/*
* 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.
*/
+#define _GNU_SOURCE
+#define _LGPL_SOURCE
#include <sys/types.h>
#include <sys/mman.h>
#include <sys/stat.h>
+#include <unistd.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 "config.h"
+#include <lttng/ringbuffer-config.h>
+#include "vatomic.h"
#include "backend.h"
#include "frontend.h"
#include "shm.h"
+#include "rb-init.h"
+#include "../liblttng-ust/compat.h" /* For ENODATA */
+
+/* Print DBG() messages about events lost only every 1048576 hits */
+#define DBG_PRINT_NR_LOST (1UL << 20)
+
+#define LTTNG_UST_RB_SIG_FLUSH SIGRTMIN
+#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;
-#ifndef max
-#define max(a, b) ((a) > (b) ? (a) : (b))
-#endif
+/*
+ * Use POSIX SHM: shm_open(3) and shm_unlink(3).
+ * close(2) to close the fd returned by shm_open.
+ * shm_unlink releases the shared memory object name.
+ * ftruncate(2) sets the size of the memory object.
+ * mmap/munmap maps the shared memory obj to a virtual address in the
+ * calling proceess (should be done both in libust and consumer).
+ * See shm_overview(7) for details.
+ * Pass file descriptor returned by shm_open(3) to ltt-sessiond through
+ * a UNIX socket.
+ *
+ * Since we don't need to access the object using its name, we can
+ * immediately shm_unlink(3) it, and only keep the handle with its file
+ * descriptor.
+ */
/*
* Internal structure representing offsets to use at a sub-buffer switch.
switch_old_end:1;
};
-__thread unsigned int lib_ring_buffer_nesting;
+DEFINE_URCU_TLS(unsigned int, lib_ring_buffer_nesting);
+
+/*
+ * wakeup_fd_mutex protects wakeup fd use by timer from concurrent
+ * close.
+ */
+static pthread_mutex_t wakeup_fd_mutex = PTHREAD_MUTEX_INITIALIZER;
static
void lib_ring_buffer_print_errors(struct channel *chan,
- struct lib_ring_buffer *buf, int cpu);
+ struct lttng_ust_lib_ring_buffer *buf, int cpu,
+ struct lttng_ust_shm_handle *handle);
/*
- * Must be called under cpu hotplug protection.
+ * Handle timer teardown race wrt memory free of private data by
+ * ring buffer signals are handled by a single thread, which permits
+ * a synchronization point between handling of each signal.
+ * Protected by the lock within the structure.
*/
-void lib_ring_buffer_free(struct lib_ring_buffer *buf)
-{
- struct channel *chan = shmp(buf->backend.chan);
+struct timer_signal_data {
+ pthread_t tid; /* thread id managing signals */
+ int setup_done;
+ int qs_done;
+ pthread_mutex_t lock;
+};
+
+static struct timer_signal_data timer_signal = {
+ .tid = 0,
+ .setup_done = 0,
+ .qs_done = 0,
+ .lock = PTHREAD_MUTEX_INITIALIZER,
+};
- lib_ring_buffer_print_errors(chan, buf, buf->backend.cpu);
- /* buf->commit_hot will be freed by shm teardown */
- /* buf->commit_cold will be freed by shm teardown */
+int lttng_ust_blocking_retry_timeout =
+ CONFIG_LTTNG_UST_DEFAULT_BLOCKING_RETRY_TIMEOUT_MS;
- lib_ring_buffer_backend_free(&buf->backend);
+void lttng_ust_ringbuffer_set_retry_timeout(int timeout)
+{
+ lttng_ust_blocking_retry_timeout = timeout;
}
/**
* should not be using the iterator concurrently with reset. The previous
* current iterator record is reset.
*/
-void lib_ring_buffer_reset(struct lib_ring_buffer *buf)
+void lib_ring_buffer_reset(struct lttng_ust_lib_ring_buffer *buf,
+ struct lttng_ust_shm_handle *handle)
{
- struct channel *chan = shmp(buf->backend.chan);
- const struct 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(buf->commit_hot)[i].cc, 0);
- v_set(config, &shmp(buf->commit_hot)[i].seq, 0);
- v_set(config, &shmp(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);
v_set(config, &buf->last_tsc, 0);
- lib_ring_buffer_backend_reset(&buf->backend);
+ lib_ring_buffer_backend_reset(&buf->backend, handle);
/* Don't reset number of active readers */
v_set(config, &buf->records_lost_full, 0);
v_set(config, &buf->records_lost_wrap, 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.
*/
-int lib_ring_buffer_create(struct lib_ring_buffer *buf,
+int lib_ring_buffer_create(struct lttng_ust_lib_ring_buffer *buf,
struct channel_backend *chanb, int cpu,
- struct shm_header *shm_header)
+ struct lttng_ust_shm_handle *handle,
+ struct shm_object *shmobj)
{
- const struct lib_ring_buffer_config *config = chanb->config;
+ const struct lttng_ust_lib_ring_buffer_config *config = &chanb->config;
struct channel *chan = caa_container_of(chanb, struct channel, backend);
- void *priv = chanb->priv;
- unsigned int num_subbuf;
+ 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;
- u64 tsc;
+ uint64_t tsc;
int ret;
/* Test for cpu hotplug */
if (buf->backend.allocated)
return 0;
- ret = lib_ring_buffer_backend_create(&buf->backend, &chan->backend,
- cpu, shm_header);
- if (ret)
- return ret;
-
- align_shm(shm_header,
- max(__alignof__(struct commit_counters_hot),
- __alignof__(struct commit_counters_cold)));
- set_shmp(&buf->commit_hot,
- zalloc_shm(shm_header,
- sizeof(*buf->commit_hot) * chan->backend.num_subbuf));
- if (!shmp(buf->commit_hot)) {
- ret = -ENOMEM;
- goto free_chanbuf;
+ 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)) {
+ return -ENOMEM;
}
- align_shm(shm_header, __alignof__(struct commit_counters_cold));
- set_shmp(&buf->commit_cold,
- zalloc_shm(shm_header,
- sizeof(*buf->commit_cold) * chan->backend.num_subbuf));
- if (!shmp(buf->commit_cold)) {
+ align_shm(shmobj, __alignof__(struct commit_counters_cold));
+ set_shmp(buf->commit_cold,
+ zalloc_shm(shmobj,
+ sizeof(struct commit_counters_cold) * chan->backend.num_subbuf));
+ if (!shmp(handle, buf->commit_cold)) {
ret = -ENOMEM;
goto free_commit;
}
- num_subbuf = chan->backend.num_subbuf;
- //init_waitqueue_head(&buf->read_wait);
+ 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
*/
subbuf_header_size = config->cb.subbuffer_header_size();
v_set(config, &buf->offset, subbuf_header_size);
- subbuffer_id_clear_noref(config, &shmp(buf->backend.buf_wsb)[0].id);
- tsc = config->cb.ring_buffer_clock_read(shmp(buf->backend.chan));
- config->cb.buffer_begin(buf, tsc, 0);
- v_add(config, subbuf_header_size, &shmp(buf->commit_hot)[0].cc);
+ 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);
+ 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);
+ 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;
free_commit:
/* commit_hot will be freed by shm teardown */
free_chanbuf:
- lib_ring_buffer_backend_free(&buf->backend);
return ret;
}
-static void switch_buffer_timer(unsigned long data)
+static
+void lib_ring_buffer_channel_switch_timer(int sig, siginfo_t *si, void *uc)
{
- struct lib_ring_buffer *buf = (struct lib_ring_buffer *)data;
- struct channel *chan = shmp(buf->backend.chan);
- const struct lib_ring_buffer_config *config = chan->backend.config;
+ const struct lttng_ust_lib_ring_buffer_config *config;
+ struct lttng_ust_shm_handle *handle;
+ struct channel *chan;
+ int cpu;
+
+ assert(CMM_LOAD_SHARED(timer_signal.tid) == pthread_self());
+
+ chan = si->si_value.sival_ptr;
+ handle = chan->handle;
+ config = &chan->backend.config;
+
+ DBG("Switch timer for channel %p\n", chan);
/*
* Only flush buffers periodically if readers are active.
