*/
#define _GNU_SOURCE
+#define _LGPL_SOURCE
#include <sys/types.h>
#include <sys/mman.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <signal.h>
#include <time.h>
+#include <stdbool.h>
#include <urcu/compiler.h>
#include <urcu/ref.h>
#include <urcu/tls-compat.h>
+#include <poll.h>
#include <helper.h>
#include "smp.h"
#include "backend.h"
#include "frontend.h"
#include "shm.h"
-#include "tlsfixup.h"
+#include "rb-init.h"
#include "../liblttng-ust/compat.h" /* For ENODATA */
-#ifndef max
-#define max(a, b) ((a) > (b) ? (a) : (b))
-#endif
-
/* Print DBG() messages about events lost only every 1048576 hits */
#define DBG_PRINT_NR_LOST (1UL << 20)
#define LTTNG_UST_RB_SIG_READ SIGRTMIN + 1
#define LTTNG_UST_RB_SIG_TEARDOWN SIGRTMIN + 2
#define CLOCKID CLOCK_MONOTONIC
+#define LTTNG_UST_RING_BUFFER_GET_RETRY 10
+#define LTTNG_UST_RING_BUFFER_RETRY_DELAY_MS 10
+#define RETRY_DELAY_MS 100 /* 100 ms. */
+
+/*
+ * Non-static to ensure the compiler does not optimize away the xor.
+ */
+uint8_t lttng_crash_magic_xor[] = RB_CRASH_DUMP_ABI_MAGIC_XOR;
/*
* Use POSIX SHM: shm_open(3) and shm_unlink(3).
.lock = PTHREAD_MUTEX_INITIALIZER,
};
+int lttng_ust_blocking_retry_timeout =
+ CONFIG_LTTNG_UST_DEFAULT_BLOCKING_RETRY_TIMEOUT_MS;
+
+void lttng_ust_ringbuffer_set_retry_timeout(int timeout)
+{
+ lttng_ust_blocking_retry_timeout = timeout;
+}
+
/**
* lib_ring_buffer_reset - Reset ring buffer to initial values.
* @buf: Ring buffer.
void lib_ring_buffer_reset(struct lttng_ust_lib_ring_buffer *buf,
struct lttng_ust_shm_handle *handle)
{
- struct channel *chan = shmp(handle, buf->backend.chan);
- const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config;
+ struct channel *chan;
+ const struct lttng_ust_lib_ring_buffer_config *config;
unsigned int i;
+ chan = shmp(handle, buf->backend.chan);
+ if (!chan)
+ return;
+ config = &chan->backend.config;
/*
* Reset iterator first. It will put the subbuffer if it currently holds
* it.
*/
v_set(config, &buf->offset, 0);
for (i = 0; i < chan->backend.num_subbuf; i++) {
- v_set(config, &shmp_index(handle, buf->commit_hot, i)->cc, 0);
- v_set(config, &shmp_index(handle, buf->commit_hot, i)->seq, 0);
- v_set(config, &shmp_index(handle, buf->commit_cold, i)->cc_sb, 0);
+ struct commit_counters_hot *cc_hot;
+ struct commit_counters_cold *cc_cold;
+
+ cc_hot = shmp_index(handle, buf->commit_hot, i);
+ if (!cc_hot)
+ return;
+ cc_cold = shmp_index(handle, buf->commit_cold, i);
+ if (!cc_cold)
+ return;
+ v_set(config, &cc_hot->cc, 0);
+ v_set(config, &cc_hot->seq, 0);
+ v_set(config, &cc_cold->cc_sb, 0);
}
uatomic_set(&buf->consumed, 0);
uatomic_set(&buf->record_disabled, 0);
/* Don't reset reader reference count */
}
+static
+void init_crash_abi(const struct lttng_ust_lib_ring_buffer_config *config,
+ struct lttng_crash_abi *crash_abi,
+ struct lttng_ust_lib_ring_buffer *buf,
+ struct channel_backend *chanb,
+ struct shm_object *shmobj,
+ struct lttng_ust_shm_handle *handle)
+{
+ int i;
+
+ for (i = 0; i < RB_CRASH_DUMP_ABI_MAGIC_LEN; i++)
+ crash_abi->magic[i] = lttng_crash_magic_xor[i] ^ 0xFF;
+ crash_abi->mmap_length = shmobj->memory_map_size;
+ crash_abi->endian = RB_CRASH_ENDIAN;
+ crash_abi->major = RB_CRASH_DUMP_ABI_MAJOR;
+ crash_abi->minor = RB_CRASH_DUMP_ABI_MINOR;
+ crash_abi->word_size = sizeof(unsigned long);
+ crash_abi->layout_type = LTTNG_CRASH_TYPE_UST;
+
+ /* Offset of fields */
+ crash_abi->offset.prod_offset =
+ (uint32_t) ((char *) &buf->offset - (char *) buf);
+ crash_abi->offset.consumed_offset =
+ (uint32_t) ((char *) &buf->consumed - (char *) buf);
+ crash_abi->offset.commit_hot_array =
+ (uint32_t) ((char *) shmp(handle, buf->commit_hot) - (char *) buf);
+ crash_abi->offset.commit_hot_seq =
+ offsetof(struct commit_counters_hot, seq);
+ crash_abi->offset.buf_wsb_array =
+ (uint32_t) ((char *) shmp(handle, buf->backend.buf_wsb) - (char *) buf);
+ crash_abi->offset.buf_wsb_id =
+ offsetof(struct lttng_ust_lib_ring_buffer_backend_subbuffer, id);
+ crash_abi->offset.sb_array =
+ (uint32_t) ((char *) shmp(handle, buf->backend.array) - (char *) buf);
+ crash_abi->offset.sb_array_shmp_offset =
+ offsetof(struct lttng_ust_lib_ring_buffer_backend_pages_shmp,
+ shmp._ref.offset);
+ crash_abi->offset.sb_backend_p_offset =
+ offsetof(struct lttng_ust_lib_ring_buffer_backend_pages,
+ p._ref.offset);
+
+ /* Field length */
+ crash_abi->length.prod_offset = sizeof(buf->offset);
+ crash_abi->length.consumed_offset = sizeof(buf->consumed);
+ crash_abi->length.commit_hot_seq =
