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

/* SPDX-License-Identifier: (GPL-2.0-only OR LGPL-2.1-only)
 *
 * lib/ringbuffer/backend_internal.h
 *
 * Ring buffer backend (internal helpers).
 *
 * Copyright (C) 2008-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
 */

#ifndef _LIB_RING_BUFFER_BACKEND_INTERNAL_H
#define _LIB_RING_BUFFER_BACKEND_INTERNAL_H

#include <wrapper/compiler.h>
#include <wrapper/ringbuffer/config.h>
#include <wrapper/ringbuffer/backend_types.h>
#include <wrapper/ringbuffer/frontend_types.h>
#include <linux/string.h>
#include <linux/uaccess.h>

/* Ring buffer backend API presented to the frontend */

/* Ring buffer and channel backend create/free */

int lib_ring_buffer_backend_create(struct lib_ring_buffer_backend *bufb,
				   struct channel_backend *chan, int cpu);
void channel_backend_unregister_notifiers(struct channel_backend *chanb);
void lib_ring_buffer_backend_free(struct lib_ring_buffer_backend *bufb);
int channel_backend_init(struct channel_backend *chanb,
			 const char *name,
			 const struct lib_ring_buffer_config *config,
			 void *priv, size_t subbuf_size,
			 size_t num_subbuf);
void channel_backend_free(struct channel_backend *chanb);

void lib_ring_buffer_backend_reset(struct lib_ring_buffer_backend *bufb);
void channel_backend_reset(struct channel_backend *chanb);

int lib_ring_buffer_backend_init(void);
void lib_ring_buffer_backend_exit(void);

extern void _lib_ring_buffer_write(struct lib_ring_buffer_backend *bufb,
				   size_t offset, const void *src, size_t len,
				   size_t pagecpy);
extern void _lib_ring_buffer_memset(struct lib_ring_buffer_backend *bufb,
				    size_t offset, int c, size_t len,
				    size_t pagecpy);
extern void _lib_ring_buffer_strcpy(struct lib_ring_buffer_backend *bufb,
				   size_t offset, const char *src, size_t len,
				   size_t pagecpy, int pad);
extern void _lib_ring_buffer_copy_from_user_inatomic(struct lib_ring_buffer_backend *bufb,
					    size_t offset, const void *src,
					    size_t len, size_t pagecpy);
extern void _lib_ring_buffer_strcpy_from_user_inatomic(struct lib_ring_buffer_backend *bufb,
		size_t offset, const char __user *src, size_t len,
		size_t pagecpy, int pad);

/*
 * Subbuffer ID bits for overwrite mode. Need to fit within a single word to be
 * exchanged atomically.
 *
 * Top half word, except lowest bit, belongs to "offset", which is used to keep
 * to count the produced buffers.  For overwrite mode, this provides the
 * consumer with the capacity to read subbuffers in order, handling the
 * situation where producers would write up to 2^15 buffers (or 2^31 for 64-bit
 * systems) concurrently with a single execution of get_subbuf (between offset
 * sampling and subbuffer ID exchange).
 */

#define HALF_ULONG_BITS		(BITS_PER_LONG >> 1)

#define SB_ID_OFFSET_SHIFT	(HALF_ULONG_BITS + 1)
#define SB_ID_OFFSET_COUNT	(1UL << SB_ID_OFFSET_SHIFT)
#define SB_ID_OFFSET_MASK	(~(SB_ID_OFFSET_COUNT - 1))
/*
 * Lowest bit of top word half belongs to noref. Used only for overwrite mode.
 */
#define SB_ID_NOREF_SHIFT	(SB_ID_OFFSET_SHIFT - 1)
#define SB_ID_NOREF_COUNT	(1UL << SB_ID_NOREF_SHIFT)
#define SB_ID_NOREF_MASK	SB_ID_NOREF_COUNT
/*
 * In overwrite mode: lowest half of word is used for index.
 * Limit of 2^16 subbuffers per buffer on 32-bit, 2^32 on 64-bit.
 * In producer-consumer mode: whole word used for index.
 */
#define SB_ID_INDEX_SHIFT	0
#define SB_ID_INDEX_COUNT	(1UL << SB_ID_INDEX_SHIFT)
#define SB_ID_INDEX_MASK	(SB_ID_NOREF_COUNT - 1)

/*
 * Construct the subbuffer id from offset, index and noref. Use only the index
 * for producer-consumer mode (offset and noref are only used in overwrite
 * mode).
 */
static inline
unsigned long subbuffer_id(const struct lib_ring_buffer_config *config,
			   unsigned long offset, unsigned long noref,
			   unsigned long index)
{
	if (config->mode == RING_BUFFER_OVERWRITE)
		return (offset << SB_ID_OFFSET_SHIFT)
		       | (noref << SB_ID_NOREF_SHIFT)
		       | index;
	else
		return index;
}

