[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_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
void lib_ring_buffer_vmcore_check_deliver(const struct lib_ring_buffer_config *config,
					  struct lib_ring_buffer *buf,
				          unsigned long commit_count,
				          unsigned long idx)
{
	if (config->oops == RING_BUFFER_OOPS_CONSISTENCY)
		v_set(config, &buf->commit_hot[idx].seq, commit_count);
}

static inline
int lib_ring_buffer_poll_deliver(const struct lib_ring_buffer_config *config,
				 struct lib_ring_buffer *buf,
			         struct channel *chan)
{
	unsigned long consumed_old, consumed_idx, commit_count, write_offset;

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

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

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

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

	return 1;

}

static inline
int lib_ring_buffer_pending_data(const struct lib_ring_buffer_config *config,
				 struct lib_ring_buffer *buf,
				 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)) {
		/*
		 * If we succeeded at updating cc_sb below, we are the subbuffer
		 * writer delivering the subbuffer. Deals with concurrent
		 * updates of the "cc" value without adding a add_return atomic
		 * operation to the fast path.
		 *
		 * We are doing the delivery in two steps:
		 * - First, we cmpxchg() cc_sb to the new value
		 *   old_commit_count + 1. This ensures that we are the only
		 *   subbuffer user successfully filling the subbuffer, but we
		 *   do _not_ set the cc_sb value to "commit_count" yet.
		 *   Therefore, other writers that would wrap around the ring
		 *   buffer and try to start writing to our subbuffer would
		 *   have to drop records, because it would appear as
		 *   non-filled.
		 *   We therefore have exclusive access to the subbuffer control
		 *   structures.  This mutual exclusion with other writers is
		 *   crucially important to perform record overruns count in
		 *   flight recorder mode locklessly.
		 * - When we are ready to release the subbuffer (either for
		 *   reading or for overrun by other writers), we simply set the
		 *   cc_sb value to "commit_count" and perform delivery.
		 *
		 * The subbuffer size is least 2 bytes (minimum size: 1 page).
		 * This guarantees that old_commit_count + 1 != commit_count.
		 */

		/*
		 * Order prior updates to reserve count prior to the
		 * commit_cold cc_sb update.
		 */
		smp_wmb();
		if (likely(v_cmpxchg(config, &buf->commit_cold[idx].cc_sb,
					 old_commit_count, old_commit_count + 1)
			   == old_commit_count)) {
			/*
			 * Start of exclusive subbuffer access. We are
			 * guaranteed to be the last writer in this subbuffer
			 * and any other writer trying to access this subbuffer
			 * in this state is required to drop records.
			 */
			v_add(config,
			      subbuffer_get_records_count(config,
							  &buf->backend, idx),
			      &buf->records_count);
			v_add(config,
			      subbuffer_count_records_overrun(config,
							      &buf->backend,
							      idx),
			      &buf->records_overrun);
			config->cb.buffer_end(buf, tsc, idx,
					      lib_ring_buffer_get_data_size(config,
									buf,
									idx));

			/*
			 * Increment the packet counter while we have exclusive
			 * access.
			 */
			subbuffer_inc_packet_count(config, &buf->backend, idx);

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

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

			/*
			 * RING_BUFFER_WAKEUP_BY_WRITER wakeup is not lock-free.
			 */
			if (config->wakeup == RING_BUFFER_WAKEUP_BY_WRITER
			    && atomic_long_read(&buf->active_readers)
			    && lib_ring_buffer_poll_deliver(config, buf, chan)) {
				wake_up_interruptible(&buf->read_wait);
				wake_up_interruptible(&chan->read_wait);
			}

