From: Jeremy Fitzhardinge <jeremy.fitzhardinge@xxxxxxxxxx>
Hi all,
This series does two major things:
1. It converts the bulk of the implementation to C, and makes the
"small ticket" and "large ticket" code common. Only the actual
size-dependent asm instructions are specific to the ticket size.
The resulting generated asm is very similar to the current
hand-written code.
This results in a very large reduction in lines of code.
2. Get rid of pv spinlocks, and replace them with pv ticket locks.
Currently we have the notion of "pv spinlocks" where a pv-ops
backend can completely replace the spinlock implementation with a
new one. This has two disadvantages:
- its a completely separate spinlock implementation, and
- there's an extra layer of indirection in front of every
spinlock operation.
To replace this, this series introduces the notion of pv ticket
locks. In this design, the ticket lock fastpath is the standard
ticketlock algorithm. However, after an iteration threshold it
falls into a slow path which invokes a pv-op to block the spinning
CPU. Conversely, on unlock it does the normal unlock, and then
checks to see if it needs to do a special "kick" to wake the next
CPU.
The net result is that the pv-op calls are restricted to the slow
paths, and the normal fast-paths are largely unaffected.
There are still some overheads, however:
- When locking, there's some extra tests to count the spin iterations.
There are no extra instructions in the uncontended case though.
- When unlocking, there are two ways to detect when it is necessary
to kick a blocked CPU:
- with an unmodified struct spinlock, it can check to see if
head == tail after unlock; if not, then there's someone else
trying to lock, and we can do a kick. Unfortunately this
generates very high level of redundant kicks, because the
waiting CPU might not have blocked yet (which is the common
case)
- With a struct spinlock modified to include a "waiters" field,
to keep count of blocked CPUs, which is a much tighter test.
But it does increase the size of each spinlock by 50%
(doubled with padding).
The series is very fine-grained, and I've left a lightly cleaned up
history of the various options I evaluated, since they're not all
obvious.
Jeremy Fitzhardinge (20):
x86/ticketlock: clean up types and accessors
x86/ticketlock: convert spin loop to C
x86/ticketlock: Use C for __ticket_spin_unlock
x86/ticketlock: make large and small ticket versions of spin_lock the
same
x86/ticketlock: make __ticket_spin_lock common
x86/ticketlock: make __ticket_spin_trylock common
x86/spinlocks: replace pv spinlocks with pv ticketlocks
x86/ticketlock: collapse a layer of functions
xen/pvticketlock: Xen implementation for PV ticket locks
x86/pvticketlock: keep count of blocked cpus
x86/pvticketlock: use callee-save for lock_spinning
x86/pvticketlock: use callee-save for unlock_kick as well
x86/pvticketlock: make sure unlock is seen by everyone before
checking waiters
x86/ticketlock: loosen ordering restraints on unlock
x86/ticketlock: prevent compiler reordering into locked region
x86/ticketlock: don't inline _spin_unlock when using paravirt
spinlocks
x86/ticketlock: clarify barrier in arch_spin_lock
x86/ticketlock: remove .slock
x86/ticketlocks: use overlapping read to eliminate mb()
x86/ticketlock: rename ticketpair to head_tail
arch/x86/Kconfig | 3 +
arch/x86/include/asm/paravirt.h | 30 +---
arch/x86/include/asm/paravirt_types.h | 8 +-
arch/x86/include/asm/spinlock.h | 250 +++++++++++++++--------------
arch/x86/include/asm/spinlock_types.h | 41 +++++-
arch/x86/kernel/paravirt-spinlocks.c | 15 +--
arch/x86/xen/spinlock.c | 282 +++++----------------------------
kernel/Kconfig.locks | 2 +-
8 files changed, 221 insertions(+), 410 deletions(-)
--
1.7.2.3
_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel
|
Home