Implement urcu workqueues internal API
authorMathieu Desnoyers <mathieu.desnoyers@efficios.com>
Tue, 30 May 2017 19:14:58 +0000 (15:14 -0400)
committerMathieu Desnoyers <mathieu.desnoyers@efficios.com>
Wed, 7 Jun 2017 21:58:28 +0000 (17:58 -0400)
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
src/Makefile.am
src/workqueue.c [new file with mode: 0644]
src/workqueue.h [new file with mode: 0644]

index a801020..60b833d 100644 (file)
@@ -54,7 +54,7 @@ liburcu_bp_la_SOURCES = urcu-bp.c urcu-pointer.c $(COMPAT)
 liburcu_bp_la_LIBADD = liburcu-common.la
 
 liburcu_cds_la_SOURCES = rculfqueue.c rculfstack.c lfstack.c \
 liburcu_bp_la_LIBADD = liburcu-common.la
 
 liburcu_cds_la_SOURCES = rculfqueue.c rculfstack.c lfstack.c \
-       $(RCULFHASH) $(COMPAT)
+       workqueue.c workqueue.h $(RCULFHASH) $(COMPAT)
 liburcu_cds_la_LIBADD = liburcu-common.la
 
 pkgconfigdir = $(libdir)/pkgconfig
 liburcu_cds_la_LIBADD = liburcu-common.la
 
 pkgconfigdir = $(libdir)/pkgconfig
diff --git a/src/workqueue.c b/src/workqueue.c
new file mode 100644 (file)
index 0000000..17ea835
--- /dev/null
@@ -0,0 +1,472 @@
+/*
+ * workqueue.c
+ *
+ * Userspace RCU library - Userspace workqeues
+ *
+ * Copyright (c) 2010 Paul E. McKenney <paulmck@linux.vnet.ibm.com>
+ * Copyright (c) 2017 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with this library; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#define _LGPL_SOURCE
+#include <stdio.h>
+#include <pthread.h>
+#include <signal.h>
+#include <assert.h>
+#include <stdlib.h>
+#include <stdint.h>
+#include <string.h>
+#include <errno.h>
+#include <poll.h>
+#include <sys/time.h>
+#include <unistd.h>
+#include <sched.h>
+
+#include "compat-getcpu.h"
+#include "urcu/wfcqueue.h"
+#include "urcu-call-rcu.h"
+#include "urcu-pointer.h"
+#include "urcu/list.h"
+#include "urcu/futex.h"
+#include "urcu/tls-compat.h"
+#include "urcu/ref.h"
+#include "urcu-die.h"
+
+#include "workqueue.h"
+
+#define SET_AFFINITY_CHECK_PERIOD              (1U << 8)       /* 256 */
+#define SET_AFFINITY_CHECK_PERIOD_MASK         (SET_AFFINITY_CHECK_PERIOD - 1)
+
+/* Data structure that identifies a workqueue. */
+
+struct urcu_workqueue {
+       /*
+        * We do not align head on a different cache-line than tail
+        * mainly because call_rcu callback-invocation threads use
+        * batching ("splice") to get an entire list of callbacks, which
+        * effectively empties the queue, and requires to touch the tail
+        * anyway.
+        */
+       struct cds_wfcq_tail cbs_tail;
+       struct cds_wfcq_head cbs_head;
+       unsigned long flags;
+       int32_t futex;
+       unsigned long qlen; /* maintained for debugging. */
+       pthread_t tid;
+       int cpu_affinity;
+       unsigned long loop_count;
+       void *priv;
+       void (*grace_period_fct)(struct urcu_workqueue *workqueue, void *priv);
+       void (*initialize_worker_fct)(struct urcu_workqueue *workqueue, void *priv);
+       void (*finalize_worker_fct)(struct urcu_workqueue *workqueue, void *priv);
+       void (*worker_before_pause_fct)(struct urcu_workqueue *workqueue, void *priv);
+       void (*worker_after_resume_fct)(struct urcu_workqueue *workqueue, void *priv);
+       void (*worker_before_wait_fct)(struct urcu_workqueue *workqueue, void *priv);
+       void (*worker_after_wake_up_fct)(struct urcu_workqueue *workqueue, void *priv);
+} __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
+
+struct urcu_workqueue_completion {
+       int barrier_count;
+       int32_t futex;
+       struct urcu_ref ref;
+};
+
+struct urcu_workqueue_completion_work {
+       struct urcu_work work;
+       struct urcu_workqueue_completion *completion;
+};
+
+/*
+ * Periodically retry setting CPU affinity if we migrate.
+ * Losing affinity can be caused by CPU hotunplug/hotplug, or by
+ * cpuset(7).
