[lttng-modules.git] / lib / ringbuffer / frontend_internal.h

#ifndef _LIB_RING_BUFFER_FRONTEND_INTERNAL_H
#define _LIB_RING_BUFFER_FRONTEND_INTERNAL_H

/*
 * linux/ringbuffer/frontend_internal.h
 *
 * Ring Buffer Library Synchronization Header (internal helpers).
 *
 * Copyright (C) 2005-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
 *
 * This library is free software; you can redistribute it and/or
 * modify it under the terms of the GNU Lesser General Public
 * License as published by the Free Software Foundation; only
 * version 2.1 of the License.
 *
 * This library is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
 * Lesser General Public License for more details.
 *
 * You should have received a copy of the GNU Lesser General Public
 * License along with this library; if not, write to the Free Software
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
 *
 * Author:
 *	Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
 *
 * See ring_buffer_frontend.c for more information on wait-free algorithms.
 */

#include <wrapper/ringbuffer/config.h>
#include <wrapper/ringbuffer/backend_types.h>
#include <wrapper/ringbuffer/frontend_types.h>
#include <lib/prio_heap/lttng_prio_heap.h>	/* For per-CPU read-side iterator */

/* 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 (BITS_PER_LONG == 32)
static inline
void save_last_tsc(const struct lib_ring_buffer_config *config,
		   struct 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 lib_ring_buffer_config *config,
		      struct 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 (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 lib_ring_buffer_config *config,
		   struct 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 lib_ring_buffer_config *config,
		      struct lib_ring_buffer *buf, u64 tsc)
{
	if (config->tsc_bits == 0 || config->tsc_bits == 64)
		return 0;

	if (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 lib_ring_buffer_ctx *ctx);

extern
void lib_ring_buffer_switch_slow(struct lib_ring_buffer *buf,
				 enum switch_mode mode);

extern
void lib_ring_buffer_check_deliver_slow(const struct lib_ring_buffer_config *config,
				   struct lib_ring_buffer *buf,
			           struct channel *chan,
			           unsigned long offset,
				   unsigned long commit_count,
			           unsigned long idx,
				   u64 tsc);

extern
void lib_ring_buffer_switch_remote(struct lib_ring_buffer *buf);
extern
void lib_ring_buffer_switch_remote_empty(struct lib_ring_buffer *buf);

/* Buffer write helpers */

static inline
void lib_ring_buffer_reserve_push_reader(struct lib_ring_buffer *buf,
					 struct channel *chan,
					 unsigned long offset)
{
	unsigned 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 (unlikely(subbuf_trunc(offset, chan)
			      - subbuf_trunc(consumed_old, chan)
			     >= chan->backend.buf_size))
			consumed_new = subbuf_align(consumed_old, chan);
		else
			return;
	} while (unlikely(atomic_long_cmpxchg(&buf->consumed, consumed_old,
					      consumed_new) != consumed_old));
}

static inline
int lib_ring_buffer_pending_data(const struct lib_ring_buffer_config *config,
				 struct 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 lib_ring_buffer_config *config,
					    struct lib_ring_buffer *buf,
					    unsigned long idx)
{
	return subbuffer_get_data_size(config, &buf->backend, idx);
}

/*
 * 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 lib_ring_buffer_config *config,
				      struct lib_ring_buffer *buf,
				      struct channel *chan)
{
	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, &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);
}

/*
 * Receive end of subbuffer TSC as parameter. It has been read in the
 * space reservation loop of either reserve or switch, which ensures it
 * progresses monotonically with event records in the buffer. Therefore,
 * it ensures that the end timestamp of a subbuffer is <= begin
 * timestamp of the following subbuffers.
 */
static inline
void lib_ring_buffer_check_deliver(const struct lib_ring_buffer_config *config,
				   struct lib_ring_buffer *buf,
			           struct channel *chan,
			           unsigned long offset,
				   unsigned long commit_count,
			           unsigned long idx,
				   u64 tsc)
{
	unsigned long old_commit_count = commit_count
					 - chan->backend.subbuf_size;

	/* Check if all commits have been done */
	if (unlikely((buf_trunc(offset, chan) >> chan->backend.num_subbuf_order)
		     - (old_commit_count & chan->commit_count_mask) == 0))
		lib_ring_buffer_check_deliver_slow(config, buf, chan, offset,
			commit_count, idx, tsc);
}

/*
 * 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 lib_ring_buffer_config *config,
					  struct lib_ring_buffer *buf,
				          struct channel *chan,
				          unsigned long buf_offset,
				          unsigned long commit_count,
					  struct commit_counters_hot *cc_hot)
{
	unsigned long commit_seq_old;

