[lttng-ust.git] / libringbuffer / frontend_internal.h

#ifndef _LINUX_RING_BUFFER_FRONTEND_INTERNAL_H
#define _LINUX_RING_BUFFER_FRONTEND_INTERNAL_H

/*
 * linux/ringbuffer/frontend_internal.h
 *
 * (C) Copyright 2005-2010 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
 *
 * Ring Buffer Library Synchronization Header (internal helpers).
 *
 * Author:
 *	Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
 *
 * See ring_buffer_frontend.c for more information on wait-free algorithms.
 *
 * Dual LGPL v2.1/GPL v2 license.
 */

#include <urcu/compiler.h>

#include <lttng/ringbuffer-config.h>
#include "backend_types.h"
#include "frontend_types.h"
#include "shm.h"

/* Buffer offset macros */

/* buf_trunc mask selects only the buffer number. */
static inline
unsigned long buf_trunc(unsigned long offset, struct channel *chan)
{
	return offset & ~(chan->backend.buf_size - 1);

}

/* Select the buffer number value (counter). */
static inline
unsigned long buf_trunc_val(unsigned long offset, struct channel *chan)
{
	return buf_trunc(offset, chan) >> chan->backend.buf_size_order;
}

/* buf_offset mask selects only the offset within the current buffer. */
static inline
unsigned long buf_offset(unsigned long offset, struct channel *chan)
{
	return offset & (chan->backend.buf_size - 1);
}

/* subbuf_offset mask selects the offset within the current subbuffer. */
static inline
unsigned long subbuf_offset(unsigned long offset, struct channel *chan)
{
	return offset & (chan->backend.subbuf_size - 1);
}

/* subbuf_trunc mask selects the subbuffer number. */
static inline
unsigned long subbuf_trunc(unsigned long offset, struct channel *chan)
{
	return offset & ~(chan->backend.subbuf_size - 1);
}

/* subbuf_align aligns the offset to the next subbuffer. */
static inline
unsigned long subbuf_align(unsigned long offset, struct channel *chan)
{
	return (offset + chan->backend.subbuf_size)
	       & ~(chan->backend.subbuf_size - 1);
}

/* subbuf_index returns the index of the current subbuffer within the buffer. */
static inline
unsigned long subbuf_index(unsigned long offset, struct channel *chan)
{
	return buf_offset(offset, chan) >> chan->backend.subbuf_size_order;
}

/*
 * Last TSC comparison functions. Check if the current TSC overflows tsc_bits
 * bits from the last TSC read. When overflows are detected, the full 64-bit
 * timestamp counter should be written in the record header. Reads and writes
 * last_tsc atomically.
 */

#if (CAA_BITS_PER_LONG == 32)
static inline
void save_last_tsc(const struct lttng_ust_lib_ring_buffer_config *config,
		   struct lttng_ust_lib_ring_buffer *buf, u64 tsc)
{
	if (config->tsc_bits == 0 || config->tsc_bits == 64)
		return;

	/*
	 * Ensure the compiler performs this update in a single instruction.
	 */
	v_set(config, &buf->last_tsc, (unsigned long)(tsc >> config->tsc_bits));
}

static inline
int last_tsc_overflow(const struct lttng_ust_lib_ring_buffer_config *config,
		      struct lttng_ust_lib_ring_buffer *buf, u64 tsc)
{
	unsigned long tsc_shifted;

	if (config->tsc_bits == 0 || config->tsc_bits == 64)
		return 0;

	tsc_shifted = (unsigned long)(tsc >> config->tsc_bits);
	if (caa_unlikely(tsc_shifted
		     - (unsigned long)v_read(config, &buf->last_tsc)))
		return 1;
	else
		return 0;
}
#else
static inline
void save_last_tsc(const struct lttng_ust_lib_ring_buffer_config *config,
		   struct lttng_ust_lib_ring_buffer *buf, u64 tsc)
{
	if (config->tsc_bits == 0 || config->tsc_bits == 64)
		return;

	v_set(config, &buf->last_tsc, (unsigned long)tsc);
}

static inline
int last_tsc_overflow(const struct lttng_ust_lib_ring_buffer_config *config,
		      struct lttng_ust_lib_ring_buffer *buf, u64 tsc)
{
	if (config->tsc_bits == 0 || config->tsc_bits == 64)
		return 0;

	if (caa_unlikely((tsc - v_read(config, &buf->last_tsc))
		     >> config->tsc_bits))
		return 1;
	else
		return 0;
}
#endif

extern
int lib_ring_buffer_reserve_slow(struct lttng_ust_lib_ring_buffer_ctx *ctx);

extern
void lib_ring_buffer_switch_slow(struct lttng_ust_lib_ring_buffer *buf,
				 enum switch_mode mode,
				 struct lttng_ust_shm_handle *handle);

