/*
* 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
* Returns the success/failure of the operation. (%NOTIFY_OK, %NOTIFY_BAD)
*/
static
-int __cpuinit lib_ring_buffer_cpu_hp_callback(struct notifier_block *nb,
+int lib_ring_buffer_cpu_hp_callback(struct notifier_block *nb,
unsigned long action,
void *hcpu)
{
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;
{
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();
*/
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);
static void channel_free(struct channel *chan)
{
+ if (chan->backend.release_priv_ops) {
+ chan->backend.release_priv_ops(chan->backend.priv_ops);
+ }
channel_iterator_free(chan);
channel_backend_free(&chan->backend);
kfree(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) {
#if defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER)
}
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;
+ if (buf->get_subbuf) {
+ /*
+ * Reader is trying to get a subbuffer twice.
+ */
+ CHAN_WARN_ON(chan, 1);
+ return -EBUSY;
+ }
retry:
finalized = ACCESS_ONCE(buf->finalized);
/*
*/
if (((commit_count - chan->backend.subbuf_size)
& chan->commit_count_mask)
- - (buf_trunc(consumed_cur, chan)
+ - (buf_trunc(consumed, chan)
>> chan->backend.num_subbuf_order)
!= 0)
goto nodata;
* Check that we are not about to read the same subbuffer in
* which the writer head is.
*/
- if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_cur, chan)
+ if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed, chan)
== 0)
goto nodata;
{
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;
commit_count = v_read(config, &buf->commit_hot[oldidx].cc);
/* Check if the written buffer has to be delivered */
lib_ring_buffer_check_deliver(config, buf, chan, offsets->old,
- commit_count, oldidx);
+ commit_count, oldidx, tsc);
lib_ring_buffer_write_commit_counter(config, buf, chan, oldidx,
- offsets->old, commit_count,
- config->cb.subbuffer_header_size());
+ offsets->old + config->cb.subbuffer_header_size(),
+ 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;
v_add(config, padding_size, &buf->commit_hot[oldidx].cc);
commit_count = v_read(config, &buf->commit_hot[oldidx].cc);
lib_ring_buffer_check_deliver(config, buf, chan, offsets->old - 1,
- commit_count, oldidx);
+ commit_count, oldidx, tsc);
lib_ring_buffer_write_commit_counter(config, buf, chan, oldidx,
- offsets->old, commit_count,
- padding_size);
+ offsets->old + padding_size, 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 beginidx = subbuf_index(offsets->begin, chan);
unsigned long commit_count;
commit_count = v_read(config, &buf->commit_hot[beginidx].cc);
/* Check if the written buffer has to be delivered */
lib_ring_buffer_check_deliver(config, buf, chan, offsets->begin,
- commit_count, beginidx);
+ commit_count, beginidx, tsc);
lib_ring_buffer_write_commit_counter(config, buf, chan, beginidx,
- offsets->begin, commit_count,
- config->cb.subbuffer_header_size());
+ offsets->begin + config->cb.subbuffer_header_size(),
+ commit_count);
}
/*
* lib_ring_buffer_switch_new_end: finish switching current subbuffer
*
- * The only remaining threads could be the ones with pending commits. They will
- * have to do the deliver themselves.
+ * Calls subbuffer_set_data_size() to set the data size of the current
+ * sub-buffer. We do not need to perform check_deliver nor commit here,
+ * since this task will be done by the "commit" of the event for which
+ * we are currently doing the space reservation.
*/
static
void lib_ring_buffer_switch_new_end(struct lib_ring_buffer *buf,
struct switch_offsets *offsets,
u64 tsc)
{
- 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;
+ const struct lib_ring_buffer_config *config = &chan->backend.config;
+ unsigned long endidx, data_size;
+ endidx = subbuf_index(offsets->end - 1, chan);
data_size = subbuf_offset(offsets->end - 1, chan) + 1;
- padding_size = chan->backend.subbuf_size - data_size;
subbuffer_set_data_size(config, &buf->backend, endidx, data_size);
-
- /*
- * Order all writes to buffer before the commit count update that will
- * determine that the subbuffer is full.