*/
- if (uatomic_read(&buf->active_readers))
- lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE);
-
- //TODO timers
- //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
- // mod_timer_pinned(&buf->switch_timer,
- // jiffies + chan->switch_timer_interval);
- //else
- // mod_timer(&buf->switch_timer,
- // jiffies + chan->switch_timer_interval);
+ pthread_mutex_lock(&wakeup_fd_mutex);
+ if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) {
+ 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);
+ }
+ } else {
+ 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 void lib_ring_buffer_start_switch_timer(struct lib_ring_buffer *buf)
+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)
{
- struct channel *chan = shmp(buf->backend.chan);
- const struct lib_ring_buffer_config *config = chan->backend.config;
+ unsigned long consumed_old, consumed_idx, commit_count, write_offset;
+ struct commit_counters_cold *cc_cold;
- if (!chan->switch_timer_interval || buf->switch_timer_enabled)
- return;
- //TODO
- //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)
- // add_timer_on(&buf->switch_timer, buf->backend.cpu);
- //else
- // add_timer(&buf->switch_timer);
- buf->switch_timer_enabled = 1;
+ 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_stop_switch_timer(struct lib_ring_buffer *buf)
+static
+void lib_ring_buffer_wakeup(struct lttng_ust_lib_ring_buffer *buf,
+ struct lttng_ust_shm_handle *handle)
{
- struct channel *chan = shmp(buf->backend.chan);
+ 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 (!chan->switch_timer_interval || !buf->switch_timer_enabled)
+ if (wakeup_fd < 0)
return;
- //TODO
- //del_timer_sync(&buf->switch_timer);
- buf->switch_timer_enabled = 0;
+ /*
+ * 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)
+{
+ const struct lttng_ust_lib_ring_buffer_config *config;
+ struct lttng_ust_shm_handle *handle;
+ int cpu;
+
+ handle = chan->handle;
+ config = &chan->backend.config;
+
+ /*
+ * Only flush buffers periodically if readers are active.
+ */
+ pthread_mutex_lock(&wakeup_fd_mutex);
+ if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) {
+ 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_poll_deliver(config, buf,
+ chan, handle)) {
+ lib_ring_buffer_wakeup(buf, handle);
+ }
+ }
+ } else {
+ 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);
+}
+
+static
+void lib_ring_buffer_channel_read_timer(int sig, siginfo_t *si, void *uc)
+{
+ struct channel *chan;
+
+ assert(CMM_LOAD_SHARED(timer_signal.tid) == pthread_self());
+ chan = si->si_value.sival_ptr;
+ DBG("Read timer for channel %p\n", chan);
+ lib_ring_buffer_channel_do_read(chan);
+ return;
+}
+
+static
+void rb_setmask(sigset_t *mask)
+{
+ int ret;
+
+ ret = sigemptyset(mask);
+ if (ret) {
+ PERROR("sigemptyset");
+ }
+ ret = sigaddset(mask, LTTNG_UST_RB_SIG_FLUSH);
+ if (ret) {
+ PERROR("sigaddset");
+ }
+ ret = sigaddset(mask, LTTNG_UST_RB_SIG_READ);
+ if (ret) {
+ PERROR("sigaddset");
+ }
+ ret = sigaddset(mask, LTTNG_UST_RB_SIG_TEARDOWN);
+ if (ret) {
+ PERROR("sigaddset");
+ }
+}
+
+static
+void *sig_thread(void *arg)
+{
+ sigset_t mask;
+ siginfo_t info;
+ int signr;
+
+ /* Only self thread will receive signal mask. */
+ rb_setmask(&mask);
+ CMM_STORE_SHARED(timer_signal.tid, pthread_self());
+
+ for (;;) {
+ signr = sigwaitinfo(&mask, &info);
+ if (signr == -1) {
+ if (errno != EINTR)
+ PERROR("sigwaitinfo");
+ continue;
+ }
+ if (signr == LTTNG_UST_RB_SIG_FLUSH) {
+ lib_ring_buffer_channel_switch_timer(info.si_signo,
+ &info, NULL);
+ } else if (signr == LTTNG_UST_RB_SIG_READ) {
+ lib_ring_buffer_channel_read_timer(info.si_signo,
+ &info, NULL);
+ } else if (signr == LTTNG_UST_RB_SIG_TEARDOWN) {
+ cmm_smp_mb();
+ CMM_STORE_SHARED(timer_signal.qs_done, 1);
+ cmm_smp_mb();
+ } else {
+ ERR("Unexptected signal %d\n", info.si_signo);
+ }
+ }
+ return NULL;
}
/*
- * Polling timer to check the channels for data.
+ * Ensure only a single thread listens on the timer signal.
*/
-static void read_buffer_timer(unsigned long data)
+static
+void lib_ring_buffer_setup_timer_thread(void)
{
- struct lib_ring_buffer *buf = (struct lib_ring_buffer *)data;
- struct channel *chan = shmp(buf->backend.chan);
- const struct lib_ring_buffer_config *config = chan->backend.config;
+ pthread_t thread;
+ int ret;
- CHAN_WARN_ON(chan, !buf->backend.allocated);
+ pthread_mutex_lock(&timer_signal.lock);
+ if (timer_signal.setup_done)
+ goto end;
- if (uatomic_read(&buf->active_readers)
- && lib_ring_buffer_poll_deliver(config, buf, chan)) {
- //TODO
- //wake_up_interruptible(&buf->read_wait);
- //wake_up_interruptible(&chan->read_wait);
+ ret = pthread_create(&thread, NULL, &sig_thread, NULL);
+ if (ret) {
+ errno = ret;
+ PERROR("pthread_create");
+ }
+ ret = pthread_detach(thread);
+ if (ret) {
+ errno = ret;
+ PERROR("pthread_detach");
}
+ timer_signal.setup_done = 1;
+end:
+ pthread_mutex_unlock(&timer_signal.lock);
+}
+
+/*
+ * Wait for signal-handling thread quiescent state.
+ */
+static
+void lib_ring_buffer_wait_signal_thread_qs(unsigned int signr)
+{
+ sigset_t pending_set;
+ int ret;
- //TODO
- //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
- // mod_timer_pinned(&buf->read_timer,
- // jiffies + chan->read_timer_interval);
- //else
- // mod_timer(&buf->read_timer,
- // jiffies + chan->read_timer_interval);
+ /*
+ * We need to be the only thread interacting with the thread
+ * that manages signals for teardown synchronization.
+ */
+ pthread_mutex_lock(&timer_signal.lock);
+
+ /*
+ * Ensure we don't have any signal queued for this channel.