/*
 * Compare offset with the offset contained within id. Return 1 if the offset
 * bits are identical, else 0.
 */
static inline
int subbuffer_id_compare_offset(const struct lib_ring_buffer_config *config,
				unsigned long id, unsigned long offset)
{
	return (id & SB_ID_OFFSET_MASK) == (offset << SB_ID_OFFSET_SHIFT);
}

static inline
unsigned long subbuffer_id_get_index(const struct lib_ring_buffer_config *config,
				     unsigned long id)
{
	if (config->mode == RING_BUFFER_OVERWRITE)
		return id & SB_ID_INDEX_MASK;
	else
		return id;
}

static inline
unsigned long subbuffer_id_is_noref(const struct lib_ring_buffer_config *config,
				    unsigned long id)
{
	if (config->mode == RING_BUFFER_OVERWRITE)
		return !!(id & SB_ID_NOREF_MASK);
	else
		return 1;
}

/*
 * Only used by reader on subbuffer ID it has exclusive access to. No volatile
 * needed.
 */
static inline
void subbuffer_id_set_noref(const struct lib_ring_buffer_config *config,
			    unsigned long *id)
{
	if (config->mode == RING_BUFFER_OVERWRITE)
		*id |= SB_ID_NOREF_MASK;
}

static inline
void subbuffer_id_set_noref_offset(const struct lib_ring_buffer_config *config,
				   unsigned long *id, unsigned long offset)
{
	unsigned long tmp;

	if (config->mode == RING_BUFFER_OVERWRITE) {
		tmp = *id;
		tmp &= ~SB_ID_OFFSET_MASK;
		tmp |= offset << SB_ID_OFFSET_SHIFT;
		tmp |= SB_ID_NOREF_MASK;
		/* Volatile store, read concurrently by readers. */
		WRITE_ONCE(*id, tmp);
	}
}

/* No volatile access, since already used locally */
static inline
void subbuffer_id_clear_noref(const struct lib_ring_buffer_config *config,
			      unsigned long *id)
{
	if (config->mode == RING_BUFFER_OVERWRITE)
		*id &= ~SB_ID_NOREF_MASK;
}

/*
 * For overwrite mode, cap the number of subbuffers per buffer to:
 * 2^16 on 32-bit architectures
 * 2^32 on 64-bit architectures
 * This is required to fit in the index part of the ID. Return 0 on success,
 * -EPERM on failure.
 */
static inline
int subbuffer_id_check_index(const struct lib_ring_buffer_config *config,
			     unsigned long num_subbuf)
{
	if (config->mode == RING_BUFFER_OVERWRITE)
		return (num_subbuf > (1UL << HALF_ULONG_BITS)) ? -EPERM : 0;
	else
		return 0;
}

static inline
void lib_ring_buffer_backend_get_pages(const struct lib_ring_buffer_config *config,
			struct lib_ring_buffer_ctx *ctx,
			struct lib_ring_buffer_backend_pages **backend_pages)
{
	struct lib_ring_buffer_backend *bufb = &ctx->buf->backend;
	struct channel_backend *chanb = &ctx->chan->backend;
	size_t sbidx, offset = ctx->buf_offset;
	unsigned long sb_bindex, id;
	struct lib_ring_buffer_backend_pages *rpages;

	offset &= chanb->buf_size - 1;
	sbidx = offset >> chanb->subbuf_size_order;
	id = bufb->buf_wsb[sbidx].id;
	sb_bindex = subbuffer_id_get_index(config, id);
	rpages = bufb->array[sb_bindex];
	CHAN_WARN_ON(ctx->chan,
		     config->mode == RING_BUFFER_OVERWRITE
		     && subbuffer_id_is_noref(config, id));
	*backend_pages = rpages;
}

/* Get backend pages from cache. */
static inline
struct lib_ring_buffer_backend_pages *
	lib_ring_buffer_get_backend_pages_from_ctx(const struct lib_ring_buffer_config *config,
		struct lib_ring_buffer_ctx *ctx)
{
	return ctx->backend_pages;
}

/*
 * The ring buffer can count events recorded and overwritten per buffer,
 * but it is disabled by default due to its performance overhead.
 */
#ifdef LTTNG_RING_BUFFER_COUNT_EVENTS
static inline
void subbuffer_count_record(const struct lib_ring_buffer_config *config,
			    struct lib_ring_buffer_backend *bufb,
			    unsigned long idx)
{
	unsigned long sb_bindex;

	sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
	v_inc(config, &bufb->array[sb_bindex]->records_commit);
}
#else /* LTTNG_RING_BUFFER_COUNT_EVENTS */
static inline
void subbuffer_count_record(const struct lib_ring_buffer_config *config,
			    struct lib_ring_buffer_backend *bufb,
			    unsigned long idx)
{
}
#endif /* #else LTTNG_RING_BUFFER_COUNT_EVENTS */

/*
 * Reader has exclusive subbuffer access for record consumption. No need to
 * perform the decrement atomically.
 */
static inline
void subbuffer_consume_record(const struct lib_ring_buffer_config *config,
			      struct lib_ring_buffer_backend *bufb)
{
	unsigned long sb_bindex;

	sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id);
	CHAN_WARN_ON(bufb->chan,
		     !v_read(config, &bufb->array[sb_bindex]->records_unread));
	/* Non-atomic decrement protected by exclusive subbuffer access */
	_v_dec(config, &bufb->array[sb_bindex]->records_unread);
	v_inc(config, &bufb->records_read);
}

static inline
unsigned long subbuffer_get_records_count(
				const struct lib_ring_buffer_config *config,
				struct lib_ring_buffer_backend *bufb,
				unsigned long idx)
{
	unsigned long sb_bindex;

	sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
	return v_read(config, &bufb->array[sb_bindex]->records_commit);
}