+ sizeof(((struct commit_counters_hot *) NULL)->seq);
+ crash_abi->length.buf_wsb_id =
+ sizeof(((struct lttng_ust_lib_ring_buffer_backend_subbuffer *) NULL)->id);
+ crash_abi->length.sb_array_shmp_offset =
+ sizeof(((struct lttng_ust_lib_ring_buffer_backend_pages_shmp *) NULL)->shmp._ref.offset);
+ crash_abi->length.sb_backend_p_offset =
+ sizeof(((struct lttng_ust_lib_ring_buffer_backend_pages *) NULL)->p._ref.offset);
+
+ /* Array stride */
+ crash_abi->stride.commit_hot_array =
+ sizeof(struct commit_counters_hot);
+ crash_abi->stride.buf_wsb_array =
+ sizeof(struct lttng_ust_lib_ring_buffer_backend_subbuffer);
+ crash_abi->stride.sb_array =
+ sizeof(struct lttng_ust_lib_ring_buffer_backend_pages_shmp);
+
+ /* Buffer constants */
+ crash_abi->buf_size = chanb->buf_size;
+ crash_abi->subbuf_size = chanb->subbuf_size;
+ crash_abi->num_subbuf = chanb->num_subbuf;
+ crash_abi->mode = (uint32_t) chanb->config.mode;
+
+ if (config->cb.content_size_field) {
+ size_t offset, length;
+
+ config->cb.content_size_field(config, &offset, &length);
+ crash_abi->offset.content_size = offset;
+ crash_abi->length.content_size = length;
+ } else {
+ crash_abi->offset.content_size = 0;
+ crash_abi->length.content_size = 0;
+ }
+ if (config->cb.packet_size_field) {
+ size_t offset, length;
+
+ config->cb.packet_size_field(config, &offset, &length);
+ crash_abi->offset.packet_size = offset;
+ crash_abi->length.packet_size = length;
+ } else {
+ crash_abi->offset.packet_size = 0;
+ crash_abi->length.packet_size = 0;
+ }
+}
+
/*
* Must be called under cpu hotplug protection.
*/
{
const struct lttng_ust_lib_ring_buffer_config *config = &chanb->config;
struct channel *chan = caa_container_of(chanb, struct channel, backend);
+ struct lttng_ust_lib_ring_buffer_backend_subbuffer *wsb;
+ struct channel *shmp_chan;
+ struct commit_counters_hot *cc_hot;
void *priv = channel_get_private(chan);
size_t subbuf_header_size;
uint64_t tsc;
if (buf->backend.allocated)
return 0;
- ret = lib_ring_buffer_backend_create(&buf->backend, &chan->backend,
- cpu, handle, shmobj);
- if (ret)
- return ret;
-
align_shm(shmobj, __alignof__(struct commit_counters_hot));
set_shmp(buf->commit_hot,
zalloc_shm(shmobj,
sizeof(struct commit_counters_hot) * chan->backend.num_subbuf));
if (!shmp(handle, buf->commit_hot)) {
- ret = -ENOMEM;
- goto free_chanbuf;
+ return -ENOMEM;
}
align_shm(shmobj, __alignof__(struct commit_counters_cold));
goto free_commit;
}
+ ret = lib_ring_buffer_backend_create(&buf->backend, &chan->backend,
+ cpu, handle, shmobj);
+ if (ret) {
+ goto free_init;
+ }
+
/*
* Write the subbuffer header for first subbuffer so we know the total
* duration of data gathering.
*/
subbuf_header_size = config->cb.subbuffer_header_size();
v_set(config, &buf->offset, subbuf_header_size);
- subbuffer_id_clear_noref(config, &shmp_index(handle, buf->backend.buf_wsb, 0)->id);
- tsc = config->cb.ring_buffer_clock_read(shmp(handle, buf->backend.chan));
+ wsb = shmp_index(handle, buf->backend.buf_wsb, 0);
+ if (!wsb) {
+ ret = -EPERM;
+ goto free_chanbuf;
+ }
+ subbuffer_id_clear_noref(config, &wsb->id);
+ shmp_chan = shmp(handle, buf->backend.chan);
+ if (!shmp_chan) {
+ ret = -EPERM;
+ goto free_chanbuf;
+ }
+ tsc = config->cb.ring_buffer_clock_read(shmp_chan);
config->cb.buffer_begin(buf, tsc, 0, handle);
- v_add(config, subbuf_header_size, &shmp_index(handle, buf->commit_hot, 0)->cc);
+ cc_hot = shmp_index(handle, buf->commit_hot, 0);
+ if (!cc_hot) {
+ ret = -EPERM;
+ goto free_chanbuf;
+ }
+ v_add(config, subbuf_header_size, &cc_hot->cc);
+ v_add(config, subbuf_header_size, &cc_hot->seq);
if (config->cb.buffer_create) {
ret = config->cb.buffer_create(buf, priv, cpu, chanb->name, handle);
if (ret)
- goto free_init;
+ goto free_chanbuf;
}
+
+ init_crash_abi(config, &buf->crash_abi, buf, chanb, shmobj, handle);
+
buf->backend.allocated = 1;
return 0;
for_each_possible_cpu(cpu) {
struct lttng_ust_lib_ring_buffer *buf =
shmp(handle, chan->backend.buf[cpu].shmp);
+
+ if (!buf)
+ goto end;
if (uatomic_read(&buf->active_readers))
lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE,
chan->handle);
struct lttng_ust_lib_ring_buffer *buf =
shmp(handle, chan->backend.buf[0].shmp);
+ if (!buf)
+ goto end;
if (uatomic_read(&buf->active_readers))
lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE,
chan->handle);
}
+end:
pthread_mutex_unlock(&wakeup_fd_mutex);
return;
}
+static
+int lib_ring_buffer_poll_deliver(const struct lttng_ust_lib_ring_buffer_config *config,
+ struct lttng_ust_lib_ring_buffer *buf,