/* Buffer write helpers */

static inline
void lib_ring_buffer_reserve_push_reader(struct lttng_ust_lib_ring_buffer *buf,
					 struct channel *chan,
					 unsigned long offset)
{
	unsigned long consumed_old, consumed_new;

	do {
		consumed_old = uatomic_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 (caa_unlikely(subbuf_trunc(offset, chan)
			      - subbuf_trunc(consumed_old, chan)
			     >= chan->backend.buf_size))
			consumed_new = subbuf_align(consumed_old, chan);
		else
			return;
	} while (caa_unlikely(uatomic_cmpxchg(&buf->consumed, consumed_old,
					      consumed_new) != consumed_old));
}

static inline
void lib_ring_buffer_vmcore_check_deliver(const struct lttng_ust_lib_ring_buffer_config *config,
					  struct lttng_ust_lib_ring_buffer *buf,
				          unsigned long commit_count,
				          unsigned long idx,
					  struct lttng_ust_shm_handle *handle)
{
	if (config->oops == RING_BUFFER_OOPS_CONSISTENCY)
		v_set(config, &shmp_index(handle, buf->commit_hot, idx)->seq, commit_count);
}

static inline
int lib_ring_buffer_poll_deliver(const struct lttng_ust_lib_ring_buffer_config *config,
				 struct lttng_ust_lib_ring_buffer *buf,
			         struct channel *chan,
				 struct lttng_ust_shm_handle *handle)
{
	unsigned long consumed_old, consumed_idx, commit_count, write_offset;

	consumed_old = uatomic_read(&buf->consumed);
	consumed_idx = subbuf_index(consumed_old, chan);
	commit_count = v_read(config, &shmp_index(handle, buf->commit_cold, consumed_idx)->cc_sb);
	/*
	 * No memory barrier here, since we are only interested
	 * in a statistically correct polling result. The next poll will
	 * get the data is we are racing. The mb() that ensures correct
	 * memory order is in get_subbuf.
	 */
	write_offset = v_read(config, &buf->offset);

	/*
	 * Check that the subbuffer we are trying to consume has been
	 * already fully committed.
	 */

	if (((commit_count - chan->backend.subbuf_size)
	     & chan->commit_count_mask)
	    - (buf_trunc(consumed_old, chan)
	       >> chan->backend.num_subbuf_order)
	    != 0)
		return 0;

	/*
	 * Check that we are not about to read the same subbuffer in
	 * which the writer head is.
	 */
	if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_old, chan)
	    == 0)
		return 0;

	return 1;

}

static inline
int lib_ring_buffer_pending_data(const struct lttng_ust_lib_ring_buffer_config *config,
				 struct lttng_ust_lib_ring_buffer *buf,
				 struct channel *chan)
{
	return !!subbuf_offset(v_read(config, &buf->offset), chan);
}

static inline
unsigned long lib_ring_buffer_get_data_size(const struct lttng_ust_lib_ring_buffer_config *config,
					    struct lttng_ust_lib_ring_buffer *buf,
					    unsigned long idx,
					    struct lttng_ust_shm_handle *handle)
{
	return subbuffer_get_data_size(config, &buf->backend, idx, handle);
}

/*
 * Check if all space reservation in a buffer have been committed. This helps
 * knowing if an execution context is nested (for per-cpu buffers only).
 * This is a very specific ftrace use-case, so we keep this as "internal" API.
 */
static inline
int lib_ring_buffer_reserve_committed(const struct lttng_ust_lib_ring_buffer_config *config,
				      struct lttng_ust_lib_ring_buffer *buf,
				      struct channel *chan,
				      struct lttng_ust_shm_handle *handle)
{
	unsigned long offset, idx, commit_count;

	CHAN_WARN_ON(chan, config->alloc != RING_BUFFER_ALLOC_PER_CPU);
	CHAN_WARN_ON(chan, config->sync != RING_BUFFER_SYNC_PER_CPU);

	/*
	 * Read offset and commit count in a loop so they are both read
	 * atomically wrt interrupts. By deal with interrupt concurrency by
	 * restarting both reads if the offset has been pushed. Note that given
	 * we only have to deal with interrupt concurrency here, an interrupt
	 * modifying the commit count will also modify "offset", so it is safe
	 * to only check for offset modifications.
	 */
	do {
		offset = v_read(config, &buf->offset);
		idx = subbuf_index(offset, chan);
		commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, idx)->cc);
	} while (offset != v_read(config, &buf->offset));

	return ((buf_trunc(offset, chan) >> chan->backend.num_subbuf_order)
		     - (commit_count & chan->commit_count_mask) == 0);
}

static inline
void lib_ring_buffer_check_deliver(const struct lttng_ust_lib_ring_buffer_config *config,
				   struct lttng_ust_lib_ring_buffer *buf,
			           struct channel *chan,
			           unsigned long offset,
				   unsigned long commit_count,
			           unsigned long idx,
				   struct lttng_ust_shm_handle *handle)
{
	unsigned long old_commit_count = commit_count
					 - chan->backend.subbuf_size;
	u64 tsc;