/*
 * Must be executed at subbuffer delivery when the writer has _exclusive_
 * subbuffer access. See lib_ring_buffer_check_deliver() for details.
 * lib_ring_buffer_get_records_count() must be called to get the records
 * count before this function, because it resets the records_commit
 * count.
 */
static inline
unsigned long subbuffer_count_records_overrun(
				const struct lib_ring_buffer_config *config,
				struct lib_ring_buffer_backend *bufb,
				unsigned long idx)
{
	struct lib_ring_buffer_backend_pages *pages;
	unsigned long overruns, sb_bindex;

	sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
	pages = bufb->array[sb_bindex];
	overruns = v_read(config, &pages->records_unread);
	v_set(config, &pages->records_unread,
	      v_read(config, &pages->records_commit));
	v_set(config, &pages->records_commit, 0);

	return overruns;
}

static inline
void subbuffer_set_data_size(const struct lib_ring_buffer_config *config,
			     struct lib_ring_buffer_backend *bufb,
			     unsigned long idx,
			     unsigned long data_size)
{
	struct lib_ring_buffer_backend_pages *pages;
	unsigned long sb_bindex;

	sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
	pages = bufb->array[sb_bindex];
	pages->data_size = data_size;
}

static inline
unsigned long subbuffer_get_read_data_size(
				const struct lib_ring_buffer_config *config,
				struct lib_ring_buffer_backend *bufb)
{
	struct lib_ring_buffer_backend_pages *pages;
	unsigned long sb_bindex;

	sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id);
	pages = bufb->array[sb_bindex];
	return pages->data_size;
}

static inline
unsigned long subbuffer_get_data_size(
				const struct lib_ring_buffer_config *config,
				struct lib_ring_buffer_backend *bufb,
				unsigned long idx)
{
	struct lib_ring_buffer_backend_pages *pages;
	unsigned long sb_bindex;

	sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
	pages = bufb->array[sb_bindex];
	return pages->data_size;
}

static inline
void subbuffer_inc_packet_count(const struct lib_ring_buffer_config *config,
				struct lib_ring_buffer_backend *bufb,
				unsigned long idx)
{
	bufb->buf_cnt[idx].seq_cnt++;
}

/**
 * lib_ring_buffer_clear_noref - Clear the noref subbuffer flag, called by
 *                               writer.
 */
static inline
void lib_ring_buffer_clear_noref(const struct lib_ring_buffer_config *config,
				 struct lib_ring_buffer_backend *bufb,
				 unsigned long idx)
{
	unsigned long id, new_id;

	if (config->mode != RING_BUFFER_OVERWRITE)
		return;

	/*
	 * Performing a volatile access to read the sb_pages, because we want to
	 * read a coherent version of the pointer and the associated noref flag.
	 */
	id = READ_ONCE(bufb->buf_wsb[idx].id);
	for (;;) {
		/* This check is called on the fast path for each record. */
		if (likely(!subbuffer_id_is_noref(config, id))) {
			/*
			 * Store after load dependency ordering the writes to
			 * the subbuffer after load and test of the noref flag
			 * matches the memory barrier implied by the cmpxchg()
			 * in update_read_sb_index().
			 */
			return;	/* Already writing to this buffer */
		}
		new_id = id;
		subbuffer_id_clear_noref(config, &new_id);
		new_id = cmpxchg(&bufb->buf_wsb[idx].id, id, new_id);
		if (likely(new_id == id))
			break;
		id = new_id;
	}
}

/**
 * lib_ring_buffer_set_noref_offset - Set the noref subbuffer flag and offset,
 *                                    called by writer.
 */
static inline
void lib_ring_buffer_set_noref_offset(const struct lib_ring_buffer_config *config,
				      struct lib_ring_buffer_backend *bufb,
				      unsigned long idx, unsigned long offset)
{
	if (config->mode != RING_BUFFER_OVERWRITE)
		return;

	/*
	 * Because ring_buffer_set_noref() is only called by a single thread
	 * (the one which updated the cc_sb value), there are no concurrent
	 * updates to take care of: other writers have not updated cc_sb, so
	 * they cannot set the noref flag, and concurrent readers cannot modify
	 * the pointer because the noref flag is not set yet.
	 * The smp_wmb() in ring_buffer_commit() takes care of ordering writes
	 * to the subbuffer before this set noref operation.
	 * subbuffer_set_noref() uses a volatile store to deal with concurrent
	 * readers of the noref flag.
	 */
	CHAN_WARN_ON(bufb->chan,
		     subbuffer_id_is_noref(config, bufb->buf_wsb[idx].id));
	/*
	 * Memory barrier that ensures counter stores are ordered before set
	 * noref and offset.
	 */
	smp_mb();
	subbuffer_id_set_noref_offset(config, &bufb->buf_wsb[idx].id, offset);
}

