X-Git-Url: http://git.lttng.org/?a=blobdiff_plain;f=libust%2Fbuffers.c;h=12753c1961345f4430779ef8d1a69f48ce2aaf21;hb=b73a4c471dc987ea8548632dffb3c7050de77dd0;hp=241ca74d4172160000c3dbed8840acf617a90bde;hpb=058a68cbea07bb2b276c12212ea6129f166b2893;p=ust.git diff --git a/libust/buffers.c b/libust/buffers.c index 241ca74..12753c1 100644 --- a/libust/buffers.c +++ b/libust/buffers.c @@ -33,6 +33,13 @@ #include "tracercore.h" #include "usterr.h" +struct ltt_reserve_switch_offsets { + long begin, end, old; + long begin_switch, end_switch_current, end_switch_old; + size_t before_hdr_pad, size; +}; + + static DEFINE_MUTEX(ust_buffers_channels_mutex); static LIST_HEAD(ust_buffers_channels); @@ -63,7 +70,31 @@ static int get_n_cpus(void) return result; } -static int ust_buffers_init_buffer(struct ltt_trace_struct *trace, +/* _ust_buffers_write() + * + * @buf: destination buffer + * @offset: offset in destination + * @src: source buffer + * @len: length of source + * @cpy: already copied + */ + +void _ust_buffers_write(struct ust_buffer *buf, size_t offset, + const void *src, size_t len, ssize_t cpy) +{ + do { + len -= cpy; + src += cpy; + offset += cpy; + + WARN_ON(offset >= buf->buf_size); + + cpy = min_t(size_t, len, buf->buf_size - offset); + ust_buffers_do_copy(buf->buf_data + offset, src, cpy); + } while (unlikely(len != cpy)); +} + +static int ust_buffers_init_buffer(struct ust_trace *trace, struct ust_channel *ltt_chan, struct ust_buffer *buf, unsigned int n_subbufs); @@ -194,11 +225,20 @@ int ust_buffers_channel_open(struct ust_channel *chan, size_t subbuf_size, size_ if(subbuf_size == 0 || subbuf_cnt == 0) return -1; + /* Check that the subbuffer size is larger than a page. */ + WARN_ON_ONCE(subbuf_size < PAGE_SIZE); + + /* + * Make sure the number of subbuffers and subbuffer size are power of 2. + */ + WARN_ON_ONCE(hweight32(subbuf_size) != 1); + WARN_ON(hweight32(subbuf_cnt) != 1); + chan->version = UST_CHANNEL_VERSION; chan->subbuf_cnt = subbuf_cnt; chan->subbuf_size = subbuf_size; chan->subbuf_size_order = get_count_order(subbuf_size); - chan->alloc_size = FIX_SIZE(subbuf_size * subbuf_cnt); + chan->alloc_size = subbuf_size * subbuf_cnt; kref_init(&chan->kref); @@ -245,86 +285,10 @@ void ust_buffers_channel_close(struct ust_channel *chan) mutex_unlock(&ust_buffers_channels_mutex); } -/* _ust_buffers_write() - * - * @buf: destination buffer - * @offset: offset in destination - * @src: source buffer - * @len: length of source - * @cpy: already copied - */ - -void _ust_buffers_write(struct ust_buffer *buf, size_t offset, - const void *src, size_t len, ssize_t cpy) -{ - do { - len -= cpy; - src += cpy; - offset += cpy; - - WARN_ON(offset >= buf->buf_size); - - cpy = min_t(size_t, len, buf->buf_size - offset); - ust_buffers_do_copy(buf->buf_data + offset, src, cpy); - } while (unlikely(len != cpy)); -} - -void *ltt_buffers_offset_address(struct ust_buffer *buf, size_t offset) -{ - return ((char *)buf->buf_data)+offset; -} - /* * ------- */ -/* - * Last TSC comparison functions. Check if the current TSC overflows - * LTT_TSC_BITS bits from the last TSC read. Reads and writes last_tsc - * atomically. - */ - -/* FIXME: does this test work properly? */ -#if (BITS_PER_LONG == 32) -static inline void save_last_tsc(struct ust_buffer *ltt_buf, - u64 tsc) -{ - ltt_buf->last_tsc = (unsigned long)(tsc >> LTT_TSC_BITS); -} - -static inline int last_tsc_overflow(struct ust_buffer *ltt_buf, - u64 tsc) -{ - unsigned long tsc_shifted = (unsigned long)(tsc >> LTT_TSC_BITS); - - if (unlikely((tsc_shifted - ltt_buf->last_tsc))) - return 1; - else - return 0; -} -#else -static inline void save_last_tsc(struct ust_buffer *ltt_buf, - u64 tsc) -{ - ltt_buf->last_tsc = (unsigned long)tsc; -} - -static inline int last_tsc_overflow(struct ust_buffer *ltt_buf, - u64 tsc) -{ - if (unlikely((tsc - ltt_buf->last_tsc) >> LTT_TSC_BITS)) - return 1; - else - return 0; -} -#endif - -/* - * A switch is done during tracing or as a final flush after tracing (so it - * won't write in the new sub-buffer). - */ -enum force_switch_mode { FORCE_ACTIVE, FORCE_FLUSH }; - static void ust_buffers_destroy_buffer(struct ust_channel *ltt_chan, int cpu); static void ltt_force_switch(struct ust_buffer *buf, @@ -333,13 +297,13 @@ static void ltt_force_switch(struct ust_buffer *buf, /* * Trace callbacks */ -static void ltt_buffer_begin_callback(struct ust_buffer *buf, +static void ltt_buffer_begin(struct ust_buffer *buf, u64 tsc, unsigned int subbuf_idx) { struct ust_channel *channel = buf->chan; struct ltt_subbuffer_header *header = (struct ltt_subbuffer_header *) - ltt_buffers_offset_address(buf, + ust_buffers_offset_address(buf, subbuf_idx * buf->chan->subbuf_size); header->cycle_count_begin = tsc; @@ -352,12 +316,12 @@ static void ltt_buffer_begin_callback(struct ust_buffer *buf, * offset is assumed to never be 0 here : never deliver a completely empty * subbuffer. The lost size is between 0 and subbuf_size-1. */ -static notrace void ltt_buffer_end_callback(struct ust_buffer *buf, +static notrace void ltt_buffer_end(struct ust_buffer *buf, u64 tsc, unsigned int offset, unsigned int subbuf_idx) { struct ltt_subbuffer_header *header = (struct ltt_subbuffer_header *) - ltt_buffers_offset_address(buf, + ust_buffers_offset_address(buf, subbuf_idx * buf->chan->subbuf_size); header->lost_size = SUBBUF_OFFSET((buf->chan->subbuf_size - offset), @@ -365,38 +329,6 @@ static notrace void ltt_buffer_end_callback(struct ust_buffer *buf, header->cycle_count_end = tsc; header->events_lost = local_read(&buf->events_lost); header->subbuf_corrupt = local_read(&buf->corrupted_subbuffers); - -} - -void (*wake_consumer)(void *, int) = NULL; - -void relay_set_wake_consumer(void (*wake)(void *, int)) -{ - wake_consumer = wake; -} - -void relay_wake_consumer(void *arg, int finished) -{ - if(wake_consumer) - wake_consumer(arg, finished); -} - -static notrace void ltt_deliver(struct ust_buffer *buf, unsigned int subbuf_idx, - long commit_count) -{ - int result; - -//ust// #ifdef CONFIG_LTT_VMCORE - local_set(&buf->commit_seq[subbuf_idx], commit_count); -//ust// #endif - - /* wakeup consumer */ - result = write(buf->data_ready_fd_write, "1", 1); - if(result == -1) { - PERROR("write (in ltt_relay_buffer_flush)"); - ERR("this should never happen!"); - } -//ust// atomic_set(<t_buf->wakeup_readers, 1); } /* @@ -406,27 +338,79 @@ static notrace void ltt_buf_unfull(struct ust_buffer *buf, unsigned int subbuf_idx, long offset) { -//ust// struct ltt_channel_struct *ltt_channel = -//ust// (struct ltt_channel_struct *)buf->chan->private_data; -//ust// struct ltt_channel_buf_struct *ltt_buf = ltt_channel->buf; -//ust// -//ust// ltt_relay_wake_writers(ltt_buf); } -int ust_buffers_do_get_subbuf(struct ust_buffer *buf, long *pconsumed_old) +/* + * Promote compiler barrier to a smp_mb(). + * For the specific LTTng case, this IPI call should be removed if the + * architecture does not reorder writes. This should eventually be provided by + * a separate architecture-specific infrastructure. + */ +static void remote_mb(void *info) +{ + smp_mb(); +} + +int ust_buffers_get_subbuf(struct ust_buffer *buf, long *consumed) { struct ust_channel *channel = buf->chan; long consumed_old, consumed_idx, commit_count, write_offset; +//ust// int retval; + consumed_old = atomic_long_read(&buf->consumed); consumed_idx = SUBBUF_INDEX(consumed_old, buf->chan); - commit_count = local_read(&buf->commit_count[consumed_idx]); + commit_count = local_read(&buf->commit_count[consumed_idx].cc_sb); /* * Make sure we read the commit count before reading the buffer * data and the write offset. Correct consumed offset ordering * wrt commit count is insured by the use of cmpxchg to update * the consumed offset. + * smp_call_function_single can fail if the remote CPU is offline, + * this is OK because then there is no wmb to execute there. + * If our thread is executing on the same CPU as the on the buffers + * belongs to, we don't have to synchronize it at all. If we are + * migrated, the scheduler will take care of the memory barriers. + * Normally, smp_call_function_single() should ensure program order when + * executing the remote function, which implies that it surrounds the + * function