	/* Check if all commits have been done */
	if (caa_unlikely((buf_trunc(offset, chan) >> chan->backend.num_subbuf_order)
		     - (old_commit_count & chan->commit_count_mask) == 0)) {
		/*
		 * If we succeeded at updating cc_sb below, we are the subbuffer
		 * writer delivering the subbuffer. Deals with concurrent
		 * updates of the "cc" value without adding a add_return atomic
		 * operation to the fast path.
		 *
		 * We are doing the delivery in two steps:
		 * - First, we cmpxchg() cc_sb to the new value
		 *   old_commit_count + 1. This ensures that we are the only
		 *   subbuffer user successfully filling the subbuffer, but we
		 *   do _not_ set the cc_sb value to "commit_count" yet.
		 *   Therefore, other writers that would wrap around the ring
		 *   buffer and try to start writing to our subbuffer would
		 *   have to drop records, because it would appear as
		 *   non-filled.
		 *   We therefore have exclusive access to the subbuffer control
		 *   structures.  This mutual exclusion with other writers is
		 *   crucially important to perform record overruns count in
		 *   flight recorder mode locklessly.
		 * - When we are ready to release the subbuffer (either for
		 *   reading or for overrun by other writers), we simply set the
		 *   cc_sb value to "commit_count" and perform delivery.
		 *
		 * The subbuffer size is least 2 bytes (minimum size: 1 page).
		 * This guarantees that old_commit_count + 1 != commit_count.
		 */
		if (caa_likely(v_cmpxchg(config, &shmp_index(handle, buf->commit_cold, idx)->cc_sb,
					 old_commit_count, old_commit_count + 1)
			   == old_commit_count)) {
			/*
			 * Start of exclusive subbuffer access. We are
			 * guaranteed to be the last writer in this subbuffer
			 * and any other writer trying to access this subbuffer
			 * in this state is required to drop records.
			 */
			tsc = config->cb.ring_buffer_clock_read(chan);
			v_add(config,
			      subbuffer_get_records_count(config,
							  &buf->backend,
							  idx, handle),
			      &buf->records_count);
			v_add(config,
			      subbuffer_count_records_overrun(config,
							      &buf->backend,
							      idx, handle),
			      &buf->records_overrun);
			config->cb.buffer_end(buf, tsc, idx,
					      lib_ring_buffer_get_data_size(config,
									buf,
									idx,
									handle),
					      handle);

			/*
			 * Set noref flag and offset for this subbuffer id.
			 * Contains a memory barrier that ensures counter stores
			 * are ordered before set noref and offset.
			 */
			lib_ring_buffer_set_noref_offset(config, &buf->backend, idx,
							 buf_trunc_val(offset, chan), handle);

			/*
			 * Order set_noref and record counter updates before the
			 * end of subbuffer exclusive access. Orders with
			 * respect to writers coming into the subbuffer after
			 * wrap around, and also order wrt concurrent readers.
			 */
			cmm_smp_mb();
			/* End of exclusive subbuffer access */
			v_set(config, &shmp_index(handle, buf->commit_cold, idx)->cc_sb,
			      commit_count);
			lib_ring_buffer_vmcore_check_deliver(config, buf,
						 commit_count, idx, handle);

			/*
			 * RING_BUFFER_WAKEUP_BY_WRITER wakeup is not lock-free.
			 */
			if (config->wakeup == RING_BUFFER_WAKEUP_BY_WRITER
			    && (uatomic_read(&buf->active_readers)
				|| uatomic_read(&buf->active_shadow_readers))
			    && lib_ring_buffer_poll_deliver(config, buf, chan, handle)) {
				int wakeup_fd = shm_get_wakeup_fd(handle, &buf->self._ref);

				if (wakeup_fd >= 0) {
					int ret;
					/*
					 * Wake-up the other end by
					 * writing a null byte in the
					 * pipe (non-blocking).
					 * Important note: Because
					 * writing into the pipe is
					 * non-blocking (and therefore
					 * we allow dropping wakeup
					 * data, as long as there is
					 * wakeup data present in the
					 * pipe buffer to wake up the
					 * consumer), the consumer
					 * should perform the following
					 * sequence for waiting:
					 * 1) empty the pipe (reads).
					 * 2) check if there is data in
					 *    the buffer.
					 * 3) wait on the pipe (poll).
					 */
					do {
						ret = write(wakeup_fd, "", 1);
					} while (ret == -1L && errno == EINTR);
				}
			}