		}
	}
}

/*
 * 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 idx,
				          unsigned long buf_offset,
				          unsigned long commit_count)
{
	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, &buf->commit_hot[idx].seq);
	while ((long) (commit_seq_old - commit_count) < 0)
		commit_seq_old = v_cmpxchg(config, &buf->commit_hot[idx].seq,
					   commit_seq_old, 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
5e391252 MD	159	extern
	160	void lib_ring_buffer_switch_remote(struct lib_ring_buffer *buf);
	161
7cf44d03 MD	162	extern
	163	void lib_ring_buffer_switch_remote_empty(struct lib_ring_buffer *buf);
	164
f3bc08c5 MD	165	/* Buffer write helpers */
	166
	167	static inline
	168	void lib_ring_buffer_reserve_push_reader(struct lib_ring_buffer *buf,
	169	struct channel *chan,
	170	unsigned long offset)
	171	{
	172	unsigned long consumed_old, consumed_new;
	173
	174	do {
	175	consumed_old = atomic_long_read(&buf->consumed);
	176	/*
	177	* If buffer is in overwrite mode, push the reader consumed
	178	* count if the write position has reached it and we are not
	179	* at the first iteration (don't push the reader farther than
	180	* the writer). This operation can be done concurrently by many
	181	* writers in the same buffer, the writer being at the farthest
	182	* write position sub-buffer index in the buffer being the one
	183	* which will win this loop.
	184	*/
	185	if (unlikely(subbuf_trunc(offset, chan)
	186	- subbuf_trunc(consumed_old, chan)
	187	>= chan->backend.buf_size))
	188	consumed_new = subbuf_align(consumed_old, chan);
	189	else
	190	return;
	191	} while (unlikely(atomic_long_cmpxchg(&buf->consumed, consumed_old,
	192	consumed_new) != consumed_old));
	193	}
	194
	195	static inline
	196	void lib_ring_buffer_vmcore_check_deliver(const struct lib_ring_buffer_config *config,
	197	struct lib_ring_buffer *buf,
	198	unsigned long commit_count,
	199	unsigned long idx)
	200	{
	201	if (config->oops == RING_BUFFER_OOPS_CONSISTENCY)
	202	v_set(config, &buf->commit_hot[idx].seq, commit_count);
	203	}
	204
	205	static inline
	206	int lib_ring_buffer_poll_deliver(const struct lib_ring_buffer_config *config,
	207	struct lib_ring_buffer *buf,
	208	struct channel *chan)
	209	{
	210	unsigned long consumed_old, consumed_idx, commit_count, write_offset;
	211
	212	consumed_old = atomic_long_read(&buf->consumed);
	213	consumed_idx = subbuf_index(consumed_old, chan);
	214	commit_count = v_read(config, &buf->commit_cold[consumed_idx].cc_sb);
	215	/*
	216	* No memory barrier here, since we are only interested
	217	* in a statistically correct polling result. The next poll will
	218	* get the data is we are racing. The mb() that ensures correct
	219	* memory order is in get_subbuf.
	220	*/
	221	write_offset = v_read(config, &buf->offset);
	222
	223	/*
	224	* Check that the subbuffer we are trying to consume has been
	225	* already fully committed.
	226	*/
	227
	228	if (((commit_count - chan->backend.subbuf_size)
229	& chan->commit_count_mask)
230	- (buf_trunc(consumed_old, chan)
231	>> chan->backend.num_subbuf_order)
232	!= 0)
233	return 0;
234
235	/*
236	* Check that we are not about to read the same subbuffer in
237	* which the writer head is.
238	*/
239	if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_old, chan)
240	== 0)
241	return 0;
242
243	return 1;
244
245	}
246
247	static inline
248	int lib_ring_buffer_pending_data(const struct lib_ring_buffer_config *config,
249	struct lib_ring_buffer *buf,
250	struct channel *chan)
251	{
252	return !!subbuf_offset(v_read(config, &buf->offset), chan);
253	}
254
255	static inline
256	unsigned long lib_ring_buffer_get_data_size(const struct lib_ring_buffer_config *config,
257	struct lib_ring_buffer *buf,
258	unsigned long idx)
259	{
