wfcqueue documentation: hint at for_each iterators

Reported-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Fix urcu-call-rcu-impl.h: false-sharing

> >  struct call_rcu_data {
> > -   struct cds_wfq_queue cbs;
> > +   /*
> > +    * Align the tail on cache line size to eliminate false-sharing
> > +    * with head.
> > +    */
> > +   struct cds_wfcq_tail __attribute__((aligned(CAA_CACHE_LINE_SIZE))) cbs_tail;
> > +   /* Alignment on cache line size will add padding here */
> > +
> > +   struct cds_wfcq_head cbs_head;
>
>
> wrong here. In this code, cbs_tail and cbs_head are in the same cache line.

Reported-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

call_rcu: use wfcqueue, eliminate false-sharing

Eliminate false-sharing between call_rcu (enqueuer) and worker threads
on the queue head and tail.

Acked-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

wfcqueue test

Acked-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

wfcqueue: implement concurrency-efficient queue

This new API simplify the wfqueue implementation, and brings a 2.3x to
2.6x performance boost due to the ability to eliminate false-sharing
between enqueue and dequeue.

This work is derived from the patch from Lai Jiangshan submitted as
"urcu: new wfqueue implementation"
(http://lists.lttng.org/pipermail/lttng-dev/2012-August/018379.html)

Its changelog:

> Some guys would be surprised by this fact:
> There are already TWO implementations of wfqueue in urcu.
>
> The first one is in urcu/static/wfqueue.h:
> 1) enqueue: exchange the tail and then update previous->next
> 2) dequeue: wait for first node's next pointer and them shift, a dummy node
>  is introduced to avoid the queue->tail become NULL when shift.
>
> The second one shares some code with the first one, and the left code
> are spreading in urcu-call-rcu-impl.h:
> 1) enqueue: share with the first one
> 2) no dequeue operation: and no shift, so it don't need dummy node,
>  Although the dummy node is queued when initialization, but it is removed
>  after the first dequeue_all operation in call_rcu_thread().
>  call_rcu_data_free() forgets to handle the dummy node if it is not removed.
> 3)dequeue_all: record the old head and tail, and queue->head become the special
>  tail node.(atomic record the tail and change the tail).
>
> The second implementation's code are spreading, bad for review, and it is not
> tested by tests/test_urcu_wfq.
>
> So we need a better implementation avoid the dummy node dancing and can service
> both generic wfqueue APIs and dequeue_all API for call rcu.
>
> The new implementation:
> 1) enqueue: share with the first one/original implementation.
> 2) dequeue: shift when node count >= 2, cmpxchg when node count = 1.
>  no dummy node, save memory.
> 3) dequeue_all: simply set queue->head.next to NULL, xchg the tail
>  and return the old head.next.
>
> More implementation details are in the code.
> tests/test_urcu_wfq will be update in future for testing new APIs.

The patch proposed by Lai brings a very interesting simplification to
the single-node handling (which is kept here), and moves all queue
handling code away from call_rcu implementation, back into the wfqueue
code. This has the benefit to allow testing enhancements.

I modified it so the API does not expose implementation details to the
user (e.g. ___cds_wfq_node_sync_next). I added a "splice" operation and
a for loop iterator which should allow wfqueue users to use the list
very efficiently both from LGPL/GPL code and from non-LGPL-compatible
code.

I also changed the API so the queue head and tail are now two separate
structures: it allows the queue user to place these as they like, either
on different cache lines (to eliminate false-sharing), or close one to
another (on same cache-line) in case a queue is spliced onto the stack
and not concurrently accessed.