+ */
+#if HAVE_SCHED_SETAFFINITY
+static int set_thread_cpu_affinity(struct urcu_workqueue *workqueue)
+{
+       cpu_set_t mask;
+       int ret;
+
+       if (workqueue->cpu_affinity < 0)
+               return 0;
+       if (++workqueue->loop_count & SET_AFFINITY_CHECK_PERIOD_MASK)
+               return 0;
+       if (urcu_sched_getcpu() == workqueue->cpu_affinity)
+               return 0;
+
+       CPU_ZERO(&mask);
+       CPU_SET(workqueue->cpu_affinity, &mask);
+#if SCHED_SETAFFINITY_ARGS == 2
+       ret = sched_setaffinity(0, &mask);
+#else
+       ret = sched_setaffinity(0, sizeof(mask), &mask);
+#endif
+       /*
+        * EINVAL is fine: can be caused by hotunplugged CPUs, or by
+        * cpuset(7). This is why we should always retry if we detect
+        * migration.
+        */
+       if (ret && errno == EINVAL) {
+               ret = 0;
+               errno = 0;
+       }
+       return ret;
+}
+#else
+static int set_thread_cpu_affinity(struct urcu_workqueue *workqueue)
+{
+       return 0;
+}
+#endif
+
+static void futex_wait(int32_t *futex)
+{
+       /* Read condition before read futex */
+       cmm_smp_mb();
+       if (uatomic_read(futex) != -1)
+               return;
+       while (futex_async(futex, FUTEX_WAIT, -1, NULL, NULL, 0)) {
+               switch (errno) {
+               case EWOULDBLOCK:
+                       /* Value already changed. */
+                       return;
+               case EINTR:
+                       /* Retry if interrupted by signal. */
+                       break;  /* Get out of switch. */
+               default:
+                       /* Unexpected error. */
+                       urcu_die(errno);
+               }
+       }
+}
+
+static void futex_wake_up(int32_t *futex)
+{
+       /* Write to condition before reading/writing futex */
+       cmm_smp_mb();
+       if (caa_unlikely(uatomic_read(futex) == -1)) {
+               uatomic_set(futex, 0);
+               if (futex_async(futex, FUTEX_WAKE, 1,
+                               NULL, NULL, 0) < 0)
+                       urcu_die(errno);
+       }
+}
+
+/* This is the code run by each worker thread. */
+
+static void *workqueue_thread(void *arg)
+{
+       unsigned long cbcount;
+       struct urcu_workqueue *workqueue = (struct urcu_workqueue *) arg;
+       int rt = !!(uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_RT);
+
+       if (set_thread_cpu_affinity(workqueue))
+               urcu_die(errno);
+
+       if (workqueue->initialize_worker_fct)
+               workqueue->initialize_worker_fct(workqueue, workqueue->priv);
+
+       if (!rt) {
+               uatomic_dec(&workqueue->futex);
+               /* Decrement futex before reading workqueue */
+               cmm_smp_mb();
+       }
+       for (;;) {
+               struct cds_wfcq_head cbs_tmp_head;
+               struct cds_wfcq_tail cbs_tmp_tail;
+               struct cds_wfcq_node *cbs, *cbs_tmp_n;
+               enum cds_wfcq_ret splice_ret;
+
+               if (set_thread_cpu_affinity(workqueue))
+                       urcu_die(errno);
+
+               if (uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_PAUSE) {
+                       /*
+                        * Pause requested. Become quiescent: remove
+                        * ourself from all global lists, and don't
+                        * process any callback. The callback lists may
+                        * still be non-empty though.
