X-Git-Url: http://git.lttng.org/?a=blobdiff_plain;f=lib%2Fringbuffer%2Fring_buffer_frontend.c;h=28525a691ec677ae14e028f3aeb065f3580a2445;hb=0348b08949345b263015bb95bcaae8f7a607fb77;hp=db33d04cba8dc7ba8f23d4b90ee0658cbe1888c4;hpb=b265a505ba25418104f3425e4475f45419844e8a;p=lttng-modules.git diff --git a/lib/ringbuffer/ring_buffer_frontend.c b/lib/ringbuffer/ring_buffer_frontend.c index db33d04c..28525a69 100644 --- a/lib/ringbuffer/ring_buffer_frontend.c +++ b/lib/ringbuffer/ring_buffer_frontend.c @@ -1,7 +1,22 @@ /* * ring_buffer_frontend.c * - * (C) Copyright 2005-2010 - Mathieu Desnoyers + * Copyright (C) 2005-2012 Mathieu Desnoyers + * + * 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: @@ -34,8 +49,6 @@ * - 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. */ #include @@ -103,7 +116,7 @@ void lib_ring_buffer_free(struct lib_ring_buffer *buf) void lib_ring_buffer_reset(struct lib_ring_buffer *buf) { struct channel *chan = buf->backend.chan; - const struct lib_ring_buffer_config *config = chan->backend.config; + const struct lib_ring_buffer_config *config = &chan->backend.config; unsigned int i; /* @@ -161,7 +174,7 @@ EXPORT_SYMBOL_GPL(channel_reset); int lib_ring_buffer_create(struct lib_ring_buffer *buf, struct channel_backend *chanb, int cpu) { - const struct lib_ring_buffer_config *config = chanb->config; + const struct lib_ring_buffer_config *config = &chanb->config; struct channel *chan = container_of(chanb, struct channel, backend); void *priv = chanb->priv; size_t subbuf_header_size; @@ -203,6 +216,7 @@ int lib_ring_buffer_create(struct lib_ring_buffer *buf, } init_waitqueue_head(&buf->read_wait); + init_waitqueue_head(&buf->write_wait); raw_spin_lock_init(&buf->raw_tick_nohz_spinlock); /* @@ -252,7 +266,7 @@ static void switch_buffer_timer(unsigned long data) { struct lib_ring_buffer *buf = (struct lib_ring_buffer *)data; struct channel *chan = buf->backend.chan; - const struct lib_ring_buffer_config *config = chan->backend.config; + const struct lib_ring_buffer_config *config = &chan->backend.config; /* * Only flush buffers periodically if readers are active. @@ -274,7 +288,7 @@ static void switch_buffer_timer(unsigned long data) static void lib_ring_buffer_start_switch_timer(struct lib_ring_buffer *buf) { struct channel *chan = buf->backend.chan; - const struct lib_ring_buffer_config *config = chan->backend.config; + const struct lib_ring_buffer_config *config = &chan->backend.config; if (!chan->switch_timer_interval || buf->switch_timer_enabled) return; @@ -310,7 +324,7 @@ static void read_buffer_timer(unsigned long data) { struct lib_ring_buffer *buf = (struct lib_ring_buffer *)data; struct channel *chan = buf->backend.chan; - const struct lib_ring_buffer_config *config = chan->backend.config; + const struct lib_ring_buffer_config *config = &chan->backend.config; CHAN_WARN_ON(chan, !buf->backend.allocated); @@ -334,7 +348,7 @@ static void read_buffer_timer(unsigned long data) static void lib_ring_buffer_start_read_timer(struct lib_ring_buffer *buf) { struct channel *chan = buf->backend.chan; - const struct lib_ring_buffer_config *config = chan->backend.config; + const struct lib_ring_buffer_config *config = &chan->backend.config; if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER || !chan->read_timer_interval @@ -359,7 +373,7 @@ static void lib_ring_buffer_start_read_timer(struct lib_ring_buffer *buf) static void lib_ring_buffer_stop_read_timer(struct lib_ring_buffer *buf) { struct channel *chan = buf->backend.chan; - const struct lib_ring_buffer_config *config = chan->backend.config; + const struct lib_ring_buffer_config *config = &chan->backend.config; if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER || !chan->read_timer_interval @@ -396,7 +410,7 @@ int __cpuinit lib_ring_buffer_cpu_hp_callback(struct notifier_block *nb, struct channel *chan = container_of(nb, struct