execution with : + * smp_mb() + * send IPI + * csd_lock_wait + * recv IPI + * smp_mb() + * exec. function + * smp_mb() + * csd unlock + * smp_mb() + * + * However, smp_call_function_single() does not seem to clearly execute + * such barriers. It depends on spinlock semantic to provide the barrier + * before executing the IPI and, when busy-looping, csd_lock_wait only + * executes smp_mb() when it has to wait for the other CPU. + * + * I don't trust this code. Therefore, let's add the smp_mb() sequence + * required ourself, even if duplicated. It has no performance impact + * anyway. + * + * smp_mb() is needed because smp_rmb() and smp_wmb() only order read vs + * read and write vs write. They do not ensure core synchronization. We + * really have to ensure total order between the 3 barriers running on + * the 2 CPUs. + */ +//ust// #ifdef LTT_NO_IPI_BARRIER + /* + * Local rmb to match the remote wmb to read the commit count before the + * buffer data and the write offset. */ smp_rmb(); +//ust// #else +//ust// if (raw_smp_processor_id() != buf->cpu) { +//ust// smp_mb(); /* Total order with IPI handler smp_mb() */ +//ust// smp_call_function_single(buf->cpu, remote_mb, NULL, 1); +//ust// smp_mb(); /* Total order with IPI handler smp_mb() */ +//ust// } +//ust// #endif + write_offset = local_read(&buf->offset); /* * Check that the subbuffer we are trying to consume has been @@ -449,11 +433,17 @@ int ust_buffers_do_get_subbuf(struct ust_buffer *buf, long *pconsumed_old) return -EAGAIN; } - *pconsumed_old = consumed_old; + /* FIXME: is this ok to disable the reading feature? */ +//ust// retval = update_read_sb_index(buf, consumed_idx); +//ust// if (retval) +//ust// return retval; + + *consumed = consumed_old; + return 0; } -int ust_buffers_do_put_subbuf(struct ust_buffer *buf, u32 uconsumed_old) +int ust_buffers_put_subbuf(struct ust_buffer *buf, unsigned long uconsumed_old) { long consumed_new, consumed_old; @@ -483,36 +473,125 @@ int ust_buffers_do_put_subbuf(struct ust_buffer *buf, u32 uconsumed_old) return 0; } +//ust// static void switch_buffer(unsigned long data) +//ust// { +//ust// struct ltt_channel_buf_struct *ltt_buf = +//ust// (struct ltt_channel_buf_struct *)data; +//ust// struct rchan_buf *buf = ltt_buf->rbuf; +//ust// +//ust// if (buf) +//ust// ltt_force_switch(buf, FORCE_ACTIVE); +//ust// +//ust// ltt_buf->switch_timer.expires += ltt_buf->switch_timer_interval; +//ust// add_timer_on(<t_buf->switch_timer, smp_processor_id()); +//ust// } +//ust// +//ust// static void start_switch_timer(struct ltt_channel_struct *ltt_channel) +//ust// { +//ust// struct rchan *rchan = ltt_channel->trans_channel_data; +//ust// int cpu; +//ust// +//ust// if (!ltt_channel->switch_timer_interval) +//ust// return; +//ust// +//ust// // TODO : hotplug +//ust// for_each_online_cpu(cpu) { +//ust// struct ltt_channel_buf_struct *ltt_buf; +//ust// struct rchan_buf *buf; +//ust// +//ust// buf = rchan->buf[cpu]; +//ust// ltt_buf = buf->chan_private; +//ust// buf->random_access = 1; +//ust// ltt_buf->switch_timer_interval = +//ust// ltt_channel->switch_timer_interval; +//ust// init_timer(<t_buf->switch_timer); +//ust// ltt_buf->switch_timer.function = switch_buffer; +//ust// ltt_buf->switch_timer.expires = jiffies + +//ust// ltt_buf->switch_timer_interval; +//ust// ltt_buf->switch_timer.data = (unsigned long)ltt_buf; +//ust// add_timer_on(<t_buf->switch_timer, cpu); +//ust// } +//ust// } +//ust// +//ust// /* +//ust// * Cannot use del_timer_sync with add_timer_on, so use an IPI to locally +//ust// * delete the timer. +//ust// */ +//ust// static void stop_switch_timer_ipi(void *info) +//ust// { +//ust// struct ltt_channel_buf_struct *ltt_buf = +//ust// (struct ltt_channel_buf_struct *)info; +//ust// +//ust// del_timer(<t_buf->switch_timer); +//ust// } +//ust// +//ust// static void stop_switch_timer(struct ltt_channel_struct *ltt_channel) +//ust// { +//ust// struct rchan *rchan = ltt_channel->trans_channel_data; +//ust// int cpu; +//ust// +//ust// if (!ltt_channel->switch_timer_interval) +//ust// return; +//ust// +//ust// // TODO : hotplug +//ust// for_each_online_cpu(cpu) { +//ust// struct ltt_channel_buf_struct *ltt_buf; +//ust// struct rchan_buf *buf; +//ust// +//ust// buf = rchan->buf[cpu]; +//ust// ltt_buf = buf->chan_private; +//ust// smp_call_function(stop_switch_timer_ipi, ltt_buf, 1); +//ust// buf->random_access = 0; +//ust// } +//ust// } + +static void ust_buffers_print_written(struct ust_channel *chan, + long cons_off, unsigned int cpu) +{ + struct ust_buffer *buf = chan->buf[cpu]; + long cons_idx, events_count; + + cons_idx = SUBBUF_INDEX(cons_off, chan); + events_count = local_read(&buf->commit_count[cons_idx].events); + + if (events_count) + printk(KERN_INFO + "channel %s: %lu events written (cpu %u, index %lu)\n", + chan->channel_name, events_count, cpu, cons_idx); +} + static void ltt_relay_print_subbuffer_errors( struct ust_channel *channel, long cons_off, int cpu) { struct ust_buffer *ltt_buf = channel->buf[cpu]; - long cons_idx, commit_count, write_offset; + long cons_idx, commit_count, commit_count_sb, write_offset; cons_idx = SUBBUF_INDEX(cons_off, channel); - commit_count = local_read(<t_buf->commit_count[cons_idx]); + commit_count = local_read(<t_buf->commit_count[cons_idx].cc); + commit_count_sb = local_read(<t_buf->commit_count[cons_idx].cc_sb); + /* * No need to order commit_count and write_offset reads because we * execute after trace is stopped when there are no readers left. */ write_offset = local_read(<t_buf->offset); WARN( "LTT : unread channel %s offset is %ld " - "and cons_off : %ld\n", - channel->channel_name, write_offset, cons_off); + "and cons_off : %ld (cpu %d)\n", + channel->channel_name, write_offset, cons_off, cpu); /* Check each sub-buffer for non filled commit count */ if (((commit_count - channel->subbuf_size) & channel->commit_count_mask) - (BUFFER_TRUNC(cons_off, channel) >> channel->n_subbufs_order) != 0) { ERR("LTT : %s : subbuffer %lu has non filled " - "commit count %lu.\n", - channel->channel_name, cons_idx, commit_count); + "commit count [cc, cc_sb] [%lu,%lu].\n", + channel->channel_name, cons_idx, commit_count, commit_count_sb); } ERR("LTT : %s : commit count : %lu, subbuf size %zd\n", channel->channel_name, commit_count, channel->subbuf_size); } -static void ltt_relay_print_errors(struct ltt_trace_struct *trace, +static void ltt_relay_print_errors(struct ust_trace *trace, struct ust_channel *channel, int cpu) { struct ust_buffer *ltt_buf = channel->buf[cpu]; @@ -525,6 +604,9 @@ static void ltt_relay_print_errors(struct ltt_trace_struct *trace, if (!channel) return; + for (cons_off = 0; cons_off < rchan->alloc_size; + cons_off = SUBBUF_ALIGN(cons_off, rchan)) + ust_buffers_print_written(ltt_chan, cons_off, cpu); for (cons_off = atomic_long_read(<t_buf->consumed); (SUBBUF_TRUNC(local_read(<t_buf->offset), channel) @@ -535,17 +617,17 @@ static void ltt_relay_print_errors(struct ltt_trace_struct *trace, static void ltt_relay_print_buffer_errors(struct ust_channel *channel, int cpu) { - struct ltt_trace_struct *trace = channel->trace; + struct ust_trace *trace = channel->trace; struct ust_buffer *ltt_buf = channel->buf[cpu]; if (local_read(<t_buf->events_lost)) - ERR("channel %s: %ld events lost", + ERR("channel %s: %ld events lost (cpu %d)", channel->channel_name, - local_read(<t_buf->events_lost)); + local_read(<t_buf->events_lost), cpu); if (local_read(<t_buf->corrupted_subbuffers)) - ERR("channel %s : %ld corrupted subbuffers", + ERR("channel %s : %ld corrupted subbuffers (cpu %d)", channel->channel_name, - local_read(<t_buf->corrupted_subbuffers)); + local_read(<t_buf->corrupted_subbuffers), cpu); ltt_relay_print_errors(trace, channel, cpu); } @@ -560,14 +642,14 @@ static void ltt_relay_release_channel(struct kref *kref) /* * Create ltt buffer. */ -//ust// static int ltt_relay_create_buffer(struct ltt_trace_struct *trace, +//ust// static int ltt_relay_create_buffer(struct ust_trace *trace, //ust// struct ltt_channel_struct *ltt_chan, struct rchan_buf *buf, //ust// unsigned int cpu, unsigned int n_subbufs) //ust// { //ust// struct ltt_channel_buf_struct *ltt_buf = //ust// percpu_ptr(ltt_chan->buf, cpu); //ust// unsigned int j; -//ust// +//ust// //ust// ltt_buf->commit_count = //ust// kzalloc_node(sizeof(ltt_buf->commit_count) * n_subbufs, //ust// GFP_KERNEL, cpu_to_node(cpu)); @@ -584,19 +666,19 @@ static void ltt_relay_release_channel(struct kref *kref) //ust// init_waitqueue_head(<t_buf->write_wait); //ust// atomic_set(<t_buf->wakeup_readers, 0); //ust// spin_lock_init(<t_buf->full_lock); -//ust// +//ust// //ust// ltt_buffer_begin_callback(buf, trace->start_tsc, 0); //ust// /* atomic_add made on local variable on data that belongs to //ust// * various CPUs : ok because tracing not started (for this cpu). */ //ust// local_add(ltt_subbuffer_header_size(), <t_buf->commit_count[0]); -//ust// +//ust// //ust// local_set(<t_buf->events_lost, 0); //ust// local_set(<t_buf->corrupted_subbuffers, 0); -//ust// +//ust// //ust// return 0; //ust// } -static int ust_buffers_init_buffer(struct ltt_trace_struct *trace, +static int ust_buffers_init_buffer(struct ust_trace *trace, struct ust_channel *ltt_chan, struct ust_buffer *buf, unsigned int n_subbufs) { @@ -605,7 +687,7 @@ static int ust_buffers_init_buffer(struct ltt_trace_struct *trace, int result; buf->commit_count = - zmalloc(sizeof(buf->commit_count) * n_subbufs); + zmalloc(sizeof(*buf->commit_count) * n_subbufs); if (!buf->commit_count) return -ENOMEM; kref_get(&trace->kref); @@ -614,15 +696,17 @@ static int ust_buffers_init_buffer(struct ltt_trace_struct *trace, local_set(&buf->offset, ltt_subbuffer_header_size()); atomic_long_set(&buf->consumed, 0); atomic_long_set(&buf->active_readers, 0); - for (j = 0; j < n_subbufs; j++) - local_set(&buf->commit_count[j], 0); + for (j = 0; j < n_subbufs; j++) { + local_set(&buf->commit_count[j].cc, 0); + local_set(&buf->commit_count[j].cc_sb, 0); + } //ust// init_waitqueue_head(&buf->write_wait); //ust// atomic_set(&buf->wakeup_readers, 0); //ust// spin_lock_init(&buf->full_lock); - ltt_buffer_begin_callback(buf, trace->start_tsc, 0); + ltt_buffer_begin(buf, trace->start_tsc, 0); - local_add(ltt_subbuffer_header_size(), &buf->commit_count[0]); + local_add(ltt_subbuffer_header_size(), &buf->commit_count[0].cc); local_set(&buf->events_lost, 0); local_set(&buf->corrupted_subbuffers, 0); @@ -654,7 +738,7 @@ static int ust_buffers_init_buffer(struct ltt_trace_struct *trace, /* FIXME: use this function */ static void ust_buffers_destroy_buffer(struct ust_channel *ltt_chan, int cpu) { - struct ltt_trace_struct *trace = ltt_chan->trace; + struct ust_trace *trace = ltt_chan->trace; struct ust_buffer *ltt_buf = ltt_chan->buf[cpu]; kref_put(<t_chan->trace->ltt_transport_kref, @@ -726,7 +810,7 @@ static int ust_buffers_alloc_channel_buf_structs(struct ust_channel *chan) /* * Create channel. */ -static int ust_buffers_create_channel(const char *trace_name, struct ltt_trace_struct *trace, +static int ust_buffers_create_channel(const char *trace_name, struct ust_trace *trace, const char *channel_name, struct ust_channel *ltt_chan, unsigned int subbuf_size, unsigned int n_subbufs, int overwrite) { @@ -735,8 +819,6 @@ static int ust_buffers_create_channel(const char *trace_name, struct ltt_trace_s kref_init(<t_chan->kref); ltt_chan->trace = trace; - ltt_chan->buffer_begin = ltt_buffer_begin_callback; - ltt_chan->buffer_end = ltt_buffer_end_callback; ltt_chan->overwrite = overwrite; ltt_chan->n_subbufs_order = get_count_order(n_subbufs); ltt_chan->commit_count_mask = (~0UL >> ltt_chan->n_subbufs_order); @@ -847,137 +929,512 @@ static void ltt_relay_remove_channel(struct ust_channel *channel) kref_put(&channel->kref, ltt_relay_release_channel); } -struct ltt_reserve_switch_offsets { - long begin, end, old; - long begin_switch, end_switch_current, end_switch_old; - long commit_count, reserve_commit_diff; - size_t before_hdr_pad, size; -}; +//ust// /* +//ust// * Returns : +//ust// * 0 if ok +//ust// * !0 if execution must be aborted. +//ust// */ +//ust// static inline int ltt_relay_try_reserve( +//ust// struct ust_channel *channel, struct ust_buffer *buf, +//ust// struct ltt_reserve_switch_offsets *offsets, size_t data_size, +//ust// u64 *tsc, unsigned int *rflags, int largest_align) +//ust// { +//ust// offsets->begin = local_read(&buf->offset); +//ust// offsets->old = offsets->begin; +//ust// offsets->begin_switch = 0; +//ust// offsets->end_switch_current = 0; +//ust// offsets->end_switch_old = 0; +//ust// +//ust// *tsc = trace_clock_read64(); +//ust// if (last_tsc_overflow(buf, *tsc)) +//ust// *rflags = LTT_RFLAG_ID_SIZE_TSC; +//ust// +//ust// if (SUBBUF_OFFSET(offsets->begin, buf->chan) == 0) { +//ust// offsets->begin_switch = 1; /* For offsets->begin */ +//ust// } else { +//ust// offsets->size = ust_get_header_size(channel, +//ust// offsets->begin, data_size, +//ust// &offsets->before_hdr_pad, *rflags); +//ust// offsets->size += ltt_align(offsets->begin + offsets->size, +//ust// largest_align) +//ust// + data_size; +//ust// if ((SUBBUF_OFFSET(offsets->begin, buf->chan) + offsets->size) +//ust// > buf->chan->subbuf_size) { +//ust// offsets->end_switch_old = 1; /* For offsets->old */ +//ust// offsets->begin_switch = 1; /* For offsets->begin */ +//ust// } +//ust// } +//ust// if (offsets->begin_switch) { +//ust// long subbuf_index; +//ust// +//ust// if (offsets->end_switch_old) +//ust// offsets->begin = SUBBUF_ALIGN(offsets->begin, +//ust// buf->chan); +//ust// offsets->begin = offsets->begin + ltt_subbuffer_header_size(); +//ust// /* Test new buffer integrity */ +//ust// subbuf_index = SUBBUF_INDEX(offsets->begin, buf->chan); +//ust// offsets->reserve_commit_diff = +//ust// (BUFFER_TRUNC(offsets->begin, buf->chan) +//ust// >> channel->n_subbufs_order) +//ust// - (local_read(&buf->commit_count[subbuf_index]) +//ust// & channel->commit_count_mask); +//ust// if (offsets->reserve_commit_diff == 0) { +//ust// long consumed; +//ust// +//ust// consumed = atomic_long_read(&buf->consumed); +//ust// +//ust// /* Next buffer not corrupted. */ +//ust// if (!channel->overwrite && +//ust// (SUBBUF_TRUNC(offsets->begin, buf->chan) +//ust// - SUBBUF_TRUNC(consumed, buf->chan)) +//ust// >= channel->alloc_size) { +//ust// +//ust// long consumed_idx = SUBBUF_INDEX(consumed, buf->chan); +//ust// long commit_count = local_read(&buf->commit_count[consumed_idx]); +//ust// if(((commit_count - buf->chan->subbuf_size) & channel->commit_count_mask) - (BUFFER_TRUNC(consumed, buf->chan) >> channel->n_subbufs_order) != 0) { +//ust// WARN("Event dropped. Caused by non-committed event."); +//ust// } +//ust// else { +//ust// WARN("Event dropped. Caused by non-consumed buffer."); +//ust// } +//ust// /* +//ust// * We do not overwrite non consumed buffers +//ust// * and we are full : event is lost. +//ust// */ +//ust// local_inc(&buf->events_lost); +//ust// return -1; +//ust// } else { +//ust// /* +//ust// * next buffer not corrupted, we are either in +//ust// * overwrite mode or the buffer is not full. +//ust// * It's safe to write in this new subbuffer. +//ust// */ +//ust// } +//ust// } else { +//ust// /* +//ust// * Next subbuffer corrupted. Force pushing reader even +//ust// * in normal mode. It's safe to write in this new +//ust// * subbuffer. +//ust// */ +//ust// } +//ust// offsets->size = ust_get_header_size(channel, +//ust// offsets->begin, data_size, +//ust// &offsets->before_hdr_pad, *rflags); +//ust// offsets->size += ltt_align(offsets->begin + offsets->size, +//ust// largest_align) +//ust// + data_size; +//ust// if ((SUBBUF_OFFSET(offsets->begin, buf->chan) + offsets->size) +//ust// > buf->chan->subbuf_size) { +//ust// /* +//ust// * Event too big for subbuffers, report error, don't +//ust// * complete the sub-buffer switch. +//ust// */ +//ust// local_inc(&buf->events_lost); +//ust// return -1; +//ust// } else { +//ust// /* +//ust// * We just made a successful buffer switch and the event +//ust// * fits in the new subbuffer. Let's write. +//ust// */ +//ust// } +//ust// } else { +//ust// /* +//ust// * Event fits in the current buffer and we are not on a switch +//ust// * boundary. It's safe to write. +//ust// */ +//ust// } +//ust// offsets->end = offsets->begin + offsets->size; +//ust// +//ust// if ((SUBBUF_OFFSET(offsets->end, buf->chan)) == 0) { +//ust// /* +//ust// * The offset_end will fall at the very beginning of the next +//ust// * subbuffer. +//ust// */ +//ust// offsets->end_switch_current = 1; /* For offsets->begin */ +//ust// } +//ust// return 0; +//ust// } +//ust// +//ust// /* +//ust// * Returns : +//ust// * 0 if ok +//ust// * !0 if execution must be aborted. +//ust// */ +//ust// static inline int ltt_relay_try_switch( +//ust// enum force_switch_mode mode, +//ust// struct ust_channel *channel, +//ust// struct ust_buffer *buf, +//ust// struct ltt_reserve_switch_offsets *offsets, +//ust// u64 *tsc) +//ust// { +//ust// long subbuf_index; +//ust// +//ust// offsets->begin = local_read(&buf->offset); +//ust// offsets->old = offsets->begin; +//ust// offsets->begin_switch = 0; +//ust// offsets->end_switch_old = 0; +//ust// +//ust// *tsc = trace_clock_read64(); +//ust// +//ust// if (SUBBUF_OFFSET(offsets->begin, buf->chan) != 0) { +//ust// offsets->begin = SUBBUF_ALIGN(offsets->begin, buf->chan); +//ust// offsets->end_switch_old = 1; +//ust// } else { +//ust// /* we do not have to switch : buffer is empty */ +//ust// return -1; +//ust// } +//ust// if (mode == FORCE_ACTIVE) +//ust// offsets->begin += ltt_subbuffer_header_size(); +//ust// /* +//ust// * Always begin_switch in FORCE_ACTIVE mode. +//ust// * Test new buffer integrity +//ust// */ +//ust// subbuf_index = SUBBUF_INDEX(offsets->begin, buf->chan); +//ust// offsets->reserve_commit_diff = +//ust// (BUFFER_TRUNC(offsets->begin, buf->chan) +//ust// >> channel->n_subbufs_order) +//ust// - (local_read(&buf->commit_count[subbuf_index]) +//ust// & channel->commit_count_mask); +//ust// if (offsets->reserve_commit_diff == 0) { +//ust// /* Next buffer not corrupted. */ +//ust// if (mode == FORCE_ACTIVE +//ust// && !channel->overwrite +//ust// && offsets->begin - atomic_long_read(&buf->consumed) +//ust// >= channel->alloc_size) { +//ust// /* +//ust// * We do not overwrite non consumed buffers and we are +//ust// * full : ignore switch while tracing is active. +//ust// */ +//ust// return -1; +//ust// } +//ust// } else { +//ust// /* +//ust// * Next subbuffer corrupted. Force pushing reader even in normal +//ust// * mode +//ust// */ +//ust// } +//ust// offsets->end = offsets->begin; +//ust// return 0; +//ust// } +//ust// +//ust// static inline void ltt_reserve_push_reader( +//ust// struct ust_channel *channel, +//ust// struct ust_buffer *buf, +//ust// struct ltt_reserve_switch_offsets *offsets) +//ust// { +//ust// long consumed_old, consumed_new; +//ust// +//ust// do { +//ust// consumed_old = atomic_long_read(&buf->consumed); +//ust// /* +//ust// * If buffer is in overwrite mode, push the reader consumed +//ust// * count if the write position has reached it and we are not +//ust// * at the first iteration (don't push the reader farther than +//ust// * the writer). This operation can be done concurrently by many +//ust// * writers in the same buffer, the writer being at the farthest +//ust// * write position sub-buffer index in the buffer being the one +//ust// * which will win this loop. +//ust// * If the buffer is not in overwrite mode, pushing the reader +//ust// * only happens if a sub-buffer is corrupted. +//ust// */ +//ust// if ((SUBBUF_TRUNC(offsets->end-1, buf->chan) +//ust// - SUBBUF_TRUNC(consumed_old, buf->chan)) +//ust// >= channel->alloc_size) +//ust// consumed_new = SUBBUF_ALIGN(consumed_old, buf->chan); +//ust// else { +//ust// consumed_new = consumed_old; +//ust// break; +//ust// } +//ust// } while (atomic_long_cmpxchg(&buf->consumed, consumed_old, +//ust// consumed_new) != consumed_old); +//ust// +//ust// if (consumed_old != consumed_new) { +//ust// /* +//ust// * Reader pushed : we are the winner of the push, we can +//ust// * therefore reequilibrate reserve and commit. Atomic increment +//ust// * of the commit count permits other writers to play around +//ust// * with this variable before us. We keep track of +//ust// * corrupted_subbuffers even in overwrite mode : +//ust// * we never want to write over a non completely committed +//ust// * sub-buffer : possible causes : the buffer size is too low +//ust// * compared to the unordered data input, or there is a writer +//ust// * that died between the reserve and the commit. +//ust// */ +//ust// if (offsets->reserve_commit_diff) { +//ust// /* +//ust// * We have to alter the sub-buffer commit count. +//ust// * We do not deliver the previous subbuffer, given it +//ust// * was either corrupted or not consumed (overwrite +//ust// * mode). +//ust// */ +//ust// local_add(offsets->reserve_commit_diff, +//ust// &buf->commit_count[ +//ust// SUBBUF_INDEX(offsets->begin, +//ust// buf->chan)]); +//ust// if (!channel->overwrite +//ust// || offsets->reserve_commit_diff +//ust// != channel->subbuf_size) { +//ust// /* +//ust// * The reserve commit diff was not subbuf_size : +//ust// * it means the subbuffer was partly written to +//ust// * and is therefore corrupted. If it is multiple +//ust// * of subbuffer size and we are in flight +//ust// * recorder mode, we are skipping over a whole +//ust// * subbuffer. +//ust// */ +//ust// local_inc(&buf->corrupted_subbuffers); +//ust// } +//ust// } +//ust// } +//ust// } +//ust// +//ust// /** +//ust// * ltt_relay_reserve_slot - Atomic slot reservation in a LTTng buffer. +//ust// * @trace: the trace structure to log to. +//ust// * @ltt_channel: channel structure +//ust// * @transport_data: data structure specific to ltt relay +//ust// * @data_size: size of the variable length data to log. +//ust// * @slot_size: pointer to total size of the slot (out) +//ust// * @buf_offset : pointer to reserved buffer offset (out) +//ust// * @tsc: pointer to the tsc at the slot reservation (out) +//ust// * @cpu: cpuid +//ust// * +//ust// * Return : -ENOSPC if not enough space, else returns 0. +//ust// * It will take care of sub-buffer switching. +//ust// */ +//ust// static notrace int ltt_relay_reserve_slot(struct ust_trace *trace, +//ust// struct ust_channel *channel, void **transport_data, +//ust// size_t data_size, size_t *slot_size, long *buf_offset, u64 *tsc, +//ust// unsigned int *rflags, int largest_align, int cpu) +//ust// { +//ust// struct ust_buffer *buf = *transport_data = channel->buf[cpu]; +//ust// struct ltt_reserve_switch_offsets offsets; +//ust// +//ust// offsets.reserve_commit_diff = 0; +//ust// offsets.size = 0; +//ust// +//ust// /* +//ust// * Perform retryable operations. +//ust// */ +//ust// if (ltt_nesting > 4) { +//ust// local_inc(&buf->events_lost); +//ust// return -EPERM; +//ust// } +//ust// do { +//ust// if (ltt_relay_try_reserve(channel, buf, &offsets, data_size, tsc, rflags, +//ust// largest_align)) +//ust// return -ENOSPC; +//ust// } while (local_cmpxchg(&buf->offset, offsets.old, +//ust// offsets.end) != offsets.old); +//ust// +//ust// /* +//ust// * Atomically update last_tsc. This update races against concurrent +//ust// * atomic updates, but the race will always cause supplementary full TSC +//ust// * events, never the opposite (missing a full TSC event when it would be +//ust// * needed). +//ust// */ +//ust// save_last_tsc(buf, *tsc); +//ust// +//ust// /* +//ust// * Push the reader if necessary +//ust// */ +//ust// ltt_reserve_push_reader(channel, buf, &offsets); +//ust// +//ust// /* +//ust// * Switch old subbuffer if needed. +//ust// */ +//ust// if (offsets.end_switch_old) +//ust// ltt_reserve_switch_old_subbuf(channel, buf, &offsets, tsc); +//ust// +//ust// /* +//ust// * Populate new subbuffer. +//ust// */ +//ust// if (offsets.begin_switch) +//ust// ltt_reserve_switch_new_subbuf(channel, buf, &offsets, tsc); +//ust// +//ust// if (offsets.end_switch_current) +//ust// ltt_reserve_end_switch_current(channel, buf, &offsets, tsc); +//ust// +//ust// *slot_size = offsets.size; +//ust// *buf_offset = offsets.begin + offsets.before_hdr_pad; +//ust// return 0; +//ust// } +//ust// +//ust// /* +//ust// * Force a sub-buffer switch for a per-cpu buffer. This operation is +//ust// * completely reentrant : can be called while tracing is active with +//ust// * absolutely no lock held. +//ust// * +//ust// * Note, however, that as a local_cmpxchg is used for some atomic +//ust// * operations, this function must be called from the CPU which owns the buffer +//ust// * for a ACTIVE flush. +//ust// */ +//ust// static notrace void ltt_force_switch(struct ust_buffer *buf, +//ust// enum force_switch_mode mode) +//ust// { +//ust// struct