/**
 * update_read_sb_index - Read-side subbuffer index update.
 */
static inline
int update_read_sb_index(const struct lib_ring_buffer_config *config,
			 struct lib_ring_buffer_backend *bufb,
			 struct channel_backend *chanb,
			 unsigned long consumed_idx,
			 unsigned long consumed_count)
{
	unsigned long old_id, new_id;

	if (config->mode == RING_BUFFER_OVERWRITE) {
		/*
		 * Exchange the target writer subbuffer with our own unused
		 * subbuffer. No need to use READ_ONCE() here to read the
		 * old_wpage, because the value read will be confirmed by the
		 * following cmpxchg().
		 */
		old_id = bufb->buf_wsb[consumed_idx].id;
		if (unlikely(!subbuffer_id_is_noref(config, old_id)))
			return -EAGAIN;
		/*
		 * Make sure the offset count we are expecting matches the one
		 * indicated by the writer.
		 */
		if (unlikely(!subbuffer_id_compare_offset(config, old_id,
							  consumed_count)))
			return -EAGAIN;
		CHAN_WARN_ON(bufb->chan,
			     !subbuffer_id_is_noref(config, bufb->buf_rsb.id));
		subbuffer_id_set_noref_offset(config, &bufb->buf_rsb.id,
					      consumed_count);
		new_id = cmpxchg(&bufb->buf_wsb[consumed_idx].id, old_id,
				 bufb->buf_rsb.id);
		if (unlikely(old_id != new_id))
			return -EAGAIN;
		bufb->buf_rsb.id = new_id;
	} else {
		/* No page exchange, use the writer page directly */
		bufb->buf_rsb.id = bufb->buf_wsb[consumed_idx].id;
	}
	return 0;
}

static inline __attribute__((always_inline))
void lttng_inline_memcpy(void *dest, const void *src,
		unsigned long len)
{
	switch (len) {
	case 1:
		*(uint8_t *) dest = *(const uint8_t *) src;
		break;
	case 2:
		*(uint16_t *) dest = *(const uint16_t *) src;
		break;
	case 4:
		*(uint32_t *) dest = *(const uint32_t *) src;
		break;
	case 8:
		*(uint64_t *) dest = *(const uint64_t *) src;
		break;
	default:
		inline_memcpy(dest, src, len);
	}
}

/*
 * Use the architecture-specific memcpy implementation for constant-sized
 * inputs, but rely on an inline memcpy for length statically unknown.
 * The function call to memcpy is just way too expensive for a fast path.
 */
#define lib_ring_buffer_do_copy(config, dest, src, len)		\
do {								\
	size_t __len = (len);					\
	if (__builtin_constant_p(len))				\
		memcpy(dest, src, __len);			\
	else							\
		lttng_inline_memcpy(dest, src, __len);		\
} while (0)

/*
 * We use __copy_from_user_inatomic to copy userspace data since we already
 * did the access_ok for the whole range.
 *
 * Return 0 if OK, nonzero on error.
 */
static inline
unsigned long lib_ring_buffer_do_copy_from_user_inatomic(void *dest,
						const void __user *src,
						unsigned long len)
{
	return __copy_from_user_inatomic(dest, src, len);
}

/*
 * write len bytes to dest with c
 */
static inline
void lib_ring_buffer_do_memset(char *dest, int c,
	unsigned long len)
{
	unsigned long i;