		}
	}
}

/*
 * lib_ring_buffer_write_commit_counter
 *
 * For flight recording. must be called after commit.
 * This function increments the subbuffer's commit_seq counter each time the
 * commit count reaches back the reserve offset (modulo subbuffer size). It is
 * useful for crash dump.
 */
static inline
void lib_ring_buffer_write_commit_counter(const struct lttng_ust_lib_ring_buffer_config *config,
					  struct lttng_ust_lib_ring_buffer *buf,
				          struct channel *chan,
				          unsigned long idx,
				          unsigned long buf_offset,
				          unsigned long commit_count,
				          size_t slot_size,
					  struct lttng_ust_shm_handle *handle)
{
	unsigned long offset, commit_seq_old;

	if (config->oops != RING_BUFFER_OOPS_CONSISTENCY)
		return;

	offset = buf_offset + slot_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 record headers have non-zero length).
	 */
	if (caa_unlikely(subbuf_offset(offset - commit_count, chan)))
		return;

	commit_seq_old = v_read(config, &shmp_index(handle, buf->commit_hot, idx)->seq);
	while ((long) (commit_seq_old - commit_count) < 0)
		commit_seq_old = v_cmpxchg(config, &shmp_index(handle, buf->commit_hot, idx)->seq,
					   commit_seq_old, commit_count);
}

extern int lib_ring_buffer_create(struct lttng_ust_lib_ring_buffer *buf,
				  struct channel_backend *chanb, int cpu,
				  struct lttng_ust_shm_handle *handle,
				  struct shm_object *shmobj);
extern void lib_ring_buffer_free(struct lttng_ust_lib_ring_buffer *buf,
				 struct lttng_ust_shm_handle *handle);

/* Keep track of trap nesting inside ring buffer code */
extern __thread unsigned int lib_ring_buffer_nesting;