ust_channel *channel = buf->chan; +//ust// struct ltt_reserve_switch_offsets offsets; +//ust// u64 tsc; +//ust// +//ust// offsets.reserve_commit_diff = 0; +//ust// offsets.size = 0; +//ust// +//ust// /* +//ust// * Perform retryable operations. +//ust// */ +//ust// do { +//ust// if (ltt_relay_try_switch(mode, channel, buf, &offsets, &tsc)) +//ust// return; +//ust// } while (local_cmpxchg(&buf->offset, offsets.old, +//ust// offsets.end) != offsets.old); +//ust// +//ust// /* +//ust// * Atomically update last_tsc. This update races against concurrent +//ust// * atomic updates, but the race will always cause supplementary full TSC +//ust// * events, never the opposite (missing a full TSC event when it would be +//ust// * needed). +//ust// */ +//ust// save_last_tsc(buf, tsc); +//ust// +//ust// /* +//ust// * Push the reader if necessary +//ust// */ +//ust// if (mode == FORCE_ACTIVE) +//ust// ltt_reserve_push_reader(channel, buf, &offsets); +//ust// +//ust// /* +//ust// * Switch old subbuffer if needed. +//ust// */ +//ust// if (offsets.end_switch_old) +//ust// ltt_reserve_switch_old_subbuf(channel, buf, &offsets, &tsc); +//ust// +//ust// /* +//ust// * Populate new subbuffer. +//ust// */ +//ust// if (mode == FORCE_ACTIVE) +//ust// ltt_reserve_switch_new_subbuf(channel, buf, &offsets, &tsc); +//ust// } /* - * Returns : - * 0 if ok - * !0 if execution must be aborted. + * ltt_reserve_switch_old_subbuf: switch old subbuffer + * + * Concurrency safe because we are the last and only thread to alter this + * sub-buffer. As long as it is not delivered and read, no other thread can + * alter the offset, alter the reserve_count or call the + * client_buffer_end_callback on this sub-buffer. + * + * The only remaining threads could be the ones with pending commits. They will + * have to do the deliver themselves. Not concurrency safe in overwrite mode. + * We detect corrupted subbuffers with commit and reserve counts. We keep a + * corrupted sub-buffers count and push the readers across these sub-buffers. + * + * Not concurrency safe if a writer is stalled in a subbuffer and another writer + * switches in, finding out it's corrupted. The result will be than the old + * (uncommited) subbuffer will be declared corrupted, and that the new subbuffer + * will be declared corrupted too because of the commit count adjustment. + * + * Note : offset_old should never be 0 here. */ -static inline int ltt_relay_try_reserve( - struct ust_channel *channel, struct ust_buffer *buf, - struct ltt_reserve_switch_offsets *offsets, size_t data_size, - u64 *tsc, unsigned int *rflags, int largest_align) +static void ltt_reserve_switch_old_subbuf( + struct ust_channel *chan, struct ust_buffer *buf, + struct ltt_reserve_switch_offsets *offsets, u64 *tsc) { - offsets->begin = local_read(&buf->offset); - offsets->old = offsets->begin; - offsets->begin_switch = 0; - offsets->end_switch_current = 0; - offsets->end_switch_old = 0; + long oldidx = SUBBUF_INDEX(offsets->old - 1, chan); + long commit_count, padding_size; - *tsc = trace_clock_read64(); - if (last_tsc_overflow(buf, *tsc)) - *rflags = LTT_RFLAG_ID_SIZE_TSC; + padding_size = chan->subbuf_size + - (SUBBUF_OFFSET(offsets->old - 1, chan) + 1); + ltt_buffer_end(buf, *tsc, offsets->old, oldidx); - if (SUBBUF_OFFSET(offsets->begin, buf->chan) == 0) { - offsets->begin_switch = 1; /* For offsets->begin */ - } else { - offsets->size = ust_get_header_size(channel, - offsets->begin, data_size, - &offsets->before_hdr_pad, *rflags); - offsets->size += ltt_align(offsets->begin + offsets->size, - largest_align) - + data_size; - if ((SUBBUF_OFFSET(offsets->begin, buf->chan) + offsets->size) - > buf->chan->subbuf_size) { - offsets->end_switch_old = 1; /* For offsets->old */ - offsets->begin_switch = 1; /* For offsets->begin */ - } - } - if (offsets->begin_switch) { - long subbuf_index; + /* + * Must write slot data before incrementing commit count. + * This compiler barrier is upgraded into a smp_wmb() by the IPI + * sent by get_subbuf() when it does its smp_rmb(). + */ + barrier(); + local_add(padding_size, + &buf->commit_count[oldidx].cc); + commit_count = local_read(&buf->commit_count[oldidx].cc); + ltt_check_deliver(chan, buf, offsets->old - 1, commit_count, oldidx); + ltt_write_commit_counter(buf, oldidx, + offsets->old, commit_count, padding_size); +} - if (offsets->end_switch_old) - offsets->begin = SUBBUF_ALIGN(offsets->begin, - buf->chan); - offsets->begin = offsets->begin + ltt_subbuffer_header_size(); - /* Test new buffer integrity */ - subbuf_index = SUBBUF_INDEX(offsets->begin, buf->chan); - offsets->reserve_commit_diff = - (BUFFER_TRUNC(offsets->begin, buf->chan) - >> channel->n_subbufs_order) - - (local_read(&buf->commit_count[subbuf_index]) - & channel->commit_count_mask); - if (offsets->reserve_commit_diff == 0) { - long consumed; +/* + * ltt_reserve_switch_new_subbuf: Populate new subbuffer. + * + * This code can be executed unordered : writers may already have written to the + * sub-buffer before this code gets executed, caution. The commit makes sure + * that this code is executed before the deliver of this sub-buffer. + */ +static void ltt_reserve_switch_new_subbuf( + struct ust_channel *chan, struct ust_buffer *buf, + struct ltt_reserve_switch_offsets *offsets, u64 *tsc) +{ + long beginidx = SUBBUF_INDEX(offsets->begin, chan); + long commit_count; - consumed = atomic_long_read(&buf->consumed); + ltt_buffer_begin(buf, *tsc, beginidx); - /* Next buffer not corrupted. */ - if (!channel->overwrite && - (SUBBUF_TRUNC(offsets->begin, buf->chan) - - SUBBUF_TRUNC(consumed, buf->chan)) - >= channel->alloc_size) { - - long consumed_idx = SUBBUF_INDEX(consumed, buf->chan); - long commit_count = local_read(&buf->commit_count[consumed_idx]); - if(((commit_count - buf->chan->subbuf_size) & channel->commit_count_mask) - (BUFFER_TRUNC(consumed, buf->chan) >> channel->n_subbufs_order) != 0) { - WARN("Event dropped. Caused by non-committed event."); - } - else { - WARN("Event dropped. Caused by non-consumed buffer."); - } - /* - * We do not overwrite non consumed buffers - * and we are full : event is lost. - */ - local_inc(&buf->events_lost); - return -1; - } else { - /* - * next buffer not corrupted, we are either in - * overwrite mode or the buffer is not full. - * It's safe to write in this new subbuffer. - */ - } - } else { - /* - * Next subbuffer corrupted. Force pushing reader even - * in normal mode. It's safe to write in this new - * subbuffer. - */ - } - offsets->size = ust_get_header_size(channel, - offsets->begin, data_size, - &offsets->before_hdr_pad, *rflags); - offsets->size += ltt_align(offsets->begin + offsets->size, - largest_align) - + data_size; - if ((SUBBUF_OFFSET(offsets->begin, buf->chan) + offsets->size) - > buf->chan->subbuf_size) { - /* - * Event too big for subbuffers, report error, don't - * complete the sub-buffer switch. - */ - local_inc(&buf->events_lost); - return -1; - } else { - /* - * We just made a successful buffer switch and the event - * fits in the new subbuffer. Let's write. - */ - } - } else { - /* - * Event fits in the current buffer and we are not on a switch - * boundary. It's safe to write. - */ - } - offsets->end = offsets->begin + offsets->size; + /* + * Must write slot data before incrementing commit count. + * This compiler barrier is upgraded into a smp_wmb() by the IPI + * sent by get_subbuf() when it does its smp_rmb(). + */ + barrier(); + local_add(ltt_subbuffer_header_size(), + &buf->commit_count[beginidx].cc); + commit_count = local_read(&buf->commit_count[beginidx].cc); + /* Check if the written buffer has to be delivered */ + ltt_check_deliver(chan, buf, offsets->begin, commit_count, beginidx); + ltt_write_commit_counter(buf, beginidx, + offsets->begin, commit_count, ltt_subbuffer_header_size()); +} - if ((SUBBUF_OFFSET(offsets->end, buf->chan)) == 0) { - /* - * The offset_end will fall at the very beginning of the next - * subbuffer. - */ - offsets->end_switch_current = 1; /* For offsets->begin */ - } - return 0; +/* + * ltt_reserve_end_switch_current: finish switching current subbuffer + * + * Concurrency safe because we are the last and only thread to alter this + * sub-buffer. As long as it is not