#include <wrapper/atomic.h>
#include <wrapper/kref.h>
#include <wrapper/percpu-defs.h>
+#include <wrapper/timer.h>
/*
* Internal structure representing offsets to use at a sub-buffer switch.
static
void lib_ring_buffer_print_errors(struct channel *chan,
struct lib_ring_buffer *buf, int cpu);
+static
+void _lib_ring_buffer_switch_remote(struct lib_ring_buffer *buf,
+ enum switch_mode mode);
+
+static
+int lib_ring_buffer_poll_deliver(const struct lib_ring_buffer_config *config,
+ struct lib_ring_buffer *buf,
+ struct channel *chan)
+{
+ unsigned long consumed_old, consumed_idx, commit_count, write_offset;
+
+ consumed_old = atomic_long_read(&buf->consumed);
+ consumed_idx = subbuf_index(consumed_old, chan);
+ commit_count = v_read(config, &buf->commit_cold[consumed_idx].cc_sb);
+ /*
+ * No memory barrier here, since we are only interested
+ * in a statistically correct polling result. The next poll will
+ * get the data is we are racing. The mb() that ensures correct
+ * memory order is in get_subbuf.
+ */
+ write_offset = v_read(config, &buf->offset);
+
+ /*
+ * Check that the subbuffer we are trying to consume has been
+ * already fully committed.
+ */
+
+ if (((commit_count - chan->backend.subbuf_size)
+ & chan->commit_count_mask)
+ - (buf_trunc(consumed_old, chan)
+ >> chan->backend.num_subbuf_order)
+ != 0)
+ return 0;
+
+ /*
+ * 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_old, chan)
+ == 0)
+ return 0;
+
+ return 1;
+}
/*
* Must be called under cpu hotplug protection.
kzalloc_node(ALIGN(sizeof(*buf->commit_hot)
* chan->backend.num_subbuf,
1 << INTERNODE_CACHE_SHIFT),
- GFP_KERNEL, cpu_to_node(max(cpu, 0)));
+ GFP_KERNEL | __GFP_NOWARN,
+ cpu_to_node(max(cpu, 0)));
if (!buf->commit_hot) {
ret = -ENOMEM;
goto free_chanbuf;
kzalloc_node(ALIGN(sizeof(*buf->commit_cold)
* chan->backend.num_subbuf,
1 << INTERNODE_CACHE_SHIFT),
- GFP_KERNEL, cpu_to_node(max(cpu, 0)));
+ GFP_KERNEL | __GFP_NOWARN,
+ cpu_to_node(max(cpu, 0)));
if (!buf->commit_cold) {
ret = -ENOMEM;
goto free_commit;
lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE);
if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
- mod_timer_pinned(&buf->switch_timer,
+ lttng_mod_timer_pinned(&buf->switch_timer,
jiffies + chan->switch_timer_interval);
else
mod_timer(&buf->switch_timer,
if (!chan->switch_timer_interval || buf->switch_timer_enabled)
return;
- init_timer(&buf->switch_timer);
+
+ if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
+ lttng_init_timer_pinned(&buf->switch_timer);
+ else
+ init_timer(&buf->switch_timer);
+
buf->switch_timer.function = switch_buffer_timer;
buf->switch_timer.expires = jiffies + chan->switch_timer_interval;
buf->switch_timer.data = (unsigned long)buf;
}
if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
- mod_timer_pinned(&buf->read_timer,
+ lttng_mod_timer_pinned(&buf->read_timer,
jiffies + chan->read_timer_interval);
else
mod_timer(&buf->read_timer,
|| buf->read_timer_enabled)
return;
- init_timer(&buf->read_timer);
+ if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
+ lttng_init_timer_pinned(&buf->read_timer);
+ else
+ init_timer(&buf->read_timer);
+
buf->read_timer.function = read_buffer_timer;
buf->read_timer.expires = jiffies + chan->read_timer_interval;
buf->read_timer.data = (unsigned long)buf;
channel_backend_unregister_notifiers(&chan->backend);
}
+static void lib_ring_buffer_set_quiescent(struct lib_ring_buffer *buf)
+{
+ if (!buf->quiescent) {
+ buf->quiescent = true;
+ _lib_ring_buffer_switch_remote(buf, SWITCH_FLUSH);
+ }
+}
+
+static void lib_ring_buffer_clear_quiescent(struct lib_ring_buffer *buf)
+{
+ buf->quiescent = false;
+}
+
+void lib_ring_buffer_set_quiescent_channel(struct channel *chan)
+{
+ int cpu;
+ const struct lib_ring_buffer_config *config = &chan->backend.config;
+
+ if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) {
+ get_online_cpus();
+ for_each_channel_cpu(cpu, chan) {
+ struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf,
+ cpu);
+
+ lib_ring_buffer_set_quiescent(buf);
+ }
+ put_online_cpus();
+ } else {
+ struct lib_ring_buffer *buf = chan->backend.buf;
+
+ lib_ring_buffer_set_quiescent(buf);
+ }
+}
+EXPORT_SYMBOL_GPL(lib_ring_buffer_set_quiescent_channel);
+
+void lib_ring_buffer_clear_quiescent_channel(struct channel *chan)
+{
+ int cpu;
+ const struct lib_ring_buffer_config *config = &chan->backend.config;
+
+ if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) {
+ get_online_cpus();
+ for_each_channel_cpu(cpu, chan) {
+ struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf,
+ cpu);
+
+ lib_ring_buffer_clear_quiescent(buf);
+ }
+ put_online_cpus();
+ } else {
+ struct lib_ring_buffer *buf = chan->backend.buf;
+
+ lib_ring_buffer_clear_quiescent(buf);
+ }
+}
+EXPORT_SYMBOL_GPL(lib_ring_buffer_clear_quiescent_channel);
+
static void channel_free(struct channel *chan)
{
if (chan->backend.release_priv_ops) {
chan->backend.priv,
cpu);
if (buf->backend.allocated)
- lib_ring_buffer_switch_slow(buf, SWITCH_FLUSH);
+ lib_ring_buffer_set_quiescent(buf);
/*
* Perform flush before writing to finalized.