#endif /* _LINUX_RING_BUFFER_FRONTEND_INTERNAL_H */
Commit	Line	Data
852c2936 MD	1	#ifndef _LINUX_RING_BUFFER_FRONTEND_INTERNAL_H
	2	#define _LINUX_RING_BUFFER_FRONTEND_INTERNAL_H
	3
	4	/*
	5	* linux/ringbuffer/frontend_internal.h
	6	*
	7	* (C) Copyright 2005-2010 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
	8	*
	9	* Ring Buffer Library Synchronization Header (internal helpers).
	10	*
	11	* Author:
	12	* Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
	13	*
	14	* See ring_buffer_frontend.c for more information on wait-free algorithms.
	15	*
	16	* Dual LGPL v2.1/GPL v2 license.
	17	*/
	18
14641deb MD	19	#include <urcu/compiler.h>
14641deb MD	20
4318ae1b	21	#include <lttng/ringbuffer-config.h>
4931a13e MD	22	#include "backend_types.h"
4931a13e MD	23	#include "frontend_types.h"
a6352fd4	24	#include "shm.h"
852c2936 MD	25
	26	/* Buffer offset macros */
	27
	28	/* buf_trunc mask selects only the buffer number. */
	29	static inline
	30	unsigned long buf_trunc(unsigned long offset, struct channel *chan)
	31	{
	32	return offset & ~(chan->backend.buf_size - 1);
	33
	34	}
	35
	36	/* Select the buffer number value (counter). */
	37	static inline
	38	unsigned long buf_trunc_val(unsigned long offset, struct channel *chan)
	39	{
	40	return buf_trunc(offset, chan) >> chan->backend.buf_size_order;
	41	}
	42
	43	/* buf_offset mask selects only the offset within the current buffer. */
	44	static inline
	45	unsigned long buf_offset(unsigned long offset, struct channel *chan)
	46	{
	47	return offset & (chan->backend.buf_size - 1);
	48	}
	49
	50	/* subbuf_offset mask selects the offset within the current subbuffer. */
	51	static inline
	52	unsigned long subbuf_offset(unsigned long offset, struct channel *chan)
	53	{
	54	return offset & (chan->backend.subbuf_size - 1);
	55	}
	56
	57	/* subbuf_trunc mask selects the subbuffer number. */
	58	static inline
	59	unsigned long subbuf_trunc(unsigned long offset, struct channel *chan)
	60	{
	61	return offset & ~(chan->backend.subbuf_size - 1);
	62	}
	63
	64	/* subbuf_align aligns the offset to the next subbuffer. */
	65	static inline
	66	unsigned long subbuf_align(unsigned long offset, struct channel *chan)
	67	{
	68	return (offset + chan->backend.subbuf_size)
	69	& ~(chan->backend.subbuf_size - 1);
	70	}
	71
	72	/* subbuf_index returns the index of the current subbuffer within the buffer. */
	73	static inline
	74	unsigned long subbuf_index(unsigned long offset, struct channel *chan)
	75	{
	76	return buf_offset(offset, chan) >> chan->backend.subbuf_size_order;
	77	}
	78
	79	/*
	80	* Last TSC comparison functions. Check if the current TSC overflows tsc_bits
	81	* bits from the last TSC read. When overflows are detected, the full 64-bit
	82	* timestamp counter should be written in the record header. Reads and writes
	83	* last_tsc atomically.
	84	*/
	85
14641deb	86	#if (CAA_BITS_PER_LONG == 32)
852c2936	87	static inline
4cfec15c MD	88	void save_last_tsc(const struct lttng_ust_lib_ring_buffer_config *config,
4cfec15c MD	89	struct lttng_ust_lib_ring_buffer *buf, u64 tsc)
852c2936 MD	90	{
	91	if (config->tsc_bits == 0 \|\| config->tsc_bits == 64)
	92	return;
	93
	94	/*
	95	* Ensure the compiler performs this update in a single instruction.
	96	*/
	97	v_set(config, &buf->last_tsc, (unsigned long)(tsc >> config->tsc_bits));
	98	}
	99
	100	static inline
4cfec15c MD	101	int last_tsc_overflow(const struct lttng_ust_lib_ring_buffer_config *config,
4cfec15c MD	102	struct lttng_ust_lib_ring_buffer *buf, u64 tsc)
852c2936 MD	103	{
	104	unsigned long tsc_shifted;
	105
	106	if (config->tsc_bits == 0 \|\| config->tsc_bits == 64)
	107	return 0;
	108
	109	tsc_shifted = (unsigned long)(tsc >> config->tsc_bits);
b5a3dfa5	110	if (caa_unlikely(tsc_shifted
852c2936 MD	111	- (unsigned long)v_read(config, &buf->last_tsc)))
	112	return 1;
	113	else
	114	return 0;
	115	}
	116	#else
	117	static inline
4cfec15c MD	118	void save_last_tsc(const struct lttng_ust_lib_ring_buffer_config *config,
4cfec15c MD	119	struct lttng_ust_lib_ring_buffer *buf, u64 tsc)
852c2936 MD	120	{
	121	if (config->tsc_bits == 0 \|\| config->tsc_bits == 64)
	122	return;
	123
	124	v_set(config, &buf->last_tsc, (unsigned long)tsc);
	125	}
	126
	127	static inline
4cfec15c MD	128	int last_tsc_overflow(const struct lttng_ust_lib_ring_buffer_config *config,
4cfec15c MD	129	struct lttng_ust_lib_ring_buffer *buf, u64 tsc)
852c2936 MD	130	{
