/*
* ring_buffer_frontend.c
*
- * (C) Copyright 2005-2010 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
+ * Copyright (C) 2005-2012 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; only
+ * version 2.1 of the License.
+ *
+ * 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
+ *
*
* Ring buffer wait-free buffer synchronization. Producer-consumer and flight
* recorder (overwrite) modes. See thesis:
* - splice one subbuffer worth of data to a pipe
* - splice the data from pipe to disk/network
* - put_subbuf
- *
- * Dual LGPL v2.1/GPL v2 license.
*/
#include <linux/delay.h>
void lib_ring_buffer_reset(struct lib_ring_buffer *buf)
{
struct channel *chan = buf->backend.chan;
- const struct lib_ring_buffer_config *config = chan->backend.config;
+ const struct lib_ring_buffer_config *config = &chan->backend.config;
unsigned int i;
/*
int lib_ring_buffer_create(struct lib_ring_buffer *buf,
struct channel_backend *chanb, int cpu)
{
- const struct lib_ring_buffer_config *config = chanb->config;
+ const struct lib_ring_buffer_config *config = &chanb->config;
struct channel *chan = container_of(chanb, struct channel, backend);
void *priv = chanb->priv;
- unsigned int num_subbuf;
size_t subbuf_header_size;
u64 tsc;
int ret;
goto free_commit;
}
- num_subbuf = chan->backend.num_subbuf;
init_waitqueue_head(&buf->read_wait);
+ init_waitqueue_head(&buf->write_wait);
raw_spin_lock_init(&buf->raw_tick_nohz_spinlock);
/*
{
struct lib_ring_buffer *buf = (struct lib_ring_buffer *)data;
struct channel *chan = buf->backend.chan;
- const struct lib_ring_buffer_config *config = chan->backend.config;
+ const struct lib_ring_buffer_config *config = &chan->backend.config;
/*
* Only flush buffers periodically if readers are active.
static void lib_ring_buffer_start_switch_timer(struct lib_ring_buffer *buf)
{
struct channel *chan = buf->backend.chan;
- const struct lib_ring_buffer_config *config = chan->backend.config;
+ const struct lib_ring_buffer_config *config = &chan->backend.config;
if (!chan->switch_timer_interval || buf->switch_timer_enabled)
return;
{
struct lib_ring_buffer *buf = (struct lib_ring_buffer *)data;
struct channel *chan = buf->backend.chan;
- const struct lib_ring_buffer_config *config = chan->backend.config;
+ const struct lib_ring_buffer_config *config = &chan->backend.config;
CHAN_WARN_ON(chan, !buf->backend.allocated);
static void lib_ring_buffer_start_read_timer(struct lib_ring_buffer *buf)
{
struct channel *chan = buf->backend.chan;
- const struct lib_ring_buffer_config *config = chan->backend.config;
+ const struct lib_ring_buffer_config *config = &chan->backend.config;
if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER
|| !chan->read_timer_interval
static void lib_ring_buffer_stop_read_timer(struct lib_ring_buffer *buf)
{
struct channel *chan = buf->backend.chan;
- const struct lib_ring_buffer_config *config = chan->backend.config;
+ const struct lib_ring_buffer_config *config = &chan->backend.config;
if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER
|| !chan->read_timer_interval
struct channel *chan = container_of(nb, struct channel,
cpu_hp_notifier);
struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, cpu);
- const struct lib_ring_buffer_config *config = chan->backend.config;
+ const struct lib_ring_buffer_config *config = &chan->backend.config;
if (!chan->cpu_hp_enable)
return NOTIFY_DONE;
case CPU_DOWN_FAILED_FROZEN:
case CPU_ONLINE:
case CPU_ONLINE_FROZEN:
+ wake_up_interruptible(&chan->hp_wait);
lib_ring_buffer_start_switch_timer(buf);
lib_ring_buffer_start_read_timer(buf);
return NOTIFY_OK;
}
#endif
-#if defined(CONFIG_NO_HZ) && !defined(CONFIG_LIB_RING_BUFFER)
+#if defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER)
/*
* For per-cpu buffers, call the reader wakeups before switching the buffer, so
* that wake-up-tracing generated events are flushed before going idle (in
{
struct channel *chan = container_of(nb, struct channel,
tick_nohz_notifier);
- const struct lib_ring_buffer_config *config = chan->backend.config;
+ const struct lib_ring_buffer_config *config = &chan->backend.config;
struct lib_ring_buffer *buf;
int cpu = smp_processor_id();
atomic_notifier_call_chain(&tick_nohz_notifier, TICK_NOHZ_RESTART,
NULL);
}
-#endif /* defined(CONFIG_NO_HZ) && !defined(CONFIG_LIB_RING_BUFFER) */
+#endif /* defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) */
/*
* Holds CPU hotplug.