+ struct channel *chan,
+ struct lttng_ust_shm_handle *handle)
+{
+ unsigned long consumed_old, consumed_idx, commit_count, write_offset;
+ struct commit_counters_cold *cc_cold;
+
+ consumed_old = uatomic_read(&buf->consumed);
+ consumed_idx = subbuf_index(consumed_old, chan);
+ cc_cold = shmp_index(handle, buf->commit_cold, consumed_idx);
+ if (!cc_cold)
+ return 0;
+ commit_count = v_read(config, &cc_cold->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;
+}
+
+static
+void lib_ring_buffer_wakeup(struct lttng_ust_lib_ring_buffer *buf,
+ struct lttng_ust_shm_handle *handle)
+{
+ int wakeup_fd = shm_get_wakeup_fd(handle, &buf->self._ref);
+ sigset_t sigpipe_set, pending_set, old_set;
+ int ret, sigpipe_was_pending = 0;
+
+ if (wakeup_fd < 0)
+ return;
+
+ /*
+ * Wake-up the other end by writing a null byte in the pipe
+ * (non-blocking). Important note: Because writing into the
+ * pipe is non-blocking (and therefore we allow dropping wakeup
+ * data, as long as there is wakeup data present in the pipe
+ * buffer to wake up the consumer), the consumer should perform
+ * the following sequence for waiting:
+ * 1) empty the pipe (reads).
+ * 2) check if there is data in the buffer.
+ * 3) wait on the pipe (poll).
+ *
+ * Discard the SIGPIPE from write(), not disturbing any SIGPIPE
+ * that might be already pending. If a bogus SIGPIPE is sent to
+ * the entire process concurrently by a malicious user, it may
+ * be simply discarded.
+ */
+ ret = sigemptyset(&pending_set);
+ assert(!ret);
+ /*
+ * sigpending returns the mask of signals that are _both_
+ * blocked for the thread _and_ pending for either the thread or
+ * the entire process.
+ */
+ ret = sigpending(&pending_set);
+ assert(!ret);
+ sigpipe_was_pending = sigismember(&pending_set, SIGPIPE);
+ /*
+ * If sigpipe was pending, it means it was already blocked, so
+ * no need to block it.
+ */
+ if (!sigpipe_was_pending) {
+ ret = sigemptyset(&sigpipe_set);
+ assert(!ret);
+ ret = sigaddset(&sigpipe_set, SIGPIPE);
+ assert(!ret);
+ ret = pthread_sigmask(SIG_BLOCK, &sigpipe_set, &old_set);
+ assert(!ret);
+ }
+ do {
+ ret = write(wakeup_fd, "", 1);
+ } while (ret == -1L && errno == EINTR);
+ if (ret == -1L && errno == EPIPE && !sigpipe_was_pending) {
+ struct timespec timeout = { 0, 0 };
+ do {
+ ret = sigtimedwait(&sigpipe_set, NULL,
+ &timeout);
+ } while (ret == -1L && errno == EINTR);
+ }
+ if (!sigpipe_was_pending) {
+ ret = pthread_sigmask(SIG_SETMASK, &old_set, NULL);
+ assert(!ret);
+ }
+}
+
static
void lib_ring_buffer_channel_do_read(struct channel *chan)
{
struct lttng_ust_lib_ring_buffer *buf =
shmp(handle, chan->backend.buf[cpu].shmp);
+ if (!buf)
+ goto end;
if (uatomic_read(&buf->active_readers)
&& lib_ring_buffer_poll_deliver(config, buf,
chan, handle)) {
struct lttng_ust_lib_ring_buffer *buf =
shmp(handle, chan->backend.buf[0].shmp);
+ if (!buf)
+ goto end;
if (uatomic_read(&buf->active_readers)
&& lib_ring_buffer_poll_deliver(config, buf,
chan, handle)) {
lib_ring_buffer_wakeup(buf, handle);
}
}
+end:
pthread_mutex_unlock(&wakeup_fd_mutex);
}
}
its.it_value.tv_sec = chan->switch_timer_interval / 1000000;
- its.it_value.tv_nsec = chan->switch_timer_interval % 1000000;
+ its.it_value.tv_nsec = (chan->switch_timer_interval % 1000000) * 1000;
its.it_interval.tv_sec = its.it_value.tv_sec;
its.it_interval.tv_nsec = its.it_value.tv_nsec;
}
its.it_value.tv_sec = chan->read_timer_interval / 1000000;
- its.it_value.tv_nsec = chan->read_timer_interval % 1000000;
+ its.it_value.tv_nsec = (chan->read_timer_interval % 1000000) * 1000;
its.it_interval.tv_sec = its.it_value.tv_sec;
its.it_interval.tv_nsec = its.it_value.tv_nsec;
for_each_possible_cpu(cpu) {
struct lttng_ust_lib_ring_buffer *buf =
shmp(handle, chan->backend.buf[cpu].shmp);
- lib_ring_buffer_print_errors(chan, buf, cpu, handle);
+ if (buf)
+ lib_ring_buffer_print_errors(chan, buf, cpu, handle);
}
} else {
struct lttng_ust_lib_ring_buffer *buf =
shmp(handle, chan->backend.buf[0].shmp);
- lib_ring_buffer_print_errors(chan, buf, -1, handle);
+ if (buf)
+ lib_ring_buffer_print_errors(chan, buf, -1, handle);
}
}
static void channel_free(struct channel *chan,
- struct lttng_ust_shm_handle *handle)
+ struct lttng_ust_shm_handle *handle,
+ int consumer)
{
channel_backend_free(&chan->backend, handle);
/* chan is freed by shm teardown */
- shm_object_table_destroy(handle->table);
+ shm_object_table_destroy(handle->table, consumer);
free(handle);
}
* padding to let readers get those sub-buffers.