	for (i = 0; i < len; i++)
		dest[i] = c;
}

#endif /* _LIB_RING_BUFFER_BACKEND_INTERNAL_H */
Commit	Line	Data
b7cdc182	1	/* SPDX-License-Identifier: (GPL-2.0-only OR LGPL-2.1-only)
9f36eaed	2	*
886d51a3	3	* lib/ringbuffer/backend_internal.h
f3bc08c5 MD	4	*
	5	* Ring buffer backend (internal helpers).
	6	*
886d51a3	7	* Copyright (C) 2008-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
f3bc08c5 MD	8	*/
f3bc08c5 MD	9
9f36eaed MJ	10	#ifndef _LIB_RING_BUFFER_BACKEND_INTERNAL_H
	11	#define _LIB_RING_BUFFER_BACKEND_INTERNAL_H
	12
a8f2d0c7	13	#include <wrapper/compiler.h>
5671a661 MD	14	#include <wrapper/ringbuffer/config.h>
	15	#include <wrapper/ringbuffer/backend_types.h>
	16	#include <wrapper/ringbuffer/frontend_types.h>
f3bc08c5	17	#include <linux/string.h>
4ea00e4f	18	#include <linux/uaccess.h>
f3bc08c5 MD	19
	20	/* Ring buffer backend API presented to the frontend */
	21
	22	/* Ring buffer and channel backend create/free */
	23
	24	int lib_ring_buffer_backend_create(struct lib_ring_buffer_backend *bufb,
	25	struct channel_backend *chan, int cpu);
	26	void channel_backend_unregister_notifiers(struct channel_backend *chanb);
	27	void lib_ring_buffer_backend_free(struct lib_ring_buffer_backend *bufb);
	28	int channel_backend_init(struct channel_backend *chanb,
	29	const char *name,
	30	const struct lib_ring_buffer_config *config,
	31	void *priv, size_t subbuf_size,
	32	size_t num_subbuf);
	33	void channel_backend_free(struct channel_backend *chanb);
	34
	35	void lib_ring_buffer_backend_reset(struct lib_ring_buffer_backend *bufb);
	36	void channel_backend_reset(struct channel_backend *chanb);
	37
	38	int lib_ring_buffer_backend_init(void);
	39	void lib_ring_buffer_backend_exit(void);
	40
	41	extern void _lib_ring_buffer_write(struct lib_ring_buffer_backend *bufb,
	42	size_t offset, const void *src, size_t len,
bfe529f9	43	size_t pagecpy);
4ea00e4f JD	44	extern void _lib_ring_buffer_memset(struct lib_ring_buffer_backend *bufb,
4ea00e4f JD	45	size_t offset, int c, size_t len,
bfe529f9	46	size_t pagecpy);
16f78f3a MD	47	extern void _lib_ring_buffer_strcpy(struct lib_ring_buffer_backend *bufb,
	48	size_t offset, const char *src, size_t len,
	49	size_t pagecpy, int pad);
7b8ea3a5	50	extern void _lib_ring_buffer_copy_from_user_inatomic(struct lib_ring_buffer_backend *bufb,
4ea00e4f	51	size_t offset, const void *src,
bfe529f9	52	size_t len, size_t pagecpy);
16f78f3a MD	53	extern void _lib_ring_buffer_strcpy_from_user_inatomic(struct lib_ring_buffer_backend *bufb,
	54	size_t offset, const char __user *src, size_t len,
	55	size_t pagecpy, int pad);
f3bc08c5 MD	56
	57	/*
	58	* Subbuffer ID bits for overwrite mode. Need to fit within a single word to be
	59	* exchanged atomically.
	60	*
	61	* Top half word, except lowest bit, belongs to "offset", which is used to keep
	62	* to count the produced buffers. For overwrite mode, this provides the
	63	* consumer with the capacity to read subbuffers in order, handling the
	64	* situation where producers would write up to 2^15 buffers (or 2^31 for 64-bit
	65	* systems) concurrently with a single execution of get_subbuf (between offset
	66	* sampling and subbuffer ID exchange).
	67	*/
	68
	69	#define HALF_ULONG_BITS (BITS_PER_LONG >> 1)
	70
	71	#define SB_ID_OFFSET_SHIFT (HALF_ULONG_BITS + 1)
	72	#define SB_ID_OFFSET_COUNT (1UL << SB_ID_OFFSET_SHIFT)
	73	#define SB_ID_OFFSET_MASK (~(SB_ID_OFFSET_COUNT - 1))
	74	/*
	75	* Lowest bit of top word half belongs to noref. Used only for overwrite mode.
	76	*/
	77	#define SB_ID_NOREF_SHIFT (SB_ID_OFFSET_SHIFT - 1)
	78	#define SB_ID_NOREF_COUNT (1UL << SB_ID_NOREF_SHIFT)
	79	#define SB_ID_NOREF_MASK SB_ID_NOREF_COUNT
	80	/*
	81	* In overwrite mode: lowest half of word is used for index.
	82	* Limit of 2^16 subbuffers per buffer on 32-bit, 2^32 on 64-bit.
	83	* In producer-consumer mode: whole word used for index.
	84	*/
	85	#define SB_ID_INDEX_SHIFT 0
	86	#define SB_ID_INDEX_COUNT (1UL << SB_ID_INDEX_SHIFT)
	87	#define SB_ID_INDEX_MASK (SB_ID_NOREF_COUNT - 1)
	88
	89	/*
	90	* Construct the subbuffer id from offset, index and noref. Use only the index
	91	* for producer-consumer mode (offset and noref are only used in overwrite
	92	* mode).
	93	*/
	94	static inline
	95	unsigned long subbuffer_id(const struct lib_ring_buffer_config *config,
	96	unsigned long offset, unsigned long noref,
	97	unsigned long index)
	98	{
	99	if (config->mode == RING_BUFFER_OVERWRITE)
	100	return (offset << SB_ID_OFFSET_SHIFT)
	101	\| (noref << SB_ID_NOREF_SHIFT)
	102	\| index;
	103	else
	104	return index;
	105	}
	106
	107	/*
	108	* Compare offset with the offset contained within id. Return 1 if the offset