*/
static void channel_unregister_notifiers(struct channel *chan)
{
- const struct lib_ring_buffer_config *config = chan->backend.config;
+ const struct lib_ring_buffer_config *config = &chan->backend.config;
int cpu;
channel_iterator_unregister_notifiers(chan);
chan->commit_count_mask = (~0UL >> chan->backend.num_subbuf_order);
chan->switch_timer_interval = usecs_to_jiffies(switch_timer_interval);
chan->read_timer_interval = usecs_to_jiffies(read_timer_interval);
+ kref_init(&chan->ref);
init_waitqueue_head(&chan->read_wait);
+ init_waitqueue_head(&chan->hp_wait);
if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) {
-#ifdef CONFIG_NO_HZ
+#if defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER)
/* Only benefit from NO_HZ idle with per-cpu buffers for now. */
chan->tick_nohz_notifier.notifier_call =
ring_buffer_tick_nohz_callback;
chan->tick_nohz_notifier.priority = ~0U;
atomic_notifier_chain_register(&tick_nohz_notifier,
&chan->tick_nohz_notifier);
-#endif /* CONFIG_NO_HZ */
+#endif /* defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) */
/*
* In case of non-hotplug cpu, if the ring-buffer is allocated
}
EXPORT_SYMBOL_GPL(channel_create);
+static
+void channel_release(struct kref *kref)
+{
+ struct channel *chan = container_of(kref, struct channel, ref);
+ channel_free(chan);
+}
+
/**
* channel_destroy - Finalize, wait for q.s. and destroy channel.
* @chan: channel to destroy
*
* Holds cpu hotplug.
- * Call "destroy" callback, finalize channels, wait for readers to release their
- * reference, then destroy ring buffer data. Note that when readers have
- * completed data consumption of finalized channels, get_subbuf() will return
- * -ENODATA. They should release their handle at that point.
- * Returns the private data pointer.
+ * Call "destroy" callback, finalize channels, and then decrement the
+ * channel reference count. Note that when readers have completed data
+ * consumption of finalized channels, get_subbuf() will return -ENODATA.
+ * They should release their handle at that point. Returns the private
+ * data pointer.