+ */
+ for (;;) {
+ ret = sigemptyset(&pending_set);
+ if (ret == -1) {
+ PERROR("sigemptyset");
+ }
+ ret = sigpending(&pending_set);
+ if (ret == -1) {
+ PERROR("sigpending");
+ }
+ if (!sigismember(&pending_set, signr))
+ break;
+ caa_cpu_relax();
+ }
+
+ /*
+ * From this point, no new signal handler will be fired that
+ * would try to access "chan". However, we still need to wait
+ * for any currently executing handler to complete.
+ */
+ cmm_smp_mb();
+ CMM_STORE_SHARED(timer_signal.qs_done, 0);
+ cmm_smp_mb();
+
+ /*
+ * Kill with LTTNG_UST_RB_SIG_TEARDOWN, so signal management
+ * thread wakes up.
+ */
+ kill(getpid(), LTTNG_UST_RB_SIG_TEARDOWN);
+
+ while (!CMM_LOAD_SHARED(timer_signal.qs_done))
+ caa_cpu_relax();
+ cmm_smp_mb();
+
+ pthread_mutex_unlock(&timer_signal.lock);
}
-static void lib_ring_buffer_start_read_timer(struct lib_ring_buffer *buf)
+static
+void lib_ring_buffer_channel_switch_timer_start(struct channel *chan)
{
- struct channel *chan = shmp(buf->backend.chan);
- const struct lib_ring_buffer_config *config = chan->backend.config;
+ struct sigevent sev;
+ struct itimerspec its;
+ int ret;
+
+ if (!chan->switch_timer_interval || chan->switch_timer_enabled)
+ return;
+
+ chan->switch_timer_enabled = 1;
+
+ lib_ring_buffer_setup_timer_thread();
+
+ sev.sigev_notify = SIGEV_SIGNAL;
+ sev.sigev_signo = LTTNG_UST_RB_SIG_FLUSH;
+ sev.sigev_value.sival_ptr = chan;
+ ret = timer_create(CLOCKID, &sev, &chan->switch_timer);
+ if (ret == -1) {
+ PERROR("timer_create");
+ }
+
+ its.it_value.tv_sec = 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;
+
+ ret = timer_settime(chan->switch_timer, 0, &its, NULL);
+ if (ret == -1) {
+ PERROR("timer_settime");
+ }
+}
+
+static
+void lib_ring_buffer_channel_switch_timer_stop(struct channel *chan)
+{
+ int ret;
+
+ if (!chan->switch_timer_interval || !chan->switch_timer_enabled)
+ return;
+
+ ret = timer_delete(chan->switch_timer);
+ if (ret == -1) {
+ PERROR("timer_delete");
+ }
+
+ lib_ring_buffer_wait_signal_thread_qs(LTTNG_UST_RB_SIG_FLUSH);
+
+ chan->switch_timer = 0;
+ chan->switch_timer_enabled = 0;
+}
+
+static
+void lib_ring_buffer_channel_read_timer_start(struct channel *chan)
+{
+ const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config;
+ struct sigevent sev;
+ struct itimerspec its;
+ int ret;
if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER
- || !chan->read_timer_interval
- || buf->read_timer_enabled)
+ || !chan->read_timer_interval || chan->read_timer_enabled)
return;
- //TODO
- //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;
-
- //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
- // add_timer_on(&buf->read_timer, buf->backend.cpu);
- //else
- // add_timer(&buf->read_timer);
- buf->read_timer_enabled = 1;
+ chan->read_timer_enabled = 1;
+
+ lib_ring_buffer_setup_timer_thread();
+
+ sev.sigev_notify = SIGEV_SIGNAL;
+ sev.sigev_signo = LTTNG_UST_RB_SIG_READ;
+ sev.sigev_value.sival_ptr = chan;
+ ret = timer_create(CLOCKID, &sev, &chan->read_timer);
+ if (ret == -1) {
+ PERROR("timer_create");
+ }
+
+ its.it_value.tv_sec = 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;
+
+ ret = timer_settime(chan->read_timer, 0, &its, NULL);
+ if (ret == -1) {
+ PERROR("timer_settime");
+ }
}
-static void lib_ring_buffer_stop_read_timer(struct lib_ring_buffer *buf)
+static
+void lib_ring_buffer_channel_read_timer_stop(struct channel *chan)
{
- struct channel *chan = shmp(buf->backend.chan);
- const struct lib_ring_buffer_config *config = chan->backend.config;
+ const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config;
+ int ret;
if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER
- || !chan->read_timer_interval
- || !buf->read_timer_enabled)
+ || !chan->read_timer_interval || !chan->read_timer_enabled)
return;
- //TODO
- //del_timer_sync(&buf->read_timer);
+ ret = timer_delete(chan->read_timer);
+ if (ret == -1) {
+ PERROR("timer_delete");
+ }
+
/*
* do one more check to catch data that has been written in the last
* timer period.
*/
- if (lib_ring_buffer_poll_deliver(config, buf, chan)) {
- //TODO
- //wake_up_interruptible(&buf->read_wait);
- //wake_up_interruptible(&chan->read_wait);
- }
- buf->read_timer_enabled = 0;
+ lib_ring_buffer_channel_do_read(chan);
+
+ lib_ring_buffer_wait_signal_thread_qs(LTTNG_UST_RB_SIG_READ);
+
+ chan->read_timer = 0;
+ chan->read_timer_enabled = 0;
}
-static void channel_unregister_notifiers(struct channel *chan)
+static void channel_unregister_notifiers(struct channel *chan,
+ struct lttng_ust_shm_handle *handle)
{
- const struct lib_ring_buffer_config *config = chan->backend.config;
+ lib_ring_buffer_channel_switch_timer_stop(chan);
+ lib_ring_buffer_channel_read_timer_stop(chan);
+}
+
+static void channel_print_errors(struct channel *chan,
+ struct lttng_ust_shm_handle *handle)
+{
+ const struct lttng_ust_lib_ring_buffer_config *config =
+ &chan->backend.config;
int cpu;
if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) {
for_each_possible_cpu(cpu) {
- struct lib_ring_buffer *buf = &shmp(chan->backend.buf)[cpu];
-
- lib_ring_buffer_stop_switch_timer(buf);
- lib_ring_buffer_stop_read_timer(buf);
+ struct lttng_ust_lib_ring_buffer *buf =
+ shmp(handle, chan->backend.buf[cpu].shmp);
+ if (buf)
+ lib_ring_buffer_print_errors(chan, buf, cpu, handle);
}
} else {
- struct lib_ring_buffer *buf = shmp(chan->backend.buf);
+ struct lttng_ust_lib_ring_buffer *buf =
+ shmp(handle, chan->backend.buf[0].shmp);
- lib_ring_buffer_stop_switch_timer(buf);
- lib_ring_buffer_stop_read_timer(buf);
+ if (buf)
+ lib_ring_buffer_print_errors(chan, buf, -1, handle);
}
- channel_backend_unregister_notifiers(&chan->backend);
}
-static void channel_free(struct shm_handle *handle)
+static void channel_free(struct channel *chan,
+ struct lttng_ust_shm_handle *handle,
+ int consumer)
{
- struct shm_header *header = handle->header;
- struct channel *chan = shmp(header->chan);
- int ret;
-
- channel_backend_free(&chan->backend);
+ channel_backend_free(&chan->backend, handle);
/* chan is freed by shm teardown */
- ret = munmap(header, header->shm_size);
- if (ret) {
- PERROR("umnmap");
- assert(0);
- }
- ret = close(handle->shmfd);
- if (ret) {
- PERROR("close");
- assert(0);
- }
+ shm_object_table_destroy(handle->table, consumer);
+ free(handle);
}
/**
* channel_create - Create channel.