Benchmarks performed on Intel(R) Core(TM) i7-3520M CPU @ 2.90GHz
(dual-core, with hyperthreading)

Benchmark invoked:
for a in $(seq 1 10); do ./test_urcu_wfq 1 1 10 -a 0 -a 2; done

(using cpu number 0 and 2, which should correspond to two cores of my
Intel 2-core/hyperthread processor)

Before patch:

testdur   10 nr_enqueuers   1 wdelay      0 nr_dequeuers   1 rdur      0 nr_enqueues     97274297 nr_dequeues     80745742 successful enqueues     97274297 successful dequeues     80745321 end_dequeues 16528976 nr_ops    178020039
testdur   10 nr_enqueuers   1 wdelay      0 nr_dequeuers   1 rdur      0 nr_enqueues     92300568 nr_dequeues     75019529 successful enqueues     92300568 successful dequeues     74973237 end_dequeues 17327331 nr_ops    167320097
testdur   10 nr_enqueuers   1 wdelay      0 nr_dequeuers   1 rdur      0 nr_enqueues     93516443 nr_dequeues     75846726 successful enqueues     93516443 successful dequeues     75826578 end_dequeues 17689865 nr_ops    169363169
testdur   10 nr_enqueuers   1 wdelay      0 nr_dequeuers   1 rdur      0 nr_enqueues     94160362 nr_dequeues     77967638 successful enqueues     94160362 successful dequeues     77967638 end_dequeues 16192724 nr_ops    172128000
testdur   10 nr_enqueuers   1 wdelay      0 nr_dequeuers   1 rdur      0 nr_enqueues     97491956 nr_dequeues     81001191 successful enqueues     97491956 successful dequeues     81000247 end_dequeues 16491709 nr_ops    178493147
testdur   10 nr_enqueuers   1 wdelay      0 nr_dequeuers   1 rdur      0 nr_enqueues     94101298 nr_dequeues     75650510 successful enqueues     94101298 successful dequeues     75649318 end_dequeues 18451980 nr_ops    169751808
testdur   10 nr_enqueuers   1 wdelay      0 nr_dequeuers   1 rdur      0 nr_enqueues     94742803 nr_dequeues     75402105 successful enqueues     94742803 successful dequeues     75341859 end_dequeues 19400944 nr_ops    170144908
testdur   10 nr_enqueuers   1 wdelay      0 nr_dequeuers   1 rdur      0 nr_enqueues     92198835 nr_dequeues     75037877 successful enqueues     92198835 successful dequeues     75027605 end_dequeues 17171230 nr_ops    167236712
testdur   10 nr_enqueuers   1 wdelay      0 nr_dequeuers   1 rdur      0 nr_enqueues     94159560 nr_dequeues     77895972 successful enqueues     94159560 successful dequeues     77858442 end_dequeues 16301118 nr_ops    172055532
testdur   10 nr_enqueuers   1 wdelay      0 nr_dequeuers   1 rdur      0 nr_enqueues     96059399 nr_dequeues     80115442 successful enqueues     96059399 successful dequeues     80066843 end_dequeues 15992556 nr_ops    176174841

After patch:

testdur   10 nr_enqueuers   1 wdelay      0 nr_dequeuers   1 rdur      0 nr_enqueues    221229322 nr_dequeues    210645491 successful enqueues    221229322 successful dequeues    210645088 end_dequeues 10584234 nr_ops    431874813
testdur   10 nr_enqueuers   1 wdelay      0 nr_dequeuers   1 rdur      0 nr_enqueues    219803943 nr_dequeues    210377337 successful enqueues    219803943 successful dequeues    210368680 end_dequeues 9435263 nr_ops    430181280
testdur   10 nr_enqueuers   1 wdelay      0 nr_dequeuers   1 rdur      0 nr_enqueues    237006358 nr_dequeues    237035340 successful enqueues    237006358 successful dequeues    236997050 end_dequeues 9308 nr_ops    474041698
testdur   10 nr_enqueuers   1 wdelay      0 nr_dequeuers   1 rdur      0 nr_enqueues    235822443 nr_dequeues    235815942 successful enqueues    235822443 successful dequeues    235814020 end_dequeues 8423 nr_ops    471638385
testdur   10 nr_enqueuers   1 wdelay      0 nr_dequeuers   1 rdur      0 nr_enqueues    235825567 nr_dequeues    235811803 successful enqueues    235825567 successful dequeues    235810526 end_dequeues 15041 nr_ops    471637370
testdur   10 nr_enqueuers   1 wdelay      0 nr_dequeuers   1 rdur      0 nr_enqueues    221974953 nr_dequeues    210938190 successful enqueues    221974953 successful dequeues    210938190 end_dequeues 11036763 nr_ops    432913143
testdur   10 nr_enqueuers   1 wdelay      0 nr_dequeuers   1 rdur      0 nr_enqueues    237994492 nr_dequeues    237938119 successful enqueues    237994492 successful dequeues    237930648 end_dequeues 63844 nr_ops    475932611
testdur   10 nr_enqueuers   1 wdelay      0 nr_dequeuers   1 rdur      0 nr_enqueues    220634365 nr_dequeues    210491382 successful enqueues    220634365 successful dequeues    210490995 end_dequeues 10143370 nr_ops    431125747
testdur   10 nr_enqueuers   1 wdelay      0 nr_dequeuers   1 rdur      0 nr_enqueues    237388065 nr_dequeues    237401251 successful enqueues    237388065 successful dequeues    237380295 end_dequeues 7770 nr_ops    474789316
testdur   10 nr_enqueuers   1 wdelay      0 nr_dequeuers   1 rdur      0 nr_enqueues    221201436 nr_dequeues    210831162 successful enqueues    221201436 successful dequeues    210831162 end_dequeues 10370274 nr_ops    432032598