*/
void *channel_destroy(struct channel *chan)
{
int cpu;
- const struct lib_ring_buffer_config *config = chan->backend.config;
+ const struct lib_ring_buffer_config *config = &chan->backend.config;
void *priv;
channel_unregister_notifiers(chan);
ACCESS_ONCE(buf->finalized) = 1;
wake_up_interruptible(&buf->read_wait);
}
+ ACCESS_ONCE(chan->finalized) = 1;
+ wake_up_interruptible(&chan->hp_wait);
wake_up_interruptible(&chan->read_wait);
-
- while (atomic_long_read(&chan->read_ref) > 0)
- msleep(100);
- /* Finish waiting for refcount before free */
- smp_mb();
priv = chan->backend.priv;
- channel_free(chan);
+ kref_put(&chan->ref, channel_release);
return priv;
}
EXPORT_SYMBOL_GPL(channel_destroy);
if (!atomic_long_add_unless(&buf->active_readers, 1, 1))
return -EBUSY;
- atomic_long_inc(&chan->read_ref);
+ kref_get(&chan->ref);
smp_mb__after_atomic_inc();
return 0;
}
CHAN_WARN_ON(chan, atomic_long_read(&buf->active_readers) != 1);
smp_mb__before_atomic_dec();
- atomic_long_dec(&chan->read_ref);
atomic_long_dec(&buf->active_readers);
+ kref_put(&chan->ref, channel_release);
}
EXPORT_SYMBOL_GPL(lib_ring_buffer_release_read);
unsigned long *consumed, unsigned long *produced)
{
struct channel *chan = buf->backend.chan;
- const struct lib_ring_buffer_config *config = chan->backend.config;
+ const struct lib_ring_buffer_config *config = &chan->backend.config;
unsigned long consumed_cur, write_offset;
int finalized;
/**
* lib_ring_buffer_put_snapshot - move consumed counter forward
+ *
+ * Should only be called from consumer context.
* @buf: ring buffer
* @consumed_new: new consumed count value
*/
while ((long) consumed - (long) consumed_new < 0)
consumed = atomic_long_cmpxchg(&buf->consumed, consumed,
consumed_new);
+ /* Wake-up the metadata producer */
+ wake_up_interruptible(&buf->write_wait);
}
EXPORT_SYMBOL_GPL(lib_ring_buffer_move_consumer);
unsigned long consumed)
{
struct channel *chan = buf->backend.chan;
- const struct lib_ring_buffer_config *config = chan->backend.config;
+ const struct lib_ring_buffer_config *config = &chan->backend.config;
unsigned long consumed_cur, consumed_idx, commit_count, write_offset;
int ret;
int finalized;
{
struct lib_ring_buffer_backend *bufb = &buf->backend;
struct channel *chan = bufb->chan;
- const struct lib_ring_buffer_config *config = chan->backend.config;
+ const struct lib_ring_buffer_config *config = &chan->backend.config;
unsigned long read_sb_bindex, consumed_idx, consumed;
CHAN_WARN_ON(chan, atomic_long_read(&buf->active_readers) != 1);
unsigned long cons_offset,
int cpu)
{
- const struct lib_ring_buffer_config *config = chan->backend.config;
+ const struct lib_ring_buffer_config *config = &chan->backend.config;
unsigned long cons_idx, commit_count, commit_count_sb;
cons_idx = subbuf_index(cons_offset, chan);
struct channel *chan,
void *priv, int cpu)
{
- const struct lib_ring_buffer_config *config = chan->backend.config;
+ const struct lib_ring_buffer_config *config = &chan->backend.config;
unsigned long write_offset, cons_offset;
- /*
- * Can be called in the error path of allocation when
- * trans_channel_data is not yet set.
