Message ID | 1394115854-11709-1-git-send-email-nyushchenko@dev.rtsoft.ru (mailing list archive) |
---|---|
State | Changes Requested |
Headers | show |
On Thu, 2014-03-06 at 18:24 +0400, Nikita Yushchenko wrote: > When preparing dump-capturing kernel, kexec userspace tool needs to know > actual amount of memory used by the running kernel. This may differ from > extire available DRAM for a couple of reasons. To address this issue, > kdump kernel support code injects several attributes into device tree that > are later captured by userspace kexec tool via /proc interface. > > One such attrubute is 'chosen/linux,memory_limit' that is used to pass > memory limit of the running kernel. > > This was initialized using kernel's 'memory_limit' variable, that is set > by early init code based on mem= kernel parameter and other reasons. > > But there are cases when memory_limit variable does not contain proper > information. One such case is when !CONFIG_HIGHMEM kernel runs on system > with memory large enough not to fit into lowmem. Why doesn't the !CONFIG_HIGHMEM code update memory_limit to reflect reality. cheers
> On Thu, 2014-03-06 at 18:24 +0400, Nikita Yushchenko wrote: > > When preparing dump-capturing kernel, kexec userspace tool needs to > > know actual amount of memory used by the running kernel. This may > > differ from extire available DRAM for a couple of reasons. To address > > this issue, kdump kernel support code injects several attributes into > > device tree that are later captured by userspace kexec tool via /proc > > interface. > > > > One such attrubute is 'chosen/linux,memory_limit' that is used to pass > > memory limit of the running kernel. > > > > This was initialized using kernel's 'memory_limit' variable, that is > > set by early init code based on mem= kernel parameter and other > > reasons. > > > > But there are cases when memory_limit variable does not contain proper > > information. One such case is when !CONFIG_HIGHMEM kernel runs on > > system with memory large enough not to fit into lowmem. > > Why doesn't the !CONFIG_HIGHMEM code update memory_limit to reflect > reality. I guess because memory_limit is used for ... well, memory limit, set by mem=. And for the rest memblock is used (and it *is* updated). And code elsewhere does use memblock, see e.g. numa_enforce_memory_limit() in arch/powerpc/mm/numa.c In MMU init (MMU_init() in arch/powerpc/mm/init_32.c -which is the point where final memory configuration is set) memblock, not memory_limit, is both used and updated.
Le 07/03/2014 à 05:38, Nikita Yushchenko a écrit : >> On Thu, 2014-03-06 at 18:24 +0400, Nikita Yushchenko wrote: >>> When preparing dump-capturing kernel, kexec userspace tool needs to >>> know actual amount of memory used by the running kernel. This may >>> differ from extire available DRAM for a couple of reasons. To address >>> this issue, kdump kernel support code injects several attributes into >>> device tree that are later captured by userspace kexec tool via /proc >>> interface. >>> >>> One such attrubute is 'chosen/linux,memory_limit' that is used to pass >>> memory limit of the running kernel. >>> >>> This was initialized using kernel's 'memory_limit' variable, that is >>> set by early init code based on mem= kernel parameter and other >>> reasons. >>> >>> But there are cases when memory_limit variable does not contain proper >>> information. One such case is when !CONFIG_HIGHMEM kernel runs on >>> system with memory large enough not to fit into lowmem. >> >> Why doesn't the !CONFIG_HIGHMEM code update memory_limit to reflect >> reality. > > I guess because memory_limit is used for ... well, memory limit, set by > mem=. And for the rest memblock is used (and it *is* updated). > > And code elsewhere does use memblock, see e.g. numa_enforce_memory_limit() > in arch/powerpc/mm/numa.c > > In MMU init (MMU_init() in arch/powerpc/mm/init_32.c -which is the point > where final memory configuration is set) memblock, not memory_limit, is > both used and updated. We still have this patch as "New" in patchwork. I don't know if it is relevant but directory structure has changed so if still needed this patch needs rebase. Christophe
diff --git a/arch/powerpc/kernel/machine_kexec.c b/arch/powerpc/kernel/machine_kexec.c index 015ae55..372cda5 100644 --- a/arch/powerpc/kernel/machine_kexec.c +++ b/arch/powerpc/kernel/machine_kexec.c @@ -250,8 +250,14 @@ static void __init export_crashk_values(struct device_node *node) /* * memory_limit is required by the kexec-tools to limit the * crash regions to the actual memory used. + * + * There are cases when memory_limit variable does not hold actual + * limit, for example when memory was limited by no kernel support + * for HIGHMEM. Reliable information is known by memblock because + * memory management init adjusts it. */ - mem_limit = cpu_to_be_ulong(memory_limit); + mem_limit = cpu_to_be_ulong(memblock_end_of_DRAM() - + memblock_start_of_DRAM()); of_update_property(node, &memory_limit_prop); }
When preparing dump-capturing kernel, kexec userspace tool needs to know actual amount of memory used by the running kernel. This may differ from extire available DRAM for a couple of reasons. To address this issue, kdump kernel support code injects several attributes into device tree that are later captured by userspace kexec tool via /proc interface. One such attrubute is 'chosen/linux,memory_limit' that is used to pass memory limit of the running kernel. This was initialized using kernel's 'memory_limit' variable, that is set by early init code based on mem= kernel parameter and other reasons. But there are cases when memory_limit variable does not contain proper information. One such case is when !CONFIG_HIGHMEM kernel runs on system with memory large enough not to fit into lowmem. This patch fixes initialization of 'chosen/linux,memory_limit' to use values from memblock subsystem. These are adjusted at kernel memory management init and thus always contain values that match reality. Signed-off-by: Nikita Yushchenko <nyushchenko@dev.rtsoft.ru> --- arch/powerpc/kernel/machine_kexec.c | 8 +++++++- 1 file changed, 7 insertions(+), 1 deletion(-)