Summary: Both enqueue and dequeue speed increase: around 2.3x speedup
for enqueue, and around 2.6x for dequeue.

We can verify that:
   successful enqueues - successful dequeues = end_dequeues

For all runs (ensures correctness: no lost node).

CC: Lai Jiangshan <laijs@cn.fujitsu.com>
Acked-by: Paul McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Ensure that read-side functions meet 10-line LGPL criterion

This commit ensures that all read-side functions meet the 10-line LGPL
criterion that permits them to be expanded directly into non-LGPL code,
without function-call instructions.  It also documents this as the
intent.

[ paulmck: Spelling fixes called out by Josh Triplett and name
change called out by Mathieu Desnoyers (_rcu_read_lock_help() ->
_rcu_read_lock_update(). ]

[ Mathieu Desnoyers: _rcu_read_unlock_help renamed to
  _rcu_read_unlock_update_and_wakeup, and spelling fix for
  preced -> precede. ]

Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

tls-compat.h: document sigaltstack(2) limitation

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

urcu: add notice to URCU_TLS() for it is not strictly async-signal-safe

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Document sigaltstack(2) limitation

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Documentation: update LICENSE file

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Update version to 0.7.4

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash API documentation: document destroy RCU read-lock constraint

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Fix: rculfhash should be offline while waiting for resize to complete

Causes hang on destroy with urcu QSBR if destroy is called within a rcu
registered thread.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Add missing entry to gitignore

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

urcu: move busy-wait code and name it ___cds_wfq_node_sync_next()

This code which waits for a node's next pointer until it appears, will
be used many times, move it to a help function and name it
___cds_wfq_node_sync_next().

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

urcu: fix compat_futex_noasync()

This patch fix two critical problems in the compatibility fallback of
compact_futex_noasync():

1) compat_futex_cond is not bound to any @uaddr, it services all @uaddr,
   if you wakeup only one thread(pthread_cond_signal), the @uaddr of
   this waking thread and the @uaddr of the woken-up thread may be different.
   The woken-up thread will very probably go to sleep again
   because his own condition is not true.

   *And* this waking thread(FUTEX_WAKE) wake up NOTHING.

2) If the caller want to wake up all waiting threads, he will use INT_MAX
   for @val, and:
                for (i = 0; i < INT_MAX; i++)
                        pthread_cond_signal(&compat_futex_cond);
   becomes almost infinity loop.

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

urcu: add hint to DEFINE_URCU_TLS() for compound types

Just a hint.

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Fix: CAA_BUILD_BUG_ON should refer to CAA_BUILD_BUG_ON_ZERO

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Add MIPS support

[ Edit by Mathieu Desnoyers: add explanations about supported
MIPS architectures, extracted from conversation with Ralf Baechle:

* Supported architectures

Ralf Baechle (edited by Mathieu Desnoyers):

This code works on all MIPS architecture variants.  The memory barrier
instruction, SYNC, was introduced for MIPS II.  The original MIPS I
instruction set doesn't have SYNC nor SMP support in the processor
architecture itself so SMP required horrible kludges in the system
hardware.  I think it's safe to say that Linux/MIPS will never support
any of these MIPS I SMP systems.  In the unlikely case this happens
anyway, we have a (Linux) kernel emulation of the SYNC instruction.
Voila - full binary compatibility across all MIPS processors and the
oldest hardware pays the performance penalty.