+                        */
+                       if (workqueue->worker_before_pause_fct)
+                               workqueue->worker_before_pause_fct(workqueue, workqueue->priv);
+                       cmm_smp_mb__before_uatomic_or();
+                       uatomic_or(&workqueue->flags, URCU_WORKQUEUE_PAUSED);
+                       while ((uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_PAUSE) != 0)
+                               (void) poll(NULL, 0, 1);
+                       uatomic_and(&workqueue->flags, ~URCU_WORKQUEUE_PAUSED);
+                       cmm_smp_mb__after_uatomic_and();
+                       if (workqueue->worker_after_resume_fct)
+                               workqueue->worker_after_resume_fct(workqueue, workqueue->priv);
+               }
+
+               cds_wfcq_init(&cbs_tmp_head, &cbs_tmp_tail);
+               splice_ret = __cds_wfcq_splice_blocking(&cbs_tmp_head,
+                       &cbs_tmp_tail, &workqueue->cbs_head, &workqueue->cbs_tail);
+               assert(splice_ret != CDS_WFCQ_RET_WOULDBLOCK);
+               assert(splice_ret != CDS_WFCQ_RET_DEST_NON_EMPTY);
+               if (splice_ret != CDS_WFCQ_RET_SRC_EMPTY) {
+                       if (workqueue->grace_period_fct)
+                               workqueue->grace_period_fct(workqueue, workqueue->priv);
+                       cbcount = 0;
+                       __cds_wfcq_for_each_blocking_safe(&cbs_tmp_head,
+                                       &cbs_tmp_tail, cbs, cbs_tmp_n) {
+                               struct rcu_head *rhp;
+
+                               rhp = caa_container_of(cbs,
+                                       struct rcu_head, next);
+                               rhp->func(rhp);
+                               cbcount++;
+                       }
+                       uatomic_sub(&workqueue->qlen, cbcount);
+               }
+               if (uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_STOP)
+                       break;
+               if (workqueue->worker_before_wait_fct)
+                       workqueue->worker_before_wait_fct(workqueue, workqueue->priv);
+               if (!rt) {
+                       if (cds_wfcq_empty(&workqueue->cbs_head,
+                                       &workqueue->cbs_tail)) {
+                               futex_wait(&workqueue->futex);
+                               (void) poll(NULL, 0, 10);
+                               uatomic_dec(&workqueue->futex);
+                               /*
+                                * Decrement futex before reading
+                                * call_rcu list.
+                                */
+                               cmm_smp_mb();
+                       } else {
+                               (void) poll(NULL, 0, 10);
+                       }
+               } else {
+                       (void) poll(NULL, 0, 10);
+               }
+               if (workqueue->worker_after_wake_up_fct)
+                       workqueue->worker_after_wake_up_fct(workqueue, workqueue->priv);
+       }
+       if (!rt) {
+               /*
+                * Read call_rcu list before write futex.
+                */
+               cmm_smp_mb();
+               uatomic_set(&workqueue->futex, 0);
+       }
+       if (workqueue->finalize_worker_fct)
+               workqueue->finalize_worker_fct(workqueue, workqueue->priv);
+       return NULL;
+}
+
+struct urcu_workqueue *urcu_workqueue_create(unsigned long flags,
+               int cpu_affinity, void *priv,
+               void (*grace_period_fct)(struct urcu_workqueue *workqueue, void *priv),
+               void (*initialize_worker_fct)(struct urcu_workqueue *workqueue, void *priv),
+               void (*finalize_worker_fct)(struct urcu_workqueue *workqueue, void *priv),
+               void (*worker_before_wait_fct)(struct urcu_workqueue *workqueue, void *priv),
+               void (*worker_after_wake_up_fct)(struct urcu_workqueue *workqueue, void *priv),
+               void (*worker_before_pause_fct)(struct urcu_workqueue *workqueue, void *priv),
+               void (*worker_after_resume_fct)(struct urcu_workqueue *workqueue, void *priv))
+{
+       struct urcu_workqueue *workqueue;
+       int ret;
+
+       workqueue = malloc(sizeof(*workqueue));
+       if (workqueue == NULL)
+               urcu_die(errno);
+       memset(workqueue, '\0', sizeof(*workqueue));
+       cds_wfcq_init(&workqueue->cbs_head, &workqueue->cbs_tail);
+       workqueue->qlen = 0;
+       workqueue->futex = 0;