* @config: ring buffer instance configuration
* @name: name of the channel
- * @priv: ring buffer client private data
+ * @priv_data: ring buffer client private data area pointer (output)
+ * @priv_data_size: length, in bytes, of the private data area.
+ * @priv_data_init: initialization data for private data.
* @buf_addr: pointer the the beginning of the preallocated buffer contiguous
* address mapping. It is used only by RING_BUFFER_STATIC
* configuration. It can be set to NULL for other backends.
* padding to let readers get those sub-buffers.
* Used for live streaming.
* @read_timer_interval: Time interval (in us) to wake up pending readers.
- * @shmfd: shared memory file descriptor (output, needs to be closed by
- * the caller)
+ * @stream_fds: array of stream file descriptors.
+ * @nr_stream_fds: number of file descriptors in array.
*
* Holds cpu hotplug.
* Returns NULL on failure.
*/
-struct shm_handle *channel_create(const struct lib_ring_buffer_config *config,
- const char *name, void *priv, void *buf_addr,
- size_t subbuf_size,
+struct lttng_ust_shm_handle *channel_create(const struct lttng_ust_lib_ring_buffer_config *config,
+ const char *name,
+ void **priv_data,
+ size_t priv_data_align,
+ size_t priv_data_size,
+ 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, cpu, shmfd;
+ int ret;
+ size_t shmsize, chansize;
struct channel *chan;
- size_t shmsize, bufshmsize, bufshmalign;
- struct shm_header *shm_header;
- unsigned long num_subbuf_alloc;
- struct shm_handle *handle;
+ struct lttng_ust_shm_handle *handle;
+ struct shm_object *shmobj;
+ unsigned int nr_streams;
+
+ if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
+ nr_streams = num_possible_cpus();
+ 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;
- handle = zmalloc(sizeof(struct shm_handle));
+ handle = zmalloc(sizeof(struct lttng_ust_shm_handle));
if (!handle)
return NULL;
- /* Calculate the shm allocation layout */
- shmsize = sizeof(struct shm_header);
- shmsize += offset_align(shmsize, __alignof__(struct channel));
- shmsize += sizeof(struct channel);
-
- /* Per-cpu buffer size: control (prior to backend) */
- shmsize += offset_align(shmsize, __alignof__(struct lib_ring_buffer));
- bufshmsize = sizeof(struct lib_ring_buffer);
- shmsize += bufshmsize * num_possible_cpus();
-
- /* Per-cpu buffer size: backend */
- shmsize += offset_align(shmsize, PAGE_SIZE);
- /* num_subbuf + 1 is the worse case */
- num_subbuf_alloc = num_subbuf + 1;
- bufshmsize = sizeof(struct lib_ring_buffer_backend_pages *) * num_subbuf_alloc;
- bufshmsize += offset_align(bufshmsize, PAGE_SIZE);
- bufshmsize += subbuf_size * num_subbuf_alloc;
- bufshmsize += offset_align(bufshmsize, __alignof__(struct lib_ring_buffer_backend_pages));
- bufshmsize += sizeof(struct lib_ring_buffer_backend_pages) * num_subbuf_alloc;
- bufshmsize += offset_align(bufshmsize, __alignof__(struct lib_ring_buffer_backend_subbuffer));
- bufshmsize += sizeof(struct lib_ring_buffer_backend_subbuffer) * num_subbuf;
- bufshmsize += offset_align(bufshmsize, PAGE_SIZE);
- shmsize += bufshmsize * num_possible_cpus();
-
- /* Per-cpu buffer size: control (after backend) */
- shmsize += offset_align(shmsize,
- max(__alignof__(struct commit_counters_hot),
- __alignof__(struct commit_counters_cold)));
- bufshmsize = sizeof(struct commit_counters_hot) * num_subbuf;
- bufshmsize += offset_align(bufshmsize, __alignof__(struct commit_counters_cold));
- bufshmsize += sizeof(struct commit_counters_cold) * num_subbuf;
- shmsize += bufshmsize * num_possible_cpus();
+ /* Allocate table for channel + per-cpu buffers */
+ handle->table = shm_object_table_create(1 + num_possible_cpus());
+ if (!handle->table)
+ goto error_table_alloc;
- /*
- * Allocate shm, and immediately unlink its shm oject, keeping
- * only the file descriptor as a reference to the object. If it
- * already exists (caused by short race window during which the
- * global object exists in a concurrent shm_open), simply retry.
- */
- do {
- shmfd = shm_open("/ust-shm-tmp",
- O_CREAT | O_EXCL | O_RDWR, 0700);
- } while (shmfd < 0 && errno == EEXIST);
- if (shmfd < 0) {
- PERROR("shm_open");
- goto error_shm_open;
- }
- ret = shm_unlink("/ust-shm-tmp");
- if (ret) {
- PERROR("shm_unlink");
- goto error_unlink;
- }
- ret = ftruncate(shmfd, shmsize);
- if (ret) {
- PERROR("ftruncate");
- goto error_ftruncate;
+ /* Calculate the shm allocation layout */
+ shmsize = sizeof(struct channel);
+ shmsize += offset_align(shmsize, __alignof__(struct lttng_ust_lib_ring_buffer_shmp));
+ shmsize += sizeof(struct lttng_ust_lib_ring_buffer_shmp) * nr_streams;
+ chansize = shmsize;
+ if (priv_data_align)
+ shmsize += offset_align(shmsize, priv_data_align);
+ shmsize += priv_data_size;
+
+ /* Allocate normal memory for channel (not shared) */
+ 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) */
+ set_shmp(handle->chan, zalloc_shm(shmobj, chansize));
+ assert(handle->chan._ref.index == 0);
+ assert(handle->chan._ref.offset == 0);
+ chan = shmp(handle, handle->chan);
+ if (!chan)
+ goto error_append;
+ chan->nr_streams = nr_streams;
+
+ /* space for private data */
+ if (priv_data_size) {
+ DECLARE_SHMP(void, priv_data_alloc);
+
+ align_shm(shmobj, priv_data_align);
+ chan->priv_data_offset = shmobj->allocated_len;
+ set_shmp(priv_data_alloc, zalloc_shm(shmobj, priv_data_size));
+ if (!shmp(handle, priv_data_alloc))
+ goto error_append;
+ *priv_data = channel_get_private(chan);
+ memcpy(*priv_data, priv_data_init, priv_data_size);
+ } else {
+ chan->priv_data_offset = -1;
+ if (priv_data)
+ *priv_data = NULL;
}
- shm_header = mmap(NULL, shmsize, PROT_READ | PROT_WRITE,
- MAP_SHARED, shmfd, 0);
- if (shm_header == MAP_FAILED) {
- PERROR("mmap");
- goto error_mmap;
- }
+ ret = channel_backend_init(&chan->backend, name, config,
+ subbuf_size, num_subbuf, handle,
+ stream_fds);
+ if (ret)
+ goto error_backend_init;
- shm_header->magic = SHM_MAGIC;
- shm_header->major = SHM_MAJOR;
- shm_header->major = SHM_MINOR;
- shm_header->bits_per_long = CAA_BITS_PER_LONG;
- shm_header->shm_size = shmsize;
- shm_header->shm_allocated = sizeof(struct shm_header);
+ chan->handle = handle;
+ chan->commit_count_mask = (~0UL >> chan->backend.num_subbuf_order);
- align_shm(shm_header, __alignof__(struct channel));
- chan = zalloc_shm(shm_header, sizeof(struct channel));
- if (!chan)
- goto destroy_shmem;
- set_shmp(shm_header->chan, chan);
+ chan->switch_timer_interval = switch_timer_interval;
+ chan->read_timer_interval = read_timer_interval;
+ lib_ring_buffer_channel_switch_timer_start(chan);
+ lib_ring_buffer_channel_read_timer_start(chan);
- ret = channel_backend_init(&chan->backend, name, config, priv,
- subbuf_size, num_subbuf, shm_header);
- if (ret)
- goto destroy_shmem;
+ return handle;
- chan->commit_count_mask = (~0UL >> chan->backend.num_subbuf_order);
- //TODO
- //chan->switch_timer_interval = usecs_to_jiffies(switch_timer_interval);
- //chan->read_timer_interval = usecs_to_jiffies(read_timer_interval);
- //TODO
- //init_waitqueue_head(&chan->read_wait);
- //init_waitqueue_head(&chan->hp_wait);
+error_backend_init:
+error_append:
+ shm_object_table_destroy(handle->table, 1);
+error_table_alloc:
+ free(handle);
+ return NULL;
+}
- if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) {
- /*
- * In case of non-hotplug cpu, if the ring-buffer is allocated
- * in early initcall, it will not be notified of secondary cpus.
