Messages in this thread | | | Date | Mon, 28 Nov 2011 13:35:17 -0800 | From | Tejun Heo <> | Subject | [GIT PULL] percpu fixes for 3.2-rc3 |
| |
Hello, Linus.
Each percpu chunk contains units for each CPU which can be offset by arbitrary amount (usually determined by how early mem gets allocated for static percpu areas).
Way back, the offsets were always monotonically increasing and first/last units were assumed to have the lowest and highest offsets. These were used to determine cache flush ranges and map percpu to physical addresses.
However, with embedded first chunk allocation, the assumption is no longer true. There's no relation between unit number and its offset, which broke percpu -> physical translation. The bug is rather serious as it has remote chance of causing extremely obscure and difficult to reproduce problems on archs which require vmalloc area flushing.
This pull includes fix for the above problem, a patch explaining why phys addr translation is done the way it's done, and a trivial percpu internal function rename patch which was sitting in my queue for too long.
Please pull from the following git branch,
git://git.kernel.org/pub/scm/linux/kernel/git/tj/percpu.git for-3.2-fixes
Bob Liu (1): percpu: rename pcpu_mem_alloc to pcpu_mem_zalloc
Dave Young (1): percpu: explain why per_cpu_ptr_to_phys() is more complicated than necessary
Tejun Heo (1): percpu: fix chunk range calculation
mm/percpu-vm.c | 17 +++++++-------- mm/percpu.c | 62 ++++++++++++++++++++++++++++++++++++------------------- 2 files changed, 48 insertions(+), 31 deletions(-)
Thanks.
-- tejun
| |