*/
if (config->cb.buffer_finalize)
config->cb.buffer_finalize(buf, chan->backend.priv, -1);
if (buf->backend.allocated)
- lib_ring_buffer_switch_slow(buf, SWITCH_FLUSH);
+ lib_ring_buffer_set_quiescent(buf);
/*
* Perform flush before writing to finalized.
*/
/*
* lib_ring_buffer_switch_old_start: Populate old subbuffer header.
*
- * Only executed when the buffer is finalized, in SWITCH_FLUSH.
+ * Only executed by SWITCH_FLUSH, which can be issued while tracing is active
+ * or at buffer finalization (destroy).
*/
static
void lib_ring_buffer_switch_old_start(struct lib_ring_buffer *buf,
unsigned long sb_index, commit_count;
/*
- * We are performing a SWITCH_FLUSH. At this stage, there are no
- * concurrent writes into the buffer.
+ * We are performing a SWITCH_FLUSH. There may be concurrent
+ * writes into the buffer if e.g. invoked while performing a
+ * snapshot on an active trace.
*
- * 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 the client does not save any header information (sub-buffer
+ * header size == 0), 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;
}
EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_slow);
+struct switch_param {
+ struct lib_ring_buffer *buf;
+ enum switch_mode mode;
+};
+
static void remote_switch(void *info)
{
- struct lib_ring_buffer *buf = info;
+ struct switch_param *param = info;
+ struct lib_ring_buffer *buf = param->buf;
- lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE);
+ lib_ring_buffer_switch_slow(buf, param->mode);
}
-void lib_ring_buffer_switch_remote(struct lib_ring_buffer *buf)
+static void _lib_ring_buffer_switch_remote(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;
int ret;
+ struct switch_param param;
/*
* 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);
+ lib_ring_buffer_switch_slow(buf, mode);
return;
}
* switch.
*/
get_online_cpus();
+ param.buf = buf;
+ param.mode = mode;
ret = smp_call_function_single(buf->backend.cpu,
- remote_switch, buf, 1);
+ remote_switch, ¶m, 1);
if (ret) {
/* Remote CPU is offline, do it ourself. */
- lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE);
+ lib_ring_buffer_switch_slow(buf, mode);
}
put_online_cpus();
}
+
+void lib_ring_buffer_switch_remote(struct lib_ring_buffer *buf)
+{
+ _lib_ring_buffer_switch_remote(buf, SWITCH_ACTIVE);
+}
EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_remote);
+/* Switch sub-buffer even if current sub-buffer is empty. */
+void lib_ring_buffer_switch_remote_empty(struct lib_ring_buffer *buf)
+{
+ _lib_ring_buffer_switch_remote(buf, SWITCH_FLUSH);
+}
+EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_remote_empty);
+
/*
* Returns :
* 0 if ok
}
EXPORT_SYMBOL_GPL(lib_ring_buffer_reserve_slow);
+static
+void lib_ring_buffer_vmcore_check_deliver(const struct lib_ring_buffer_config *config,
+ struct lib_ring_buffer *buf,
+ unsigned long commit_count,
+ unsigned long idx)
+{
+ if (config->oops == RING_BUFFER_OOPS_CONSISTENCY)
+ v_set(config, &buf->commit_hot[idx].seq, commit_count);
+}
+
+/*
+ * The ring buffer can count events recorded and overwritten per buffer,
+ * but it is disabled by default due to its performance overhead.
