workqueue: keep one work item from stealing co-workers
Takes care of lack of progress issue that could theoretically occur due
to a steady work-stealing circular loop amongst co-workers. Keeping a
work item hidden from co-workers ensures forward progress.
workqueue/waitqueue: use lock-free stack for wakeup
The case for which we try to reduce latency is the wakeup, not the
"wait", since "wait" is typically done when a thread is ready to put
itself asleep awaiting for more work.
wfstack has blocking "pop" operation, which means that if a worker
thread is delayed for an extremely long amount of time during the push
operation (add to waitqueue), it could delay execution of pop (wakeup)
for that same amount of time.
lfstack does not have this downside, at the expense of having the "wait"
operation (push) being only lock-free rather than wait-free. However,
since we don't care that much about "wait" being wait-free, it makes
sense to use lfstack here.
Fix: preserve example files' timestamps when copying
This fixes an issue where examples were always being rebuilt
when performing an out of tree build since the examples were
being copied to the build directory with a timestamp more
recent than the already-built example objects.
Eric Wong [Mon, 1 Sep 2014 21:25:06 +0000 (21:25 +0000)]
wfstack: implement mutex-free wfstack with transparent union
This allows users more freedom to use alternative synchronization
mechanisms.
Changes since v1:
- Fix typos in cds_wfs_stack_ptr_t documentation.
Thanks to Mathieu for spotting.
Signed-off-by: Eric Wong <normalperson@yhbt.net> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Fix: incorrect parenthesis in cds_hlist_for_each_entry_safe_2
commit db903109f0031c831e8fdc95cb7197996e53f46d introduced a regression
in cds_hlist_for_each_entry_safe_2(): incorrect parenthesis assign "e"
to 1, rather than assign "e" to the next pointer, and evaluating the
expression to 1 (comma expression).
Reported-by: Daniel Thibault <Daniel.Thibault@drdc-rddc.gc.ca> Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Do not free the rcu_barrier() completion struct until all threads are
done with it.
It cannot reside on the waiter's stack as rcu_barrier() may return
before the call_rcu handlers have finished checking whether it needs a
futex wakeup. Instead we dynamically allocate the structure and
determine its lifetime with a reference count.
Signed-off-by: Keir Fraser <keir@cohodata.com>
[ Edit by Mathieu Desnoyers: use urcu/ref.h. Cleanup: use
uatomic_sub_return() rather than uatomic_add_return() with negative
value. ] Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
call_rcu threads should clear their PAUSED flag when they unpause
And call_rcu_after_fork_parent should spin-wait on this.
Otherwise a second fork in the parent will see the PAUSED flags
already set and call_rcu_before_fork will not correctly wait for the
call_rcu threads to quiesce on this second occasion.
Add the missing architecture specific functions to provide support for
the hppa/PA-RISC architecture:
- the processor internal time stamp counter (Control Register CR16) is
used to get high-performance/low-latency cycle counts
- gcc provides the necessary built-in atomic functions on hppa (which in
turn uses the light-weigth atomic locking syscall-interface of the
Linux kernel)
Lars Persson [Wed, 12 Mar 2014 09:36:04 +0000 (10:36 +0100)]
Use autoconf AM_MAINTAINER_MODE
Give distribution maintainers the option to skip rebuilding
autoconf and automake generated files. The default behaviour
is still to have the rebuild rules enabled.
Signed-off-by: Lars Persson <larper@axis.com> Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Currently there are two fairly recent architectures, which at the
moment can only be compiled with "gcc atomics" code path.
The two new architectures are (GNU Types):
* aarch64-linux-gnu (aka ARMv8, ARM64, AARCH64, etc)
* powerpc64le-linux-gnu
Fix: move wait loop increment before first conditional block
The fix "Fix: high cpu usage in synchronize_rcu with long RCU read-side
C.S." has an imperfection in urcu.c and urcu-qsbr.c: when incrementing
the wait loop counter for the last time, the first conditional branch is
not taken, but the following conditionals are, and they assume the first
conditional has been taken.
Within urcu.c (urcu-mb, urcu-membarrier and urcu-signal), and
urcu-qsbr.c, this will simply skip the first wait_gp() call, without any
noticeable ill side-effect.
Fix: high cpu usage in synchronize_rcu with long RCU read-side C.S.