- * In that off case, we need to allocate for all possible cpus.
- */
- for_each_possible_cpu(cpu) {
- struct lib_ring_buffer *buf = &shmp(chan->backend.buf)[cpu];
- lib_ring_buffer_start_switch_timer(buf);
- lib_ring_buffer_start_read_timer(buf);
- }
- } else {
- struct lib_ring_buffer *buf = shmp(chan->backend.buf);
+struct lttng_ust_shm_handle *channel_handle_create(void *data,
+ uint64_t memory_map_size,
+ int wakeup_fd)
+{
+ struct lttng_ust_shm_handle *handle;
+ struct shm_object *object;
- lib_ring_buffer_start_switch_timer(buf);
- lib_ring_buffer_start_read_timer(buf);
- }
+ handle = zmalloc(sizeof(struct lttng_ust_shm_handle));
+ if (!handle)
+ return NULL;
- handle->header = shm_header;
- handle->shmfd = shmfd;
+ /* Allocate table for channel + per-cpu buffers */
+ handle->table = shm_object_table_create(1 + num_possible_cpus());
+ if (!handle->table)
+ goto error_table_alloc;
+ /* Add channel object */
+ object = shm_object_table_append_mem(handle->table, data,
+ memory_map_size, wakeup_fd);
+ if (!object)
+ goto error_table_object;
+ /* struct channel is at object 0, offset 0 (hardcoded) */
+ handle->chan._ref.index = 0;
+ handle->chan._ref.offset = 0;
return handle;
-destroy_shmem:
- ret = munmap(shm_header, shmsize);
- if (ret) {
- PERROR("umnmap");
- assert(0);
- }
-error_mmap:
-error_ftruncate:
-error_unlink:
- ret = close(shmfd);
- if (ret) {
- PERROR("close");
- assert(0);
- }
-error_shm_open:
+error_table_object:
+ shm_object_table_destroy(handle->table, 0);
+error_table_alloc:
free(handle);
return NULL;
}
+int channel_handle_add_stream(struct lttng_ust_shm_handle *handle,
+ int shm_fd, int wakeup_fd, uint32_t stream_nr,
+ uint64_t memory_map_size)
+{
+ struct shm_object *object;
+
+ /* Add stream object */
+ object = shm_object_table_append_shm(handle->table,
+ shm_fd, wakeup_fd, stream_nr,
+ memory_map_size);
+ if (!object)
+ return -EINVAL;
+ return 0;
+}
+
+unsigned int channel_handle_get_nr_streams(struct lttng_ust_shm_handle *handle)
+{
+ assert(handle->table);
+ return handle->table->allocated_len - 1;
+}
+
static
-void channel_release(struct shm_handle *handle)
+void channel_release(struct channel *chan, struct lttng_ust_shm_handle *handle,
+ int consumer)
{
- channel_free(handle);
+ channel_free(chan, handle, consumer);
}
/**
* Call "destroy" callback, finalize channels, 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.
+ * They should release their handle at that point.
*/
-void *channel_destroy(struct shm_handle *handle)
+void channel_destroy(struct channel *chan, struct lttng_ust_shm_handle *handle,
+ int consumer)
{
- struct shm_header *header = handle->header;
- struct channel *chan = shmp(header->chan);
- const struct lib_ring_buffer_config *config = chan->backend.config;
- void *priv;
- int cpu;
-
- channel_unregister_notifiers(chan);
-
- if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) {
- for_each_channel_cpu(cpu, chan) {
- struct lib_ring_buffer *buf = &shmp(chan->backend.buf)[cpu];
-
- if (config->cb.buffer_finalize)
- config->cb.buffer_finalize(buf,
- chan->backend.priv,
- cpu);
- if (buf->backend.allocated)
- lib_ring_buffer_switch_slow(buf, SWITCH_FLUSH);
- /*
- * Perform flush before writing to finalized.
- */
- cmm_smp_wmb();
- CMM_ACCESS_ONCE(buf->finalized) = 1;
- //wake_up_interruptible(&buf->read_wait);
- }
- } else {
- struct lib_ring_buffer *buf = shmp(chan->backend.buf);
-
- 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);
+ if (consumer) {
/*
- * Perform flush before writing to finalized.
+ * Note: the consumer takes care of finalizing and
+ * switching the buffers.
*/
- cmm_smp_wmb();
- CMM_ACCESS_ONCE(buf->finalized) = 1;
- //wake_up_interruptible(&buf->read_wait);
+ channel_unregister_notifiers(chan, handle);
+ /*
+ * The consumer prints errors.
+ */
+ channel_print_errors(chan, handle);
}
- CMM_ACCESS_ONCE(chan->finalized) = 1;
- //wake_up_interruptible(&chan->hp_wait);
- //wake_up_interruptible(&chan->read_wait);
+
/*
* sessiond/consumer are keeping a reference on the shm file
* descriptor directly. No need to refcount.