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

	/*
	 * 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 (unlikely(subbuf_offset(buf_offset - commit_count, chan)))
		return;

	commit_seq_old = v_read(config, &cc_hot->seq);
	if (likely((long) (commit_seq_old - commit_count) < 0))
		v_set(config, &cc_hot->seq, commit_count);
}

extern int lib_ring_buffer_create(struct lib_ring_buffer *buf,
				  struct channel_backend *chanb, int cpu);
extern void lib_ring_buffer_free(struct lib_ring_buffer *buf);

/* Keep track of trap nesting inside ring buffer code */
DECLARE_PER_CPU(unsigned int, lib_ring_buffer_nesting);

#endif /* _LIB_RING_BUFFER_FRONTEND_INTERNAL_H */
Commit	Line	Data
886d51a3 MD	1	#ifndef _LIB_RING_BUFFER_FRONTEND_INTERNAL_H
886d51a3 MD	2	#define _LIB_RING_BUFFER_FRONTEND_INTERNAL_H
f3bc08c5 MD	3
	4	/*
	5	* linux/ringbuffer/frontend_internal.h
	6	*
f3bc08c5 MD	7	* Ring Buffer Library Synchronization Header (internal helpers).
f3bc08c5 MD	8	*
886d51a3 MD	9	* Copyright (C) 2005-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
	10	*
	11	* This library is free software; you can redistribute it and/or
	12	* modify it under the terms of the GNU Lesser General Public
	13	* License as published by the Free Software Foundation; only
	14	* version 2.1 of the License.
	15	*
	16	* This library is distributed in the hope that it will be useful,
	17	* but WITHOUT ANY WARRANTY; without even the implied warranty of
	18	* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
	19	* Lesser General Public License for more details.
	20	*
	21	* You should have received a copy of the GNU Lesser General Public
	22	* License along with this library; if not, write to the Free Software
	23	* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
	24	*
f3bc08c5 MD	25	* Author:
	26	* Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
	27	*
	28	* See ring_buffer_frontend.c for more information on wait-free algorithms.
f3bc08c5 MD	29	*/
f3bc08c5 MD	30
5671a661 MD	31	#include <wrapper/ringbuffer/config.h>
	32	#include <wrapper/ringbuffer/backend_types.h>
	33	#include <wrapper/ringbuffer/frontend_types.h>
	34	#include <lib/prio_heap/lttng_prio_heap.h> /* For per-CPU read-side iterator */
f3bc08c5 MD	35
	36	/* Buffer offset macros */
	37
	38	/* buf_trunc mask selects only the buffer number. */
	39	static inline
	40	unsigned long buf_trunc(unsigned long offset, struct channel *chan)
	41	{
	42	return offset & ~(chan->backend.buf_size - 1);
	43
	44	}
	45
	46	/* Select the buffer number value (counter). */
	47	static inline
	48	unsigned long buf_trunc_val(unsigned long offset, struct channel *chan)
	49	{
	50	return buf_trunc(offset, chan) >> chan->backend.buf_size_order;
	51	}
	52
	53	/* buf_offset mask selects only the offset within the current buffer. */
	54	static inline
	55	unsigned long buf_offset(unsigned long offset, struct channel *chan)
	56	{
	57	return offset & (chan->backend.buf_size - 1);
	58	}
	59
	60	/* subbuf_offset mask selects the offset within the current subbuffer. */
	61	static inline
	62	unsigned long subbuf_offset(unsigned long offset, struct channel *chan)
	63	{
	64	return offset & (chan->backend.subbuf_size - 1);
	65	}
	66
	67	/* subbuf_trunc mask selects the subbuffer number. */
	68	static inline
	69	unsigned long subbuf_trunc(unsigned long offset, struct channel *chan)
	70	{
	71	return offset & ~(chan->backend.subbuf_size - 1);
	72	}
	73
	74	/* subbuf_align aligns the offset to the next subbuffer. */
	75	static inline
	76	unsigned long subbuf_align(unsigned long offset, struct channel *chan)
	77	{
	78	return (offset + chan->backend.subbuf_size)
	79	& ~(chan->backend.subbuf_size - 1);
	80	}
	81
	82	/* subbuf_index returns the index of the current subbuffer within the buffer. */
	83	static inline
	84	unsigned long subbuf_index(unsigned long offset, struct channel *chan)
	85	{
	86	return buf_offset(offset, chan) >> chan->backend.subbuf_size_order;
	87	}
	88
	89	/*
	90	* Last TSC comparison functions. Check if the current TSC overflows tsc_bits
	91	* bits from the last TSC read. When overflows are detected, the full 64-bit