+       workqueue->flags = flags;
+       workqueue->priv = priv;
+       workqueue->grace_period_fct = grace_period_fct;
+       workqueue->initialize_worker_fct = initialize_worker_fct;
+       workqueue->finalize_worker_fct = finalize_worker_fct;
+       workqueue->worker_before_wait_fct = worker_before_wait_fct;
+       workqueue->worker_after_wake_up_fct = worker_after_wake_up_fct;
+       workqueue->worker_before_pause_fct = worker_before_pause_fct;
+       workqueue->worker_after_resume_fct = worker_after_resume_fct;
+       workqueue->cpu_affinity = cpu_affinity;
+       workqueue->loop_count = 0;
+       cmm_smp_mb();  /* Structure initialized before pointer is planted. */
+       ret = pthread_create(&workqueue->tid, NULL, workqueue_thread, workqueue);
+       if (ret) {
+               urcu_die(ret);
+       }
+       return workqueue;
+}
+
+static void wake_worker_thread(struct urcu_workqueue *workqueue)
+{
+       if (!(_CMM_LOAD_SHARED(workqueue->flags) & URCU_CALL_RCU_RT))
+               futex_wake_up(&workqueue->futex);
+}
+
+static int urcu_workqueue_destroy_worker(struct urcu_workqueue *workqueue)
+{
+       int ret;
+       void *retval;
+
+       uatomic_or(&workqueue->flags, URCU_WORKQUEUE_STOP);
+       wake_worker_thread(workqueue);
+
+       ret = pthread_join(workqueue->tid, &retval);
+       if (ret) {
+               urcu_die(ret);
+       }
+       if (retval != NULL) {
+               urcu_die(EINVAL);
+       }
+       workqueue->flags &= ~URCU_WORKQUEUE_STOP;
+       workqueue->tid = 0;
+       return 0;
+}
+
+void urcu_workqueue_destroy(struct urcu_workqueue *workqueue)
+{
+       if (workqueue == NULL) {
+               return;
+       }
+       if (urcu_workqueue_destroy_worker(workqueue)) {
+               urcu_die(errno);
+       }
+       assert(cds_wfcq_empty(&workqueue->cbs_head, &workqueue->cbs_tail));
+       free(workqueue);
+}
+
+void urcu_workqueue_queue_work(struct urcu_workqueue *workqueue,
+                     struct urcu_work *work,
+                     void (*func)(struct urcu_work *work))
+{
+       cds_wfcq_node_init(&work->next);
+       work->func = func;
+       cds_wfcq_enqueue(&workqueue->cbs_head, &workqueue->cbs_tail, &work->next);
+       uatomic_inc(&workqueue->qlen);
+       wake_worker_thread(workqueue);
+}
+
+static
+void free_completion(struct urcu_ref *ref)
+{
+       struct urcu_workqueue_completion *completion;
+
+       completion = caa_container_of(ref, struct urcu_workqueue_completion, ref);
+       free(completion);
+}
+
+static
+void _urcu_workqueue_wait_complete(struct urcu_work *work)
+{
+       struct urcu_workqueue_completion_work *completion_work;
+       struct urcu_workqueue_completion *completion;
+
+       completion_work = caa_container_of(work, struct urcu_workqueue_completion_work, work);
+       completion = completion_work->completion;
+       if (!uatomic_sub_return(&completion->barrier_count, 1))
+               futex_wake_up(&completion->futex);
+       urcu_ref_put(&completion->ref, free_completion);
+       free(completion_work);
+}
+
+struct urcu_workqueue_completion *urcu_workqueue_create_completion(void)
+{
+       struct urcu_workqueue_completion *completion;
+
+       completion = calloc(sizeof(*completion), 1);
+       if (!completion)
+               urcu_die(errno);
+       urcu_ref_set(&completion->ref, 1);
+       completion->barrier_count = 0;
+       return completion;
+}
+
+void urcu_workqueue_destroy_completion(struct urcu_workqueue_completion *completion)
+{
+       urcu_ref_put(&completion->ref, free_completion);
+}
+
+void urcu_workqueue_wait_completion(struct urcu_workqueue_completion *completion)
+{
+       /* Wait for them */
+       for (;;) {
+               uatomic_dec(&completion->futex);
+               /* Decrement futex before reading barrier_count */
+               cmm_smp_mb();
+               if (!uatomic_read(&completion->barrier_count))
+                       break;
+               futex_wait(&completion->futex);
+       }
+}
+
+void urcu_workqueue_queue_completion(struct urcu_workqueue *workqueue,
+               struct urcu_workqueue_completion *completion)
+{
+       struct urcu_workqueue_completion_work *work;
+
+       work = calloc(sizeof(*work), 1);
+       if (!work)
+               urcu_die(errno);
+       work->completion = completion;
+       urcu_ref_get(&completion->ref);
+       uatomic_inc(&completion->barrier_count);
+       urcu_workqueue_queue_work(workqueue, &work->work, _urcu_workqueue_wait_complete);
+}
+
+/*
+ * Wait for all in-flight work to complete execution.