*/
- channel_release(handle);
- priv = chan->backend.priv;
- return priv;
+ channel_release(chan, handle, consumer);
+ return;
}
-struct lib_ring_buffer *channel_get_ring_buffer(
- const struct lib_ring_buffer_config *config,
- struct channel *chan, int cpu)
+struct lttng_ust_lib_ring_buffer *channel_get_ring_buffer(
+ const struct lttng_ust_lib_ring_buffer_config *config,
+ struct channel *chan, int cpu,
+ struct lttng_ust_shm_handle *handle,
+ int *shm_fd, int *wait_fd,
+ int *wakeup_fd,
+ uint64_t *memory_map_size)
{
- if (config->alloc == RING_BUFFER_ALLOC_GLOBAL)
- return shmp(chan->backend.buf);
- else
- return &shmp(chan->backend.buf)[cpu];
+ struct shm_ref *ref;
+
+ if (config->alloc == RING_BUFFER_ALLOC_GLOBAL) {
+ cpu = 0;
+ } else {
+ if (cpu >= num_possible_cpus())
+ return NULL;
+ }
+ ref = &chan->backend.buf[cpu].shmp._ref;
+ *shm_fd = shm_get_shm_fd(handle, ref);
+ *wait_fd = shm_get_wait_fd(handle, ref);
+ *wakeup_fd = shm_get_wakeup_fd(handle, ref);
+ if (shm_get_shm_size(handle, ref, memory_map_size))
+ return NULL;
+ return shmp(handle, chan->backend.buf[cpu].shmp);
+}
+
+int ring_buffer_channel_close_wait_fd(const struct lttng_ust_lib_ring_buffer_config *config,
+ struct channel *chan,
+ struct lttng_ust_shm_handle *handle)
+{
+ struct shm_ref *ref;
+
+ ref = &handle->chan._ref;
+ return shm_close_wait_fd(handle, ref);
+}
+
+int ring_buffer_channel_close_wakeup_fd(const struct lttng_ust_lib_ring_buffer_config *config,
+ struct channel *chan,
+ struct lttng_ust_shm_handle *handle)
+{
+ struct shm_ref *ref;
+
+ ref = &handle->chan._ref;
+ return shm_close_wakeup_fd(handle, ref);
}
-int lib_ring_buffer_open_read(struct lib_ring_buffer *buf)
+int ring_buffer_stream_close_wait_fd(const struct lttng_ust_lib_ring_buffer_config *config,
+ struct channel *chan,
+ struct lttng_ust_shm_handle *handle,
+ int cpu)
{
- struct channel *chan = shmp(buf->backend.chan);
+ struct shm_ref *ref;
+ if (config->alloc == RING_BUFFER_ALLOC_GLOBAL) {
+ cpu = 0;
+ } else {
+ if (cpu >= num_possible_cpus())
+ return -EINVAL;
+ }
+ ref = &chan->backend.buf[cpu].shmp._ref;
+ return shm_close_wait_fd(handle, ref);
+}
+
+int ring_buffer_stream_close_wakeup_fd(const struct lttng_ust_lib_ring_buffer_config *config,
+ struct channel *chan,
+ struct lttng_ust_shm_handle *handle,
+ int cpu)
+{
+ struct shm_ref *ref;
+ int ret;
+
+ if (config->alloc == RING_BUFFER_ALLOC_GLOBAL) {
+ cpu = 0;
+ } else {
+ if (cpu >= num_possible_cpus())
+ return -EINVAL;
+ }
+ ref = &chan->backend.buf[cpu].shmp._ref;
+ pthread_mutex_lock(&wakeup_fd_mutex);
+ ret = shm_close_wakeup_fd(handle, ref);
+ pthread_mutex_unlock(&wakeup_fd_mutex);
+ return ret;
+}
+
+int lib_ring_buffer_open_read(struct lttng_ust_lib_ring_buffer *buf,
+ struct lttng_ust_shm_handle *handle)
+{
if (uatomic_cmpxchg(&buf->active_readers, 0, 1) != 0)
return -EBUSY;
cmm_smp_mb();
return 0;
}
-void lib_ring_buffer_release_read(struct lib_ring_buffer *buf)
+void lib_ring_buffer_release_read(struct lttng_ust_lib_ring_buffer *buf,
+ struct lttng_ust_shm_handle *handle)
{
- struct channel *chan = shmp(buf->backend.chan);
+ 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);
* data to read at consumed position, or 0 if the get operation succeeds.
*/
-int lib_ring_buffer_snapshot(struct lib_ring_buffer *buf,
- unsigned long *consumed, unsigned long *produced)
+int lib_ring_buffer_snapshot(struct lttng_ust_lib_ring_buffer *buf,
+ unsigned long *consumed, unsigned long *produced,
+ struct lttng_ust_shm_handle *handle)
{
- struct channel *chan = shmp(buf->backend.chan);
- const struct 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.
}
/**
- * lib_ring_buffer_put_snapshot - move consumed counter forward
+ * 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
* @consumed_new: new consumed count value
*/
-void lib_ring_buffer_move_consumer(struct lib_ring_buffer *buf,
- unsigned long consumed_new)
+void lib_ring_buffer_move_consumer(struct lttng_ust_lib_ring_buffer *buf,
+ unsigned long consumed_new,
+ struct lttng_ust_shm_handle *handle)
{
- struct lib_ring_buffer_backend *bufb = &buf->backend;
- struct channel *chan = shmp(bufb->chan);
+ struct lttng_ust_lib_ring_buffer_backend *bufb = &buf->backend;
+ struct channel *chan;
unsigned long consumed;
+ chan = shmp(handle, bufb->chan);
+ if (!chan)
+ return;
CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1);
/*
* Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
* data to read at consumed position, or 0 if the get operation succeeds.
*/
-int lib_ring_buffer_get_subbuf(struct lib_ring_buffer *buf,
- unsigned long consumed)
+int lib_ring_buffer_get_subbuf(struct lttng_ust_lib_ring_buffer *buf,
+ unsigned long consumed,
+ struct lttng_ust_shm_handle *handle)
{
- struct channel *chan = shmp(buf->backend.chan);
- const struct 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(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));
- if (ret)
- goto retry;
+ consumed_idx, buf_trunc_val(consumed, chan),
+ handle);
+ 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;
* lib_ring_buffer_put_subbuf - release exclusive subbuffer access
* @buf: ring buffer
*/
-void lib_ring_buffer_put_subbuf(struct lib_ring_buffer *buf)
+void lib_ring_buffer_put_subbuf(struct lttng_ust_lib_ring_buffer *buf,
+ struct lttng_ust_shm_handle *handle)
{
- struct lib_ring_buffer_backend *bufb = &buf->backend;
- struct channel *chan = shmp(bufb->chan);
- const struct lib_ring_buffer_config *config = chan->backend.config;
- unsigned long read_sb_bindex, consumed_idx, consumed;
+ struct lttng_ust_lib_ring_buffer_backend *bufb = &buf->backend;
+ 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(bufb->array)[read_sb_bindex]->records_unread),
- &bufb->records_read);
- v_set(config, &shmp(bufb->array)[read_sb_bindex]->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);
*/
consumed_idx = subbuf_index(consumed, chan);
update_read_sb_index(config, &buf->backend, &chan->backend,
- consumed_idx, buf_trunc_val(consumed, chan));
+ consumed_idx, buf_trunc_val(consumed, chan),
+ handle);
/*
* update_read_sb_index return value ignored. Don't exchange sub-buffer
* if the writer concurrently updated it.