+ */
+#ifdef LTTNG_RING_BUFFER_COUNT_EVENTS
+static
+void deliver_count_events(const struct lib_ring_buffer_config *config,
+ struct lib_ring_buffer *buf,
+ unsigned long idx)
+{
+ v_add(config, subbuffer_get_records_count(config,
+ &buf->backend, idx),
+ &buf->records_count);
+ v_add(config, subbuffer_count_records_overrun(config,
+ &buf->backend, idx),
+ &buf->records_overrun);
+}
+#else /* LTTNG_RING_BUFFER_COUNT_EVENTS */
+static
+void deliver_count_events(const struct lib_ring_buffer_config *config,
+ struct lib_ring_buffer *buf,
+ unsigned long idx)
+{
+}
+#endif /* #else LTTNG_RING_BUFFER_COUNT_EVENTS */
+
+
+void lib_ring_buffer_check_deliver_slow(const struct lib_ring_buffer_config *config,
+ struct lib_ring_buffer *buf,
+ struct channel *chan,
+ unsigned long offset,
+ unsigned long commit_count,
+ unsigned long idx,
+ u64 tsc)
+{
+ unsigned long old_commit_count = commit_count
+ - chan->backend.subbuf_size;
+
+ /*
+ * If we succeeded at updating cc_sb below, we are the subbuffer
+ * writer delivering the subbuffer. Deals with concurrent
+ * updates of the "cc" value without adding a add_return atomic
+ * operation to the fast path.
+ *
+ * We are doing the delivery in two steps:
+ * - First, we cmpxchg() cc_sb to the new value
+ * old_commit_count + 1. This ensures that we are the only
+ * subbuffer user successfully filling the subbuffer, but we
+ * do _not_ set the cc_sb value to "commit_count" yet.
+ * Therefore, other writers that would wrap around the ring
+ * buffer and try to start writing to our subbuffer would
+ * have to drop records, because it would appear as
+ * non-filled.
+ * We therefore have exclusive access to the subbuffer control
+ * structures. This mutual exclusion with other writers is
+ * crucially important to perform record overruns count in
+ * flight recorder mode locklessly.
+ * - When we are ready to release the subbuffer (either for
+ * reading or for overrun by other writers), we simply set the
+ * cc_sb value to "commit_count" and perform delivery.
+ *
+ * The subbuffer size is least 2 bytes (minimum size: 1 page).
+ * This guarantees that old_commit_count + 1 != commit_count.
+ */
+
+ /*
+ * Order prior updates to reserve count prior to the
+ * commit_cold cc_sb update.
+ */
+ smp_wmb();
+ if (likely(v_cmpxchg(config, &buf->commit_cold[idx].cc_sb,
+ old_commit_count, old_commit_count + 1)
+ == old_commit_count)) {
+ /*
+ * Start of exclusive subbuffer access. We are
+ * guaranteed to be the last writer in this subbuffer
+ * and any other writer trying to access this subbuffer
+ * in this state is required to drop records.
+ */
+ deliver_count_events(config, buf, idx);
+ config->cb.buffer_end(buf, tsc, idx,
+ lib_ring_buffer_get_data_size(config,
+ buf,
+ idx));
+
+ /*
+ * Increment the packet counter while we have exclusive
+ * access.
+ */
+ subbuffer_inc_packet_count(config, &buf->backend, idx);
+
+ /*
+ * Set noref flag and offset for this subbuffer id.
+ * Contains a memory barrier that ensures counter stores
+ * are ordered before set noref and offset.
+ */
+ lib_ring_buffer_set_noref_offset(config, &buf->backend, idx,
+ buf_trunc_val(offset, chan));
+
+ /*
+ * Order set_noref and record counter updates before the
+ * end of subbuffer exclusive access. Orders with
+ * respect to writers coming into the subbuffer after
+ * wrap around, and also order wrt concurrent readers.
+ */
+ smp_mb();
+ /* End of exclusive subbuffer access */
+ v_set(config, &buf->commit_cold[idx].cc_sb,
+ commit_count);
+ /*
+ * Order later updates to reserve count after
+ * the commit_cold cc_sb update.
+ */
+ smp_wmb();
+ lib_ring_buffer_vmcore_check_deliver(config, buf,
+ commit_count, idx);
+
+ /*
+ * RING_BUFFER_WAKEUP_BY_WRITER wakeup is not lock-free.
+ */
+ if (config->wakeup == RING_BUFFER_WAKEUP_BY_WRITER
+ && atomic_long_read(&buf->active_readers)
+ && lib_ring_buffer_poll_deliver(config, buf, chan)) {
+ wake_up_interruptible(&buf->read_wait);
+ wake_up_interruptible(&chan->read_wait);
+ }
+
+ }
+}
+EXPORT_SYMBOL_GPL(lib_ring_buffer_check_deliver_slow);
+
int __init init_lib_ring_buffer_frontend(void)
{
int cpu;
projects / lttng-modules.git / blobdiff