+ */
+void urcu_workqueue_flush_queued_work(struct urcu_workqueue *workqueue)
+{
+       struct urcu_workqueue_completion *completion;
+
+       completion = urcu_workqueue_create_completion();
+       if (!completion)
+               urcu_die(ENOMEM);
+       urcu_workqueue_queue_completion(workqueue, completion);
+       urcu_workqueue_wait_completion(completion);
+       urcu_workqueue_destroy_completion(completion);
+}
+
+/* To be used in before fork handler. */
+void urcu_workqueue_pause_worker(struct urcu_workqueue *workqueue)
+{
+       uatomic_or(&workqueue->flags, URCU_WORKQUEUE_PAUSE);
+       cmm_smp_mb__after_uatomic_or();
+       wake_worker_thread(workqueue);
+
+       while ((uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_PAUSED) == 0)
+               (void) poll(NULL, 0, 1);
+}
+
+/* To be used in after fork parent handler. */
+void urcu_workqueue_resume_worker(struct urcu_workqueue *workqueue)
+{
+       uatomic_and(&workqueue->flags, ~URCU_WORKQUEUE_PAUSE);
+       while ((uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_PAUSED) != 0)
+               (void) poll(NULL, 0, 1);
+}
+
+void urcu_workqueue_create_worker(struct urcu_workqueue *workqueue)
+{
+       int ret;
+
+       /* Clear workqueue state from parent. */
+       workqueue->flags &= ~URCU_WORKQUEUE_PAUSED;
+       workqueue->flags &= ~URCU_WORKQUEUE_PAUSE;
+       workqueue->tid = 0;
+       ret = pthread_create(&workqueue->tid, NULL, workqueue_thread, workqueue);
+       if (ret) {
+               urcu_die(ret);
+       }
+}
diff --git a/src/workqueue.h b/src/workqueue.h
new file mode 100644 (file)
index 0000000..52b6973
--- /dev/null
@@ -0,0 +1,104 @@
+#ifndef _URCU_WORKQUEUE_H
+#define _URCU_WORKQUEUE_H
+
+/*
+ * workqueue.h
+ *
+ * Userspace RCU header - Userspace workqueues
+ *
+ * Copyright (c) 2009,2017 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
+ * Copyright (c) 2009 Paul E. McKenney, IBM Corporation.
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with this library; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include <stdlib.h>
+#include <pthread.h>
+
+#include <urcu/wfcqueue.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/* Note that struct urcu_workqueue is opaque to callers. */
+
+struct urcu_workqueue;
+struct urcu_workqueue_completion;
+
+/* Flag values. */
+
+#define URCU_WORKQUEUE_RT      (1U << 0)
+#define URCU_WORKQUEUE_STOP    (1U << 1)
+#define URCU_WORKQUEUE_PAUSE   (1U << 2)
+#define URCU_WORKQUEUE_PAUSED  (1U << 3)
+
+/*
+ * The urcu_work data structure is placed in the structure to be acted
+ * upon via urcu_workqueue_queue_work().
+ */
+
+struct urcu_work {
+       struct cds_wfcq_node next;
+       void (*func)(struct urcu_work *head);
+};
+
+/*
+ * Exported functions
+ */
+
+struct urcu_workqueue *urcu_workqueue_create(unsigned long flags,
+               int cpu_affinity, void *priv,
+               void (*grace_period_fct)(struct urcu_workqueue *workqueue, void *priv),
+               void (*initialize_worker_fct)(struct urcu_workqueue *workqueue, void *priv),
+               void (*finalize_worker_fct)(struct urcu_workqueue *workqueue, void *priv),
+               void (*worker_before_wait_fct)(struct urcu_workqueue *workqueue, void *priv),
+               void (*worker_after_wake_up_fct)(struct urcu_workqueue *workqueue, void *priv),
+               void (*worker_before_pause_fct)(struct urcu_workqueue *workqueue, void *priv),
+               void (*worker_after_resume_fct)(struct urcu_workqueue *workqueue, void *priv));
+void urcu_workqueue_destroy(struct urcu_workqueue *workqueue);
+
+/*
+ * Never fails. Should not be used to enqueue work from worker threads
+ * after the application invokes urcu_workqueue_free.
+ */
+void urcu_workqueue_queue_work(struct urcu_workqueue *workqueue,
+               struct urcu_work *work,
+               void (*func)(struct urcu_work *work));
+
+struct urcu_workqueue_completion *urcu_workqueue_create_completion(void);
+void urcu_workqueue_destroy_completion(struct urcu_workqueue_completion *completion);
+
+void urcu_workqueue_queue_completion(struct urcu_workqueue *workqueue,
+               struct urcu_workqueue_completion *completion);
+void urcu_workqueue_wait_completion(struct urcu_workqueue_completion *completion);
+
+void urcu_workqueue_flush_queued_work(struct urcu_workqueue *workqueue);
+
+/*
+ * pause/resume/create worker threads. Can be used to pause worker
+ * threads across fork/clone while keeping the workqueue in place.
+ * Pause is used in parent pre-fork, resume in parent post-fork, create
+ * in child after-fork.
+ */
+void urcu_workqueue_pause_worker(struct urcu_workqueue *workqueue);
+void urcu_workqueue_resume_worker(struct urcu_workqueue *workqueue);
+void urcu_workqueue_create_worker(struct urcu_workqueue *workqueue);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* _URCU_WORKQUEUE_H */
This page took 0.042025 seconds and 4 git commands to generate.