	109	* bits are identical, else 0.
	110	*/
	111	static inline
	112	int subbuffer_id_compare_offset(const struct lib_ring_buffer_config *config,
	113	unsigned long id, unsigned long offset)
	114	{
	115	return (id & SB_ID_OFFSET_MASK) == (offset << SB_ID_OFFSET_SHIFT);
	116	}
	117
	118	static inline
	119	unsigned long subbuffer_id_get_index(const struct lib_ring_buffer_config *config,
120	unsigned long id)
121	{
122	if (config->mode == RING_BUFFER_OVERWRITE)
123	return id & SB_ID_INDEX_MASK;
124	else
125	return id;
126	}
127
128	static inline
129	unsigned long subbuffer_id_is_noref(const struct lib_ring_buffer_config *config,
130	unsigned long id)
131	{
132	if (config->mode == RING_BUFFER_OVERWRITE)
133	return !!(id & SB_ID_NOREF_MASK);
134	else
135	return 1;
136	}
137
138	/*
139	* Only used by reader on subbuffer ID it has exclusive access to. No volatile
140	* needed.
141	*/
142	static inline
143	void subbuffer_id_set_noref(const struct lib_ring_buffer_config *config,
144	unsigned long *id)
145	{
146	if (config->mode == RING_BUFFER_OVERWRITE)
147	*id \|= SB_ID_NOREF_MASK;
148	}
149
150	static inline
151	void subbuffer_id_set_noref_offset(const struct lib_ring_buffer_config *config,
152	unsigned long *id, unsigned long offset)
153	{
154	unsigned long tmp;
155
156	if (config->mode == RING_BUFFER_OVERWRITE) {
157	tmp = *id;
158	tmp &= ~SB_ID_OFFSET_MASK;
159	tmp \|= offset << SB_ID_OFFSET_SHIFT;
160	tmp \|= SB_ID_NOREF_MASK;
161	/* Volatile store, read concurrently by readers. */
a8f2d0c7	162	WRITE_ONCE(*id, tmp);
f3bc08c5 MD	163	}
	164	}
	165
	166	/* No volatile access, since already used locally */
	167	static inline
	168	void subbuffer_id_clear_noref(const struct lib_ring_buffer_config *config,
	169	unsigned long *id)
	170	{
	171	if (config->mode == RING_BUFFER_OVERWRITE)
	172	*id &= ~SB_ID_NOREF_MASK;
	173	}
	174
	175	/*
	176	* For overwrite mode, cap the number of subbuffers per buffer to:
	177	* 2^16 on 32-bit architectures
	178	* 2^32 on 64-bit architectures
	179	* This is required to fit in the index part of the ID. Return 0 on success,
	180	* -EPERM on failure.
	181	*/
	182	static inline
	183	int subbuffer_id_check_index(const struct lib_ring_buffer_config *config,
	184	unsigned long num_subbuf)
	185	{
	186	if (config->mode == RING_BUFFER_OVERWRITE)
	187	return (num_subbuf > (1UL << HALF_ULONG_BITS)) ? -EPERM : 0;
	188	else
	189	return 0;
	190	}
	191
85a07c33 MD	192	static inline
	193	void lib_ring_buffer_backend_get_pages(const struct lib_ring_buffer_config *config,
	194	struct lib_ring_buffer_ctx *ctx,
	195	struct lib_ring_buffer_backend_pages **backend_pages)
	196	{
	197	struct lib_ring_buffer_backend *bufb = &ctx->buf->backend;
	198	struct channel_backend *chanb = &ctx->chan->backend;
	199	size_t sbidx, offset = ctx->buf_offset;
	200	unsigned long sb_bindex, id;
	201	struct lib_ring_buffer_backend_pages *rpages;
	202
	203	offset &= chanb->buf_size - 1;
	204	sbidx = offset >> chanb->subbuf_size_order;
	205	id = bufb->buf_wsb[sbidx].id;
	206	sb_bindex = subbuffer_id_get_index(config, id);
	207	rpages = bufb->array[sb_bindex];
	208	CHAN_WARN_ON(ctx->chan,
	209	config->mode == RING_BUFFER_OVERWRITE
	210	&& subbuffer_id_is_noref(config, id));
	211	*backend_pages = rpages;
	212	}
	213
	214	/* Get backend pages from cache. */
	215	static inline
	216	struct lib_ring_buffer_backend_pages *
	217	lib_ring_buffer_get_backend_pages_from_ctx(const struct lib_ring_buffer_config *config,
	218	struct lib_ring_buffer_ctx *ctx)
	219	{
	220	return ctx->backend_pages;
	221	}
	222
25337cb5 MD	223	/*
	224	* The ring buffer can count events recorded and overwritten per buffer,
	225	* but it is disabled by default due to its performance overhead.
	226	*/
	227	#ifdef LTTNG_RING_BUFFER_COUNT_EVENTS
f3bc08c5 MD	228	static inline
	229	void subbuffer_count_record(const struct lib_ring_buffer_config *config,
	230	struct lib_ring_buffer_backend *bufb,
	231	unsigned long idx)
	232	{
	233	unsigned long sb_bindex;
	234
	235	sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
	236	v_inc(config, &bufb->array[sb_bindex]->records_commit);
	237	}
25337cb5 MD	238	#else /* LTTNG_RING_BUFFER_COUNT_EVENTS */
	239	static inline
	240	void subbuffer_count_record(const struct lib_ring_buffer_config *config,
	241	struct lib_ring_buffer_backend *bufb,
	242	unsigned long idx)
	243	{
	244	}
	245	#endif /* #else LTTNG_RING_BUFFER_COUNT_EVENTS */
f3bc08c5 MD	246
	247	/*
	248	* Reader has exclusive subbuffer access for record consumption. No need to
	249	* perform the decrement atomically.
	250	*/
	251	static inline
	252	void subbuffer_consume_record(const struct lib_ring_buffer_config *config,