260	return subbuffer_get_data_size(config, &buf->backend, idx);
261	}
262
263	/*
264	* Check if all space reservation in a buffer have been committed. This helps
265	* knowing if an execution context is nested (for per-cpu buffers only).
266	* This is a very specific ftrace use-case, so we keep this as "internal" API.
267	*/
268	static inline
269	int lib_ring_buffer_reserve_committed(const struct lib_ring_buffer_config *config,
270	struct lib_ring_buffer *buf,
271	struct channel *chan)
272	{
273	unsigned long offset, idx, commit_count;
274
275	CHAN_WARN_ON(chan, config->alloc != RING_BUFFER_ALLOC_PER_CPU);
276	CHAN_WARN_ON(chan, config->sync != RING_BUFFER_SYNC_PER_CPU);
277
278	/*
279	* Read offset and commit count in a loop so they are both read
280	* atomically wrt interrupts. By deal with interrupt concurrency by
281	* restarting both reads if the offset has been pushed. Note that given
282	* we only have to deal with interrupt concurrency here, an interrupt
283	* modifying the commit count will also modify "offset", so it is safe
284	* to only check for offset modifications.
285	*/
286	do {
287	offset = v_read(config, &buf->offset);
288	idx = subbuf_index(offset, chan);
289	commit_count = v_read(config, &buf->commit_hot[idx].cc);
290	} while (offset != v_read(config, &buf->offset));
291
292	return ((buf_trunc(offset, chan) >> chan->backend.num_subbuf_order)
293	- (commit_count & chan->commit_count_mask) == 0);
294	}
295
635e457c MD	296	/*
	297	* Receive end of subbuffer TSC as parameter. It has been read in the
	298	* space reservation loop of either reserve or switch, which ensures it
	299	* progresses monotonically with event records in the buffer. Therefore,
	300	* it ensures that the end timestamp of a subbuffer is <= begin
	301	* timestamp of the following subbuffers.
	302	*/
f3bc08c5 MD	303	static inline
	304	void lib_ring_buffer_check_deliver(const struct lib_ring_buffer_config *config,
	305	struct lib_ring_buffer *buf,
	306	struct channel *chan,
	307	unsigned long offset,
	308	unsigned long commit_count,
635e457c MD	309	unsigned long idx,
635e457c MD	310	u64 tsc)
f3bc08c5 MD	311	{
	312	unsigned long old_commit_count = commit_count
	313	- chan->backend.subbuf_size;
f3bc08c5 MD	314
	315	/* Check if all commits have been done */
	316	if (unlikely((buf_trunc(offset, chan) >> chan->backend.num_subbuf_order)
	317	- (old_commit_count & chan->commit_count_mask) == 0)) {
	318	/*
	319	* If we succeeded at updating cc_sb below, we are the subbuffer
	320	* writer delivering the subbuffer. Deals with concurrent
	321	* updates of the "cc" value without adding a add_return atomic
	322	* operation to the fast path.
	323	*
	324	* We are doing the delivery in two steps:
	325	* - First, we cmpxchg() cc_sb to the new value
	326	* old_commit_count + 1. This ensures that we are the only
	327	* subbuffer user successfully filling the subbuffer, but we
	328	* do _not_ set the cc_sb value to "commit_count" yet.
	329	* Therefore, other writers that would wrap around the ring
	330	* buffer and try to start writing to our subbuffer would
	331	* have to drop records, because it would appear as
	332	* non-filled.
	333	* We therefore have exclusive access to the subbuffer control
	334	* structures. This mutual exclusion with other writers is
	335	* crucially important to perform record overruns count in
	336	* flight recorder mode locklessly.
	337	* - When we are ready to release the subbuffer (either for
	338	* reading or for overrun by other writers), we simply set the
	339	* cc_sb value to "commit_count" and perform delivery.
	340	*
	341	* The subbuffer size is least 2 bytes (minimum size: 1 page).
	342	* This guarantees that old_commit_count + 1 != commit_count.
	343	*/
0fdec686 MD	344
	345	/*
	346	* Order prior updates to reserve count prior to the