- */
- if (!chan)
- return;
/*
* No need to order commit_count, write_offset and cons_offset reads
* because we execute at teardown when no more writer nor reader
write_offset = v_read(config, &buf->offset);
cons_offset = atomic_long_read(&buf->consumed);
if (write_offset != cons_offset)
- printk(KERN_WARNING
+ printk(KERN_DEBUG
"ring buffer %s, cpu %d: "
"non-consumed data\n"
" [ %lu bytes written, %lu bytes read ]\n",
void lib_ring_buffer_print_errors(struct channel *chan,
struct lib_ring_buffer *buf, int cpu)
{
- const struct lib_ring_buffer_config *config = chan->backend.config;
+ const struct lib_ring_buffer_config *config = &chan->backend.config;
void *priv = chan->backend.priv;
- printk(KERN_DEBUG "ring buffer %s, cpu %d: %lu records written, "
- "%lu records overrun\n",
- chan->backend.name, cpu,
- v_read(config, &buf->records_count),
- v_read(config, &buf->records_overrun));
-
- if (v_read(config, &buf->records_lost_full)
- || v_read(config, &buf->records_lost_wrap)
- || v_read(config, &buf->records_lost_big))
- printk(KERN_WARNING
- "ring buffer %s, cpu %d: records were lost. Caused by:\n"
- " [ %lu buffer full, %lu nest buffer wrap-around, "
- "%lu event too big ]\n",
- chan->backend.name, cpu,
- v_read(config, &buf->records_lost_full),
- v_read(config, &buf->records_lost_wrap),
- v_read(config, &buf->records_lost_big));
-
+ if (!strcmp(chan->backend.name, "relay-metadata")) {
+ printk(KERN_DEBUG "ring buffer %s: %lu records written, "
+ "%lu records overrun\n",
+ chan->backend.name,
+ v_read(config, &buf->records_count),
+ v_read(config, &buf->records_overrun));
+ } else {
+ printk(KERN_DEBUG "ring buffer %s, cpu %d: %lu records written, "
+ "%lu records overrun\n",
+ chan->backend.name, cpu,
+ v_read(config, &buf->records_count),
+ v_read(config, &buf->records_overrun));
+
+ if (v_read(config, &buf->records_lost_full)
+ || v_read(config, &buf->records_lost_wrap)
+ || v_read(config, &buf->records_lost_big))
+ printk(KERN_WARNING
+ "ring buffer %s, cpu %d: records were lost. Caused by:\n"
+ " [ %lu buffer full, %lu nest buffer wrap-around, "
+ "%lu event too big ]\n",
+ chan->backend.name, cpu,
+ v_read(config, &buf->records_lost_full),
+ v_read(config, &buf->records_lost_wrap),
+ v_read(config, &buf->records_lost_big));
+ }
lib_ring_buffer_print_buffer_errors(buf, chan, priv, cpu);
}
struct switch_offsets *offsets,
u64 tsc)
{
- const struct lib_ring_buffer_config *config = chan->backend.config;
+ const struct lib_ring_buffer_config *config = &chan->backend.config;
unsigned long oldidx = subbuf_index(offsets->old, chan);
unsigned long commit_count;
struct switch_offsets *offsets,
u64 tsc)
{
- const struct lib_ring_buffer_config *config = chan->backend.config;
+ const struct lib_ring_buffer_config *config = &chan->backend.config;
unsigned long oldidx = subbuf_index(offsets->old - 1, chan);
unsigned long commit_count, padding_size, data_size;
struct switch_offsets *offsets,
u64 tsc)
{
- const struct lib_ring_buffer_config *config = chan->backend.config;
+ const struct lib_ring_buffer_config *config = &chan->backend.config;
unsigned long beginidx = subbuf_index(offsets->begin, chan);
unsigned long commit_count;
struct switch_offsets *offsets,
u64 tsc)
{
- const struct lib_ring_buffer_config *config = chan->backend.config;
+ const struct lib_ring_buffer_config *config = &chan->backend.config;
unsigned long endidx = subbuf_index(offsets->end - 1, chan);
unsigned long commit_count, padding_size, data_size;
struct switch_offsets *offsets,
u64 *tsc)
{
- const struct lib_ring_buffer_config *config = chan->backend.config;
+ const struct lib_ring_buffer_config *config = &chan->backend.config;
unsigned long off;
offsets->begin = v_read(config, &buf->offset);
void lib_ring_buffer_switch_slow(struct lib_ring_buffer *buf, enum switch_mode mode)
{
struct channel *chan = buf->backend.chan;
- const struct lib_ring_buffer_config *config = chan->backend.config;
+ const struct lib_ring_buffer_config *config = &chan->backend.config;
struct switch_offsets offsets;
unsigned long oldidx;
u64 tsc;
/*
* Returns :
* 0 if ok
- * !0 if execution must be aborted.
+ * -ENOSPC if event size is too large for packet.
+ * -ENOBUFS if there is currently not enough space in buffer for the event.
+ * -EIO if data cannot be written into the buffer for any other reason.