	253	struct lib_ring_buffer_backend *bufb)
	254	{
	255	unsigned long sb_bindex;
	256
	257	sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id);
	258	CHAN_WARN_ON(bufb->chan,
	259	!v_read(config, &bufb->array[sb_bindex]->records_unread));
	260	/* Non-atomic decrement protected by exclusive subbuffer access */
	261	_v_dec(config, &bufb->array[sb_bindex]->records_unread);
	262	v_inc(config, &bufb->records_read);
	263	}
	264
	265	static inline
	266	unsigned long subbuffer_get_records_count(
	267	const struct lib_ring_buffer_config *config,
	268	struct lib_ring_buffer_backend *bufb,
	269	unsigned long idx)
	270	{
	271	unsigned long sb_bindex;
	272
	273	sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
	274	return v_read(config, &bufb->array[sb_bindex]->records_commit);
	275	}
	276
	277	/*
	278	* Must be executed at subbuffer delivery when the writer has _exclusive_
c68be968 MD	279	* subbuffer access. See lib_ring_buffer_check_deliver() for details.
	280	* lib_ring_buffer_get_records_count() must be called to get the records
	281	* count before this function, because it resets the records_commit
	282	* count.
f3bc08c5 MD	283	*/
	284	static inline
	285	unsigned long subbuffer_count_records_overrun(
	286	const struct lib_ring_buffer_config *config,
	287	struct lib_ring_buffer_backend *bufb,
	288	unsigned long idx)
	289	{
	290	struct lib_ring_buffer_backend_pages *pages;
	291	unsigned long overruns, sb_bindex;
	292
	293	sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
	294	pages = bufb->array[sb_bindex];
	295	overruns = v_read(config, &pages->records_unread);
	296	v_set(config, &pages->records_unread,
	297	v_read(config, &pages->records_commit));
	298	v_set(config, &pages->records_commit, 0);
	299
	300	return overruns;
	301	}
	302
	303	static inline
	304	void subbuffer_set_data_size(const struct lib_ring_buffer_config *config,
	305	struct lib_ring_buffer_backend *bufb,
	306	unsigned long idx,
	307	unsigned long data_size)
	308	{
	309	struct lib_ring_buffer_backend_pages *pages;
	310	unsigned long sb_bindex;
	311
	312	sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
	313	pages = bufb->array[sb_bindex];
	314	pages->data_size = data_size;
	315	}
	316
	317	static inline
	318	unsigned long subbuffer_get_read_data_size(
	319	const struct lib_ring_buffer_config *config,
	320	struct lib_ring_buffer_backend *bufb)
	321	{
	322	struct lib_ring_buffer_backend_pages *pages;
	323	unsigned long sb_bindex;
	324
	325	sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id);
	326	pages = bufb->array[sb_bindex];
	327	return pages->data_size;
	328	}
	329
	330	static inline
	331	unsigned long subbuffer_get_data_size(
	332	const struct lib_ring_buffer_config *config,
	333	struct lib_ring_buffer_backend *bufb,
	334	unsigned long idx)
	335	{
	336	struct lib_ring_buffer_backend_pages *pages;
	337	unsigned long sb_bindex;
	338
	339	sb_bindex = subbuffer_id_get_index(config, bufb->buf_wsb[idx].id);
	340	pages = bufb->array[sb_bindex];
	341	return pages->data_size;
	342	}
	343
5b3cf4f9 JD	344	static inline
	345	void subbuffer_inc_packet_count(const struct lib_ring_buffer_config *config,
	346	struct lib_ring_buffer_backend *bufb,
	347	unsigned long idx)
	348	{
	349	bufb->buf_cnt[idx].seq_cnt++;
	350	}
	351
f3bc08c5 MD	352	/**
	353	* lib_ring_buffer_clear_noref - Clear the noref subbuffer flag, called by
	354	* writer.
	355	*/
	356	static inline
	357	void lib_ring_buffer_clear_noref(const struct lib_ring_buffer_config *config,
	358	struct lib_ring_buffer_backend *bufb,
	359	unsigned long idx)
	360	{
	361	unsigned long id, new_id;
	362
	363	if (config->mode != RING_BUFFER_OVERWRITE)
	364	return;
	365
	366	/*
	367	* Performing a volatile access to read the sb_pages, because we want to
	368	* read a coherent version of the pointer and the associated noref flag.
	369	*/
a8f2d0c7	370	id = READ_ONCE(bufb->buf_wsb[idx].id);
f3bc08c5 MD	371	for (;;) {
	372	/* This check is called on the fast path for each record. */
	373	if (likely(!subbuffer_id_is_noref(config, id))) {
	374	/*
	375	* Store after load dependency ordering the writes to
	376	* the subbuffer after load and test of the noref flag
	377	* matches the memory barrier implied by the cmpxchg()
	378	* in update_read_sb_index().
	379	*/
	380	return; /* Already writing to this buffer */
	381	}
	382	new_id = id;
	383	subbuffer_id_clear_noref(config, &new_id);
	384	new_id = cmpxchg(&bufb->buf_wsb[idx].id, id, new_id);
	385	if (likely(new_id == id))
	386	break;
	387	id = new_id;
	388	}
	389	}
	390
	391	/**
	392	* lib_ring_buffer_set_noref_offset - Set the noref subbuffer flag and offset,
	393	* called by writer.