We noticed that with this kind of scenario:
- application using urcu-mb, urcu-membarrier, urcu-signal, or urcu-bp,
- long RCU read-side critical sections, caused by e.g. long network I/O
system calls,
- other short lived RCU critical sections running in other threads,
- very frequent invocation of call_rcu to enqueue callbacks,
lead to abnormally high CPU usage within synchronize_rcu() in the
call_rcu worker threads.
Inspection of the code gives us the answer: in urcu.c, we expect that if
we need to wait on a futex (wait_gp()), we expect to be able to end the
grace period within the next loop, having been notified by a
rcu_read_unlock(). However, this is not always the case: we can very
well be awakened by a rcu_read_unlock() executed on a thread running
short-lived RCU read-side critical sections, while the long-running RCU
read-side C.S. is still active. We end up in a situation where we
busy-wait for a very long time, because the counter is !=
RCU_QS_ACTIVE_ATTEMPTS until a 32-bit overflow happens (or more likely,
until we complete the grace period). We need to change the wait_loops ==
RCU_QS_ACTIVE_ATTEMPTS check into an inequality to use wait_gp() for
every attempts beyond RCU_QS_ACTIVE_ATTEMPTS loops.
urcu-bp.c also has this issue. Moreover, it uses usleep() rather than
poll() when dealing with long-running RCU read-side critical sections.
Turn the usleep 1000us (1ms) into a poll of 10ms. One of the advantage
of using poll() rather than usleep() is that it does not interact with
SIGALRM.
urcu-qsbr.c already checks for wait_loops >= RCU_QS_ACTIVE_ATTEMPTS, so
it is not affected by this issue.
Looking into these loops, however, shows that overflow of the loop
counter, although unlikely, would bring us back to a situation of high
cpu usage (a negative value well below RCU_QS_ACTIVE_ATTEMPTS).
Therefore, change the counter behavior so it stops incrementing when it
reaches RCU_QS_ACTIVE_ATTEMPTS, to eliminate overflow.
Fix: urcu-bp interaction with threads vs constructors/destructors
Add a reference counter for threads using urcu-bp, thus ensuring that
even if the urcu destructor is executed before each thread using RCU
read-side critical sections exit, those threads will not see a corrupted
thread list.
Also, don't use URCU_TLS() within urcu_bp_thread_exit_notifier(). It
appears that this is racy (although this was probably due to the issue
fixed by reference counting). Anyway, play safe, and pass the rcu_key
received as parameter instead.
Those issues only reproduce when threads are still active when the
urcu-bp destructor is called.
Clang 3.3 with -O2 optimisations is especially picky about arithmetic on
NULL pointers. This undefined behavior is turned into optimized out NULL
checks by clang 3.3. Fix the undefined behavior by checking against the
pointer directly, without going back and forth around NULL with
pointer arithmetic.
Charles Briere [Mon, 2 Dec 2013 02:00:50 +0000 (18:00 -0800)]
Cleanup: Check for pthread in Libc
Some libc like Googles's Bionic for Android include pThread. Instead of
checking for Android as in commit 58a052effb2305655ff1b7233c78ab24fc25257e this patch checks for pthread
in libc as proposed by Thomas Petazzoni.
Reported-by: Thomas Petazzoni <thomas.petazzoni@free-electrons.com> Signed-off-by: Charles Briere <charlesbriere.flatzo@gmail.com> Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Charles Briere [Mon, 2 Dec 2013 02:00:49 +0000 (18:00 -0800)]
Cleanup: Android: Do not redefine gettid
When compiling with Android Toolchain, gettid is already defined in
<unistd.h> and therefore doesn't need to be reimplemented. This patch
checks for gettid instead of looking for Android in case other
environements behave the same way.
Signed-off-by: Charles Briere <charlesbriere.flatzo@gmail.com> Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Clang 3.3 with -O2 optimisations is especially picky about arithmetic
on NULL pointers. This undefined behavior is turned into optimized out
NULL checks by clang 3.3. Fix the undefined behavior by checking against
the pointer directly, without going back and forth around NULL with
pointer arithmetic.
Library major version number (soname) increment to 3
Due to the urcu/tls-compat.h TLS fallback symbol clash fix, we need to
bump the Userspace RCU soname major version number so we don't have to
keep erroneous usage detection code around.