- */
- if (config->ipi == RING_BUFFER_IPI_BARRIER) {
- /*
- * Must write slot data before incrementing commit count. This
- * compiler barrier is upgraded into a smp_mb() by the IPI sent
- * by get_subbuf().
- */
- barrier();
- } else
- smp_wmb();
- v_add(config, padding_size, &buf->commit_hot[endidx].cc);
- commit_count = v_read(config, &buf->commit_hot[endidx].cc);
- lib_ring_buffer_check_deliver(config, buf, chan, offsets->end - 1,
- commit_count, endidx);
- lib_ring_buffer_write_commit_counter(config, buf, chan, endidx,
- offsets->end, commit_count,
- padding_size);
}
/*
struct switch_offsets *offsets,
u64 *tsc)
{
- const struct lib_ring_buffer_config *config = chan->backend.config;
- unsigned long off;
+ const struct lib_ring_buffer_config *config = &chan->backend.config;
+ unsigned long off, reserve_commit_diff;
offsets->begin = v_read(config, &buf->offset);
offsets->old = offsets->begin;
* (records and header timestamps) are visible to the reader. This is
* required for quiescence guarantees for the fusion merge.
*/
- if (mode == SWITCH_FLUSH || off > 0) {
- if (unlikely(off == 0)) {
- /*
- * The client does not save any header information.
- * Don't switch empty subbuffer on finalize, because it
- * is invalid to deliver a completely empty subbuffer.
- */
- if (!config->cb.subbuffer_header_size())
+ if (mode != SWITCH_FLUSH && !off)
+ return -1; /* we do not have to switch : buffer is empty */
+
+ if (unlikely(off == 0)) {
+ unsigned long sb_index, commit_count;
+
+ /*
+ * We are performing a SWITCH_FLUSH. At this stage, there are no
+ * concurrent writes into the buffer.
+ *
+ * The client does not save any header information. Don't
+ * switch empty subbuffer on finalize, because it is invalid to
+ * deliver a completely empty subbuffer.
+ */
+ if (!config->cb.subbuffer_header_size())
+ return -1;
+
+ /* Test new buffer integrity */
+ sb_index = subbuf_index(offsets->begin, chan);
+ commit_count = v_read(config,
+ &buf->commit_cold[sb_index].cc_sb);
+ reserve_commit_diff =
+ (buf_trunc(offsets->begin, chan)
+ >> chan->backend.num_subbuf_order)
+ - (commit_count & chan->commit_count_mask);
+ if (likely(reserve_commit_diff == 0)) {
+ /* Next subbuffer not being written to. */
+ if (unlikely(config->mode != RING_BUFFER_OVERWRITE &&
+ subbuf_trunc(offsets->begin, chan)
+ - subbuf_trunc((unsigned long)
+ atomic_long_read(&buf->consumed), chan)
+ >= chan->backend.buf_size)) {
+ /*
+ * We do not overwrite non consumed buffers
+ * and we are full : don't switch.
+ */
return -1;
+ } else {
+ /*
+ * Next subbuffer not being written to, and we
+ * are either in overwrite mode or the buffer is
+ * not full. It's safe to write in this new
+ * subbuffer.
+ */
+ }
+ } else {
/*
- * Need to write the subbuffer start header on finalize.
+ * Next subbuffer reserve offset does not match the
+ * commit offset. Don't perform switch in
+ * producer-consumer and overwrite mode. Caused by
+ * either a writer OOPS or too many nested writes over a
+ * reserve/commit pair.
*/
- offsets->switch_old_start = 1;
+ return -1;
}
- offsets->begin = subbuf_align(offsets->begin, chan);
- } else
- return -1; /* we do not have to switch : buffer is empty */
+
+ /*
+ * Need to write the subbuffer start header on finalize.
