2 * SPDX-License-Identifier: LGPL-2.1-only
4 * Copyright (C) 2005-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
6 * See ring_buffer_frontend.c for more information on wait-free
8 * See frontend.h for channel allocation and read-side API.
11 #ifndef _LTTNG_RING_BUFFER_FRONTEND_API_H
12 #define _LTTNG_RING_BUFFER_FRONTEND_API_H
16 #include <urcu/compiler.h>
21 * lib_ring_buffer_nesting_inc - Ring buffer recursive use protection.
23 * The rint buffer buffer nesting count is a safety net to ensure tracer
24 * client code will never trigger an endless recursion.
25 * Returns a nesting level >= 0 on success, -EPERM on failure (nesting
28 * asm volatile and "memory" clobber prevent the compiler from moving
29 * instructions out of the ring buffer nesting count. This is required to ensure
30 * that probe side-effects which can cause recursion (e.g. unforeseen traps,
31 * divisions by 0, ...) are triggered within the incremented nesting count
35 int lib_ring_buffer_nesting_inc(
36 const struct lttng_ust_lib_ring_buffer_config
*config
__attribute__((unused
)))
40 nesting
= ++URCU_TLS(lib_ring_buffer_nesting
);
42 if (caa_unlikely(nesting
>= LIB_RING_BUFFER_MAX_NESTING
)) {
44 URCU_TLS(lib_ring_buffer_nesting
)--;
51 int lib_ring_buffer_nesting_count(
52 const struct lttng_ust_lib_ring_buffer_config
*config
__attribute__((unused
)))
54 return URCU_TLS(lib_ring_buffer_nesting
);
58 void lib_ring_buffer_nesting_dec(
59 const struct lttng_ust_lib_ring_buffer_config
*config
__attribute__((unused
)))
62 URCU_TLS(lib_ring_buffer_nesting
)--; /* TLS */
66 * lib_ring_buffer_try_reserve is called by lib_ring_buffer_reserve(). It is not
67 * part of the API per se.
69 * returns 0 if reserve ok, or 1 if the slow path must be taken.
72 int lib_ring_buffer_try_reserve(const struct lttng_ust_lib_ring_buffer_config
*config
,
73 struct lttng_ust_lib_ring_buffer_ctx
*ctx
,
75 unsigned long *o_begin
, unsigned long *o_end
,
76 unsigned long *o_old
, size_t *before_hdr_pad
)
78 struct lttng_ust_lib_ring_buffer_ctx_private
*ctx_private
= ctx
->priv
;
79 struct lttng_ust_lib_ring_buffer_channel
*chan
= ctx_private
->chan
;
80 struct lttng_ust_lib_ring_buffer
*buf
= ctx_private
->buf
;
81 *o_begin
= v_read(config
, &buf
->offset
);
84 ctx_private
->tsc
= lib_ring_buffer_clock_read(chan
);
85 if ((int64_t) ctx_private
->tsc
== -EIO
)
89 * Prefetch cacheline for read because we have to read the previous
90 * commit counter to increment it and commit seq value to compare it to
93 //prefetch(&buf->commit_hot[subbuf_index(*o_begin, chan)]);
95 if (last_tsc_overflow(config
, buf
, ctx_private
->tsc
))
96 ctx_private
->rflags
|= RING_BUFFER_RFLAG_FULL_TSC
;
98 if (caa_unlikely(subbuf_offset(*o_begin
, chan
) == 0))
101 ctx_private
->slot_size
= record_header_size(config
, chan
, *o_begin
,
102 before_hdr_pad
, ctx
, client_ctx
);
103 ctx_private
->slot_size
+=
104 lttng_ust_lib_ring_buffer_align(*o_begin
+ ctx_private
->slot_size
,
105 ctx
->largest_align
) + ctx
->data_size
;
106 if (caa_unlikely((subbuf_offset(*o_begin
, chan
) + ctx_private
->slot_size
)
107 > chan
->backend
.subbuf_size
))
111 * Record fits in the current buffer and we are not on a switch
112 * boundary. It's safe to write.
114 *o_end
= *o_begin
+ ctx_private
->slot_size
;
116 if (caa_unlikely((subbuf_offset(*o_end
, chan
)) == 0))
118 * The offset_end will fall at the very beginning of the next
127 * lib_ring_buffer_reserve - Reserve space in a ring buffer.
128 * @config: ring buffer instance configuration.
129 * @ctx: ring buffer context. (input and output) Must be already initialized.
131 * Atomic wait-free slot reservation. The reserved space starts at the context
132 * "pre_offset". Its length is "slot_size". The associated time-stamp is "tsc".
136 * -EAGAIN if channel is disabled.