	131	if (config->tsc_bits == 0 \|\| config->tsc_bits == 64)
	132	return 0;
	133
b5a3dfa5	134	if (caa_unlikely((tsc - v_read(config, &buf->last_tsc))
852c2936 MD	135	>> config->tsc_bits))
	136	return 1;
	137	else
	138	return 0;
	139	}
	140	#endif
	141
	142	extern
4cfec15c	143	int lib_ring_buffer_reserve_slow(struct lttng_ust_lib_ring_buffer_ctx *ctx);
852c2936 MD	144
852c2936 MD	145	extern
4cfec15c	146	void lib_ring_buffer_switch_slow(struct lttng_ust_lib_ring_buffer *buf,
1d498196	147	enum switch_mode mode,
38fae1d3	148	struct lttng_ust_shm_handle *handle);
852c2936 MD	149
	150	/* Buffer write helpers */
	151
	152	static inline
4cfec15c	153	void lib_ring_buffer_reserve_push_reader(struct lttng_ust_lib_ring_buffer *buf,
852c2936 MD	154	struct channel *chan,
	155	unsigned long offset)
	156	{
	157	unsigned long consumed_old, consumed_new;
	158
	159	do {
14641deb	160	consumed_old = uatomic_read(&buf->consumed);
852c2936 MD	161	/*
	162	* If buffer is in overwrite mode, push the reader consumed
	163	* count if the write position has reached it and we are not
	164	* at the first iteration (don't push the reader farther than
	165	* the writer). This operation can be done concurrently by many
	166	* writers in the same buffer, the writer being at the farthest
	167	* write position sub-buffer index in the buffer being the one
	168	* which will win this loop.
	169	*/
b5a3dfa5	170	if (caa_unlikely(subbuf_trunc(offset, chan)
852c2936 MD	171	- subbuf_trunc(consumed_old, chan)
	172	>= chan->backend.buf_size))
	173	consumed_new = subbuf_align(consumed_old, chan);
	174	else
	175	return;
b5a3dfa5	176	} while (caa_unlikely(uatomic_cmpxchg(&buf->consumed, consumed_old,
852c2936 MD	177	consumed_new) != consumed_old));
	178	}
	179
	180	static inline
4cfec15c MD	181	void lib_ring_buffer_vmcore_check_deliver(const struct lttng_ust_lib_ring_buffer_config *config,
4cfec15c MD	182	struct lttng_ust_lib_ring_buffer *buf,
852c2936	183	unsigned long commit_count,
1d498196	184	unsigned long idx,
38fae1d3	185	struct lttng_ust_shm_handle *handle)
852c2936 MD	186	{
852c2936 MD	187	if (config->oops == RING_BUFFER_OOPS_CONSISTENCY)
4746ae29	188	v_set(config, &shmp_index(handle, buf->commit_hot, idx)->seq, commit_count);
852c2936 MD	189	}
	190
	191	static inline
4cfec15c MD	192	int lib_ring_buffer_poll_deliver(const struct lttng_ust_lib_ring_buffer_config *config,
4cfec15c MD	193	struct lttng_ust_lib_ring_buffer *buf,
1d498196	194	struct channel *chan,
38fae1d3	195	struct lttng_ust_shm_handle *handle)
852c2936 MD	196	{
	197	unsigned long consumed_old, consumed_idx, commit_count, write_offset;
	198
14641deb	199	consumed_old = uatomic_read(&buf->consumed);
852c2936	200	consumed_idx = subbuf_index(consumed_old, chan);
4746ae29	201	commit_count = v_read(config, &shmp_index(handle, buf->commit_cold, consumed_idx)->cc_sb);
852c2936 MD	202	/*
	203	* No memory barrier here, since we are only interested
	204	* in a statistically correct polling result. The next poll will
	205	* get the data is we are racing. The mb() that ensures correct
	206	* memory order is in get_subbuf.
	207	*/
	208	write_offset = v_read(config, &buf->offset);
	209
	210	/*
	211	* Check that the subbuffer we are trying to consume has been
	212	* already fully committed.
	213	*/
	214
	215	if (((commit_count - chan->backend.subbuf_size)
	216	& chan->commit_count_mask)
	217	- (buf_trunc(consumed_old, chan)
	218	>> chan->backend.num_subbuf_order)
	219	!= 0)
	220	return 0;
	221
	222	/*
	223	* Check that we are not about to read the same subbuffer in
	224	* which the writer head is.
	225	*/
	226	if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_old, chan)
	227	== 0)
	228	return 0;
	229
	230	return 1;
	231
	232	}
	233
	234	static inline
4cfec15c MD	235	int lib_ring_buffer_pending_data(const struct lttng_ust_lib_ring_buffer_config *config,
4cfec15c MD	236	struct lttng_ust_lib_ring_buffer *buf,
852c2936 MD	237	struct channel *chan)
	238	{
	239	return !!subbuf_offset(v_read(config, &buf->offset), chan);
	240	}
	241
	242	static inline
4cfec15c MD	243	unsigned long lib_ring_buffer_get_data_size(const struct lttng_ust_lib_ring_buffer_config *config,
4cfec15c MD	244	struct lttng_ust_lib_ring_buffer *buf,
1d498196	245	unsigned long idx,
38fae1d3	246	struct lttng_ust_shm_handle *handle)
852c2936	247	{
1d498196	248	return subbuffer_get_data_size(config, &buf->backend, idx, handle);