Vladimir Nikulichev noticed crashes when using this setup. The problem
can be pinpointed to a missing macro expansion in urcu/tls-compat.h:
looking at the output of
nm tests/unit/.libs/test_urcu_multiflavor :
U __tls_access_rcu_reader
this seems to be the issue. We're missing macro expansion in
tls-compat.h. With this commit, it becomes:
U __tls_access_rcu_reader_bp
U __tls_access_rcu_reader_mb
U __tls_access_rcu_reader_memb
U __tls_access_rcu_reader_sig
Please note that this affects an unusual configuration of userspace RCU
(with TLS pthread key fallback), needed for some BSD that don't support
compiler TLS. Strictly speaking, this requires bumping the URCU library
soname version major number, because it breaks the ABI presented to
applications on those unusual configurations.
A following commit will handle the ABI migration: for stable releases
(stable-0.7 and stable-0.8 branches), the ABI is kept compatible, and
bogus usage are detected. For the upcoming stable-0.9, the soname will
simply be bumped.
Reported-by: Vladimir Nikulichev <nvs@tbricks.com> Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
When compiling code using the rcu_xchg_pointer() family of functions,
with the following define:
#define URCU_INLINE_SMALL_FUNCTIONS
prior to including urcu headers, when compiling with gcc with
-Wsign-compare and -Wextra, gcc warns about:
urcu-xchg.c: In function ‘reload’:
urcu-xchg.c:19:1: warning: ordered comparison of pointer with integer zero [-Wextra]
urcu-xchg.c:19:1: warning: signed and unsigned type in conditional expression [-Wsign-compare]
For the "ordered comparison of pointer with integer zero" warning, fix
this by comparing (type) -1 against (type) 0 instead of just 0, so if
"type" is a pointer type, this pointer type will be applied to the right
operand too, thus fixing the warning.
For the "signed and unsigned type in conditional expression" warning, we
need caa_cast_long_keep_sign() to always evaluate to the same type
signedness. In order to do so, when we need to sign-extend the value,
cast it to unsigned long after first casting it to long.
Reported-by: Stephen Hemminger <stephen@networkplumber.org> Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
* Dmitri Shubin <sbn@tbricks.com> wrote:
> Shouldn't the condition in line 94 actually be
>
> 94 #if (!defined(BUILD_QSBR_LIB) && !defined(RCU_DEBUG))
>
> So when RCU_DEBUG is _not_ defined we get static inlines for
> rcu_read_{,un}lock() ?
Indeed!
Reported-by: Dmitri Shubin <sbn@tbricks.com> Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
This fixes an issue that appears after this recent urcu-bp fix is
applied:
Fix: urcu-bp: Bulletproof RCU arena resize bug
Prior to this fix, on Linux at least, the behavior was to allocate
(and leak) one memory map region per reader thread. It worked, except
for the unfortunate leak. The fact that it worked, even though not the
way we had intended it to, is is why testing did not raise any red flag.
That state of affairs has prevailed for a long time, but it was
side-tracking some issues. After fixing the underlying bug that was
causing the memory map leak, another issue appears.
The garbage collection scheme reclaiming the thread tracking structures
in urcu-bp fails in stress tests to due a bug in glibc (tested against
glibc 2.13 and 2.17). Under this workload, on a 2-core/hyperthreaded i7:
./test_urcu_bp 40 4 10
we can easily trigger a segmentation fault in the pthread_kill() code.
Program terminated with signal 11, Segmentation fault.
Backtrace:
#0 __pthread_kill (threadid=140723681437440, signo=0) at ../nptl/sysdeps/unix/sysv/linux/pthread_kill.c:42
42 ../nptl/sysdeps/unix/sysv/linux/pthread_kill.c: No such file or directory.
