Performance: split check deliver fast/slow paths
authorMathieu Desnoyers <mathieu.desnoyers@efficios.com>
Sat, 16 Jul 2016 03:22:08 +0000 (23:22 -0400)
committerMathieu Desnoyers <mathieu.desnoyers@efficios.com>
Sun, 21 Aug 2016 20:18:56 +0000 (16:18 -0400)
On ARMv7l (Cubietruck), the compiler generates a function call for each
lib_ring_buffer_check_deliver, even though it typically only do an
unlikely check. Split it into an inline fast path, and a function call
for the slow path. This brings a performance gain of about 500ns/event
on the Cubietruck.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
libringbuffer/frontend_internal.h
libringbuffer/ring_buffer_frontend.c

index 00b9508d4df54768767fbe0aaff401cb0f965b0d..43a2d0a2639b29d319ddb7b9d1d417855f9c96bc 100644 (file)
@@ -165,6 +165,15 @@ void lib_ring_buffer_switch_slow(struct lttng_ust_lib_ring_buffer *buf,
                                 enum switch_mode mode,
                                 struct lttng_ust_shm_handle *handle);
 
+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);
+
 /* Buffer write helpers */
 
 static inline
@@ -195,60 +204,6 @@ void lib_ring_buffer_reserve_push_reader(struct lttng_ust_lib_ring_buffer *buf,
                                              consumed_new) != consumed_old));
 }
 
-static inline
-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)
-{
-       if (config->oops == RING_BUFFER_OOPS_CONSISTENCY)
-               v_set(config, &shmp_index(handle, buf->commit_hot, idx)->seq, commit_count);
-}
-
-static inline
-int lib_ring_buffer_poll_deliver(const struct lttng_ust_lib_ring_buffer_config *config,
-                                struct lttng_ust_lib_ring_buffer *buf,
-                                struct channel *chan,
-                                struct lttng_ust_shm_handle *handle)
-{
-       unsigned long consumed_old, consumed_idx, commit_count, write_offset;
-
-       consumed_old = uatomic_read(&buf->consumed);
-       consumed_idx = subbuf_index(consumed_old, chan);
-       commit_count = v_read(config, &shmp_index(handle, buf->commit_cold, consumed_idx)->cc_sb);
-       /*
-        * No memory barrier here, since we are only interested
-        * in a statistically correct polling result. The next poll will
-        * get the data is we are racing. The mb() that ensures correct
-        * memory order is in get_subbuf.
-        */
-       write_offset = v_read(config, &buf->offset);
-
-       /*
-        * Check that the subbuffer we are trying to consume has been
-        * already fully committed.
-        */
-
-       if (((commit_count - chan->backend.subbuf_size)
-            & chan->commit_count_mask)
-           - (buf_trunc(consumed_old, chan)
-              >> chan->backend.num_subbuf_order)
-           != 0)
-               return 0;
-
-       /*
-        * Check that we are not about to read the same subbuffer in
-        * which the writer head is.
-        */
-       if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_old, chan)
-           == 0)
-               return 0;
-
-       return 1;
-
-}
-
 static inline
 int lib_ring_buffer_pending_data(const struct lttng_ust_lib_ring_buffer_config *config,
                                 struct lttng_ust_lib_ring_buffer *buf,
@@ -300,71 +255,6 @@ int lib_ring_buffer_reserve_committed(const struct lttng_ust_lib_ring_buffer_con
                     - (commit_count & chan->commit_count_mask) == 0);
 }
 