* Choice of barrier for cmm_mb()/cmm_rmb()/cmm_wmb()

Ralf Baechle:
"RMI (aka Netlogic and now Broadcom) XLR processor cores can be
configured to permit LD-LD, LD-ST, ST-LD and ST-ST reordering; default
is only ST-ST reordering.  To allow Linux to eventually enable full
reordering cmm_mb(), cmm_rmb() and cmm_wmb() all should perform SYNC
and a compiler barrier."

* No-op choice for cmm_read_barrier_depends():

Ralf Baechle:
"Technically there is nothing in the MIPS architecture spec that would
keep a MIPS implementation from reordering as freely as an Alpha or
even more liberally.  In practice most do strong ordering.  However
there is no MIPS implementation that makes full use of all the rope
provided.  So in theory a paranoid implementation of
cmm_read_barrier_depends() for MIPS should perform a SYNC.  In reality
it's not necessary and no sane MIPS core designer would implement
something that would design a core that need a non-empty
cmm_read_barrier_depends().  The reason why my patch had an empty one
is that I was using the Alpha code as a template."

Mathieu Desnoyers:
Moreover, the Linux kernel chooses a no-op for MIPS
read_barrier_depends() implementation, so any MIPS architecture that
would be as weak as Alpha would break the Linux kernel before breaking
the userspace RCU library.

* No need to put ".set noreorder" in cmm_mb() inline assembly:

Ralf Baechle:
"Certain instructions such as SYNC won't get reordered." ]

Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
CC: Paul McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Compatibility: remove bash-ismsm from test scripts

+= is not supported by all shells.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Fix inappropriate lib behavior: don't call exit()

Use abort() (implemented through the new urcu_die()) instead of exit(-1)
for unrecoverable errors.

Fixes #152

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Fix: re-enable compatibility with autoconf < 2.64

> I tried to build the latest urcu (git master e51500) on a Centos 6.2 box, and got:
>
> jscott@dxi0-62:~/src/userspace-rcu$ make -j4
> CDPATH="${ZSH_VERSION+.}:" && cd . && /bin/sh /users/jscott/src/userspace-rcu/config/missing --run aclocal-1.11 -I
> +config
> CDPATH="${ZSH_VERSION+.}:" && cd . && /bin/sh /users/jscott/src/userspace-rcu/config/missing --run autoconf
>  cd . && /bin/sh /users/jscott/src/userspace-rcu/config/missing --run automake-1.11 --foreign
> configure:4010: error: possibly undefined macro: m4_ifnblank
>       If this token and others are legitimate, please use m4_pattern_allow.
>       See the Autoconf documentation.
> make: *** [configure] Error 1
> make: *** Waiting for unfinished jobs....
>
> Some digging showed that the macro m4_ifnblank requires autoconf 2.64. Centos 6.2 has autoconf 2.63. :(
>
> I just worked around it by reverting commit a767fd locally, then I can build fine.

Reported-by: John Steele Scott <toojays@toojays.net>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Fix c99 compatibility: use __asm__ and __volatile__ in public headers

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Fix c99 compatibility: use __typeof__ instead of typeof in public headers

Reported-by: John Steele Scott <toojays@toojays.net>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

warning fix: tests urcutorture for NetBSD 5

>   CC     rcutorture_urcu-urcutorture.o
> In file included from urcutorture.c:9:
> api.h: In function '__smp_thread_id':
> api.h:160: warning: cast from pointer to integer of different size
> api.h:160: warning: cast from pointer to integer of different size
> api.h: In function 'wait_thread':
> api.h:210: warning: cast from pointer to integer of different size
> api.h:210: warning: cast from pointer to integer of different size

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Update version to 0.7.3

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Fix tests: make dist lib dependency

Some test programs were depending in SOURCES on the CDS library. Change
this for a LDADD, which makes "make dist" work after a make clean.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Update README for OS supported, tests dependency