channel, cpu_hp_notifier); struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, cpu); - const struct lib_ring_buffer_config *config = chan->backend.config; + const struct lib_ring_buffer_config *config = &chan->backend.config; if (!chan->cpu_hp_enable) return NOTIFY_DONE; @@ -451,7 +465,7 @@ static int notrace ring_buffer_tick_nohz_callback(struct notifier_block *nb, { struct channel *chan = container_of(nb, struct channel, tick_nohz_notifier); - const struct lib_ring_buffer_config *config = chan->backend.config; + const struct lib_ring_buffer_config *config = &chan->backend.config; struct lib_ring_buffer *buf; int cpu = smp_processor_id(); @@ -523,7 +537,7 @@ void notrace lib_ring_buffer_tick_nohz_restart(void) */ static void channel_unregister_notifiers(struct channel *chan) { - const struct lib_ring_buffer_config *config = chan->backend.config; + const struct lib_ring_buffer_config *config = &chan->backend.config; int cpu; channel_iterator_unregister_notifiers(chan); @@ -707,7 +721,7 @@ void channel_release(struct kref *kref) void *channel_destroy(struct channel *chan) { int cpu; - const struct lib_ring_buffer_config *config = chan->backend.config; + const struct lib_ring_buffer_config *config = &chan->backend.config; void *priv; channel_unregister_notifiers(chan); @@ -751,8 +765,8 @@ void *channel_destroy(struct channel *chan) ACCESS_ONCE(chan->finalized) = 1; wake_up_interruptible(&chan->hp_wait); wake_up_interruptible(&chan->read_wait); - kref_put(&chan->ref, channel_release); priv = chan->backend.priv; + kref_put(&chan->ref, channel_release); return priv; } EXPORT_SYMBOL_GPL(channel_destroy); @@ -817,7 +831,7 @@ int lib_ring_buffer_snapshot(struct lib_ring_buffer *buf, unsigned long *consumed, unsigned long *produced) { struct channel *chan = buf->backend.chan; - const struct lib_ring_buffer_config *config = chan->backend.config; + const struct lib_ring_buffer_config *config = &chan->backend.config; unsigned long consumed_cur, write_offset; int finalized; @@ -867,6 +881,8 @@ EXPORT_SYMBOL_GPL(lib_ring_buffer_snapshot); /** * lib_ring_buffer_put_snapshot - move consumed counter forward + * + * Should only be called from consumer context. * @buf: ring buffer * @consumed_new: new consumed count value */ @@ -888,6 +904,8 @@ void lib_ring_buffer_move_consumer(struct lib_ring_buffer *buf, while ((long) consumed - (long) consumed_new < 0) consumed = atomic_long_cmpxchg(&buf->consumed, consumed, consumed_new); + /* Wake-up the metadata producer */ + wake_up_interruptible(&buf->write_wait); } EXPORT_SYMBOL_GPL(lib_ring_buffer_move_consumer); @@ -904,7 +922,7 @@ int lib_ring_buffer_get_subbuf(struct lib_ring_buffer *buf, unsigned long consumed) { struct channel *chan = buf->backend.chan; - const struct lib_ring_buffer_config *config = chan->backend.config; + const struct lib_ring_buffer_config *config = &chan->backend.config; unsigned long consumed_cur, consumed_idx, commit_count, write_offset; int ret; int finalized; @@ -1050,7 +1068,7 @@ void lib_ring_buffer_put_subbuf(struct lib_ring_buffer *buf) { struct lib_ring_buffer_backend *bufb = &buf->backend; struct channel *chan = bufb->chan; - const struct lib_ring_buffer_config *config = chan->backend.config; + const struct lib_ring_buffer_config *config = &chan->backend.config; unsigned long read_sb_bindex, consumed_idx, consumed; CHAN_WARN_ON(chan, atomic_long_read(&buf->active_readers) != 1); @@ -1109,7 +1127,7 @@ void lib_ring_buffer_print_subbuffer_errors(struct lib_ring_buffer *buf, unsigned long cons_offset, int cpu) { - const struct lib_ring_buffer_config *config = chan->backend.config; + const struct lib_ring_buffer_config *config = &chan->backend.config; unsigned long cons_idx, commit_count, commit_count_sb; cons_idx = subbuf_index(cons_offset, chan); @@ -1135,15 +1153,9 @@ void lib_ring_buffer_print_buffer_errors(struct lib_ring_buffer *buf, struct