delivered and read, no other thread can + * alter the offset, alter the reserve_count or call the + * client_buffer_end_callback on this sub-buffer. + * + * The only remaining threads could be the ones with pending commits. They will + * have to do the deliver themselves. Not concurrency safe in overwrite mode. + * We detect corrupted subbuffers with commit and reserve counts. We keep a + * corrupted sub-buffers count and push the readers across these sub-buffers. + * + * Not concurrency safe if a writer is stalled in a subbuffer and another writer + * switches in, finding out it's corrupted. The result will be than the old + * (uncommited) subbuffer will be declared corrupted, and that the new subbuffer + * will be declared corrupted too because of the commit count adjustment. + */ +static void ltt_reserve_end_switch_current( + struct ust_channel *chan, + struct ust_buffer *buf, + struct ltt_reserve_switch_offsets *offsets, u64 *tsc) +{ + long endidx = SUBBUF_INDEX(offsets->end - 1, chan); + long commit_count, padding_size; + + padding_size = chan->subbuf_size + - (SUBBUF_OFFSET(offsets->end - 1, chan) + 1); + + ltt_buffer_end(buf, *tsc, offsets->end, endidx); + + /* + * Must write slot data before incrementing commit count. + * This compiler barrier is upgraded into a smp_wmb() by the IPI + * sent by get_subbuf() when it does its smp_rmb(). + */ + barrier(); + local_add(padding_size, + &buf->commit_count[endidx].cc); + commit_count = local_read(&buf->commit_count[endidx].cc); + ltt_check_deliver(chan, buf, + offsets->end - 1, commit_count, endidx); + ltt_write_commit_counter(buf, endidx, + offsets->end, commit_count, padding_size); } /* @@ -985,14 +1442,15 @@ static inline int ltt_relay_try_reserve( * 0 if ok * !0 if execution must be aborted. */ -static inline int ltt_relay_try_switch( +static int ltt_relay_try_switch_slow( enum force_switch_mode mode, - struct ust_channel *channel, + struct ust_channel *chan, struct ust_buffer *buf, struct ltt_reserve_switch_offsets *offsets, u64 *tsc) { long subbuf_index; + long reserve_commit_diff; offsets->begin = local_read(&buf->offset); offsets->old = offsets->begin; @@ -1015,17 +1473,17 @@ static inline int ltt_relay_try_switch( * Test new buffer integrity */ subbuf_index = SUBBUF_INDEX(offsets->begin, buf->chan); - offsets->reserve_commit_diff = + reserve_commit_diff = (BUFFER_TRUNC(offsets->begin, buf->chan) - >> channel->n_subbufs_order) - - (local_read(&buf->commit_count[subbuf_index]) - & channel->commit_count_mask); - if (offsets->reserve_commit_diff == 0) { + >> chan->n_subbufs_order) + - (local_read(&buf->commit_count[subbuf_index].cc_sb) + & chan->commit_count_mask); + if (reserve_commit_diff == 0) { /* Next buffer not corrupted. */ if (mode == FORCE_ACTIVE - && !channel->overwrite + && !chan->overwrite && offsets->begin - atomic_long_read(&buf->consumed) - >= channel->alloc_size) { + >= chan->alloc_size) { /* * We do not overwrite non consumed buffers and we are * full : ignore switch while tracing is active. @@ -1042,224 +1500,31 @@ static inline int ltt_relay_try_switch( return 0; } -static inline void ltt_reserve_push_reader( - struct ust_channel *channel, - struct ust_buffer *buf, - struct ltt_reserve_switch_offsets *offsets) -{ - long consumed_old, consumed_new; - - do { - consumed_old = atomic_long_read(&buf->consumed); - /* - * If buffer is in overwrite mode, push the reader consumed - * count if the write position has reached it and we are not - * at the first iteration (don't push the reader farther than - * the writer). This operation can be done concurrently by many - * writers in the same buffer, the writer being at the farthest - * write position sub-buffer index in the buffer being the one - * which will win this loop. - * If the buffer is not in overwrite mode, pushing the reader - * only happens if a sub-buffer is corrupted. - */ - if ((SUBBUF_TRUNC(offsets->end-1, buf->chan) - - SUBBUF_TRUNC(consumed_old, buf->chan)) - >= channel->alloc_size) - consumed_new = SUBBUF_ALIGN(consumed_old, buf->chan); - else { - consumed_new = consumed_old; - break; - } - } while (atomic_long_cmpxchg(&buf->consumed, consumed_old, - consumed_new) != consumed_old); - - if (consumed_old != consumed_new) { - /* - * Reader pushed : we are the winner of the push, we can - * therefore reequilibrate reserve and commit. Atomic increment - * of the commit count permits other writers to play around - * with this variable before us. We keep track of - * corrupted_subbuffers even in overwrite mode : - * we never want to write over a non completely committed - * sub-buffer : possible causes : the buffer size is too low - * compared to the unordered data input, or there is a writer - * that died between the reserve and the commit. - */ - if (offsets->reserve_commit_diff) { - /* - * We have to alter the sub-buffer commit count. - * We do not deliver the previous subbuffer, given it - * was either corrupted or not consumed (overwrite - * mode). - */ - local_add(offsets->reserve_commit_diff, - &buf->commit_count[ - SUBBUF_INDEX(offsets->begin, - buf->chan)]); - if (!channel->overwrite - || offsets->reserve_commit_diff - != channel->subbuf_size) { - /* - * The reserve commit diff was not subbuf_size : - * it means the subbuffer was partly written to - * and is therefore corrupted. If it is multiple - * of subbuffer size and we are in flight - * recorder mode, we are skipping over a whole - * subbuffer. - */ - local_inc(&buf->corrupted_subbuffers); - } - } - } -} - - -/* - * ltt_reserve_switch_old_subbuf: switch old subbuffer - * - * Concurrency safe because we are the last and only thread to alter this - * sub-buffer. As long as it is not delivered and read, no other thread can - * alter the offset, alter the reserve_count or call the - * client_buffer_end_callback on this sub-buffer. - * - * The only remaining threads could be the ones with pending commits. They will - * have to do the deliver themselves. Not concurrency safe in overwrite mode. - * We detect corrupted subbuffers with commit and reserve counts. We keep a - * corrupted sub-buffers count and push the readers across these sub-buffers. - * - * Not concurrency safe if a writer is stalled in a subbuffer and another writer - * switches in, finding out it's corrupted. The result will be than the old - * (uncommited) subbuffer will be declared corrupted, and that the new subbuffer - * will be declared corrupted too because of the commit count adjustment. - * - * Note : offset_old should never be 0 here. - */ -static inline void ltt_reserve_switch_old_subbuf( - struct ust_channel *channel, - struct ust_buffer *buf, - struct ltt_reserve_switch_offsets *offsets, u64 *tsc) -{ - long oldidx = SUBBUF_INDEX(offsets->old - 1, channel); - - channel->buffer_end(buf, *tsc, offsets->old, oldidx); - /* Must write buffer end before incrementing commit count */ - smp_wmb(); - offsets->commit_count = - local_add_return(channel->subbuf_size - - (SUBBUF_OFFSET(offsets->old - 1, channel) - + 1), - &buf->commit_count[oldidx]); - if ((BUFFER_TRUNC(offsets->old - 1, channel) - >> channel->n_subbufs_order) - - ((offsets->commit_count - channel->subbuf_size) - & channel->commit_count_mask) == 0) - ltt_deliver(buf, oldidx, offsets->commit_count); -} - /* - * ltt_reserve_switch_new_subbuf: Populate new subbuffer. - * - * This code can be executed unordered : writers may already have written to the - * sub-buffer before this code gets executed, caution. The commit makes sure - * that this code is executed before the deliver of this sub-buffer. - */ -static /*inline*/ void ltt_reserve_switch_new_subbuf( - struct ust_channel *channel, - struct ust_buffer *buf, - struct ltt_reserve_switch_offsets *offsets, u64 *tsc) -{ - long beginidx = SUBBUF_INDEX(offsets->begin, channel); - - channel->buffer_begin(buf, *tsc, beginidx); - /* Must write buffer end before incrementing commit count */ - smp_wmb(); - offsets->commit_count = local_add_return(ltt_subbuffer_header_size(), - &buf->commit_count[beginidx]); - /* Check if the written buffer has to be delivered */ - if ((BUFFER_TRUNC(offsets->begin, channel) - >> channel->n_subbufs_order) - - ((offsets->commit_count - channel->subbuf_size) - & channel->commit_count_mask) == 