	92	* timestamp counter should be written in the record header. Reads and writes
	93	* last_tsc atomically.
	94	*/
	95
	96	#if (BITS_PER_LONG == 32)
	97	static inline
	98	void save_last_tsc(const struct lib_ring_buffer_config *config,
99	struct lib_ring_buffer *buf, u64 tsc)
100	{
101	if (config->tsc_bits == 0 \|\| config->tsc_bits == 64)
102	return;
103
104	/*
105	* Ensure the compiler performs this update in a single instruction.
106	*/
107	v_set(config, &buf->last_tsc, (unsigned long)(tsc >> config->tsc_bits));
108	}
109
110	static inline
111	int last_tsc_overflow(const struct lib_ring_buffer_config *config,
112	struct lib_ring_buffer *buf, u64 tsc)
113	{
114	unsigned long tsc_shifted;
115
116	if (config->tsc_bits == 0 \|\| config->tsc_bits == 64)
117	return 0;
118
119	tsc_shifted = (unsigned long)(tsc >> config->tsc_bits);
120	if (unlikely(tsc_shifted
121	- (unsigned long)v_read(config, &buf->last_tsc)))
122	return 1;
123	else
124	return 0;
125	}
126	#else
127	static inline
128	void save_last_tsc(const struct lib_ring_buffer_config *config,
129	struct lib_ring_buffer *buf, u64 tsc)
130	{
131	if (config->tsc_bits == 0 \|\| config->tsc_bits == 64)
132	return;
133
134	v_set(config, &buf->last_tsc, (unsigned long)tsc);
135	}
136
137	static inline
138	int last_tsc_overflow(const struct lib_ring_buffer_config *config,
139	struct lib_ring_buffer *buf, u64 tsc)
140	{
141	if (config->tsc_bits == 0 \|\| config->tsc_bits == 64)
142	return 0;
143
144	if (unlikely((tsc - v_read(config, &buf->last_tsc))
145	>> config->tsc_bits))
146	return 1;
147	else
148	return 0;
149	}
150	#endif
151
152	extern
153	int lib_ring_buffer_reserve_slow(struct lib_ring_buffer_ctx *ctx);
154
155	extern
156	void lib_ring_buffer_switch_slow(struct lib_ring_buffer *buf,
157	enum switch_mode mode);
158
aece661f MD	159	extern
	160	void lib_ring_buffer_check_deliver_slow(const struct lib_ring_buffer_config *config,
	161	struct lib_ring_buffer *buf,
	162	struct channel *chan,
	163	unsigned long offset,
	164	unsigned long commit_count,
	165	unsigned long idx,
	166	u64 tsc);
	167
5e391252 MD	168	extern
5e391252 MD	169	void lib_ring_buffer_switch_remote(struct lib_ring_buffer *buf);
e1cdd6e9 MD	170	extern
e1cdd6e9 MD	171	void lib_ring_buffer_switch_remote_empty(struct lib_ring_buffer *buf);
5e391252	172
f3bc08c5 MD	173	/* Buffer write helpers */
	174
	175	static inline
	176	void lib_ring_buffer_reserve_push_reader(struct lib_ring_buffer *buf,
	177	struct channel *chan,
	178	unsigned long offset)
	179	{
	180	unsigned long consumed_old, consumed_new;
	181
	182	do {
	183	consumed_old = atomic_long_read(&buf->consumed);
	184	/*
	185	* If buffer is in overwrite mode, push the reader consumed
	186	* count if the write position has reached it and we are not
	187	* at the first iteration (don't push the reader farther than
	188	* the writer). This operation can be done concurrently by many
	189	* writers in the same buffer, the writer being at the farthest
	190	* write position sub-buffer index in the buffer being the one
	191	* which will win this loop.
	192	*/
	193	if (unlikely(subbuf_trunc(offset, chan)
	194	- subbuf_trunc(consumed_old, chan)
	195	>= chan->backend.buf_size))
	196	consumed_new = subbuf_align(consumed_old, chan);
	197	else
	198	return;
	199	} while (unlikely(atomic_long_cmpxchg(&buf->consumed, consumed_old,
	200	consumed_new) != consumed_old));
	201	}
	202
f3bc08c5 MD	203	static inline
	204	int lib_ring_buffer_pending_data(const struct lib_ring_buffer_config *config,
	205	struct lib_ring_buffer *buf,
	206	struct channel *chan)
	207	{
	208	return !!subbuf_offset(v_read(config, &buf->offset), chan);
	209	}
	210
	211	static inline
	212	unsigned long lib_ring_buffer_get_data_size(const struct lib_ring_buffer_config *config,
	213	struct lib_ring_buffer *buf,
	214	unsigned long idx)
	215	{
	216	return subbuffer_get_data_size(config, &buf->backend, idx);
	217	}
	218
	219	/*
	220	* Check if all space reservation in a buffer have been committed. This helps
	221	* knowing if an execution context is nested (for per-cpu buffers only).