-static inline
-void lib_ring_buffer_wakeup(struct lttng_ust_lib_ring_buffer *buf,
-               struct lttng_ust_shm_handle *handle)
-{
-       int wakeup_fd = shm_get_wakeup_fd(handle, &buf->self._ref);
-       sigset_t sigpipe_set, pending_set, old_set;
-       int ret, sigpipe_was_pending = 0;
-
-       if (wakeup_fd < 0)
-               return;
-
-       /*
-        * Wake-up the other end by writing a null byte in the pipe
-        * (non-blocking).  Important note: Because writing into the
-        * pipe is non-blocking (and therefore we allow dropping wakeup
-        * data, as long as there is wakeup data present in the pipe
-        * buffer to wake up the consumer), the consumer should perform
-        * the following sequence for waiting:
-        * 1) empty the pipe (reads).
-        * 2) check if there is data in the buffer.
-        * 3) wait on the pipe (poll).
-        *
-        * Discard the SIGPIPE from write(), not disturbing any SIGPIPE
-        * that might be already pending. If a bogus SIGPIPE is sent to
-        * the entire process concurrently by a malicious user, it may
-        * be simply discarded.
-        */
-       ret = sigemptyset(&pending_set);
-       assert(!ret);
-       /*
-        * sigpending returns the mask of signals that are _both_
-        * blocked for the thread _and_ pending for either the thread or
-        * the entire process.
-        */
-       ret = sigpending(&pending_set);
-       assert(!ret);
-       sigpipe_was_pending = sigismember(&pending_set, SIGPIPE);
-       /*
-        * If sigpipe was pending, it means it was already blocked, so
-        * no need to block it.
-        */
-       if (!sigpipe_was_pending) {
-               ret = sigemptyset(&sigpipe_set);
-               assert(!ret);
-               ret = sigaddset(&sigpipe_set, SIGPIPE);
-               assert(!ret);
-               ret = pthread_sigmask(SIG_BLOCK, &sigpipe_set, &old_set);
-               assert(!ret);
-       }
-       do {
-               ret = write(wakeup_fd, "", 1);
-       } while (ret == -1L && errno == EINTR);
-       if (ret == -1L && errno == EPIPE && !sigpipe_was_pending) {
-               struct timespec timeout = { 0, 0 };
-               do {
-                       ret = sigtimedwait(&sigpipe_set, NULL,
-                               &timeout);
-               } while (ret == -1L && errno == EINTR);
-       }
-       if (!sigpipe_was_pending) {
-               ret = pthread_sigmask(SIG_SETMASK, &old_set, NULL);
-               assert(!ret);
-       }
-}
-
 /*
  * Receive end of subbuffer TSC as parameter. It has been read in the
  * space reservation loop of either reserve or switch, which ensures it
@@ -387,107 +277,9 @@ void lib_ring_buffer_check_deliver(const struct lttng_ust_lib_ring_buffer_config
 
        /* Check if all commits have been done */
        if (caa_unlikely((buf_trunc(offset, chan) >> chan->backend.num_subbuf_order)
-                    - (old_commit_count & chan->commit_count_mask) == 0)) {
-               /*
-                * 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();
-               if (caa_likely(v_cmpxchg(config, &shmp_index(handle, buf->commit_cold, idx)->cc_sb,
-                                        old_commit_count, old_commit_count + 1)
-                          == old_commit_count)) {
-                       /*
-                        * Start of exclusive subbuffer access. We are
-                        * guaranteed to be the last writer in this subbuffer
-                        * and any other writer trying to access this subbuffer
-                        * in this state is required to drop records.
-                        */
-                       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);
-                       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, &shmp_index(handle, buf->commit_cold, idx)->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);
-                       }
-               }
-       }
+                    - (old_commit_count & chan->commit_count_mask) == 0))
+               lib_ring_buffer_check_deliver_slow(config, buf, chan, offset,
+                       commit_count, idx, handle, tsc);
 }
 
 /*
index 908539a8663973fde0a8e51384907b4690b0e30d..bda0e1f497fecc786883726b6329ba51498240c9 100644 (file)
@@ -421,6 +421,113 @@ void lib_ring_buffer_channel_switch_timer(int sig, siginfo_t *si, void *uc)
        return;
 }
 