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Add CodingStyle to tarball

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Add coding style document

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Test fix: test_perthreadlock uninitialized mutex

- Initialize the per thread mutexes. (fix)
- Remove unused count_reader/count_writer variables. (cleanup)

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

tests: support FreeBSD short "time" args

time(1) in FreeBSD does not have long argument name: change --append to
-a and --output to -o

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

freebsd 8.2 fix: define MAP_ANONYMOUS for compatibility

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Update version to 0.7.2

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Fix library compatibility

Commit 4d0d66bb795d1ed938e11a97a4e5f71326e20c71, implementing
tls-compat.h for pthread TLS compatibility, adds a prefix in front of
each TLS symbol (__tls_*). However, some of these symbols are exported
by the URCU library (e.g. rcu_reader_mb, defined in urcu.c as
"rcu_reader", which is overloaded by the urcu/map/urcu.h) to
applications. Therefore, this breaks binary compatibility with 0.6.x
versions of the library. This is not intended, and therefore is a bug,
so we remove this __tls_* prefix from the variables declared, defined
and referenced to through the tls-compat.h API for compilers supporting
"__thread".

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Update version to 0.7.1

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

fix: uatomic_set return value compile fix for non-x86 arch.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Update version to 0.7.0

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Cleanup: header comments coding style

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Document uatomic operations

Document each atomic operation provided by urcu/uatomic.h, along with
their memory barrier guarantees.

Acked-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Update return value of "set" operations

To follow the way the Linux kernel implements atomic_set(), we change
some API functions so they don't return any value anymore.

This is now the case for:

uatomic_set()
rcu_set_pointer()
rcu_assign_pointer()

This API change is very minor. In all instances of the Linux kernel
using rcu_assign_pointer(), none currently care about its return value.

However, we keep ABI compatibility: rcu_set_pointer_sym() still returns
the "v" value, even though it is not used by its wrapper macro anymore.

Reviewed-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Fix mremap wrapper for NetBSD 5

NetBSD 5 implements a mremap with a different semantic. Rename our
wrapper symbol name so it does not clash with the NetBSD 5 symbol.
Eventually, we could envision doing a special-case that uses the NetBSD
5 version instead of the fallback, but let's first get it working before
going into optimization land.

Suggested-by: Marek Vavruša <marek.vavrusa@nic.cz>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Use urcu/tls-compat.h

Provides compatibility for OpenBSD, NetBSD and Darwin.

Suggested-by: Marek Vavruša <marek.vavrusa@nic.cz>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Implement urcu/tls-compat.h

Suggested-by: Marek Vavruša <marek.vavrusa@nic.cz>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Add TLS detection m4 macro

Will allow urcu to support OSes that require to use pthread TLS (and do
not provide __thread TLS support).

Suggested-by: Marek Vavruša <marek.vavrusa@nic.cz>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

document concurrent data structures

Document the concurrent data structures provided by the userspace RCU
library.

Reviewed-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

documentation: refer to rcu-api.txt

API.txt moved to userspace-rcu documentation rcu-api.txt.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Move API.txt to doc/rcu-api.txt, install in system doc/

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: document implied memory barriers

We choose to provide full memory barriers before and after successful
hash table update operations. Eventually, new API with weaker semantic
can be added, but let's make the basic API as fool-proof as possible.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Reviewed-by: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>

rculfhash: Ensure future-proof memory barrier semantic consistency

Use cmm_smp_mb__before_uatomic_or() prior to the uatomic_or() in
_rcu_lfht_del() to ensure correct memory barrier semantic when we relax
(in the future) the barrier implementation of some architectures.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

API cleanup: use "uatomic_*" in cmm_smp_mb__ API

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

uatomic: add memory barrier API for and/or/add/sub/inc/sub

Implement:
cmm_smp_mb__before_and, cmm_smp_mb__after_and
cmm_smp_mb__before_or, cmm_smp_mb__after_or
cmm_smp_mb__before_add, cmm_smp_mb__after_add
cmm_smp_mb__before_sub, cmm_smp_mb__after_sub
cmm_smp_mb__before_inc, cmm_smp_mb__after_inc
cmm_smp_mb__before_dec, cmm_smp_mb__after_dec

For generic and x86.