137 * -ENOSPC if event size is too large for packet.
138 * -ENOBUFS if there is currently not enough space in buffer for the event.
139 * -EIO if data cannot be written into the buffer for any other reason.
143 int lib_ring_buffer_reserve(const struct lttng_ust_lib_ring_buffer_config
*config
,
144 struct lttng_ust_lib_ring_buffer_ctx
*ctx
,
147 struct lttng_ust_lib_ring_buffer_ctx_private
*ctx_private
= ctx
->priv
;
148 struct lttng_ust_lib_ring_buffer_channel
*chan
= ctx_private
->chan
;
149 struct lttng_ust_shm_handle
*handle
= chan
->handle
;
150 struct lttng_ust_lib_ring_buffer
*buf
;
151 unsigned long o_begin
, o_end
, o_old
;
152 size_t before_hdr_pad
= 0;
154 if (caa_unlikely(uatomic_read(&chan
->record_disabled
)))
157 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
158 ctx_private
->reserve_cpu
= lttng_ust_get_cpu();
159 buf
= shmp(handle
, chan
->backend
.buf
[ctx_private
->reserve_cpu
].shmp
);
161 buf
= shmp(handle
, chan
->backend
.buf
[0].shmp
);
163 if (caa_unlikely(!buf
))
165 if (caa_unlikely(uatomic_read(&buf
->record_disabled
)))
167 ctx_private
->buf
= buf
;
170 * Perform retryable operations.
172 if (caa_unlikely(lib_ring_buffer_try_reserve(config
, ctx
, client_ctx
, &o_begin
,
173 &o_end
, &o_old
, &before_hdr_pad
)))
176 if (caa_unlikely(v_cmpxchg(config
, &buf
->offset
, o_old
, o_end
)
181 * Atomically update last_tsc. This update races against concurrent
182 * atomic updates, but the race will always cause supplementary full TSC
183 * record headers, never the opposite (missing a full TSC record header
184 * when it would be needed).
186 save_last_tsc(config
, buf
, ctx_private
->tsc
);
189 * Push the reader if necessary
191 lib_ring_buffer_reserve_push_reader(buf
, chan
, o_end
- 1);
194 * Clear noref flag for this subbuffer.
196 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
197 subbuf_index(o_end
- 1, chan
), handle
);
199 ctx_private
->pre_offset
= o_begin
;
200 ctx_private
->buf_offset
= o_begin
+ before_hdr_pad
;
203 return lib_ring_buffer_reserve_slow(ctx
, client_ctx
);
207 * lib_ring_buffer_switch - Perform a sub-buffer switch for a per-cpu buffer.
208 * @config: ring buffer instance configuration.
210 * @mode: buffer switch mode (SWITCH_ACTIVE or SWITCH_FLUSH)
212 * This operation is completely reentrant : can be called while tracing is
213 * active with absolutely no lock held.
215 * Note, however, that as a v_cmpxchg is used for some atomic operations and
216 * requires to be executed locally for per-CPU buffers, this function must be
217 * called from the CPU which owns the buffer for a ACTIVE flush, with preemption
218 * disabled, for RING_BUFFER_SYNC_PER_CPU configuration.
221 void lib_ring_buffer_switch(
222 const struct lttng_ust_lib_ring_buffer_config
*config
__attribute__((unused
)),
223 struct lttng_ust_lib_ring_buffer
*buf
, enum switch_mode mode
,
224 struct lttng_ust_shm_handle
*handle
)
226 lib_ring_buffer_switch_slow(buf
, mode
, handle
);
229 /* See ring_buffer_frontend_api.h for lib_ring_buffer_reserve(). */
232 * lib_ring_buffer_commit - Commit an record.
233 * @config: ring buffer instance configuration.
234 * @ctx: ring buffer context. (input arguments only)
236 * Atomic unordered slot commit. Increments the commit count in the
237 * specified sub-buffer, and delivers it if necessary.
240 void lib_ring_buffer_commit(const struct lttng_ust_lib_ring_buffer_config
*config
,
241 const struct lttng_ust_lib_ring_buffer_ctx
*ctx
)
243 struct lttng_ust_lib_ring_buffer_ctx_private
*ctx_private
= ctx
->priv
;
244 struct lttng_ust_lib_ring_buffer_channel
*chan
= ctx_private
->chan
;
245 struct lttng_ust_shm_handle
*handle
= chan
->handle
;
246 struct lttng_ust_lib_ring_buffer
*buf
= ctx_private
->buf
;
247 unsigned long offset_end
= ctx_private
->buf_offset
;
248 unsigned long endidx
= subbuf_index(offset_end
- 1, chan
);
249 unsigned long commit_count
;
250 struct commit_counters_hot
*cc_hot
= shmp_index(handle
,
251 buf
->commit_hot
, endidx
);
253 if (caa_unlikely(!cc_hot
))
257 * Must count record before incrementing the commit count.
