From: Mathieu Desnoyers Date: Sat, 16 Jul 2016 14:07:55 +0000 (-0400) Subject: Performance: split check deliver fast/slow paths X-Git-Tag: v2.9.0-rc1~31 X-Git-Url: http://git.lttng.org/?p=lttng-modules.git;a=commitdiff_plain;h=aece661f429d4ca54cf1e47bacc556679c28f342 Performance: split check deliver fast/slow paths Performance improvement changelog from lttng-ust, ported back to lttng-modules: 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 --- diff --git a/lib/ringbuffer/frontend_internal.h b/lib/ringbuffer/frontend_internal.h index 76bf20e7..32785853 100644 --- a/lib/ringbuffer/frontend_internal.h +++ b/lib/ringbuffer/frontend_internal.h @@ -156,6 +156,15 @@ extern void lib_ring_buffer_switch_slow(struct lib_ring_buffer *buf, enum switch_mode mode); +extern +void lib_ring_buffer_check_deliver_slow(const struct lib_ring_buffer_config *config, + struct lib_ring_buffer *buf, + struct channel *chan, + unsigned long offset, + unsigned long commit_count, + unsigned long idx, + u64 tsc); + extern void lib_ring_buffer_switch_remote(struct lib_ring_buffer *buf); @@ -192,58 +201,6 @@ void lib_ring_buffer_reserve_push_reader(struct lib_ring_buffer *buf, consumed_new) != consumed_old)); } -static inline -void lib_ring_buffer_vmcore_check_deliver(const struct lib_ring_buffer_config *config, - struct lib_ring_buffer *buf, - unsigned long commit_count, - unsigned long idx) -{ - if (config->oops == RING_BUFFER_OOPS_CONSISTENCY) - v_set(config, &buf->commit_hot[idx].seq, commit_count); -} - -static inline -int lib_ring_buffer_poll_deliver(const struct lib_ring_buffer_config *config, - struct lib_ring_buffer *buf, - struct channel *chan) -{ - unsigned long consumed_old, consumed_idx, commit_count, write_offset; - - consumed_old = atomic_long_read(&buf->consumed); - consumed_idx = subbuf_index(consumed_old, chan); - commit_count = v_read(config, &buf->commit_cold[consumed_idx].cc_sb); - /* - * No memory barrier here, since we are only interested - * in a statistically correct polling result. The next poll will - * get the data is we are racing. The mb() that ensures correct - * memory order is in get_subbuf. - */ - write_offset = v_read(config, &buf->offset); - - /* - * Check that the subbuffer we are trying to consume has been - * already fully committed. - */ - - if (((commit_count - chan->backend.subbuf_size) - & chan->commit_count_mask) - - (buf_trunc(consumed_old, chan) - >> chan->backend.num_subbuf_order) - != 0) - return 0; - - /* - * Check that we are not about to read the same subbuffer in - * which the writer head is. - */ - if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_old, chan) - == 0) - return 0; - - return 1; - -} - static inline int lib_ring_buffer_pending_data(const struct lib_ring_buffer_config *config, struct lib_ring_buffer *buf, @@ -314,106 +271,9 @@ void lib_ring_buffer_check_deliver(const struct lib_ring_buffer_config *config, /* Check if all commits have been done */ if (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. - */ - smp_wmb(); - if (likely(v_cmpxchg(config, &buf->commit_cold[idx].cc_sb, - old_commit_count, old_commit_count + 1) - == old_commit_count)) { - /* - * Start of exclusive subbuffer access. We are - * guaranteed to be the last writer in this subbuffer - * and any other writer trying to access this subbuffer - * in this state is required to drop records. - */ - v_add(config, - subbuffer_get_records_count(config, - &buf->backend, idx), - &buf->records_count); - v_add(config, - subbuffer_count_records_overrun(config, - &buf->backend, - idx), - &buf->records_overrun); - config->cb.buffer_end(buf, tsc, idx, - lib_ring_buffer_get_data_size(config, - buf, - idx)); - - /* - * Increment the packet counter while we have exclusive - * access. - */ - subbuffer_inc_packet_count(config, &buf->backend, idx); - - /* - * Set noref flag and offset for this subbuffer id. - * Contains a memory barrier that ensures counter stores - * are ordered before set noref and offset. - */ - lib_ring_buffer_set_noref_offset(config, &buf->backend, idx, - buf_trunc_val(offset, chan)); - - /* - * Order set_noref and record counter updates before the - * end of subbuffer exclusive access. Orders with - * respect to writers coming into the subbuffer after - * wrap around, and also order wrt concurrent readers. - */ - smp_mb(); - /* End of exclusive subbuffer access */ - v_set(config, &buf->commit_cold[idx].cc_sb, - commit_count); - /* - * Order later updates to reserve count after - * the commit_cold cc_sb update. - */ - smp_wmb(); - lib_ring_buffer_vmcore_check_deliver(config, buf, - commit_count, idx); - - /* - * RING_BUFFER_WAKEUP_BY_WRITER wakeup is not lock-free. - */ - if (config->wakeup == RING_BUFFER_WAKEUP_BY_WRITER - && atomic_long_read(&buf->active_readers) - && lib_ring_buffer_poll_deliver(config, buf, chan)) { - wake_up_interruptible(&buf->read_wait); - wake_up_interruptible(&chan->read_wait); - } - - } - } + - (old_commit_count & chan->commit_count_mask) == 0)) + lib_ring_buffer_check_deliver_slow(config, buf, chan, offset, + commit_count, idx, tsc); } /* diff --git a/lib/ringbuffer/ring_buffer_frontend.c b/lib/ringbuffer/ring_buffer_frontend.c index 89eefe96..c5123a64 100644 --- a/lib/ringbuffer/ring_buffer_frontend.c +++ b/lib/ringbuffer/ring_buffer_frontend.c @@ -97,6 +97,47 @@ static void _lib_ring_buffer_switch_remote(struct