(gdb) bt full
#0 __pthread_kill (threadid=140723681437440, signo=0) at ../nptl/sysdeps/unix/sysv/linux/pthread_kill.c:42
__x = <optimized out>
pd = 0x7ffcc90b2700
tid = <optimized out>
val = <optimized out>
#1 0x0000000000403009 in rcu_gc_registry () at ../../urcu-bp.c:437
tid = 140723681437440
ret = 0
chunk = 0x7ffcca0b8000
rcu_reader_reg = 0x7ffcca0b8120
__PRETTY_FUNCTION__ = "rcu_gc_registry"
#2 0x0000000000402b9c in synchronize_rcu_bp () at ../../urcu-bp.c:230
cur_snap_readers = {next = 0x7ffcb4888cc0, prev = 0x7ffcb4888cc0}
qsreaders = {next = 0x7ffcb4888cd0, prev = 0x7ffcb4888cd0}
newmask = {__val = {18446744067267100671, 18446744073709551615 <repeats 15 times>}}
oldmask = {__val = {0, 140723337334144, 0, 0, 0, 140723690351643, 0, 140723127058464, 4, 0, 140723698253920, 140723693868864, 4096, 140723690370432, 140723698253920, 140723059951840}}
ret = 0
__PRETTY_FUNCTION__ = "synchronize_rcu_bp"
#3 0x0000000000401803 in thr_writer (_count=0x76b2f0) at test_urcu_bp.c:223
count = 0x76b2f0
new = 0x7ffca80008c0
old = 0x7ffca40008c0
#4 0x00007ffcc9c83f8e in start_thread (arg=0x7ffcb4889700) at pthread_create.c:311
__res = <optimized out>
pd = 0x7ffcb4889700
now = <optimized out>
unwind_buf = {cancel_jmp_buf = {{jmp_buf = {140723337336576, 6546223316613858487, 0, 140723698253920, 140723693868864, 4096, -6547756131873848137,
-6547872135220034377}, mask_was_saved = 0}}, priv = {pad = {0x0, 0x0, 0x0, 0x0}, data = {prev = 0x0, cleanup = 0x0, canceltype = 0}}}
not_first_call = 0
pagesize_m1 = <optimized out>
sp = <optimized out>
freesize = <optimized out>
__PRETTY_FUNCTION__ = "start_thread"
#5 0x00007ffcc99ade1d in clone () at ../sysdeps/unix/sysv/linux/x86_64/clone.S:113
It appears that the memory backing the thread information can be
relinquished by NPTL concurrently with execution of pthread_kill()
targeting an already joined thread and cause this segfault. We were
using pthread_kill(tid, 0) to discover if the target thread was alive or
not, as documented in pthread_kill(3):
If sig is 0, then no signal is sent, but error checking is still per‐
formed; this can be used to check for the existence of a thread ID.
but it appears that the glibc implementation is racy.
Instead of using the racy pthread_kill implementation, implement cleanup
using a pthread_key destroy notifier for a dummy key. This notifier is
called for each thread exit and destroy.
It is not correct to move the registry address range, since there are
external references from reader threads. This will trigger on workloads
with many threads.
Typically, on Linux, mremap can expand the existing range, which is OK.
However, if there is not enough space around the existing range, it may
try to map it at a different address, which is incorrect.
It is more likely that this bug will be observed on operating systems
where urcu uses the mmap/munmap fallback instead of mremap.
Moreover, prior to commit:
"Fix: urcu-bp: Bulletproof RCU arena resize bug"
this issue was hidden by the fact that each thread ended up with their
own memory mapping (leaked), on Linux at least.
compat_futex.c has one instance included in each urcu shared object, as
well as within some of the test applications. However, it is expected
that an entire program interact with the same lock and completion
variables. Therefore, define them as globally visible, but weak, so the
entire program agree on which object should be used.
Reported-by: Vladimir Nikulichev <nvs@tbricks.com> Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
compat_arch_x86.c is linked into many .so and even into test programs.
The basic problem with this is that it contains a statically defined
mutex, which will fail to protect concurrent use of this compat code by
different shared objects.
Fix this by defining both the mutex (now called __urcu_x86_compat_mutex)
and __rcu_cas_avail as weak symbols. Therefore, the first symbol that
gets loaded in a program will by used by everyone.
Reported-by: Vladimir Nikulichev <nvs@tbricks.com> Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
> While trying to use the BP flavor of RCU I ran into random crashes. I
> tracked it down to issues with resizing of the BP RCU memory pool.
>
> The problem is in the urcu-bp.c file in the resize_arena() function.
> On successful allocation / remapping the len member of the
> registry_arena struct is never set anywhere function. On the second
> resize of the arena the code in resize_arena() still thinks the
> previous size is equal to the original mapping size. I've fixed this
> issue locally by just adding the following code at the bottom of
> resize_arena().
Good catch !!
However, I think your fix misses one case: if we happen to re-use the
same region, we want to update the length too.
Fix: hash table growth (for small tables) should be limited
Buckets with many entries encountered in a hash table could cause it to
grow to a large size, beyond the scope for which this mechanism is
expected to play a role when node accounting is available. Indeed, when
the hash table grows to larger size, split-counter node accounting is
expected to deal with resize/shrink rather than relying on an heuristic
based on the largest bucket size.
This is fixing an issue where we see hash tables sometimes reaching 65k
entries index (65536*8 = 524288 bytes) for a workload limited to adding
1000 entries and then removing all of them, done in a loop (random
keys).
projects / userspace-rcu.git / log