+static
+int lib_ring_buffer_poll_deliver(const struct lttng_ust_lib_ring_buffer_config *config,
+                                struct lttng_ust_lib_ring_buffer *buf,
+                                struct channel *chan,
+                                struct lttng_ust_shm_handle *handle)
+{
+       unsigned long consumed_old, consumed_idx, commit_count, write_offset;
+
+       consumed_old = uatomic_read(&buf->consumed);
+       consumed_idx = subbuf_index(consumed_old, chan);
+       commit_count = v_read(config, &shmp_index(handle, buf->commit_cold, consumed_idx)->cc_sb);
+       /*
+        * No memory barrier here, since we are only interested
+        * in a statistically correct polling result. The next poll will
+        * get the data is we are racing. The mb() that ensures correct
+        * memory order is in get_subbuf.
+        */
+       write_offset = v_read(config, &buf->offset);
+
+       /*
+        * Check that the subbuffer we are trying to consume has been
+        * already fully committed.
+        */
+
+       if (((commit_count - chan->backend.subbuf_size)
+            & chan->commit_count_mask)
+           - (buf_trunc(consumed_old, chan)
+              >> chan->backend.num_subbuf_order)
+           != 0)
+               return 0;
+
+       /*
+        * Check that we are not about to read the same subbuffer in
+        * which the writer head is.
+        */
+       if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_old, chan)
+           == 0)
+               return 0;
+
+       return 1;
+}
+
+static
+void lib_ring_buffer_wakeup(struct lttng_ust_lib_ring_buffer *buf,
+               struct lttng_ust_shm_handle *handle)
+{
+       int wakeup_fd = shm_get_wakeup_fd(handle, &buf->self._ref);
+       sigset_t sigpipe_set, pending_set, old_set;
+       int ret, sigpipe_was_pending = 0;
+
+       if (wakeup_fd < 0)
+               return;
+
+       /*
+        * Wake-up the other end by writing a null byte in the pipe
+        * (non-blocking).  Important note: Because writing into the
+        * pipe is non-blocking (and therefore we allow dropping wakeup
+        * data, as long as there is wakeup data present in the pipe
+        * buffer to wake up the consumer), the consumer should perform
+        * the following sequence for waiting:
+        * 1) empty the pipe (reads).
+        * 2) check if there is data in the buffer.
+        * 3) wait on the pipe (poll).
+        *
+        * Discard the SIGPIPE from write(), not disturbing any SIGPIPE
+        * that might be already pending. If a bogus SIGPIPE is sent to
+        * the entire process concurrently by a malicious user, it may
+        * be simply discarded.
+        */
+       ret = sigemptyset(&pending_set);
+       assert(!ret);
+       /*
+        * sigpending returns the mask of signals that are _both_
+        * blocked for the thread _and_ pending for either the thread or
+        * the entire process.
+        */
+       ret = sigpending(&pending_set);
+       assert(!ret);
+       sigpipe_was_pending = sigismember(&pending_set, SIGPIPE);
+       /*
+        * If sigpipe was pending, it means it was already blocked, so
+        * no need to block it.
+        */
+       if (!sigpipe_was_pending) {
+               ret = sigemptyset(&sigpipe_set);
+               assert(!ret);
+               ret = sigaddset(&sigpipe_set, SIGPIPE);
+               assert(!ret);
+               ret = pthread_sigmask(SIG_BLOCK, &sigpipe_set, &old_set);
+               assert(!ret);
+       }
+       do {
+               ret = write(wakeup_fd, "", 1);
+       } while (ret == -1L && errno == EINTR);
+       if (ret == -1L && errno == EPIPE && !sigpipe_was_pending) {
+               struct timespec timeout = { 0, 0 };
+               do {
+                       ret = sigtimedwait(&sigpipe_set, NULL,
+                               &timeout);
+               } while (ret == -1L && errno == EINTR);
+       }
+       if (!sigpipe_was_pending) {
+               ret = pthread_sigmask(SIG_SETMASK, &old_set, NULL);
+               assert(!ret);
+       }
+}
+
 static
 void lib_ring_buffer_channel_do_read(struct channel *chan)
 {
@@ -2049,6 +2156,130 @@ int lib_ring_buffer_reserve_slow(struct lttng_ust_lib_ring_buffer_ctx *ctx)
        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)
+{
+       if (config->oops == RING_BUFFER_OOPS_CONSISTENCY)
+               v_set(config, &shmp_index(handle, buf->commit_hot, idx)->seq, commit_count);
+}
+
+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;
+
+       /*
+        * 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();
+       if (caa_likely(v_cmpxchg(config, &shmp_index(handle, buf->commit_cold, idx)->cc_sb,
+                                old_commit_count, old_commit_count + 1)
+                  == old_commit_count)) {
+               /*
+                * Start of exclusive subbuffer access. We are
+                * guaranteed to be the last writer in this subbuffer
+                * and any other writer trying to access this subbuffer
+                * in this state is required to drop records.
+                */
+               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);
+               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, &shmp_index(handle, buf->commit_cold, idx)->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.
  */
This page took 0.030928 seconds and 4 git commands to generate.