852c2936 MD	249	}
	250
	251	/*
	252	* Check if all space reservation in a buffer have been committed. This helps
	253	* knowing if an execution context is nested (for per-cpu buffers only).
	254	* This is a very specific ftrace use-case, so we keep this as "internal" API.
	255	*/
	256	static inline
4cfec15c MD	257	int lib_ring_buffer_reserve_committed(const struct lttng_ust_lib_ring_buffer_config *config,
4cfec15c MD	258	struct lttng_ust_lib_ring_buffer *buf,
1d498196	259	struct channel *chan,
38fae1d3	260	struct lttng_ust_shm_handle *handle)
852c2936 MD	261	{
	262	unsigned long offset, idx, commit_count;
	263
	264	CHAN_WARN_ON(chan, config->alloc != RING_BUFFER_ALLOC_PER_CPU);
	265	CHAN_WARN_ON(chan, config->sync != RING_BUFFER_SYNC_PER_CPU);
	266
	267	/*
	268	* Read offset and commit count in a loop so they are both read
	269	* atomically wrt interrupts. By deal with interrupt concurrency by
	270	* restarting both reads if the offset has been pushed. Note that given
	271	* we only have to deal with interrupt concurrency here, an interrupt
	272	* modifying the commit count will also modify "offset", so it is safe
	273	* to only check for offset modifications.
	274	*/
	275	do {
	276	offset = v_read(config, &buf->offset);
	277	idx = subbuf_index(offset, chan);
4746ae29	278	commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, idx)->cc);
852c2936 MD	279	} while (offset != v_read(config, &buf->offset));
	280
	281	return ((buf_trunc(offset, chan) >> chan->backend.num_subbuf_order)
	282	- (commit_count & chan->commit_count_mask) == 0);
	283	}
	284
	285	static inline
4cfec15c MD	286	void lib_ring_buffer_check_deliver(const struct lttng_ust_lib_ring_buffer_config *config,
4cfec15c MD	287	struct lttng_ust_lib_ring_buffer *buf,
852c2936 MD	288	struct channel *chan,
	289	unsigned long offset,
	290	unsigned long commit_count,
1d498196	291	unsigned long idx,
38fae1d3	292	struct lttng_ust_shm_handle *handle)
852c2936 MD	293	{
	294	unsigned long old_commit_count = commit_count
	295	- chan->backend.subbuf_size;
	296	u64 tsc;
	297
	298	/* Check if all commits have been done */
b5a3dfa5	299	if (caa_unlikely((buf_trunc(offset, chan) >> chan->backend.num_subbuf_order)
852c2936 MD	300	- (old_commit_count & chan->commit_count_mask) == 0)) {
	301	/*
	302	* If we succeeded at updating cc_sb below, we are the subbuffer
	303	* writer delivering the subbuffer. Deals with concurrent
	304	* updates of the "cc" value without adding a add_return atomic
	305	* operation to the fast path.
	306	*
	307	* We are doing the delivery in two steps:
	308	* - First, we cmpxchg() cc_sb to the new value
	309	* old_commit_count + 1. This ensures that we are the only
	310	* subbuffer user successfully filling the subbuffer, but we
	311	* do _not_ set the cc_sb value to "commit_count" yet.
	312	* Therefore, other writers that would wrap around the ring
	313	* buffer and try to start writing to our subbuffer would
	314	* have to drop records, because it would appear as
	315	* non-filled.
	316	* We therefore have exclusive access to the subbuffer control
	317	* structures. This mutual exclusion with other writers is
	318	* crucially important to perform record overruns count in
	319	* flight recorder mode locklessly.
	320	* - When we are ready to release the subbuffer (either for
	321	* reading or for overrun by other writers), we simply set the
	322	* cc_sb value to "commit_count" and perform delivery.
	323	*
	324	* The subbuffer size is least 2 bytes (minimum size: 1 page).
	325	* This guarantees that old_commit_count + 1 != commit_count.
	326	*/
b5a3dfa5	327	if (caa_likely(v_cmpxchg(config, &shmp_index(handle, buf->commit_cold, idx)->cc_sb,
852c2936 MD	328	old_commit_count, old_commit_count + 1)
	329	== old_commit_count)) {
	330	/*
	331	* Start of exclusive subbuffer access. We are
	332	* guaranteed to be the last writer in this subbuffer
	333	* and any other writer trying to access this subbuffer
	334	* in this state is required to drop records.
	335	*/
	336	tsc = config->cb.ring_buffer_clock_read(chan);
	337	v_add(config,
	338	subbuffer_get_records_count(config,
1d498196 MD	339	&buf->backend,
1d498196 MD	340	idx, handle),
852c2936 MD	341	&buf->records_count);
	342	v_add(config,
	343	subbuffer_count_records_overrun(config,
	344	&buf->backend,
1d498196	345	idx, handle),
852c2936 MD	346	&buf->records_overrun);
	347	config->cb.buffer_end(buf, tsc, idx,
