Messages in this thread | | | From | Jason Low <> | Subject | [PATCH 0/4] MCS spinlocks: Cancellable MCS spinlock rework | Date | Mon, 7 Jul 2014 11:50:15 -0700 |
| |
The main purpose of this patchset is to reduce the size of the cancellable MCS spinlock and reduce the overhead of rwsem (currently the largest lock in the kernel).
The overhead of the cancellable MCS lock is a pointer to a per-cpu node structure which requires 64 bits on 64 bit systems. Instead of a pointer to the per-cpu node, we can instead store the CPU # corresponding to the node in atomic_t. This reduces the overhead by 32 bits on 64 bit systems.
This then opens the opportunity to reduce the size of the rw_semaphore structure (one of the current users of this MCS lock) by moving around some of its fields. Due to padding, we would reduce the size of that structure by 64 bits (on 64 bit systems). This makes it more in line with the size of the mutex structure.
Jason Low (4): MCS spinlocks: Rename optimistic_spin_queue to optimistic_spin_node MCS spinlocks: Convert osq lock to atomic_t to reduce overhead MCS spinlocks: Micro-optimize osq_unlock() rwsem: Reduce the size of struct rw_semaphore
include/linux/mutex.h | 4 +- include/linux/osq_lock.h | 19 ++++++++++++ include/linux/rwsem.h | 11 +++---- kernel/locking/mcs_spinlock.c | 62 ++++++++++++++++++++++++++++++---------- kernel/locking/mcs_spinlock.h | 9 +++-- kernel/locking/mutex.c | 2 +- kernel/locking/rwsem-xadd.c | 2 +- 7 files changed, 79 insertions(+), 30 deletions(-) create mode 100644 include/linux/osq_lock.h
| |