Initial benchmarks show they're a net loss:
$ for i in /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor ; do echo performance > $i; done
$ echo 4096 32000 64 128 > /proc/sys/kernel/sem
$ ./sembench -t 2048 -w 1900 -o 0
Pre:
run time 30 seconds 778936 worker burns per second
run time 30 seconds 912190 worker burns per second
run time 30 seconds 817506 worker burns per second
run time 30 seconds 830870 worker burns per second
run time 30 seconds 845056 worker burns per second
Post:
run time 30 seconds 905920 worker burns per second
run time 30 seconds 849046 worker burns per second
run time 30 seconds 886286 worker burns per second
run time 30 seconds 822320 worker burns per second
run time 30 seconds 900283 worker burns per second
So about 4% faster. (!)
cpu_relax() stalls the pipeline, therefore, when used in a tight loop
it has the following benefits:
- allows SMT siblings to have a go;
- reduces pressure on the CPU interconnect.
However, cmpxchg loops are unfair and thus have unbounded completion
time, therefore we should avoid getting in such heavily contended
situations where the above benefits make any difference.
A typical cmpxchg loop should not go round more than a handfull of
times at worst, therefore adding extra delays just slows things down.
Since the llist primitives are new, there aren't any bad users yet,
and we should avoid growing them. Heavily contended sites should
generally be better off using the ticket locks for serialization since
they provide bounded completion times (fifo-fair over the cpus).
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Link: http://lkml.kernel.org/r/1315836358.26517.43.camel@twins
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Extend the llist_add*() functions to return a success indicator, this
allows us in the scheduler code to send an IPI if the queue was empty.
( There's no effect on existing users, because the list_add_xxx() functions
are inline, thus this will be optimized out by the compiler if not used
by callers. )
Signed-off-by: Huang Ying <ying.huang@intel.com>
Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1315461646-1379-5-git-send-email-ying.huang@intel.com
Signed-off-by: Ingo Molnar <mingo@elte.hu>
If in llist_add()/etc. functions the first cmpxchg() call succeeds, it is
not necessary to use cpu_relax() before the cmpxchg(). So cpu_relax() in
a busy loop involving cmpxchg() should go after cmpxchg() instead of before
that.
This patch fixes this for all involved llist functions.
Signed-off-by: Huang Ying <ying.huang@intel.com>
Acked-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1315461646-1379-4-git-send-email-ying.huang@intel.com
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Remove the nmi() checks spread around the code. in_nmi() is not available
on every architecture and it's a pretty obscure and ugly check in any case.
Cc: Huang Ying <ying.huang@intel.com>
Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1315461646-1379-3-git-send-email-ying.huang@intel.com
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Because llist code will be used in performance critical scheduler
code path, make llist_add() and llist_del_all() inline to avoid
function calling overhead and related 'glue' overhead.
Signed-off-by: Huang Ying <ying.huang@intel.com>
Acked-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1315461646-1379-2-git-send-email-ying.huang@intel.com
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Cmpxchg is used to implement adding new entry to the list, deleting
all entries from the list, deleting first entry of the list and some
other operations.
Because this is a single list, so the tail can not be accessed in O(1).
If there are multiple producers and multiple consumers, llist_add can
be used in producers and llist_del_all can be used in consumers. They
can work simultaneously without lock. But llist_del_first can not be
used here. Because llist_del_first depends on list->first->next does
not changed if list->first is not changed during its operation, but
llist_del_first, llist_add, llist_add (or llist_del_all, llist_add,
llist_add) sequence in another consumer may violate that.
If there are multiple producers and one consumer, llist_add can be
used in producers and llist_del_all or llist_del_first can be used in
the consumer.
This can be summarized as follow:
| add | del_first | del_all
add | - | - | -
del_first | | L | L
del_all | | | -
Where "-" stands for no lock is needed, while "L" stands for lock is
needed.
The list entries deleted via llist_del_all can be traversed with
traversing function such as llist_for_each etc. But the list entries
can not be traversed safely before deleted from the list. The order
of deleted entries is from the newest to the oldest added one. If you
want to traverse from the oldest to the newest, you must reverse the
order by yourself before traversing.
The basic atomic operation of this list is cmpxchg on long. On
architectures that don't have NMI-safe cmpxchg implementation, the
list can NOT be used in NMI handler. So code uses the list in NMI
handler should depend on CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG.
Signed-off-by: Huang Ying <ying.huang@intel.com>
Reviewed-by: Andi Kleen <ak@linux.intel.com>
Reviewed-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Len Brown <len.brown@intel.com>