	348	lib_ring_buffer_get_data_size(config,
	349	buf,
1d498196 MD	350	idx,
	351	handle),
	352	handle);
852c2936 MD	353
	354	/*
	355	* Set noref flag and offset for this subbuffer id.
	356	* Contains a memory barrier that ensures counter stores
	357	* are ordered before set noref and offset.
	358	*/
	359	lib_ring_buffer_set_noref_offset(config, &buf->backend, idx,
1d498196	360	buf_trunc_val(offset, chan), handle);
852c2936 MD	361
	362	/*
	363	* Order set_noref and record counter updates before the
	364	* end of subbuffer exclusive access. Orders with
	365	* respect to writers coming into the subbuffer after
	366	* wrap around, and also order wrt concurrent readers.
	367	*/
14641deb	368	cmm_smp_mb();
852c2936	369	/* End of exclusive subbuffer access */
4746ae29	370	v_set(config, &shmp_index(handle, buf->commit_cold, idx)->cc_sb,
852c2936 MD	371	commit_count);
852c2936 MD	372	lib_ring_buffer_vmcore_check_deliver(config, buf,
1d498196	373	commit_count, idx, handle);
852c2936 MD	374
	375	/*
	376	* RING_BUFFER_WAKEUP_BY_WRITER wakeup is not lock-free.
	377	*/
	378	if (config->wakeup == RING_BUFFER_WAKEUP_BY_WRITER
824f40b8 MD	379	&& (uatomic_read(&buf->active_readers)
824f40b8 MD	380	\|\| uatomic_read(&buf->active_shadow_readers))
1d498196	381	&& lib_ring_buffer_poll_deliver(config, buf, chan, handle)) {
5d61a504 MD	382	int wakeup_fd = shm_get_wakeup_fd(handle, &buf->self._ref);
	383
	384	if (wakeup_fd >= 0) {
	385	int ret;
	386	/*
	387	* Wake-up the other end by
	388	* writing a null byte in the
	389	* pipe (non-blocking).
49deb559 MD	390	* Important note: Because
	391	* writing into the pipe is
	392	* non-blocking (and therefore
	393	* we allow dropping wakeup
	394	* data, as long as there is
	395	* wakeup data present in the
	396	* pipe buffer to wake up the
	397	* consumer), the consumer
	398	* should perform the following
	399	* sequence for waiting:
	400	* 1) empty the pipe (reads).
	401	* 2) check if there is data in
	402	* the buffer.
	403	* 3) wait on the pipe (poll).
5d61a504 MD	404	*/
	405	do {
	406	ret = write(wakeup_fd, "", 1);
	407	} while (ret == -1L && errno == EINTR);
	408	}
852c2936 MD	409	}
	410
	411	}
	412	}
	413	}
	414
	415	/*
	416	* lib_ring_buffer_write_commit_counter
	417	*
	418	* For flight recording. must be called after commit.
	419	* This function increments the subbuffer's commit_seq counter each time the
	420	* commit count reaches back the reserve offset (modulo subbuffer size). It is
	421	* useful for crash dump.
	422	*/
	423	static inline
4cfec15c MD	424	void lib_ring_buffer_write_commit_counter(const struct lttng_ust_lib_ring_buffer_config *config,
4cfec15c MD	425	struct lttng_ust_lib_ring_buffer *buf,
852c2936 MD	426	struct channel *chan,
	427	unsigned long idx,
	428	unsigned long buf_offset,
	429	unsigned long commit_count,
1d498196	430	size_t slot_size,
38fae1d3	431	struct lttng_ust_shm_handle *handle)
852c2936 MD	432	{
	433	unsigned long offset, commit_seq_old;
	434
	435	if (config->oops != RING_BUFFER_OOPS_CONSISTENCY)
	436	return;
	437
	438	offset = buf_offset + slot_size;
	439
	440	/*
	441	* subbuf_offset includes commit_count_mask. We can simply
	442	* compare the offsets within the subbuffer without caring about
	443	* buffer full/empty mismatch because offset is never zero here
	444	* (subbuffer header and record headers have non-zero length).
	445	*/
b5a3dfa5	446	if (caa_unlikely(subbuf_offset(offset - commit_count, chan)))
852c2936 MD	447	return;
852c2936 MD	448
4746ae29	449	commit_seq_old = v_read(config, &shmp_index(handle, buf->commit_hot, idx)->seq);
852c2936	450	while ((long) (commit_seq_old - commit_count) < 0)
4746ae29	451	commit_seq_old = v_cmpxchg(config, &shmp_index(handle, buf->commit_hot, idx)->seq,
852c2936 MD	452	commit_seq_old, commit_count);
	453	}
	454
4cfec15c	455	extern int lib_ring_buffer_create(struct lttng_ust_lib_ring_buffer *buf,
a6352fd4	456	struct channel_backend *chanb, int cpu,
38fae1d3	457	struct lttng_ust_shm_handle *handle,
1d498196	458	struct shm_object *shmobj);
4cfec15c	459	extern void lib_ring_buffer_free(struct lttng_ust_lib_ring_buffer *buf,
38fae1d3	460	struct lttng_ust_shm_handle *handle);
852c2936 MD	461
852c2936 MD	462	/* Keep track of trap nesting inside ring buffer code */
14641deb	463	extern __thread unsigned int lib_ring_buffer_nesting;
852c2936 MD	464
852c2936 MD	465	#endif /* _LINUX_RING_BUFFER_FRONTEND_INTERNAL_H */