* position and the writer position. (inclusive)
*/
static
-void lib_ring_buffer_print_subbuffer_errors(struct lib_ring_buffer *buf,
+void lib_ring_buffer_print_subbuffer_errors(struct lttng_ust_lib_ring_buffer *buf,
struct channel *chan,
unsigned long cons_offset,
- int cpu)
+ int cpu,
+ struct lttng_ust_shm_handle *handle)
{
- const struct lib_ring_buffer_config *config = chan->backend.config;
+ 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(buf->commit_hot)[cons_idx].cc);
- commit_count_sb = v_read(config, &shmp(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)
- ERRMSG("ring buffer %s, cpu %d: "
+ DBG("ring buffer %s, cpu %d: "
"commit count in subbuffer %lu,\n"
"expecting multiples of %lu bytes\n"
" [ %lu bytes committed, %lu bytes reader-visible ]\n",
chan->backend.subbuf_size,
commit_count, commit_count_sb);
- ERRMSG("ring buffer: %s, cpu %d: %lu bytes committed\n",
+ DBG("ring buffer: %s, cpu %d: %lu bytes committed\n",
chan->backend.name, cpu, commit_count);
}
static
-void lib_ring_buffer_print_buffer_errors(struct lib_ring_buffer *buf,
+void lib_ring_buffer_print_buffer_errors(struct lttng_ust_lib_ring_buffer *buf,
struct channel *chan,
- void *priv, int cpu)
+ void *priv, int cpu,
+ struct lttng_ust_shm_handle *handle)
{
- const struct lib_ring_buffer_config *config = chan->backend.config;
+ const struct lttng_ust_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 = uatomic_read(&buf->consumed);
if (write_offset != cons_offset)
- ERRMSG("ring buffer %s, cpu %d: "
+ DBG("ring buffer %s, cpu %d: "
"non-consumed data\n"
" [ %lu bytes written, %lu bytes read ]\n",
chan->backend.name, cpu, write_offset, cons_offset);
- cons_offset) > 0;
cons_offset = subbuf_align(cons_offset, chan))
lib_ring_buffer_print_subbuffer_errors(buf, chan, cons_offset,
- cpu);
+ cpu, handle);
}
static
void lib_ring_buffer_print_errors(struct channel *chan,
- struct lib_ring_buffer *buf, int cpu)
+ struct lttng_ust_lib_ring_buffer *buf, int cpu,
+ struct lttng_ust_shm_handle *handle)
{
- const struct lib_ring_buffer_config *config = chan->backend.config;
- void *priv = chan->backend.priv;
-
- ERRMSG("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))
- ERRMSG("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);
+ const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config;
+ void *priv = channel_get_private(chan);
+
+ if (!strcmp(chan->backend.name, "relay-metadata-mmap")) {
+ DBG("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 {
+ DBG("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))
+ DBG("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, handle);
}
/*
* 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,
+void lib_ring_buffer_switch_old_start(struct lttng_ust_lib_ring_buffer *buf,
struct channel *chan,
struct switch_offsets *offsets,
- u64 tsc)
+ uint64_t tsc,
+ struct lttng_ust_shm_handle *handle)
{
- const struct lib_ring_buffer_config *config = chan->backend.config;
+ 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);
+ config->cb.buffer_begin(buf, tsc, oldidx, handle);
/*
* Order all writes to buffer before the commit count update that will
* 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(buf->commit_hot)[oldidx].cc);
- commit_count = v_read(config, &shmp(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);
- lib_ring_buffer_write_commit_counter(config, buf, chan, oldidx,
- offsets->old, commit_count,
- config->cb.subbuffer_header_size());
+ 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);
}
/*
* subbuffer.
*/
static
-void lib_ring_buffer_switch_old_end(struct lib_ring_buffer *buf,
+void lib_ring_buffer_switch_old_end(struct lttng_ust_lib_ring_buffer *buf,
struct channel *chan,
struct switch_offsets *offsets,
- u64 tsc)
+ uint64_t tsc,
+ struct lttng_ust_shm_handle *handle)
{
- const struct lib_ring_buffer_config *config = chan->backend.config;
+ 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;
- subbuffer_set_data_size(config, &buf->backend, oldidx, data_size);
+ subbuffer_set_data_size(config, &buf->backend, oldidx, 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(buf->commit_hot)[oldidx].cc);
- commit_count = v_read(config, &shmp(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);
- lib_ring_buffer_write_commit_counter(config, buf, chan, oldidx,
- offsets->old, commit_count,
- padding_size);
+ commit_count, oldidx, handle, tsc);
+ lib_ring_buffer_write_commit_counter(config, buf, chan,
+ offsets->old + padding_size, commit_count, handle,
+ cc_hot);
}
/*
* that this code is executed before the deliver of this sub-buffer.
*/
static
-void lib_ring_buffer_switch_new_start(struct lib_ring_buffer *buf,
+void lib_ring_buffer_switch_new_start(struct lttng_ust_lib_ring_buffer *buf,
struct channel *chan,
struct switch_offsets *offsets,
- u64 tsc)
+ uint64_t tsc,
+ struct lttng_ust_shm_handle *handle)
{
- const struct lib_ring_buffer_config *config = chan->backend.config;
+ 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);
+ config->cb.buffer_begin(buf, tsc, beginidx, 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, config->cb.subbuffer_header_size(),
- &shmp(buf->commit_hot)[beginidx].cc);
- commit_count = v_read(config, &shmp(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);
- lib_ring_buffer_write_commit_counter(config, buf, chan, beginidx,
- offsets->begin, commit_count,
- config->cb.subbuffer_header_size());
+ 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 lib_ring_buffer *buf,
- struct channel *chan,
- struct switch_offsets *offsets,
- u64 tsc)
+void lib_ring_buffer_switch_new_end(struct lttng_ust_lib_ring_buffer *buf,
+ struct channel *chan,
+ struct switch_offsets *offsets,
+ uint64_t tsc,
+ struct lttng_ust_shm_handle *handle)
{
- 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 lttng_ust_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.
- */
- cmm_smp_wmb();
- v_add(config, padding_size, &shmp(buf->commit_hot)[endidx].cc);
- commit_count = v_read(config, &shmp(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);
+ subbuffer_set_data_size(config, &buf->backend, endidx, data_size,
+ handle);
}
/*
*/
static
int lib_ring_buffer_try_switch_slow(enum switch_mode mode,
- struct lib_ring_buffer *buf,
+ struct lttng_ust_lib_ring_buffer *buf,
struct channel *chan,
struct switch_offsets *offsets,
- u64 *tsc)
+ uint64_t *tsc,
+ struct lttng_ust_shm_handle *handle)
{
- const struct lib_ring_buffer_config *config = chan->backend.config;
- unsigned long off;
+ const struct lttng_ust_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;
* 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 (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;
* operations, this function must be called from the CPU which owns the buffer
* for a ACTIVE flush.
*/
-void lib_ring_buffer_switch_slow(struct lib_ring_buffer *buf, enum switch_mode mode)
+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(buf->backend.chan);
- const struct 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;
- u64 tsc;
+ 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_reserve_push_reader(buf, chan, offsets.old);
oldidx = subbuf_index(offsets.old, chan);
- lib_ring_buffer_clear_noref(config, &buf->backend, oldidx);
+ lib_ring_buffer_clear_noref(config, &buf->backend, oldidx, handle);
/*
* May need to populate header start on SWITCH_FLUSH.
*/
if (offsets.switch_old_start) {
- lib_ring_buffer_switch_old_start(buf, chan, &offsets, tsc);
+ lib_ring_buffer_switch_old_start(buf, chan, &offsets, tsc, handle);
offsets.old += config->cb.subbuffer_header_size();
}
/*
* Switch old subbuffer.
