#include <urcu/compiler.h>
#include <urcu/ref.h>
#include <urcu/tls-compat.h>
+#include <poll.h>
#include <helper.h>
#include "smp.h"
#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
/*
* Use POSIX SHM: shm_open(3) and shm_unlink(3).
struct channel *chan = shmp(handle, buf->backend.chan);
const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.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;
retry:
finalized = CMM_ACCESS_ONCE(buf->finalized);
/*
* 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;
commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, oldidx)->cc);
/* Check if the written buffer has to be delivered */
lib_ring_buffer_check_deliver(config, buf, chan, offsets->old,
- commit_count, oldidx, handle);
+ commit_count, oldidx, handle, tsc);
lib_ring_buffer_write_commit_counter(config, buf, chan, oldidx,
- offsets->old, commit_count,
- config->cb.subbuffer_header_size(),
- handle);
+ offsets->old + config->cb.subbuffer_header_size(),
+ commit_count, handle);
}
/*
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);
lib_ring_buffer_check_deliver(config, buf, chan, offsets->old - 1,
- commit_count, oldidx, handle);
+ commit_count, oldidx, handle, tsc);
lib_ring_buffer_write_commit_counter(config, buf, chan, oldidx,
- offsets->old, commit_count,
- padding_size, handle);
+ offsets->old + padding_size, commit_count, handle);
}
/*
commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, beginidx)->cc);
/* Check if the written buffer has to be delivered */
lib_ring_buffer_check_deliver(config, buf, chan, offsets->begin,
- commit_count, beginidx, handle);
+ commit_count, beginidx, handle, tsc);
lib_ring_buffer_write_commit_counter(config, buf, chan, beginidx,
- offsets->begin, commit_count,
- config->cb.subbuffer_header_size(),
- handle);
+ offsets->begin + config->cb.subbuffer_header_size(),
+ commit_count, 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;
+
+ /*
+ * We are performing a SWITCH_FLUSH. At this stage, there are no
+ * concurrent writes into the buffer.
+ *
+ * The client does not save any header information. Don't
+ * switch empty subbuffer on finalize, because it is invalid to
+ * deliver a completely empty subbuffer.
+ */
+ if (!config->cb.subbuffer_header_size())
+ return -1;
+
+ /* Test new buffer integrity */
+ sb_index = subbuf_index(offsets->begin, chan);
+ commit_count = v_read(config,
+ &shmp_index(handle, buf->commit_cold,
+ sb_index)->cc_sb);
+ reserve_commit_diff =
+ (buf_trunc(offsets->begin, chan)
+ >> chan->backend.num_subbuf_order)
+ - (commit_count & chan->commit_count_mask);
+ if (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;
*/
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);
{
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;
- 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 (caa_unlikely(offsets->switch_new_start)) {
- unsigned long sb_index;
+ unsigned long sb_index, commit_count;
/*
* We are typically not filling the previous buffer completely.
+ config->cb.subbuffer_header_size();
/* Test new buffer integrity */
sb_index = subbuf_index(offsets->begin, chan);
+ /*
+ * Read buf->offset before buf->commit_cold[sb_index].cc_sb.
+ * lib_ring_buffer_check_deliver() has the matching
+ * memory barriers required around commit_cold cc_sb
+ * updates to ensure reserve and commit counter updates
+ * are not seen reordered when updated by another CPU.
+ */
+ cmm_smp_rmb();
+ commit_count = v_read(config,
+ &shmp_index(handle, buf->commit_cold,
+ sb_index)->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 &&
/*
* 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.
*/
projects / lttng-ust.git / blobdiff