	394	*/
	395	static inline
	396	void lib_ring_buffer_set_noref_offset(const struct lib_ring_buffer_config *config,
	397	struct lib_ring_buffer_backend *bufb,
	398	unsigned long idx, unsigned long offset)
	399	{
	400	if (config->mode != RING_BUFFER_OVERWRITE)
	401	return;
	402
	403	/*
	404	* Because ring_buffer_set_noref() is only called by a single thread
	405	* (the one which updated the cc_sb value), there are no concurrent
	406	* updates to take care of: other writers have not updated cc_sb, so
	407	* they cannot set the noref flag, and concurrent readers cannot modify
	408	* the pointer because the noref flag is not set yet.
	409	* The smp_wmb() in ring_buffer_commit() takes care of ordering writes
	410	* to the subbuffer before this set noref operation.
	411	* subbuffer_set_noref() uses a volatile store to deal with concurrent
	412	* readers of the noref flag.
	413	*/
	414	CHAN_WARN_ON(bufb->chan,
	415	subbuffer_id_is_noref(config, bufb->buf_wsb[idx].id));
	416	/*
	417	* Memory barrier that ensures counter stores are ordered before set
	418	* noref and offset.
	419	*/
	420	smp_mb();
	421	subbuffer_id_set_noref_offset(config, &bufb->buf_wsb[idx].id, offset);
	422	}
	423
	424	/**
	425	* update_read_sb_index - Read-side subbuffer index update.
	426	*/
	427	static inline
	428	int update_read_sb_index(const struct lib_ring_buffer_config *config,
	429	struct lib_ring_buffer_backend *bufb,
	430	struct channel_backend *chanb,
	431	unsigned long consumed_idx,
	432	unsigned long consumed_count)
	433	{
	434	unsigned long old_id, new_id;
435
436	if (config->mode == RING_BUFFER_OVERWRITE) {
437	/*
438	* Exchange the target writer subbuffer with our own unused
a8f2d0c7	439	* subbuffer. No need to use READ_ONCE() here to read the
f3bc08c5 MD	440	* old_wpage, because the value read will be confirmed by the
	441	* following cmpxchg().
	442	*/
	443	old_id = bufb->buf_wsb[consumed_idx].id;
	444	if (unlikely(!subbuffer_id_is_noref(config, old_id)))
	445	return -EAGAIN;
	446	/*
	447	* Make sure the offset count we are expecting matches the one
	448	* indicated by the writer.
	449	*/
	450	if (unlikely(!subbuffer_id_compare_offset(config, old_id,
	451	consumed_count)))
	452	return -EAGAIN;
	453	CHAN_WARN_ON(bufb->chan,
	454	!subbuffer_id_is_noref(config, bufb->buf_rsb.id));
	455	subbuffer_id_set_noref_offset(config, &bufb->buf_rsb.id,
	456	consumed_count);
	457	new_id = cmpxchg(&bufb->buf_wsb[consumed_idx].id, old_id,
	458	bufb->buf_rsb.id);
	459	if (unlikely(old_id != new_id))
	460	return -EAGAIN;
	461	bufb->buf_rsb.id = new_id;
	462	} else {
	463	/* No page exchange, use the writer page directly */
	464	bufb->buf_rsb.id = bufb->buf_wsb[consumed_idx].id;
	465	}
	466	return 0;
	467	}
	468
327177e6	469	static inline __attribute__((always_inline))
04707312	470	void lttng_inline_memcpy(void dest, const void src,
327177e6 MD	471	unsigned long len)
	472	{
	473	switch (len) {
	474	case 1:
	475	(uint8_t ) dest = (const uint8_t ) src;
	476	break;
	477	case 2:
	478	(uint16_t ) dest = (const uint16_t ) src;
	479	break;
	480	case 4:
	481	(uint32_t ) dest = (const uint32_t ) src;
	482	break;
	483	case 8:
	484	(uint64_t ) dest = (const uint64_t ) src;
	485	break;
	486	default:
	487	inline_memcpy(dest, src, len);
	488	}
	489	}
	490
f3bc08c5 MD	491	/*
	492	* Use the architecture-specific memcpy implementation for constant-sized
	493	* inputs, but rely on an inline memcpy for length statically unknown.
	494	* The function call to memcpy is just way too expensive for a fast path.
	495	*/
	496	#define lib_ring_buffer_do_copy(config, dest, src, len) \
	497	do { \
	498	size_t __len = (len); \
	499	if (__builtin_constant_p(len)) \
	500	memcpy(dest, src, __len); \
	501	else \
327177e6	502	lttng_inline_memcpy(dest, src, __len); \
f3bc08c5 MD	503	} while (0)
f3bc08c5 MD	504
4ea00e4f	505	/*
7b8ea3a5	506	* We use __copy_from_user_inatomic to copy userspace data since we already
4ea00e4f	507	* did the access_ok for the whole range.
d87a9f03 MD	508	*
d87a9f03 MD	509	* Return 0 if OK, nonzero on error.
4ea00e4f JD	510	*/
4ea00e4f JD	511	static inline
7b8ea3a5	512	unsigned long lib_ring_buffer_do_copy_from_user_inatomic(void *dest,
4ea00e4f JD	513	const void __user *src,
	514	unsigned long len)
	515	{
7b8ea3a5	516	return __copy_from_user_inatomic(dest, src, len);
4ea00e4f JD	517	}
	518
	519	/*
	520	* write len bytes to dest with c
	521	*/
	522	static inline
	523	void lib_ring_buffer_do_memset(char *dest, int c,
	524	unsigned long len)
	525	{
	526	unsigned long i;
	527
	528	for (i = 0; i < len; i++)
	529	dest[i] = c;
	530	}
	531
886d51a3	532	#endif /* _LIB_RING_BUFFER_BACKEND_INTERNAL_H */