	347	* commit_cold cc_sb update.
	348	*/
	349	smp_wmb();
f3bc08c5 MD	350	if (likely(v_cmpxchg(config, &buf->commit_cold[idx].cc_sb,
	351	old_commit_count, old_commit_count + 1)
	352	== old_commit_count)) {
	353	/*
	354	* Start of exclusive subbuffer access. We are
	355	* guaranteed to be the last writer in this subbuffer
	356	* and any other writer trying to access this subbuffer
	357	* in this state is required to drop records.
	358	*/
f3bc08c5 MD	359	v_add(config,
	360	subbuffer_get_records_count(config,
	361	&buf->backend, idx),
	362	&buf->records_count);
	363	v_add(config,
	364	subbuffer_count_records_overrun(config,
	365	&buf->backend,
	366	idx),
	367	&buf->records_overrun);
	368	config->cb.buffer_end(buf, tsc, idx,
	369	lib_ring_buffer_get_data_size(config,
	370	buf,
	371	idx));
	372
5b3cf4f9 JD	373	/*
	374	* Increment the packet counter while we have exclusive
	375	* access.
	376	*/
	377	subbuffer_inc_packet_count(config, &buf->backend, idx);
	378
f3bc08c5 MD	379	/*
	380	* Set noref flag and offset for this subbuffer id.
	381	* Contains a memory barrier that ensures counter stores
	382	* are ordered before set noref and offset.
	383	*/
	384	lib_ring_buffer_set_noref_offset(config, &buf->backend, idx,
	385	buf_trunc_val(offset, chan));
	386
	387	/*
	388	* Order set_noref and record counter updates before the
	389	* end of subbuffer exclusive access. Orders with
	390	* respect to writers coming into the subbuffer after
	391	* wrap around, and also order wrt concurrent readers.
	392	*/
	393	smp_mb();
	394	/* End of exclusive subbuffer access */
	395	v_set(config, &buf->commit_cold[idx].cc_sb,
	396	commit_count);
0fdec686 MD	397	/*
	398	* Order later updates to reserve count after
	399	* the commit_cold cc_sb update.
	400	*/
	401	smp_wmb();
f3bc08c5 MD	402	lib_ring_buffer_vmcore_check_deliver(config, buf,
	403	commit_count, idx);
	404
	405	/*
	406	* RING_BUFFER_WAKEUP_BY_WRITER wakeup is not lock-free.
	407	*/
	408	if (config->wakeup == RING_BUFFER_WAKEUP_BY_WRITER
	409	&& atomic_long_read(&buf->active_readers)
	410	&& lib_ring_buffer_poll_deliver(config, buf, chan)) {
	411	wake_up_interruptible(&buf->read_wait);
	412	wake_up_interruptible(&chan->read_wait);
	413	}
	414
	415	}
	416	}
	417	}
	418
	419	/*
	420	* lib_ring_buffer_write_commit_counter
	421	*
	422	* For flight recording. must be called after commit.
	423	* This function increments the subbuffer's commit_seq counter each time the
	424	* commit count reaches back the reserve offset (modulo subbuffer size). It is
	425	* useful for crash dump.
	426	*/
	427	static inline
	428	void lib_ring_buffer_write_commit_counter(const struct lib_ring_buffer_config *config,
	429	struct lib_ring_buffer *buf,
	430	struct channel *chan,
	431	unsigned long idx,
	432	unsigned long buf_offset,
7915e163	433	unsigned long commit_count)
f3bc08c5	434	{
7915e163	435	unsigned long commit_seq_old;
f3bc08c5 MD	436
	437	if (config->oops != RING_BUFFER_OOPS_CONSISTENCY)
	438	return;
	439
f3bc08c5 MD	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	*/
7915e163	446	if (unlikely(subbuf_offset(buf_offset - commit_count, chan)))
f3bc08c5 MD	447	return;
	448
	449	commit_seq_old = v_read(config, &buf->commit_hot[idx].seq);
	450	while ((long) (commit_seq_old - commit_count) < 0)
	451	commit_seq_old = v_cmpxchg(config, &buf->commit_hot[idx].seq,
	452	commit_seq_old, commit_count);
	453	}
	454
	455	extern int lib_ring_buffer_create(struct lib_ring_buffer *buf,
	456	struct channel_backend *chanb, int cpu);
	457	extern void lib_ring_buffer_free(struct lib_ring_buffer *buf);
	458
	459	/* Keep track of trap nesting inside ring buffer code */
	460	DECLARE_PER_CPU(unsigned int, lib_ring_buffer_nesting);
	461
886d51a3	462	#endif /* _LIB_RING_BUFFER_FRONTEND_INTERNAL_H */