*/
- lib_ring_buffer_switch_old_end(buf, chan, &offsets, tsc);
+ 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. */
}
/*
* -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,
+int lib_ring_buffer_try_reserve_slow(struct lttng_ust_lib_ring_buffer *buf,
struct channel *chan,
struct switch_offsets *offsets,
- struct lib_ring_buffer_ctx *ctx)
+ struct lttng_ust_lib_ring_buffer_ctx *ctx,
+ void *client_ctx)
{
- const struct lib_ring_buffer_config *config = chan->backend.config;
- unsigned long reserve_commit_diff;
+ 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, 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;
if (last_tsc_overflow(config, buf, ctx->tsc))
ctx->rflags |= RING_BUFFER_RFLAG_FULL_TSC;
- if (unlikely(subbuf_offset(offsets->begin, ctx->chan) == 0)) {
+ if (caa_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,
&offsets->pre_header_padding,
- ctx);
+ ctx, client_ctx);
offsets->size +=
lib_ring_buffer_align(offsets->begin + offsets->size,
ctx->largest_align)
+ ctx->data_size;
- if (unlikely(subbuf_offset(offsets->begin, chan) +
+ if (caa_unlikely(subbuf_offset(offsets->begin, chan) +
offsets->size > chan->backend.subbuf_size)) {
offsets->switch_old_end = 1; /* For offsets->old */
offsets->switch_new_start = 1; /* For offsets->begin */
}
}
- if (unlikely(offsets->switch_new_start)) {
- unsigned long sb_index;
+ if (caa_unlikely(offsets->switch_new_start)) {
+ unsigned long sb_index, commit_count;
+ struct commit_counters_cold *cc_cold;
/*
* We are typically not filling the previous buffer completely.
*/
- if (likely(offsets->switch_old_end))
+ if (caa_likely(offsets->switch_old_end))
offsets->begin = subbuf_align(offsets->begin, chan);
offsets->begin = offsets->begin
+ 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(buf->commit_cold)[sb_index].cc_sb)
- & chan->commit_count_mask);
- if (likely(reserve_commit_diff == 0)) {
+ - (commit_count & chan->commit_count_mask);
+ if (caa_likely(reserve_commit_diff == 0)) {
/* Next subbuffer not being written to. */
- if (unlikely(config->mode != RING_BUFFER_OVERWRITE &&
+ 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)) {
+ 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.
*/
+ nr_lost = v_read(config, &buf->records_lost_full);
v_inc(config, &buf->records_lost_full);
+ if ((nr_lost & (DBG_PRINT_NR_LOST - 1)) == 0) {
+ DBG("%lu or more records lost in (%s:%d) (buffer full)\n",
+ nr_lost + 1, chan->backend.name,
+ buf->backend.cpu);
+ }
return -ENOBUFS;
} else {
/*
*/
}
} else {
+ unsigned long nr_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.
*/
+ nr_lost = v_read(config, &buf->records_lost_wrap);
v_inc(config, &buf->records_lost_wrap);
+ if ((nr_lost & (DBG_PRINT_NR_LOST - 1)) == 0) {
+ DBG("%lu or more records lost in (%s:%d) (wrap-around)\n",
+ nr_lost + 1, chan->backend.name,
+ buf->backend.cpu);
+ }
return -EIO;
}
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)
+ ctx->data_size;
- if (unlikely(subbuf_offset(offsets->begin, chan)
+ if (caa_unlikely(subbuf_offset(offsets->begin, chan)
+ offsets->size > chan->backend.subbuf_size)) {
+ unsigned long nr_lost;
+
/*
* Record too big for subbuffers, report error, don't
* complete the sub-buffer switch.
*/
+ nr_lost = v_read(config, &buf->records_lost_big);
v_inc(config, &buf->records_lost_big);
+ if ((nr_lost & (DBG_PRINT_NR_LOST - 1)) == 0) {
+ DBG("%lu or more records lost in (%s:%d) record size "
+ " of %zu bytes is too large for buffer\n",
+ nr_lost + 1, chan->backend.name,
+ buf->backend.cpu, offsets->size);
+ }
return -ENOSPC;
} else {
/*
}
offsets->end = offsets->begin + offsets->size;
- if (unlikely(subbuf_offset(offsets->end, chan) == 0)) {
+ if (caa_unlikely(subbuf_offset(offsets->end, chan) == 0)) {
/*
* The offset_end will fall at the very beginning of the next
* subbuffer.
* -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)
+int lib_ring_buffer_reserve_slow(struct lttng_ust_lib_ring_buffer_ctx *ctx,
+ void *client_ctx)
{
struct channel *chan = ctx->chan;
- const struct lib_ring_buffer_config *config = chan->backend.config;
- struct lib_ring_buffer *buf;
+ struct lttng_ust_shm_handle *handle = ctx->handle;
+ const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config;
+ struct lttng_ust_lib_ring_buffer *buf;
struct switch_offsets offsets;
int ret;
if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
- buf = &shmp(chan->backend.buf)[ctx->cpu];
+ buf = shmp(handle, chan->backend.buf[ctx->cpu].shmp);
else
- buf = shmp(chan->backend.buf);
+ 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);
- if (unlikely(ret))
+ ctx, client_ctx);
+ if (caa_unlikely(ret))
return ret;
- } while (unlikely(v_cmpxchg(config, &buf->offset, offsets.old,
+ } while (caa_unlikely(v_cmpxchg(config, &buf->offset, offsets.old,
offsets.end)
!= offsets.old));
* Clear noref flag for this subbuffer.
*/
lib_ring_buffer_clear_noref(config, &buf->backend,
- subbuf_index(offsets.end - 1, chan));
+ subbuf_index(offsets.end - 1, chan),
+ handle);
/*
* Switch old subbuffer if needed.
*/
- if (unlikely(offsets.switch_old_end)) {
+ if (caa_unlikely(offsets.switch_old_end)) {
lib_ring_buffer_clear_noref(config, &buf->backend,
- subbuf_index(offsets.old - 1, chan));
- lib_ring_buffer_switch_old_end(buf, chan, &offsets, ctx->tsc);
+ subbuf_index(offsets.old - 1, chan),
+ handle);
+ lib_ring_buffer_switch_old_end(buf, chan, &offsets, ctx->tsc, handle);
}
/*
* Populate new subbuffer.
*/
- if (unlikely(offsets.switch_new_start))
- lib_ring_buffer_switch_new_start(buf, chan, &offsets, ctx->tsc);
+ if (caa_unlikely(offsets.switch_new_start))
+ lib_ring_buffer_switch_new_start(buf, chan, &offsets, ctx->tsc, handle);
- if (unlikely(offsets.switch_new_end))
- lib_ring_buffer_switch_new_end(buf, chan, &offsets, ctx->tsc);
+ if (caa_unlikely(offsets.switch_new_end))
+ lib_ring_buffer_switch_new_end(buf, chan, &offsets, ctx->tsc, handle);
ctx->slot_size = offsets.size;
ctx->pre_offset = offsets.begin;
ctx->buf_offset = offsets.begin + offsets.pre_header_padding;
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.
+ */
+void lttng_fixup_ringbuffer_tls(void)
+{
+ asm volatile ("" : : "m" (URCU_TLS(lib_ring_buffer_nesting)));
+}
+
+void lib_ringbuffer_signal_init(void)
+{
+ sigset_t mask;
+ int ret;
+
+ /*
+ * Block signal for entire process, so only our thread processes
+ * it.
+ */
+ rb_setmask(&mask);
+ ret = pthread_sigmask(SIG_BLOCK, &mask, NULL);
+ if (ret) {
+ errno = ret;
+ PERROR("pthread_sigmask");
+ }
+}
projects / lttng-ust.git / blobdiff