These currently translate into simple compiler barriers on all
architectures, but the and/or/add/sub/inc/dec uatomics do not provide
memory ordering guarantees (only uatomic_add_return, uatomic_sub_return,
uatomic_xchg, and uatomic_cmpxchg provides full memory barrier
guarantees before and after the atomic operations).

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: add runhash.sh test script

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash tests: add missing check

We need to check if test_ht is NULL.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: fix: race between replace and del operations

Bug introduced by commit db00ccc36e7fb04ce8044fb1be7964acd1de6ae0

Here is the race:

Initially in hash table:  A

T0                                         T1
replace A by B
                                           del A
  read A->next
  -> check REMOVED flag, not set yet.
                                           read A->next
                                           -> check REMOVED flag, not set yet.
  cmpxchg A->next to set REMOVED flag
  -> cmpxchg succeeds
                                           uatomic_or to set REMOVED flag
                                           uatomic_xchg to atomically set the REMOVAL_OWNER flag
                                           -> first to set the flag.
  Replace returns node -> free(A)          Del success -> free(A)

With this race, we have a double-free.

The problem with the replace code is that it does not set the
"REMOVAL_OWNER" flag.

Test case to reproduce the bug:

test_urcu_hash 0 2 20 -A -s -M 1 -N 1 -O 1

(2 threads, doing replace/del, with a hash table that has only a single
key for all values). After just a couple of seconds, either the program
hangs, or, more often, it does:

*** glibc detected ***
/media/truecrypt1/compudj/doc/userspace-rcu/tests/.libs/test_urcu_hash:
malloc(): memory corruption (fast): 0x00007ffff3a29e25 ***

Program received signal SIGSEGV, Segmentation fault.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Tested-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: replace unneeded rcu_dereference by CMM_LOAD_SHARED

The difference between the two is that CMM_LOAD_SHARED() does not imply
a read barrier between the read and following uses of the data pointed
to by the pointer read.

All sites that only use the pointer load for its bits (never
dereference) don't need the read barrier implied by rcu_dereference.

Reviewed-by: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: use do {} while (0) for dbg_printf()

found by clang(make CC=clang).

avoid empty statement.
-------------------------
if (condition)
        dbg_printf()  /* forget ";", but compiler say nothing if dbg_printf() is empty */
statement;
-------------------------

also add printf format check.
(we can use gcc extention "__printf(1, 2)" to declare a dummy inline function
to do the check, but I use "printf()" directly here)

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: cleanup typo

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: update API comments

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: update comments in implementation

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash tests: add long hash chains tests

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash tests: add uniqueness test

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash test: print test name

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: stress requirement in documentation

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: fix typo

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash tests: use array of callbacks to modularize

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash tests: modularize tests

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: document ordering guarantees

What we actually provide are ordering guarantees.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: document linearizability guarantees

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: update removal comment

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Add missing files to .gitignore

Fixes #93

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Fix out-of-tree build

Wildcards in automake seems to require that we prepent $(top_srcdir) to
get out-of-tree build to work.

Other non-wildcards work fine without the $(top_srcdir) prefix, so
leaving as-is.

Fixes #94

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: check malloc NULL pointer

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Fix uatomic sign cast

Passing an unsigned int to uatomic_sub does not honor sign extend to
long, as we should be allowed by assume.

Fix this by introducing caa_cast_long_keep_sign(), which casts either to
long or unsigned long depending on the signedness of the argument
received. It is used in uatomic_sub before applying the "-" operator,
since this operator needs to operate on the "long" type size (since sign
extension might not be performed if the argument received is unsigned).

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Remove unused jhash.h file

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Fix: use known license text, fix incorrect FSF address

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Fix: add missing sched.h include in rculfhash.c

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Allow tests to run on architectures without per-cpu call_rcu support

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Hash table test: FreeBSD compatibility fix

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

tests: define _GNU_SOURCE instead of __USE_GNU

Fixes the following warning:

In file included from urcutorture.c:9:0:
api.h:67:0: warning: "__USE_GNU" redefined [enabled by default]
/usr/include/features.h:304:0: note: this is the location of the previous definition

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Merge branch 'urcu/ht-shrink'

RCU lock-free hash table: implement cds_lfht_is_node_deleted()

Some thoughts on how to use the RCU lock-free hash table brought me to
figure out that using the lock-free hash table along with a per-node
mutex is a quite interesting way to deal with lookup vs teardown
coherency.