0) - ltt_deliver(buf, beginidx, offsets->commit_count); -} - - -/* - * ltt_reserve_end_switch_current: finish switching current subbuffer - * - * Concurrency safe because we are the last and only thread to alter this - * sub-buffer. As long as it is not delivered and read, no other thread can - * alter the offset, alter the reserve_count or call the - * client_buffer_end_callback on this sub-buffer. - * - * The only remaining threads could be the ones with pending commits. They will - * have to do the deliver themselves. Not concurrency safe in overwrite mode. - * We detect corrupted subbuffers with commit and reserve counts. We keep a - * corrupted sub-buffers count and push the readers across these sub-buffers. - * - * Not concurrency safe if a writer is stalled in a subbuffer and another writer - * switches in, finding out it's corrupted. The result will be than the old - * (uncommited) subbuffer will be declared corrupted, and that the new subbuffer - * will be declared corrupted too because of the commit count adjustment. - */ -static inline void ltt_reserve_end_switch_current( - struct ust_channel *channel, - struct ust_buffer *buf, - struct ltt_reserve_switch_offsets *offsets, u64 *tsc) -{ - long endidx = SUBBUF_INDEX(offsets->end - 1, channel); - - channel->buffer_end(buf, *tsc, offsets->end, endidx); - /* Must write buffer begin before incrementing commit count */ - smp_wmb(); - offsets->commit_count = - local_add_return(channel->subbuf_size - - (SUBBUF_OFFSET(offsets->end - 1, channel) - + 1), - &buf->commit_count[endidx]); - if ((BUFFER_TRUNC(offsets->end - 1, channel) - >> channel->n_subbufs_order) - - ((offsets->commit_count - channel->subbuf_size) - & channel->commit_count_mask) == 0) - ltt_deliver(buf, endidx, offsets->commit_count); -} - -/** - * ltt_relay_reserve_slot - Atomic slot reservation in a LTTng buffer. - * @trace: the trace structure to log to. - * @ltt_channel: channel structure - * @transport_data: data structure specific to ltt relay - * @data_size: size of the variable length data to log. - * @slot_size: pointer to total size of the slot (out) - * @buf_offset : pointer to reserved buffer offset (out) - * @tsc: pointer to the tsc at the slot reservation (out) - * @cpu: cpuid + * Force a sub-buffer switch for a per-cpu buffer. This operation is + * completely reentrant : can be called while tracing is active with + * absolutely no lock held. * - * Return : -ENOSPC if not enough space, else returns 0. - * It will take care of sub-buffer switching. + * Note, however, that as a local_cmpxchg is used for some atomic + * operations, this function must be called from the CPU which owns the buffer + * for a ACTIVE flush. */ -static notrace int ltt_relay_reserve_slot(struct ltt_trace_struct *trace, - struct ust_channel *channel, void **transport_data, - size_t data_size, size_t *slot_size, long *buf_offset, u64 *tsc, - unsigned int *rflags, int largest_align, int cpu) +void ltt_force_switch_lockless_slow(struct ust_buffer *buf, + enum force_switch_mode mode) { - struct ust_buffer *buf = *transport_data = channel->buf[cpu]; + struct ust_channel *chan = buf->chan; struct ltt_reserve_switch_offsets offsets; + u64 tsc; - offsets.reserve_commit_diff = 0; offsets.size = 0; /* * Perform retryable operations. */ - if (ltt_nesting > 4) { - local_inc(&buf->events_lost); - return -EPERM; - } do { - if (ltt_relay_try_reserve(channel, buf, &offsets, data_size, tsc, rflags, - largest_align)) - return -ENOSPC; + if (ltt_relay_try_switch_slow(mode, chan, buf, + &offsets, &tsc)) + return; } while (local_cmpxchg(&buf->offset, offsets.old, offsets.end) != offsets.old); @@ -1269,60 +1534,182 @@ static notrace int ltt_relay_reserve_slot(struct ltt_trace_struct *trace, * events, never the opposite (missing a full TSC event when it would be * needed). */ - save_last_tsc(buf, *tsc); + save_last_tsc(buf, tsc); /* * Push the reader if necessary */ - ltt_reserve_push_reader(channel, buf, &offsets); + if (mode == FORCE_ACTIVE) { + ltt_reserve_push_reader(chan, buf, offsets.end - 1); +//ust// ltt_clear_noref_flag(chan, buf, SUBBUF_INDEX(offsets.end - 1, chan)); + } /* * Switch old subbuffer if needed. */ - if (offsets.end_switch_old) - ltt_reserve_switch_old_subbuf(channel, buf, &offsets, tsc); + if (offsets.end_switch_old) { +//ust// ltt_clear_noref_flag(rchan, buf, SUBBUF_INDEX(offsets.old - 1, rchan)); + ltt_reserve_switch_old_subbuf(chan, buf, &offsets, &tsc); + } /* * Populate new subbuffer. */ - if (offsets.begin_switch) - ltt_reserve_switch_new_subbuf(channel, buf, &offsets, tsc); + if (mode == FORCE_ACTIVE) + ltt_reserve_switch_new_subbuf(chan, buf, &offsets, &tsc); +} - if (offsets.end_switch_current) - ltt_reserve_end_switch_current(channel, buf, &offsets, tsc); +/* + * Returns : + * 0 if ok + * !0 if execution must be aborted. + */ +static int ltt_relay_try_reserve_slow(struct ust_channel *chan, struct ust_buffer *buf, + struct ltt_reserve_switch_offsets *offsets, size_t data_size, + u64 *tsc, unsigned int *rflags, int largest_align) +{ + long reserve_commit_diff; - *slot_size = offsets.size; - *buf_offset = offsets.begin + offsets.before_hdr_pad; + offsets->begin = local_read(&buf->offset); + offsets->old = offsets->begin; + offsets->begin_switch = 0; + offsets->end_switch_current = 0; + offsets->end_switch_old = 0; + + *tsc = trace_clock_read64(); + if (last_tsc_overflow(buf, *tsc)) + *rflags = LTT_RFLAG_ID_SIZE_TSC; + + if (unlikely(SUBBUF_OFFSET(offsets->begin, buf->chan) == 0)) { + offsets->begin_switch = 1; /* For offsets->begin */ + } else { + offsets->size = ust_get_header_size(chan, + offsets->begin, data_size, + &offsets->before_hdr_pad, *rflags); + offsets->size += ltt_align(offsets->begin + offsets->size, + largest_align) + + data_size; + if (unlikely((SUBBUF_OFFSET(offsets->begin, buf->chan) + + offsets->size) > buf->chan->subbuf_size)) { + offsets->end_switch_old = 1; /* For offsets->old */ + offsets->begin_switch = 1; /* For offsets->begin */ + } + } + if (unlikely(offsets->begin_switch)) { + long subbuf_index; + + /* + * We are typically not filling the previous buffer completely. + */ + if (likely(offsets->end_switch_old)) + offsets->begin = SUBBUF_ALIGN(offsets->begin, + buf->chan); + offsets->begin = offsets->begin + ltt_subbuffer_header_size(); + /* Test new buffer integrity */ + subbuf_index = SUBBUF_INDEX(offsets->begin, buf->chan); + reserve_commit_diff = + (BUFFER_TRUNC(offsets->begin, buf->chan) + >> chan->n_subbufs_order) + - (local_read(&buf->commit_count[subbuf_index].cc_sb) + & chan->commit_count_mask); + if (likely(reserve_commit_diff == 0)) { + /* Next buffer not corrupted. */ + if (unlikely(!chan->overwrite && + (SUBBUF_TRUNC(offsets->begin, buf->chan) + - SUBBUF_TRUNC(atomic_long_read( + &buf->consumed), + buf->chan)) + >= chan->alloc_size)) { + /* + * We do not overwrite non consumed buffers + * and we are full : event is lost. + */ + local_inc(&buf->events_lost); + return -1; + } else { + /* + * next buffer not corrupted, we are either in + * overwrite mode or the buffer is not full. + * It's safe to write in this new subbuffer. + */ + } + } else { + /* + * Next subbuffer corrupted. Drop event in normal and + * overwrite mode. Caused by either a writer OOPS or + * too many nested writes over a reserve/commit pair. + */ + local_inc(&buf->events_lost); + return -1; + } + offsets->size = ust_get_header_size(chan, + offsets->begin, data_size, + &offsets->before_hdr_pad, *rflags); + offsets->size += ltt_align(offsets->begin + offsets->size, + largest_align) + + data_size; + if (unlikely((SUBBUF_OFFSET(offsets->begin, buf->chan) + + offsets->size) > buf->chan->subbuf_size)) { + /* + * Event too big for subbuffers, report error, don't + * complete the sub-buffer switch. + */ + local_inc(&buf->events_lost); + return -1; + } else { + /* + * We just made a successful buffer switch and the event + * fits in the new subbuffer. Let's write. + */ + } + } else { + /* + * Event fits in the current buffer and we are not on a switch + * boundary. It's safe to write. + */ + } + offsets->end = offsets->begin + offsets->size; + + if (unlikely((SUBBUF_OFFSET(offsets->end, buf->chan)) == 0)) { + /* + * The offset_end will fall at the very beginning