* Used for live streaming.
* @read_timer_interval: Time interval (in us) to wake up pending readers.
+ * @stream_fds: array of stream file descriptors.
+ * @nr_stream_fds: number of file descriptors in array.
*
* Holds cpu hotplug.
* Returns NULL on failure.
void *priv_data_init,
void *buf_addr, size_t subbuf_size,
size_t num_subbuf, unsigned int switch_timer_interval,
- unsigned int read_timer_interval)
+ unsigned int read_timer_interval,
+ const int *stream_fds, int nr_stream_fds)
{
int ret;
size_t shmsize, chansize;
else
nr_streams = 1;
+ if (nr_stream_fds != nr_streams)
+ return NULL;
+
if (lib_ring_buffer_check_config(config, switch_timer_interval,
read_timer_interval))
return NULL;
shmsize += priv_data_size;
/* Allocate normal memory for channel (not shared) */
- shmobj = shm_object_table_alloc(handle->table, shmsize, SHM_OBJECT_MEM);
+ shmobj = shm_object_table_alloc(handle->table, shmsize, SHM_OBJECT_MEM,
+ -1);
if (!shmobj)
goto error_append;
/* struct channel is at object 0, offset 0 (hardcoded) */
}
ret = channel_backend_init(&chan->backend, name, config,
- subbuf_size, num_subbuf, handle);
+ subbuf_size, num_subbuf, handle,
+ stream_fds);
if (ret)
goto error_backend_init;
error_backend_init:
error_append:
- shm_object_table_destroy(handle->table);
+ shm_object_table_destroy(handle->table, 1);
error_table_alloc:
free(handle);
return NULL;
return handle;
error_table_object:
- shm_object_table_destroy(handle->table);
+ shm_object_table_destroy(handle->table, 0);
error_table_alloc:
free(handle);
return NULL;
}
static
-void channel_release(struct channel *chan, struct lttng_ust_shm_handle *handle)
+void channel_release(struct channel *chan, struct lttng_ust_shm_handle *handle,
+ int consumer)
{
- channel_free(chan, handle);
+ channel_free(chan, handle, consumer);
}
/**
* sessiond/consumer are keeping a reference on the shm file
* descriptor directly. No need to refcount.
*/
- channel_release(chan, handle);
+ channel_release(chan, handle, consumer);
return;
}
{
struct channel *chan = shmp(handle, buf->backend.chan);
+ if (!chan)
+ return;
CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1);
cmm_smp_mb();
uatomic_dec(&buf->active_readers);
unsigned long *consumed, unsigned long *produced,
struct lttng_ust_shm_handle *handle)
{
- struct channel *chan = shmp(handle, buf->backend.chan);
- const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config;
+ struct channel *chan;
+ const struct lttng_ust_lib_ring_buffer_config *config;
unsigned long consumed_cur, write_offset;
int finalized;
+ chan = shmp(handle, buf->backend.chan);
+ if (!chan)
+ return -EPERM;
+ config = &chan->backend.config;
finalized = CMM_ACCESS_ONCE(buf->finalized);
/*
* Read finalized before counters.
return -EAGAIN;
}
+/**
+ * Performs the same function as lib_ring_buffer_snapshot(), but the positions
+ * are saved regardless of whether the consumed and produced positions are
+ * in the same subbuffer.
+ * @buf: ring buffer
+ * @consumed: consumed byte count indicating the last position read
+ * @produced: produced byte count indicating the last position written
+ *
+ * This function is meant to provide information on the exact producer and
+ * consumer positions without regard for the "snapshot" feature.
+ */
+int lib_ring_buffer_snapshot_sample_positions(
+ struct lttng_ust_lib_ring_buffer *buf,
+ unsigned long *consumed, unsigned long *produced,
+ struct lttng_ust_shm_handle *handle)
+{
+ struct channel *chan;
+ const struct lttng_ust_lib_ring_buffer_config *config;
+
+ chan = shmp(handle, buf->backend.chan);
+ if (!chan)
+ return -EPERM;
+ config = &chan->backend.config;
+ cmm_smp_rmb();
+ *consumed = uatomic_read(&buf->consumed);
+ /*
+ * No need to issue a memory barrier between consumed count read and
+ * write offset read, because consumed count can only change
+ * concurrently in overwrite mode, and we keep a sequence counter
+ * identifier derived from the write offset to check we are getting
+ * the same sub-buffer we are expecting (the sub-buffers are atomically
+ * "tagged" upon writes, tags are checked upon read).