+ */
+ offsets->switch_old_start = 1;
+ }
+ offsets->begin = subbuf_align(offsets->begin, chan);
/* Note: old points to the next subbuf at offset 0 */
offsets->end = offsets->begin;
return 0;
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;
}
EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_slow);
+static void remote_switch(void *info)
+{
+ struct lib_ring_buffer *buf = info;
+
+ lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE);
+}
+
+void lib_ring_buffer_switch_remote(struct lib_ring_buffer *buf)
+{
+ struct channel *chan = buf->backend.chan;
+ const struct lib_ring_buffer_config *config = &chan->backend.config;
+ int ret;
+
+ /*
+ * With global synchronization we don't need to use the IPI scheme.
+ */
+ if (config->sync == RING_BUFFER_SYNC_GLOBAL) {
+ lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE);
+ return;
+ }
+
+ /*
+ * Taking lock on CPU hotplug to ensure two things: first, that the
+ * target cpu is not taken concurrently offline while we are within
+ * smp_call_function_single() (I don't trust that get_cpu() on the
+ * _local_ CPU actually inhibit CPU hotplug for the _remote_ CPU (to be
+ * confirmed)). Secondly, if it happens that the CPU is not online, our
+ * own call to lib_ring_buffer_switch_slow() needs to be protected from
+ * CPU hotplug handlers, which can also perform a remote subbuffer
+ * switch.
+ */
+ get_online_cpus();
+ ret = smp_call_function_single(buf->backend.cpu,
+ remote_switch, buf, 1);
+ if (ret) {
+ /* Remote CPU is offline, do it ourself. */
+ lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE);
+ }
+ put_online_cpus();
+}
+EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_remote);
+
/*
* 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;
- unsigned long reserve_commit_diff;
+ const struct lib_ring_buffer_config *config = &chan->backend.config;
+ unsigned long reserve_commit_diff, offset_cmp;
- offsets->begin = v_read(config, &buf->offset);
+retry:
+ offsets->begin = offset_cmp = v_read(config, &buf->offset);
offsets->old = offsets->begin;
offsets->switch_new_start = 0;
offsets->switch_new_end = 0;
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)
}
}
if (unlikely(offsets->switch_new_start)) {
- unsigned long sb_index;
+ unsigned long sb_index, commit_count;
/*
* We are typically not filling the previous buffer completely.
+ config->cb.subbuffer_header_size();
/* Test new buffer integrity */
sb_index = subbuf_index(offsets->begin, chan);
+ /*
+ * Read buf->offset before buf->commit_cold[sb_index].cc_sb.
+ * lib_ring_buffer_check_deliver() has the matching
+ * memory barriers required around commit_cold cc_sb
+ * updates to ensure reserve and commit counter updates
+ * are not seen reordered when updated by another CPU.
+ */
+ smp_rmb();
+ commit_count = v_read(config,
+ &buf->commit_cold[sb_index].cc_sb);
+ /* Read buf->commit_cold[sb_index].cc_sb before buf->offset. */
+ smp_rmb();
+ if (unlikely(offset_cmp != v_read(config, &buf->offset))) {
+ /*
+ * The reserve counter have been concurrently updated
+ * while we read the commit counter. This means the
+ * commit counter we read might not match buf->offset
+ * due to concurrent update. We therefore need to retry.
+ */
+ goto retry;
+ }
reserve_commit_diff =
(buf_trunc(offsets->begin, chan)
>> chan->backend.num_subbuf_order)
- - ((unsigned long) v_read(config,
- &buf->commit_cold[sb_index].cc_sb)
- & chan->commit_count_mask);
+ - (commit_count & chan->commit_count_mask);
if (likely(reserve_commit_diff == 0)) {
/* Next subbuffer not being written to. */
if (unlikely(config->mode != RING_BUFFER_OVERWRITE &&
* 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
} else {
/*
* Next subbuffer reserve offset does not match the
- * commit offset. Drop record in producer-consumer and
- * overwrite mode. Caused by either a writer OOPS or too
- * many nested writes over a reserve/commit pair.
+ * commit offset, and this did not involve update to the
+ * reserve counter. Drop record in producer-consumer and
+ * overwrite mode. Caused by either a writer OOPS or
+ * too 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);