A per-node lock can be used to protect concurrent modifications of an
entry from one another, as well as concurrent read vs modification. In
addition, if we ensure that each reader/updater of the node checks if
the node has been removed right after taking the mutex, and if we
perform the node removal from the hash table with the per-node mutex
held, we can ensure that readers/updaters will never access unlinked
data.

struct mynode {
pthread_mutex_t mutex;
struct cds_lfht node;
}

CPU A (lookup destroy and free)           CPU B (lookup and read/modify)

                                          rcu_read_lock()
                                          mynode = caa_container_of(
                                                cds_lfht_lookup(...), ...);
                                          mutex_lock(&mynode->mutex);
                                          if (cds_lfht_is_node_deleted(
&mynode->node))
                                               goto unlock;

                                          read/modify structure....

                                        unlock:
                                          mutex_unlock(&mynode->mutex);
                                          rcu_read_unlock()

rcu_read_lock()
mynode = caa_container_of(
        cds_lfht_lookup(...), ...);

mutex_lock(&mynode->mutex);

cds_lfht_del(ht, &mynode->node);

- perform extra teardown operations
  with side-effects, for which call_rcu
  delay is not appropriate

mutex_unlock(&mynode->mutex);
rcu_read_unlock()
call_rcu(free, mynode);

To perform this efficiently, we need an API function to return whether
the node previously looked-up has been deleted since then.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Acked-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
CC: Lai Jiangshan <laijs@cn.fujitsu.com>
CC: Stephen Hemminger <shemminger@vyatta.com>

Define _GNU_SOURCE to access CPU_ZERO() macro in uClibc

Signed-off-by: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Install ChangeLog and README into system's doc

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Fix AC_LANG_SOURCE usage: only takes one parameter

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Merge branch 'master' into urcu/ht-shrink

Fix autoconf futex check

The check was always returning true.

Signed-off-by: David Goulet <dgoulet@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

configure.ac: Use AC_LANG_SOURCE for if else macros

Ref. http://www.flameeyes.eu/autotools-mythbuster/forwardporting/autoconf.html
"Noteworthy changes in autoconf version 2.66 through 2.68"

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Refresh autoconf files

Use portable shell macros wherever possible.
All functionality should remain the same.

Signed-off-by: Alexandre Montplaisir <alexandre.montplaisir@gmail.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Update gitignore

Signed-off-by: Alexandre Montplaisir <alexandre.montplaisir@gmail.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Merge branch 'master' into urcu/ht-shrink

rculfhash: add comment about hash seed randomness within test program

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

readme: state correct GCC dependency for ARM

If you are trying to compile liburcu for ARM and get errors like:

/usr/local/include/urcu/uatomic/generic.h:180: undefined reference to
`__sync_add_and_fetch_4'
usr/local/lib/liburcu-common.so: undefined reference to
`__sync_lock_test_and_set_4'
/usr/local/lib/liburcu.so: undefined reference to
`__sync_or_and_fetch_4'

please upgrade your GCC to 4.4.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: remove an invocation of bit_reverse_ulong() when adding

Suggested-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: remove unneeded conversion

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: remove unneeded clear_flag()

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

cds_lfht_replace: add checks for old/new node hash/value match

Also initialize reverse hash in replace.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: use node instead of iter argument for deletion

Using a node instead of an iterator as argument for deletion allows
passing a node pointer (that would have been looked up from another data
structure, thus not using the iterator) as argument for deletion.

Deletion still returns -ENOENT if asked to delete the NULL node. This
simplifies the caller usage.

Suggested-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

rculfhash: number of logically removed nodes should not appear in API

This is an implementation artefact that should not appear in the API. So
only count the non-removed nodes. Print a debug message showing the
number of logically removed nodes instead within the count nodes
function.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Document that flags could be represented on 2 bits

Suggested-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>