+ */
+ *produced = v_read(config, &buf->offset);
+ return 0;
+}
+
/**
* lib_ring_buffer_move_consumer - move consumed counter forward
* @buf: ring buffer
struct lttng_ust_shm_handle *handle)
{
struct lttng_ust_lib_ring_buffer_backend *bufb = &buf->backend;
- struct channel *chan = shmp(handle, bufb->chan);
+ struct channel *chan;
unsigned long consumed;
+ chan = shmp(handle, bufb->chan);
+ if (!chan)
+ return;
CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1);
/*
unsigned long consumed,
struct lttng_ust_shm_handle *handle)
{
- struct channel *chan = shmp(handle, buf->backend.chan);
- const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config;
+ struct channel *chan;
+ const struct lttng_ust_lib_ring_buffer_config *config;
unsigned long consumed_cur, consumed_idx, commit_count, write_offset;
- int ret;
- int finalized;
+ int ret, finalized, nr_retry = LTTNG_UST_RING_BUFFER_GET_RETRY;
+ struct commit_counters_cold *cc_cold;
+ chan = shmp(handle, buf->backend.chan);
+ if (!chan)
+ return -EPERM;
+ config = &chan->backend.config;
retry:
finalized = CMM_ACCESS_ONCE(buf->finalized);
/*
cmm_smp_rmb();
consumed_cur = uatomic_read(&buf->consumed);
consumed_idx = subbuf_index(consumed, chan);
- commit_count = v_read(config, &shmp_index(handle, buf->commit_cold, consumed_idx)->cc_sb);
+ cc_cold = shmp_index(handle, buf->commit_cold, consumed_idx);
+ if (!cc_cold)
+ return -EPERM;
+ commit_count = v_read(config, &cc_cold->cc_sb);
/*
* Make sure we read the commit count before reading the buffer
* data and the write offset. Correct consumed offset ordering
/*
* Check that the subbuffer we are trying to consume has been
- * already fully committed.
+ * already fully committed. There are a few causes that can make
+ * this unavailability situation occur:
+ *
+ * Temporary (short-term) situation:
+ * - Application is running on a different CPU, between reserve
+ * and commit ring buffer operations,
+ * - Application is preempted between reserve and commit ring
+ * buffer operations,
+ *
+ * Long-term situation:
+ * - Application is stopped (SIGSTOP) between reserve and commit
+ * ring buffer operations. Could eventually be resumed by
+ * SIGCONT.
+ * - Application is killed (SIGTERM, SIGINT, SIGKILL) between
+ * reserve and commit ring buffer operation.
+ *
+ * From a consumer perspective, handling short-term
+ * unavailability situations is performed by retrying a few
+ * times after a delay. Handling long-term unavailability
+ * situations is handled by failing to get the sub-buffer.
+ *
+ * In all of those situations, if the application is taking a
+ * long time to perform its commit after ring buffer space
+ * reservation, we can end up in a situation where the producer
+ * will fill the ring buffer and try to write into the same
+ * sub-buffer again (which has a missing commit). This is
+ * handled by the producer in the sub-buffer switch handling
+ * code of the reserve routine by detecting unbalanced
+ * reserve/commit counters and discarding all further events
+ * until the situation is resolved in those situations. Two
+ * scenarios can occur:
+ *
+ * 1) The application causing the reserve/commit counters to be
+ * unbalanced has been terminated. In this situation, all
+ * further events will be discarded in the buffers, and no
+ * further buffer data will be readable by the consumer
+ * daemon. Tearing down the UST tracing session and starting
+ * anew is a work-around for those situations. Note that this
+ * only affects per-UID tracing. In per-PID tracing, the
+ * application vanishes with the termination, and therefore
+ * no more data needs to be written to the buffers.
+ * 2) The application causing the unbalance has been delayed for
+ * a long time, but will eventually try to increment the
+ * commit counter after eventually writing to the sub-buffer.
+ * This situation can cause events to be discarded until the
+ * application resumes its operations.
*/
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;
+ != 0) {
+ if (nr_retry-- > 0) {
+ if (nr_retry <= (LTTNG_UST_RING_BUFFER_GET_RETRY >> 1))
+ (void) poll(NULL, 0, LTTNG_UST_RING_BUFFER_RETRY_DELAY_MS);
+ goto retry;
+ } else {
+ 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;
* the writer is getting access to a subbuffer we were trying to get
* access to. Also checks that the "consumed" buffer count we are
* looking for matches the one contained in the subbuffer id.
+ *
+ * The short-lived race window described here can be affected by
+ * application signals and preemption, thus requiring to bound
+ * the loop to a maximum number of retry.
*/
ret = update_read_sb_index(config, &buf->backend, &chan->backend,
consumed_idx, buf_trunc_val(consumed, chan),
handle);
- if (ret)
- goto retry;
+ if (ret) {
+ if (nr_retry-- > 0) {
+ if (nr_retry <= (LTTNG_UST_RING_BUFFER_GET_RETRY >> 1))
+ (void) poll(NULL, 0, LTTNG_UST_RING_BUFFER_RETRY_DELAY_MS);
+ goto retry;
+ } else {
+ goto nodata;
+ }
+ }
subbuffer_id_clear_noref(config, &buf->backend.buf_rsb.id);
buf->get_subbuf_consumed = consumed;
struct lttng_ust_shm_handle *handle)
{
struct lttng_ust_lib_ring_buffer_backend *bufb = &buf->backend;
- struct channel *chan = shmp(handle, bufb->chan);
- const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config;
- unsigned long read_sb_bindex, consumed_idx, consumed;
+ struct channel *chan;
+ const struct lttng_ust_lib_ring_buffer_config *config;
+ unsigned long sb_bindex, consumed_idx, consumed;
+ struct lttng_ust_lib_ring_buffer_backend_pages_shmp *rpages;
+ struct lttng_ust_lib_ring_buffer_backend_pages *backend_pages;
+ chan = shmp(handle, bufb->chan);
+ if (!chan)
+ return;
+ config = &chan->backend.config;
CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1);
if (!buf->get_subbuf) {
* Can be below zero if an iterator is used on a snapshot more than
* once.