lib_ring_buffer *buf, enum switch_mode mode); +static +int lib_ring_buffer_poll_deliver(const struct lib_ring_buffer_config *config, + struct lib_ring_buffer *buf, + struct channel *chan) +{ + unsigned long consumed_old, consumed_idx, commit_count, write_offset; + + consumed_old = atomic_long_read(&buf->consumed); + consumed_idx = subbuf_index(consumed_old, chan); + commit_count = v_read(config, &buf->commit_cold[consumed_idx].cc_sb); + /* + * No memory barrier here, since we are only interested + * in a statistically correct polling result. The next poll will + * get the data is we are racing. The mb() that ensures correct + * memory order is in get_subbuf. + */ + write_offset = v_read(config, &buf->offset); + + /* + * Check that the subbuffer we are trying to consume has been + * already fully committed. + */ + + if (((commit_count - chan->backend.subbuf_size) + & chan->commit_count_mask) + - (buf_trunc(consumed_old, chan) + >> chan->backend.num_subbuf_order) + != 0) + return 0; + + /* + * Check that we are not about to read the same subbuffer in + * which the writer head is. + */ + if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_old, chan) + == 0) + return 0; + + return 1; +} + /* * Must be called under cpu hotplug protection. */ @@ -1929,6 +1970,128 @@ int lib_ring_buffer_reserve_slow(struct lib_ring_buffer_ctx *ctx) } EXPORT_SYMBOL_GPL(lib_ring_buffer_reserve_slow); +static +void lib_ring_buffer_vmcore_check_deliver(const struct lib_ring_buffer_config *config, + struct lib_ring_buffer *buf, + unsigned long commit_count, + unsigned long idx) +{ + if (config->oops == RING_BUFFER_OOPS_CONSISTENCY) + v_set(config, &buf->commit_hot[idx].seq, commit_count); +} + +void lib_ring_buffer_check_deliver_slow(const struct lib_ring_buffer_config *config, + struct lib_ring_buffer *buf, + struct channel *chan, + unsigned long offset, + unsigned long commit_count, + unsigned long idx, + u64 tsc) +{ + unsigned long old_commit_count = commit_count + - chan->backend.subbuf_size; + + /* + * If we succeeded at updating cc_sb below, we are the subbuffer + * writer delivering the subbuffer. Deals with concurrent + * updates of the "cc" value without adding a add_return atomic + * operation to the fast path. + * + * We are doing the delivery in two steps: + * - First, we cmpxchg() cc_sb to the new value + * old_commit_count + 1. This ensures that we are the only + * subbuffer user successfully filling the subbuffer, but we + * do _not_ set the cc_sb value to "commit_count" yet. + * Therefore, other writers that would wrap around the ring + * buffer and try to start writing to our subbuffer would + * have to drop records, because it would appear as + * non-filled. + * We therefore have exclusive access to the subbuffer control + * structures. This mutual exclusion with other writers is + * crucially important to perform record overruns count in + * flight recorder mode locklessly. + * - When we are ready to release the subbuffer (either for + * reading or for overrun by other writers), we simply set the + * cc_sb value to "commit_count" and perform delivery. + * + * The subbuffer size is least 2 bytes (minimum size: 1 page). + * This guarantees that old_commit_count + 1 != commit_count. + */ + + /* + * Order prior updates to reserve count prior to the + * commit_cold cc_sb update. + */ + smp_wmb(); + if (likely(v_cmpxchg(config, &buf->commit_cold[idx].cc_sb, + old_commit_count, old_commit_count + 1) + == old_commit_count)) { + /* + * Start of exclusive subbuffer access. We are + * guaranteed to be the last writer in this subbuffer + * and any other writer trying to access this subbuffer + * in this state is required to drop records. + */ + v_add(config, + subbuffer_get_records_count(config, + &buf->backend, idx), + &buf->records_count); + v_add(config, + subbuffer_count_records_overrun(config, + &buf->backend, + idx), + &buf->records_overrun); + config->cb.buffer_end(buf, tsc, idx, + lib_ring_buffer_get_data_size(config, + buf, + idx)); + + /* + * Increment the packet counter while we have exclusive + * access. + */ + subbuffer_inc_packet_count(config, &buf->backend, idx); + + /* + * Set noref flag and offset for this subbuffer id. + * Contains a memory barrier that ensures counter stores + * are ordered before set noref and offset. + */ + lib_ring_buffer_set_noref_offset(config, &buf->backend, idx, + buf_trunc_val(offset, chan)); + + /* + * Order set_noref and record counter updates before the + * end of subbuffer exclusive access. Orders with + * respect to writers coming into the subbuffer after + * wrap around, and also order wrt concurrent readers. + */ + smp_mb(); + /* End of exclusive subbuffer access */ + v_set(config, &buf->commit_cold[idx].cc_sb, + commit_count); + /* + * Order later updates to reserve count after + * the commit_cold cc_sb update. + */ + smp_wmb(); + lib_ring_buffer_vmcore_check_deliver(config, buf, + commit_count, idx); + + /* + * RING_BUFFER_WAKEUP_BY_WRITER wakeup is not lock-free. + */ + if (config->wakeup == RING_BUFFER_WAKEUP_BY_WRITER + && atomic_long_read(&buf->active_readers) + && lib_ring_buffer_poll_deliver(config, buf, chan)) { + wake_up_interruptible(&buf->read_wait); + wake_up_interruptible(&chan->read_wait); + } + + } +} +EXPORT_SYMBOL_GPL(lib_ring_buffer_check_deliver_slow); + int __init init_lib_ring_buffer_frontend(void) { int cpu;