of the next + * subbuffer. + */ + offsets->end_switch_current = 1; /* For offsets->begin */ + } return 0; } -/* - * Force a sub-buffer switch for a per-cpu buffer. This operation is - * completely reentrant : can be called while tracing is active with - * absolutely no lock held. +/** + * ltt_relay_reserve_slot_lockless_slow - Atomic slot reservation in a buffer. + * @trace: the trace structure to log to. + * @ltt_channel: channel structure + * @transport_data: data structure specific to ltt relay + * @data_size: size of the variable length data to log. + * @slot_size: pointer to total size of the slot (out) + * @buf_offset : pointer to reserved buffer offset (out) + * @tsc: pointer to the tsc at the slot reservation (out) + * @cpu: cpuid * - * Note, however, that as a local_cmpxchg is used for some atomic - * operations, this function must be called from the CPU which owns the buffer - * for a ACTIVE flush. + * Return : -ENOSPC if not enough space, else returns 0. + * It will take care of sub-buffer switching. */ -static notrace void ltt_force_switch(struct ust_buffer *buf, - enum force_switch_mode mode) +int ltt_reserve_slot_lockless_slow(struct ust_trace *trace, + struct ust_channel *chan, void **transport_data, + size_t data_size, size_t *slot_size, long *buf_offset, u64 *tsc, + unsigned int *rflags, int largest_align, int cpu) { - struct ust_channel *channel = buf->chan; + struct ust_buffer *buf = chan->buf[cpu]; struct ltt_reserve_switch_offsets offsets; - u64 tsc; - offsets.reserve_commit_diff = 0; offsets.size = 0; - /* - * Perform retryable operations. - */ do { - if (ltt_relay_try_switch(mode, channel, buf, &offsets, &tsc)) - return; - } while (local_cmpxchg(&buf->offset, offsets.old, - offsets.end) != offsets.old); + if (unlikely(ltt_relay_try_reserve_slow(chan, buf, &offsets, + data_size, tsc, rflags, largest_align))) + return -ENOSPC; + } while (unlikely(local_cmpxchg(&buf->offset, offsets.old, + offsets.end) != offsets.old)); /* * Atomically update last_tsc. This update races against concurrent @@ -1330,25 +1717,38 @@ static notrace void ltt_force_switch(struct ust_buffer *buf, * events, never the opposite (missing a full TSC event when it would be * needed). */ - save_last_tsc(buf, tsc); + save_last_tsc(buf, *tsc); /* * Push the reader if necessary */ - if (mode == FORCE_ACTIVE) - ltt_reserve_push_reader(channel, buf, &offsets); + ltt_reserve_push_reader(chan, buf, offsets.end - 1); + + /* + * Clear noref flag for this subbuffer. + */ +//ust// ltt_clear_noref_flag(chan, buf, SUBBUF_INDEX(offsets.end - 1, chan)); /* * Switch old subbuffer if needed. */ - if (offsets.end_switch_old) - ltt_reserve_switch_old_subbuf(channel, buf, &offsets, &tsc); + if (unlikely(offsets.end_switch_old)) { +//ust// ltt_clear_noref_flag(chan, buf, SUBBUF_INDEX(offsets.old - 1, chan)); + ltt_reserve_switch_old_subbuf(chan, buf, &offsets, tsc); + } /* * Populate new subbuffer. */ - if (mode == FORCE_ACTIVE) - ltt_reserve_switch_new_subbuf(channel, buf, &offsets, &tsc); + if (unlikely(offsets.begin_switch)) + ltt_reserve_switch_new_subbuf(chan, buf, &offsets, tsc); + + if (unlikely(offsets.end_switch_current)) + ltt_reserve_end_switch_current(chan, buf, &offsets, tsc); + + *slot_size = offsets.size; + *buf_offset = offsets.begin + offsets.before_hdr_pad; + return 0; } static struct ltt_transport ust_relay_transport = { @@ -1358,83 +1758,9 @@ static struct ltt_transport ust_relay_transport = { .finish_channel = ltt_relay_finish_channel, .remove_channel = ltt_relay_remove_channel, .wakeup_channel = ltt_relay_async_wakeup_chan, -// .commit_slot = ltt_relay_commit_slot, - .reserve_slot = ltt_relay_reserve_slot, }, }; -/* - * for flight recording. must be called after relay_commit. - * This function decrements de subbuffer's lost_size each time the commit count - * reaches back the reserve offset (module subbuffer size). It is useful for - * crash dump. - */ -static /* inline */ void ltt_write_commit_counter(struct ust_buffer *buf, - struct ust_buffer *ltt_buf, - long idx, long buf_offset, long commit_count, size_t data_size) -{ - long offset; - long commit_seq_old; - - offset = buf_offset + data_size; - - /* - * SUBBUF_OFFSET includes commit_count_mask. We can simply - * compare the offsets within the subbuffer without caring about - * buffer full/empty mismatch because offset is never zero here - * (subbuffer header and event headers have non-zero length). - */ - if (unlikely(SUBBUF_OFFSET(offset - commit_count, buf->chan))) - return; - - commit_seq_old = local_read(<t_buf->commit_seq[idx]); - while (commit_seq_old < commit_count) - commit_seq_old = local_cmpxchg(<t_buf->commit_seq[idx], - commit_seq_old, commit_count); -} - -/* - * Atomic unordered slot commit. Increments the commit count in the - * specified sub-buffer, and delivers it if necessary. - * - * Parameters: - * - * @ltt_channel : channel structure - * @transport_data: transport-specific data - * @buf_offset : offset following the event header. - * @data_size : size of the event data. - * @slot_size : size of the reserved slot. - */ -/* FIXME: make this function static inline in the .h! */ -/*static*/ /* inline */ notrace void ltt_commit_slot( - struct ust_channel *channel, - void **transport_data, long buf_offset, - size_t data_size, size_t slot_size) -{ - struct ust_buffer *buf = *transport_data; - long offset_end = buf_offset; - long endidx = SUBBUF_INDEX(offset_end - 1, channel); - long commit_count; - - /* Must write slot data before incrementing commit count */ - smp_wmb(); - commit_count = local_add_return(slot_size, - &buf->commit_count[endidx]); - /* Check if all commits have been done */ - if ((BUFFER_TRUNC(offset_end - 1, channel) - >> channel->n_subbufs_order) - - ((commit_count - channel->subbuf_size) - & channel->commit_count_mask) == 0) - ltt_deliver(buf, endidx, commit_count); - /* - * Update lost_size for each commit. It's needed only for extracting - * ltt buffers from vmcore, after crash. - */ - ltt_write_commit_counter(buf, buf, endidx, - buf_offset, commit_count, data_size); -} - - static char initialized = 0; void __attribute__((constructor)) init_ustrelay_transport(void) @@ -1445,7 +1771,75 @@ void __attribute__((constructor)) init_ustrelay_transport(void) } } -static void __attribute__((destructor)) ltt_relay_exit(void) +static void __attribute__((destructor)) ust_buffers_exit(void) { ltt_transport_unregister(&ust_relay_transport); } + +size_t ltt_write_event_header_slow(struct ust_trace *trace, + struct ust_channel *channel, + struct ust_buffer *buf, long buf_offset, + u16 eID, u32 event_size, + u64 tsc, unsigned int rflags) +{ + struct ltt_event_header header; + u16 small_size; + + switch (rflags) { + case LTT_RFLAG_ID_SIZE_TSC: + header.id_time = 29 << LTT_TSC_BITS; + break; + case LTT_RFLAG_ID_SIZE: + header.id_time = 30 << LTT_TSC_BITS; + break; + case LTT_RFLAG_ID: + header.id_time = 31 << LTT_TSC_BITS; + break; + } + + header.id_time |= (u32)tsc & LTT_TSC_MASK; + ust_buffers_write(buf, buf_offset, &header, sizeof(header)); + buf_offset += sizeof(header); + + switch (rflags) { + case LTT_RFLAG_ID_SIZE_TSC: + small_size = (u16)min_t(u32, event_size, LTT_MAX_SMALL_SIZE); + ust_buffers_write(buf, buf_offset, + &eID, sizeof(u16)); + buf_offset += sizeof(u16); + ust_buffers_write(buf, buf_offset, + &small_size, sizeof(u16)); + buf_offset += sizeof(u16); + if (small_size == LTT_MAX_SMALL_SIZE) { + ust_buffers_write(buf, buf_offset, + &event_size, sizeof(u32)); + buf_offset += sizeof(u32); + } + buf_offset += ltt_align(buf_offset, sizeof(u64)); + ust_buffers_write(buf, buf_offset, + &tsc, sizeof(u64)); + buf_offset += sizeof(u64); + break; + case LTT_RFLAG_ID_SIZE: + small_size = (u16)min_t(u32, event_size, LTT_MAX_SMALL_SIZE); + ust_buffers_write(buf, buf_offset, + &eID, sizeof(u16)); + buf_offset += sizeof(u16); + ust_buffers_write(buf, buf_offset, + &small_size, sizeof(u16)); + buf_offset += sizeof(u16); + if (small_size == LTT_MAX_SMALL_SIZE) { + ust_buffers_write(buf, buf_offset, + &event_size, sizeof(u32)); + buf_offset += sizeof(u32); + } + break; + case LTT_RFLAG_ID: + ust_buffers_write(buf, buf_offset, + &eID, sizeof(u16)); + buf_offset += sizeof(u16); + break; + } + + return buf_offset; +}