*/
- read_sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id);
- v_add(config, v_read(config,
- &shmp(handle, shmp_index(handle, bufb->array, read_sb_bindex)->shmp)->records_unread),
- &bufb->records_read);
- v_set(config, &shmp(handle, shmp_index(handle, bufb->array, read_sb_bindex)->shmp)->records_unread, 0);
+ sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id);
+ rpages = shmp_index(handle, bufb->array, sb_bindex);
+ if (!rpages)
+ return;
+ backend_pages = shmp(handle, rpages->shmp);
+ if (!backend_pages)
+ return;
+ v_add(config, v_read(config, &backend_pages->records_unread),
+ &bufb->records_read);
+ v_set(config, &backend_pages->records_unread, 0);
CHAN_WARN_ON(chan, config->mode == RING_BUFFER_OVERWRITE
&& subbuffer_id_is_noref(config, bufb->buf_rsb.id));
subbuffer_id_set_noref(config, &bufb->buf_rsb.id);
{
const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config;
unsigned long cons_idx, commit_count, commit_count_sb;
+ struct commit_counters_hot *cc_hot;
+ struct commit_counters_cold *cc_cold;
cons_idx = subbuf_index(cons_offset, chan);
- commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, cons_idx)->cc);
- commit_count_sb = v_read(config, &shmp_index(handle, buf->commit_cold, cons_idx)->cc_sb);
+ cc_hot = shmp_index(handle, buf->commit_hot, cons_idx);
+ if (!cc_hot)
+ return;
+ cc_cold = shmp_index(handle, buf->commit_cold, cons_idx);
+ if (!cc_cold)
+ return;
+ commit_count = v_read(config, &cc_hot->cc);
+ commit_count_sb = v_read(config, &cc_cold->cc_sb);
if (subbuf_offset(commit_count, chan) != 0)
DBG("ring buffer %s, cpu %d: "
/*
* 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 lttng_ust_lib_ring_buffer *buf,
const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config;
unsigned long oldidx = subbuf_index(offsets->old, chan);
unsigned long commit_count;
+ struct commit_counters_hot *cc_hot;
config->cb.buffer_begin(buf, tsc, oldidx, handle);
* determine that the subbuffer is full.
*/
cmm_smp_wmb();
+ cc_hot = shmp_index(handle, buf->commit_hot, oldidx);
+ if (!cc_hot)
+ return;
v_add(config, config->cb.subbuffer_header_size(),
- &shmp_index(handle, buf->commit_hot, oldidx)->cc);
- commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, oldidx)->cc);
+ &cc_hot->cc);
+ commit_count = v_read(config, &cc_hot->cc);
/* Check if the written buffer has to be delivered */
lib_ring_buffer_check_deliver(config, buf, chan, offsets->old,
- commit_count, oldidx, handle);
- lib_ring_buffer_write_commit_counter(config, buf, chan, oldidx,
- offsets->old, commit_count,
- config->cb.subbuffer_header_size(),
- handle);
+ commit_count, oldidx, handle, tsc);
+ lib_ring_buffer_write_commit_counter(config, buf, chan,
+ offsets->old + config->cb.subbuffer_header_size(),
+ commit_count, handle, cc_hot);
}
/*
const struct lttng_ust_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 commit_counters_hot *cc_hot;
data_size = subbuf_offset(offsets->old - 1, chan) + 1;
padding_size = chan->backend.subbuf_size - data_size;
* determine that the subbuffer is full.
*/
cmm_smp_wmb();
- v_add(config, padding_size, &shmp_index(handle, buf->commit_hot, oldidx)->cc);
- commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, oldidx)->cc);
+ cc_hot = shmp_index(handle, buf->commit_hot, oldidx);
+ if (!cc_hot)
+ return;
+ v_add(config, padding_size, &cc_hot->cc);
+ commit_count = v_read(config, &cc_hot->cc);
lib_ring_buffer_check_deliver(config, buf, chan, offsets->old - 1,
- commit_count, oldidx, handle);
- lib_ring_buffer_write_commit_counter(config, buf, chan, oldidx,
- offsets->old, commit_count,
- padding_size, handle);
+ commit_count, oldidx, handle, tsc);
+ lib_ring_buffer_write_commit_counter(config, buf, chan,
+ offsets->old + padding_size, commit_count, handle,
+ cc_hot);
}
/*
const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config;
unsigned long beginidx = subbuf_index(offsets->begin, chan);
unsigned long commit_count;
+ struct commit_counters_hot *cc_hot;
config->cb.buffer_begin(buf, tsc, beginidx, handle);
* determine that the subbuffer is full.
*/
cmm_smp_wmb();
- v_add(config, config->cb.subbuffer_header_size(),
- &shmp_index(handle, buf->commit_hot, beginidx)->cc);
- commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, beginidx)->cc);
+ cc_hot = shmp_index(handle, buf->commit_hot, beginidx);
+ if (!cc_hot)
+ return;
+ v_add(config, config->cb.subbuffer_header_size(), &cc_hot->cc);
+ commit_count = v_read(config, &cc_hot->cc);
/* Check if the written buffer has to be delivered */
lib_ring_buffer_check_deliver(config, buf, chan, offsets->begin,
- commit_count, beginidx, handle);
- lib_ring_buffer_write_commit_counter(config, buf, chan, beginidx,
- offsets->begin, commit_count,
- config->cb.subbuffer_header_size(),
- handle);
+ commit_count, beginidx, handle, tsc);
+ lib_ring_buffer_write_commit_counter(config, buf, chan,
+ offsets->begin + config->cb.subbuffer_header_size(),
+ commit_count, handle, cc_hot);
}
/*
* 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 lttng_ust_lib_ring_buffer *buf,
struct lttng_ust_shm_handle *handle)
{
const struct lttng_ust_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;
+ 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,
handle);
-
- /*
- * Order all writes to buffer before the commit count update that will
- * determine that the subbuffer is full.