*/
static
int lib_ring_buffer_try_reserve_slow(struct lib_ring_buffer *buf,
struct switch_offsets *offsets,
struct lib_ring_buffer_ctx *ctx)
{
- const struct lib_ring_buffer_config *config = chan->backend.config;
+ const struct lib_ring_buffer_config *config = &chan->backend.config;
unsigned long reserve_commit_diff;
offsets->begin = v_read(config, &buf->offset);
offsets->pre_header_padding = 0;
ctx->tsc = config->cb.ring_buffer_clock_read(chan);
+ if ((int64_t) ctx->tsc == -EIO)
+ return -EIO;
if (last_tsc_overflow(config, buf, ctx->tsc))
- ctx->rflags = RING_BUFFER_RFLAG_FULL_TSC;
+ ctx->rflags |= RING_BUFFER_RFLAG_FULL_TSC;
if (unlikely(subbuf_offset(offsets->begin, ctx->chan) == 0)) {
offsets->switch_new_start = 1; /* For offsets->begin */
} else {
offsets->size = config->cb.record_header_size(config, chan,
offsets->begin,
- ctx->data_size,
&offsets->pre_header_padding,
- ctx->rflags, ctx);
+ ctx);
offsets->size +=
lib_ring_buffer_align(offsets->begin + offsets->size,
ctx->largest_align)
* and we are full : record is lost.
*/
v_inc(config, &buf->records_lost_full);
- return -1;
+ return -ENOBUFS;
} else {
/*
* Next subbuffer not being written to, and we
* many nested writes over a reserve/commit pair.
*/
v_inc(config, &buf->records_lost_wrap);
- return -1;
+ return -EIO;
}
offsets->size =
config->cb.record_header_size(config, chan,
offsets->begin,
- ctx->data_size,
&offsets->pre_header_padding,
- ctx->rflags, ctx);
+ ctx);
offsets->size +=
lib_ring_buffer_align(offsets->begin + offsets->size,
ctx->largest_align)
* complete the sub-buffer switch.
*/
v_inc(config, &buf->records_lost_big);
- return -1;
+ return -ENOSPC;
} else {
/*
* We just made a successful buffer switch and the
* lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer.
* @ctx: ring buffer context.
*
- * Return : -ENOSPC if not enough space, else returns 0.
+ * Return : -NOBUFS if not enough space, -ENOSPC if event size too large,
+ * -EIO for other errors, else returns 0.
* It will take care of sub-buffer switching.
*/
int lib_ring_buffer_reserve_slow(struct lib_ring_buffer_ctx *ctx)
{
struct channel *chan = ctx->chan;
- const struct lib_ring_buffer_config *config = chan->backend.config;
+ const struct lib_ring_buffer_config *config = &chan->backend.config;
struct lib_ring_buffer *buf;
struct switch_offsets offsets;
+ int ret;
if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
buf = per_cpu_ptr(chan->backend.buf, ctx->cpu);
offsets.size = 0;
do {
- if (unlikely(lib_ring_buffer_try_reserve_slow(buf, chan, &offsets,
- ctx)))
- return -ENOSPC;
+ ret = lib_ring_buffer_try_reserve_slow(buf, chan, &offsets,
+ ctx);
+ if (unlikely(ret))
+ return ret;
} while (unlikely(v_cmpxchg(config, &buf->offset, offsets.old,
offsets.end)
!= offsets.old));
return 0;
}
EXPORT_SYMBOL_GPL(lib_ring_buffer_reserve_slow);
+
+int __init init_lib_ring_buffer_frontend(void)
+{
+ int cpu;
+
+ for_each_possible_cpu(cpu)
+ spin_lock_init(&per_cpu(ring_buffer_nohz_lock, cpu));
+ return 0;
+}
+
+module_init(init_lib_ring_buffer_frontend);
+
+void __exit exit_lib_ring_buffer_frontend(void)
+{
+}
+
+module_exit(exit_lib_ring_buffer_frontend);