channel *chan, void *priv, int cpu) { - const struct lib_ring_buffer_config *config = chan->backend.config; + const struct 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 @@ -1152,7 +1164,7 @@ void lib_ring_buffer_print_buffer_errors(struct lib_ring_buffer *buf, write_offset = v_read(config, &buf->offset); cons_offset = atomic_long_read(&buf->consumed); if (write_offset != cons_offset) - printk(KERN_WARNING + printk(KERN_DEBUG "ring buffer %s, cpu %d: " "non-consumed data\n" " [ %lu bytes written, %lu bytes read ]\n", @@ -1171,27 +1183,34 @@ static void lib_ring_buffer_print_errors(struct channel *chan, struct lib_ring_buffer *buf, int cpu) { - const struct lib_ring_buffer_config *config = chan->backend.config; + const struct lib_ring_buffer_config *config = &chan->backend.config; void *priv = chan->backend.priv; - printk(KERN_DEBUG "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)) - printk(KERN_WARNING - "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)); - + if (!strcmp(chan->backend.name, "relay-metadata")) { + printk(KERN_DEBUG "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 { + printk(KERN_DEBUG "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)) + printk(KERN_WARNING + "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); } @@ -1206,7 +1225,7 @@ void lib_ring_buffer_switch_old_start(struct lib_ring_buffer *buf, struct switch_offsets *offsets, u64 tsc) { - const struct lib_ring_buffer_config *config = chan->backend.config; + const struct lib_ring_buffer_config *config = &chan->backend.config; unsigned long oldidx = subbuf_index(offsets->old, chan); unsigned long commit_count; @@ -1250,7 +1269,7 @@ void lib_ring_buffer_switch_old_end(struct lib_ring_buffer *buf, struct switch_offsets *offsets, u64 tsc) { - const struct lib_ring_buffer_config *config = chan->backend.config; + const struct 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; @@ -1293,7 +1312,7 @@ void lib_ring_buffer_switch_new_start(struct lib_ring_buffer *buf, struct switch_offsets *offsets, u64 tsc) { - const struct lib_ring_buffer_config *config = chan->backend.config; + const struct lib_ring_buffer_config *config = &chan->backend.config; unsigned long beginidx = subbuf_index(offsets->begin, chan); unsigned long commit_count; @@ -1335,7 +1354,7 @@ void lib_ring_buffer_switch_new_end(struct lib_ring_buffer *buf, struct switch_offsets *offsets, u64 tsc) { - const struct lib_ring_buffer_config *config = chan->backend.config; + 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; @@ -1377,7 +1396,7 @@ int lib_ring_buffer_try_switch_slow(enum switch_mode mode, struct switch_offsets *offsets, u64 *tsc) { - const struct lib_ring_buffer_config *config = chan->backend.config; + const struct lib_ring_buffer_config *config = &chan->backend.config; unsigned long off; offsets->begin = v_read(config, &buf->offset); @@ -1405,6 +1424,19 @@ int lib_ring_buffer_try_switch_slow(enum switch_mode mode, */ if (mode == SWITCH_FLUSH || off > 0) { if (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 (unlikely(subbuf_trunc(offsets->begin, chan) + - subbuf_trunc((unsigned long) + atomic_long_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 @@ -1436,7 +1468,7 @@ int lib_ring_buffer_try_switch_slow(enum switch_mode mode, void lib_ring_buffer_switch_slow(struct lib_ring_buffer *buf, enum switch_mode mode) { struct channel *chan = buf->backend.chan; - const struct lib_ring_buffer_config *config = chan->backend.config; + const struct lib_ring_buffer_config *config = &chan->backend.config; struct switch_offsets offsets; unsigned long oldidx; u64 tsc; @@ -1484,6 +1516,48 @@ void lib_ring_buffer_switch_slow(struct lib_ring_buffer *buf, enum switch_mode m } EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_slow); +static void remote_switch(void *info) +{ + struct lib_ring_buffer *buf = info; + + lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE); +} + +void lib_ring_buffer_switch_remote(struct lib_ring_buffer *buf) +{ + struct channel *chan = buf->backend.chan; + const struct lib_ring_buffer_config *config = &chan->backend.config; + int ret; + + /* + * With global synchronization we don't need to use the IPI scheme. + */ + if (config->sync == RING_BUFFER_SYNC_GLOBAL) { + lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE); + return; + } + + /* + * Taking lock on CPU hotplug to ensure two things: first, that the + * target cpu is not taken concurrently offline while we are within + * smp_call_function_single() (I don't trust that get_cpu() on the + * _local_ CPU actually inhibit CPU hotplug for the _remote_ CPU (to be + * confirmed)). Secondly, if it happens that the CPU is not online, our + * own call to lib_ring_buffer_switch_slow() needs to be protected from + * CPU hotplug handlers, which can also perform a remote subbuffer + * switch. + */ + get_online_cpus(); + ret = smp_call_function_single(buf->backend.cpu, + remote_switch, buf, 1); + if (ret) { + /* Remote CPU is offline, do it ourself. */ + lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE); + } + put_online_cpus(); +} +EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_remote); + /* * Returns : * 0 if ok @@ -1497,10 +1571,11 @@ int lib_ring_buffer_try_reserve_slow(struct lib_ring_buffer *buf, struct switch_offsets *offsets, struct lib_ring_buffer_ctx *ctx) { - const struct lib_ring_buffer_config *config = chan->backend.config; - unsigned long reserve_commit_diff; + const struct lib_ring_buffer_config *config = &chan->backend.config; + 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; @@ -1532,7 +1607,7 @@ int lib_ring_buffer_try_reserve_slow(struct lib_ring_buffer *buf, } } if (unlikely(offsets->switch_new_start)) { - unsigned long sb_index; + unsigned long sb_index, commit_count; /* * We are typically not filling the previous buffer completely. @@ -1543,12 +1618,31 @@ int lib_ring_buffer_try_reserve_slow(struct lib_ring_buffer *buf, + 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. + */ + smp_rmb(); + commit_count = v_read(config, + &buf->commit_cold[sb_index].cc_sb); + /* Read buf->commit_cold[sb_index].cc_sb before buf->offset. */ + smp_rmb(); + if (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, - &buf->commit_cold[sb_index].cc_sb) - & chan->commit_count_mask); + - (commit_count & chan->commit_count_mask); if (likely(reserve_commit_diff == 0)) { /* Next subbuffer not being written to. */ if (unlikely(config->mode != RING_BUFFER_OVERWRITE && @@ -1573,9 +1667,10 @@ int lib_ring_buffer_try_reserve_slow(struct lib_ring_buffer *buf, } else { /* * Next subbuffer reserve offset does not match the - * commit offset. Drop record in producer-consumer and - * overwrite mode. Caused by either a writer OOPS or too - * many nested writes over a reserve/commit pair. + * 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. */ v_inc(config, &buf->records_lost_wrap); return -EIO; @@ -1632,7 +1727,7 @@ int lib_ring_buffer_try_reserve_slow(struct lib_ring_buffer *buf, int lib_ring_buffer_reserve_slow(struct lib_ring_buffer_ctx *ctx) { struct channel *chan = ctx->chan; - const struct lib_ring_buffer_config *config = chan->backend.config; + const struct lib_ring_buffer_config *config = &chan->backend.config; struct lib_ring_buffer *buf; struct switch_offsets offsets; int ret; @@ -1697,3 +1792,20 @@ int lib_ring_buffer_reserve_slow(struct lib_ring_buffer_ctx *ctx) return 0; } EXPORT_SYMBOL_GPL(lib_ring_buffer_reserve_slow); + +int __init init_lib_ring_buffer_frontend(void) +{ + int cpu; + + for_each_possible_cpu(cpu) + spin_lock_init(&per_cpu(ring_buffer_nohz_lock, cpu)); + return 0; +} + +module_init(init_lib_ring_buffer_frontend); + +void __exit exit_lib_ring_buffer_frontend(void) +{ +} + +module_exit(exit_lib_ring_buffer_frontend);