- */
- cmm_smp_wmb();
- v_add(config, padding_size, &shmp_index(handle, buf->commit_hot, endidx)->cc);
- commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, endidx)->cc);
- lib_ring_buffer_check_deliver(config, buf, chan, offsets->end - 1,
- commit_count, endidx, handle);
- lib_ring_buffer_write_commit_counter(config, buf, chan, endidx,
- offsets->end, commit_count,
- padding_size, handle);
}
/*
struct lttng_ust_lib_ring_buffer *buf,
struct channel *chan,
struct switch_offsets *offsets,
- uint64_t *tsc)
+ uint64_t *tsc,
+ struct lttng_ust_shm_handle *handle)
{
const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config;
- unsigned long off;
+ unsigned long off, reserve_commit_diff;
offsets->begin = v_read(config, &buf->offset);
offsets->old = offsets->begin;
* timestamps) are visible to the reader. This is required for
* quiescence guarantees for the fusion merge.
*/
- if (mode == SWITCH_FLUSH || off > 0) {
- if (caa_unlikely(off == 0)) {
- /*
- * A final flush that encounters an empty
- * sub-buffer cannot switch buffer if a
- * reader is located within this sub-buffer.
- * Anyway, the purpose of final flushing of a
- * sub-buffer at offset 0 is to handle the case
- * of entirely empty stream.
- */
- if (caa_unlikely(subbuf_trunc(offsets->begin, chan)
- - subbuf_trunc((unsigned long)
- uatomic_read(&buf->consumed), chan)
- >= chan->backend.buf_size))
- return -1;
- /*
- * 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 (caa_unlikely(off == 0)) {
+ unsigned long sb_index, commit_count;
+ struct commit_counters_cold *cc_cold;
+
+ /*
+ * 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.
+ *
+ * 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;
+
+ /* Test new buffer integrity */
+ sb_index = subbuf_index(offsets->begin, chan);
+ cc_cold = shmp_index(handle, buf->commit_cold, sb_index);
+ if (!cc_cold)
+ return -1;
+ commit_count = v_read(config, &cc_cold->cc_sb);
+ reserve_commit_diff =
+ (buf_trunc(offsets->begin, chan)
+ >> chan->backend.num_subbuf_order)
+ - (commit_count & chan->commit_count_mask);
+ if (caa_likely(reserve_commit_diff == 0)) {
+ /* Next subbuffer not being written to. */
+ if (caa_unlikely(config->mode != RING_BUFFER_OVERWRITE &&
+ subbuf_trunc(offsets->begin, chan)
+ - subbuf_trunc((unsigned long)
+ uatomic_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 lttng_ust_lib_ring_buffer *buf, enum switch_mode mode,
struct lttng_ust_shm_handle *handle)
{
- struct channel *chan = shmp(handle, buf->backend.chan);
- const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config;
+ struct channel *chan;
+ const struct lttng_ust_lib_ring_buffer_config *config;
struct switch_offsets offsets;
unsigned long oldidx;
uint64_t tsc;
+ chan = shmp(handle, buf->backend.chan);
+ if (!chan)
+ return;
+ config = &chan->backend.config;
+
offsets.size = 0;
/*
*/
do {
if (lib_ring_buffer_try_switch_slow(mode, buf, chan, &offsets,
- &tsc))
+ &tsc, handle))
return; /* Switch not needed */
} while (v_cmpxchg(config, &buf->offset, offsets.old, offsets.end)
!= offsets.old);
lib_ring_buffer_switch_old_end(buf, chan, &offsets, tsc, handle);
}
+static
+bool handle_blocking_retry(int *timeout_left_ms)
+{
+ int timeout = *timeout_left_ms, delay;
+
+ if (caa_likely(!timeout))
+ return false; /* Do not retry, discard event. */
+ if (timeout < 0) /* Wait forever. */
+ delay = RETRY_DELAY_MS;
+ else
+ delay = min_t(int, timeout, RETRY_DELAY_MS);
+ (void) poll(NULL, 0, delay);
+ if (timeout > 0)
+ *timeout_left_ms -= delay;
+ return true; /* Retry. */
+}
+
/*
* Returns :
* 0 if ok
int lib_ring_buffer_try_reserve_slow(struct lttng_ust_lib_ring_buffer *buf,
struct channel *chan,
struct switch_offsets *offsets,
- struct lttng_ust_lib_ring_buffer_ctx *ctx)
+ struct lttng_ust_lib_ring_buffer_ctx *ctx,
+ void *client_ctx)
{
const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config;
struct lttng_ust_shm_handle *handle = ctx->handle;
- unsigned long reserve_commit_diff;
+ unsigned long reserve_commit_diff, offset_cmp;
+ int timeout_left_ms = lttng_ust_blocking_retry_timeout;
- 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->size = config->cb.record_header_size(config, chan,
offsets->begin,
&offsets->pre_header_padding,
- ctx);
+ ctx, client_ctx);
offsets->size +=
lib_ring_buffer_align(offsets->begin + offsets->size,
ctx->largest_align)
}
}
if (caa_unlikely(offsets->switch_new_start)) {
- unsigned long sb_index;
+ unsigned long sb_index, commit_count;
+ struct commit_counters_cold *cc_cold;
/*
* 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.
