this case, elfcodehdr is at the >>>> same section of the second kernel, pfn_valid will recongnize >>>> the page as valid, so ioremap will refuse to map it. >>> >>> So isn't the issue here that you're passing an incorrect mem= parameter >>> to the crash kernel? >>> >> >> mem= parameter is generated by kexec-tools according to /proc/iomem, it is the size >> of reserved memory minus 1MiB. So I think what you mean is I passing an incorrect >> crashkernel= parameter? > > Just FYI, kexec-tools doesn't seem to be implemented in proper way to support device-tree. > Once device-tree is handled correctly, we don't need to pass "mem=" parameter. > (Of course, only on machines that support device-tree.) > >> I'll explain limitations on crash kernel reserved memory in the case of SPARSEMEM >> enabled, and show how *impractical* the 'correct' crashkernel will be. >> >> Use realview board for example. >> >> Limitation 1: crash kernel reservation kernel must be aligned with 0x08000000 (128MiB). >> >> This is because zImage determine final kernel address by (pc & 0xf8000000). If, >> for example, set crashkernel=64M at 0x29000000, then the second kernel itself >> overwrites first kernel's memory. We'll lost some memory in /proc/vmcore. >> >> Limitation 2: crash kernel must resides in different section with the first kernel. >> >> This is because the second kernel use ioremap for accessing first kernel's memory, >> and arm prevent a valid pfn be ioremapped. Which means a whole section must be reserved >> for the secton kernel. On realview, which is 256MiB. >> >> Limitation 3: the last 1MiB of reserved memory must be ioremappable. >> >> This is because the second kernel depeneds kexec-tools passing an elfheader as >> 'elfcorehdr' to instructs it generating /proc/vmcore. See fs/proc/vmcore.c. Kexec-tools >> simply uses the last 1MiB for it. The second kernel use ioremap to access it, force >> the header be put in another section. > > We can avoid "Limitation 3" just by implementing arm's own elfcorehdr_read() with memcpy(). > I can submit a patch, but can't test it for now. > > -Takahiro AKASHI > However you still need pfn_valid to check whether elfcorehdr resides in a valid area. Furthermore, simply replacing ioremap to memcpy seems breaks things. Configurations work before replacement will fail. Finally you will find you still need strict pfn_valid to check whether to use ioremap or use memcpy. > >> In realview board, the only possible correct setting should be 'crashkernel=257M at 0x20000000'. >> However, realview has only 1GiB memory, crash kernel consumes a quarter plus 1MiB. In addition, even >> set this parameter, crash kernel is still unusable because: >> >> crashkernel reservation failed - memory is in use (0x20000000) >> >>> Will >>> >> >> >> >> _______________________________________________ >> linux-arm-kernel mailing list >> linux-arm-kernel at lists.infradead.org >> http://lists.infradead.org/mailman/listinfo/linux-arm-kernel >>