	222	* This is a very specific ftrace use-case, so we keep this as "internal" API.
	223	*/
	224	static inline
	225	int lib_ring_buffer_reserve_committed(const struct lib_ring_buffer_config *config,
	226	struct lib_ring_buffer *buf,
	227	struct channel *chan)
	228	{
	229	unsigned long offset, idx, commit_count;
	230
	231	CHAN_WARN_ON(chan, config->alloc != RING_BUFFER_ALLOC_PER_CPU);
	232	CHAN_WARN_ON(chan, config->sync != RING_BUFFER_SYNC_PER_CPU);
	233
	234	/*
	235	* Read offset and commit count in a loop so they are both read
	236	* atomically wrt interrupts. By deal with interrupt concurrency by
	237	* restarting both reads if the offset has been pushed. Note that given
	238	* we only have to deal with interrupt concurrency here, an interrupt
	239	* modifying the commit count will also modify "offset", so it is safe
	240	* to only check for offset modifications.
	241	*/
	242	do {
	243	offset = v_read(config, &buf->offset);
	244	idx = subbuf_index(offset, chan);
	245	commit_count = v_read(config, &buf->commit_hot[idx].cc);
	246	} while (offset != v_read(config, &buf->offset));
	247
	248	return ((buf_trunc(offset, chan) >> chan->backend.num_subbuf_order)
	249	- (commit_count & chan->commit_count_mask) == 0);
	250	}
	251
635e457c MD	252	/*
	253	* Receive end of subbuffer TSC as parameter. It has been read in the
	254	* space reservation loop of either reserve or switch, which ensures it
	255	* progresses monotonically with event records in the buffer. Therefore,
	256	* it ensures that the end timestamp of a subbuffer is <= begin
	257	* timestamp of the following subbuffers.
	258	*/
f3bc08c5 MD	259	static inline
	260	void lib_ring_buffer_check_deliver(const struct lib_ring_buffer_config *config,
	261	struct lib_ring_buffer *buf,
	262	struct channel *chan,
	263	unsigned long offset,
	264	unsigned long commit_count,
635e457c MD	265	unsigned long idx,
635e457c MD	266	u64 tsc)
f3bc08c5 MD	267	{
	268	unsigned long old_commit_count = commit_count
	269	- chan->backend.subbuf_size;
f3bc08c5 MD	270
	271	/* Check if all commits have been done */
	272	if (unlikely((buf_trunc(offset, chan) >> chan->backend.num_subbuf_order)
aece661f MD	273	- (old_commit_count & chan->commit_count_mask) == 0))
	274	lib_ring_buffer_check_deliver_slow(config, buf, chan, offset,
	275	commit_count, idx, tsc);
f3bc08c5 MD	276	}
	277
	278	/*
	279	* lib_ring_buffer_write_commit_counter
	280	*
	281	* For flight recording. must be called after commit.
	282	* This function increments the subbuffer's commit_seq counter each time the
	283	* commit count reaches back the reserve offset (modulo subbuffer size). It is
	284	* useful for crash dump.
	285	*/
	286	static inline
	287	void lib_ring_buffer_write_commit_counter(const struct lib_ring_buffer_config *config,
	288	struct lib_ring_buffer *buf,
	289	struct channel *chan,
f3bc08c5	290	unsigned long buf_offset,
8ec496cf MD	291	unsigned long commit_count,
8ec496cf MD	292	struct commit_counters_hot *cc_hot)
f3bc08c5	293	{
7915e163	294	unsigned long commit_seq_old;
f3bc08c5 MD	295
	296	if (config->oops != RING_BUFFER_OOPS_CONSISTENCY)
	297	return;
	298
f3bc08c5 MD	299	/*
	300	* subbuf_offset includes commit_count_mask. We can simply
	301	* compare the offsets within the subbuffer without caring about
	302	* buffer full/empty mismatch because offset is never zero here
	303	* (subbuffer header and record headers have non-zero length).
	304	*/
7915e163	305	if (unlikely(subbuf_offset(buf_offset - commit_count, chan)))
f3bc08c5 MD	306	return;
f3bc08c5 MD	307
8ec496cf	308	commit_seq_old = v_read(config, &cc_hot->seq);
6212b6b6	309	if (likely((long) (commit_seq_old - commit_count) < 0))
8ec496cf	310	v_set(config, &cc_hot->seq, commit_count);
f3bc08c5 MD	311	}
	312
	313	extern int lib_ring_buffer_create(struct lib_ring_buffer *buf,
	314	struct channel_backend *chanb, int cpu);
	315	extern void lib_ring_buffer_free(struct lib_ring_buffer *buf);
	316
	317	/* Keep track of trap nesting inside ring buffer code */
	318	DECLARE_PER_CPU(unsigned int, lib_ring_buffer_nesting);
	319
886d51a3	320	#endif /* _LIB_RING_BUFFER_FRONTEND_INTERNAL_H */