+ */
+ cmm_smp_rmb();
+ cc_cold = shmp_index(handle, buf->commit_cold, sb_index);
+ if (!cc_cold)
+ return -1;
+ commit_count = v_read(config, &cc_cold->cc_sb);
+ /* Read buf->commit_cold[sb_index].cc_sb before buf->offset. */
+ cmm_smp_rmb();
+ if (caa_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,
- &shmp_index(handle, buf->commit_cold, sb_index)->cc_sb)
- & chan->commit_count_mask);
+ - (commit_count & chan->commit_count_mask);
if (caa_likely(reserve_commit_diff == 0)) {
/* Next subbuffer not being written to. */
if (caa_unlikely(config->mode != RING_BUFFER_OVERWRITE &&
>= chan->backend.buf_size)) {
unsigned long nr_lost;
+ if (handle_blocking_retry(&timeout_left_ms))
+ goto retry;
+
/*
* We do not overwrite non consumed buffers
* and we are full : record is lost.
/*
* Next subbuffer reserve offset does not match the
- * commit offset. Drop record in producer-consumer and
+ * 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.
*/
config->cb.record_header_size(config, chan,
offsets->begin,
&offsets->pre_header_padding,
- ctx);
+ ctx, client_ctx);
offsets->size +=
lib_ring_buffer_align(offsets->begin + offsets->size,
ctx->largest_align)
* -EIO for other errors, else returns 0.
* It will take care of sub-buffer switching.
*/
-int lib_ring_buffer_reserve_slow(struct lttng_ust_lib_ring_buffer_ctx *ctx)
+int lib_ring_buffer_reserve_slow(struct lttng_ust_lib_ring_buffer_ctx *ctx,
+ void *client_ctx)
{
struct channel *chan = ctx->chan;
struct lttng_ust_shm_handle *handle = ctx->handle;
buf = shmp(handle, chan->backend.buf[ctx->cpu].shmp);
else
buf = shmp(handle, chan->backend.buf[0].shmp);
+ if (!buf)
+ return -EIO;
ctx->buf = buf;
offsets.size = 0;
do {
ret = lib_ring_buffer_try_reserve_slow(buf, chan, &offsets,
- ctx);
+ ctx, client_ctx);
if (caa_unlikely(ret))
return ret;
} while (caa_unlikely(v_cmpxchg(config, &buf->offset, offsets.old,
return 0;
}
+static
+void lib_ring_buffer_vmcore_check_deliver(const struct lttng_ust_lib_ring_buffer_config *config,
+ struct lttng_ust_lib_ring_buffer *buf,
+ unsigned long commit_count,
+ unsigned long idx,
+ struct lttng_ust_shm_handle *handle)
+{
+ struct commit_counters_hot *cc_hot;
+
+ if (config->oops != RING_BUFFER_OOPS_CONSISTENCY)
+ return;
+ cc_hot = shmp_index(handle, buf->commit_hot, idx);
+ if (!cc_hot)
+ return;
+ v_set(config, &cc_hot->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 lttng_ust_lib_ring_buffer_config *config,
+ struct lttng_ust_lib_ring_buffer *buf,
+ unsigned long idx,
+ struct lttng_ust_shm_handle *handle)
+{
+ v_add(config, subbuffer_get_records_count(config,
+ &buf->backend, idx, handle),
+ &buf->records_count);
+ v_add(config, subbuffer_count_records_overrun(config,
+ &buf->backend, idx, handle),
+ &buf->records_overrun);
+}
+#else /* LTTNG_RING_BUFFER_COUNT_EVENTS */
+static
+void deliver_count_events(const struct lttng_ust_lib_ring_buffer_config *config,
+ struct lttng_ust_lib_ring_buffer *buf,
+ unsigned long idx,
+ struct lttng_ust_shm_handle *handle)
+{
+}
+#endif /* #else LTTNG_RING_BUFFER_COUNT_EVENTS */
+
+void lib_ring_buffer_check_deliver_slow(const struct lttng_ust_lib_ring_buffer_config *config,
+ struct lttng_ust_lib_ring_buffer *buf,
+ struct channel *chan,
+ unsigned long offset,
+ unsigned long commit_count,
+ unsigned long idx,
+ struct lttng_ust_shm_handle *handle,
+ uint64_t tsc)
+{
+ unsigned long old_commit_count = commit_count
+ - chan->backend.subbuf_size;
+ struct commit_counters_cold *cc_cold;
+
+ /*
+ * 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.
+ */
+ cmm_smp_wmb();
+ cc_cold = shmp_index(handle, buf->commit_cold, idx);
+ if (!cc_cold)
+ return;
+ if (caa_likely(v_cmpxchg(config, &cc_cold->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, handle);
+ config->cb.buffer_end(buf, tsc, idx,
+ lib_ring_buffer_get_data_size(config,
+ buf,
+ idx,
+ handle),
+ handle);
+
+ /*
+ * Increment the packet counter while we have exclusive
+ * access.
+ */
+ subbuffer_inc_packet_count(config, &buf->backend, idx, handle);
+
+ /*
+ * 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), handle);
+
+ /*
+ * 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.
+ */
+ cmm_smp_mb();
+ /* End of exclusive subbuffer access */
+ v_set(config, &cc_cold->cc_sb, commit_count);
+ /*
+ * Order later updates to reserve count after
+ * the commit cold cc_sb update.
+ */
+ cmm_smp_wmb();
+ lib_ring_buffer_vmcore_check_deliver(config, buf,
+ commit_count, idx, handle);
+
+ /*
+ * RING_BUFFER_WAKEUP_BY_WRITER wakeup is not lock-free.
+ */
+ if (config->wakeup == RING_BUFFER_WAKEUP_BY_WRITER
+ && uatomic_read(&buf->active_readers)
+ && lib_ring_buffer_poll_deliver(config, buf, chan, handle)) {
+ lib_ring_buffer_wakeup(buf, handle);
+ }
+ }
+}
+
/*
* Force a read (imply TLS fixup for dlopen) of TLS variables.
*/
projects / lttng-ust.git / blobdiff