259 subbuffer_count_record(config
, ctx
, &buf
->backend
, endidx
, handle
);
262 * Order all writes to buffer before the commit count update that will
263 * determine that the subbuffer is full.
267 v_add(config
, ctx_private
->slot_size
, &cc_hot
->cc
);
270 * commit count read can race with concurrent OOO commit count updates.
271 * This is only needed for lib_ring_buffer_check_deliver (for
272 * non-polling delivery only) and for
273 * lib_ring_buffer_write_commit_counter. The race can only cause the
274 * counter to be read with the same value more than once, which could
276 * - Multiple delivery for the same sub-buffer (which is handled
277 * gracefully by the reader code) if the value is for a full
278 * sub-buffer. It's important that we can never miss a sub-buffer
279 * delivery. Re-reading the value after the v_add ensures this.
280 * - Reading a commit_count with a higher value that what was actually
281 * added to it for the lib_ring_buffer_write_commit_counter call
282 * (again caused by a concurrent committer). It does not matter,
283 * because this function is interested in the fact that the commit
284 * count reaches back the reserve offset for a specific sub-buffer,
285 * which is completely independent of the order.
287 commit_count
= v_read(config
, &cc_hot
->cc
);
289 lib_ring_buffer_check_deliver(config
, buf
, chan
, offset_end
- 1,
290 commit_count
, endidx
, handle
, ctx_private
->tsc
);
292 * Update used size at each commit. It's needed only for extracting
293 * ring_buffer buffers from vmcore, after crash.
295 lib_ring_buffer_write_commit_counter(config
, buf
, chan
,
296 offset_end
, commit_count
, handle
, cc_hot
);
300 * lib_ring_buffer_try_discard_reserve - Try discarding a record.
301 * @config: ring buffer instance configuration.
302 * @ctx: ring buffer context. (input arguments only)
304 * Only succeeds if no other record has been written after the record to
305 * discard. If discard fails, the record must be committed to the buffer.
307 * Returns 0 upon success, -EPERM if the record cannot be discarded.
310 int lib_ring_buffer_try_discard_reserve(const struct lttng_ust_lib_ring_buffer_config
*config
,
311 const struct lttng_ust_lib_ring_buffer_ctx
*ctx
)
313 struct lttng_ust_lib_ring_buffer_ctx_private
*ctx_private
= ctx
->priv
;
314 struct lttng_ust_lib_ring_buffer
*buf
= ctx_private
->buf
;
315 unsigned long end_offset
= ctx_private
->pre_offset
+ ctx_private
->slot_size
;
318 * We need to ensure that if the cmpxchg succeeds and discards the
319 * record, the next record will record a full TSC, because it cannot
320 * rely on the last_tsc associated with the discarded record to detect
321 * overflows. The only way to ensure this is to set the last_tsc to 0
322 * (assuming no 64-bit TSC overflow), which forces to write a 64-bit
323 * timestamp in the next record.
325 * Note: if discard fails, we must leave the TSC in the record header.
326 * It is needed to keep track of TSC overflows for the following
329 save_last_tsc(config
, buf
, 0ULL);
331 if (caa_likely(v_cmpxchg(config
, &buf
->offset
, end_offset
, ctx_private
->pre_offset
)
339 void channel_record_disable(
340 const struct lttng_ust_lib_ring_buffer_config
*config
__attribute__((unused
)),
341 struct lttng_ust_lib_ring_buffer_channel
*chan
)
343 uatomic_inc(&chan
->record_disabled
);
347 void channel_record_enable(
348 const struct lttng_ust_lib_ring_buffer_config
*config
__attribute__((unused
)),
349 struct lttng_ust_lib_ring_buffer_channel
*chan
)
351 uatomic_dec(&chan
->record_disabled
);
355 void lib_ring_buffer_record_disable(
356 const struct lttng_ust_lib_ring_buffer_config
*config
__attribute__((unused
)),
357 struct lttng_ust_lib_ring_buffer
*buf
)
359 uatomic_inc(&buf
->record_disabled
);
363 void lib_ring_buffer_record_enable(
364 const struct lttng_ust_lib_ring_buffer_config
*config
__attribute__((unused
)),
365 struct lttng_ust_lib_ring_buffer
*buf
)
367 uatomic_dec(&buf
->record_disabled
);
370 #endif /* _LTTNG_RING_BUFFER_FRONTEND_API_H */