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_get_cpu - Precedes ring buffer reserve/commit.
23 * Keeps a ring buffer nesting count as supplementary safety net to
24 * ensure tracer client code will never trigger an endless recursion.
25 * Returns the processor ID on success, -EPERM on failure (nesting count
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_get_cpu(const struct lttng_ust_lib_ring_buffer_config
*config
)
39 cpu
= lttng_ust_get_cpu();
40 nesting
= ++URCU_TLS(lib_ring_buffer_nesting
);
43 if (caa_unlikely(nesting
> 4)) {
45 URCU_TLS(lib_ring_buffer_nesting
)--;
52 * lib_ring_buffer_put_cpu - Follows ring buffer reserve/commit.
55 void lib_ring_buffer_put_cpu(const struct lttng_ust_lib_ring_buffer_config
*config
)
58 URCU_TLS(lib_ring_buffer_nesting
)--; /* TLS */
62 * lib_ring_buffer_try_reserve is called by lib_ring_buffer_reserve(). It is not
63 * part of the API per se.
65 * returns 0 if reserve ok, or 1 if the slow path must be taken.
68 int lib_ring_buffer_try_reserve(const struct lttng_ust_lib_ring_buffer_config
*config
,
69 struct lttng_ust_lib_ring_buffer_ctx
*ctx
,
71 unsigned long *o_begin
, unsigned long *o_end
,
72 unsigned long *o_old
, size_t *before_hdr_pad
)
74 struct channel
*chan
= ctx
->chan
;
75 struct lttng_ust_lib_ring_buffer
*buf
= ctx
->buf
;
76 *o_begin
= v_read(config
, &buf
->offset
);
79 ctx
->tsc
= lib_ring_buffer_clock_read(chan
);
80 if ((int64_t) ctx
->tsc
== -EIO
)
84 * Prefetch cacheline for read because we have to read the previous
85 * commit counter to increment it and commit seq value to compare it to
88 //prefetch(&buf->commit_hot[subbuf_index(*o_begin, chan)]);
90 if (last_tsc_overflow(config
, buf
, ctx
->tsc
))
91 ctx
->rflags
|= RING_BUFFER_RFLAG_FULL_TSC
;
93 if (caa_unlikely(subbuf_offset(*o_begin
, chan
) == 0))
96 ctx
->slot_size
= record_header_size(config
, chan
, *o_begin
,
97 before_hdr_pad
, ctx
, client_ctx
);
99 lib_ring_buffer_align(*o_begin
+ ctx
->slot_size
,
100 ctx
->largest_align
) + ctx
->data_size
;
101 if (caa_unlikely((subbuf_offset(*o_begin
, chan
) + ctx
->slot_size
)
102 > chan
->backend
.subbuf_size
))
106 * Record fits in the current buffer and we are not on a switch
107 * boundary. It's safe to write.
109 *o_end
= *o_begin
+ ctx
->slot_size
;
111 if (caa_unlikely((subbuf_offset(*o_end
, chan
)) == 0))
113 * The offset_end will fall at the very beginning of the next
122 * lib_ring_buffer_reserve - Reserve space in a ring buffer.
123 * @config: ring buffer instance configuration.
124 * @ctx: ring buffer context. (input and output) Must be already initialized.
126 * Atomic wait-free slot reservation. The reserved space starts at the context
127 * "pre_offset". Its length is "slot_size". The associated time-stamp is "tsc".
131 * -EAGAIN if channel is disabled.
132 * -ENOSPC if event size is too large for packet.
133 * -ENOBUFS if there is currently not enough space in buffer for the event.
134 * -EIO if data cannot be written into the buffer for any other reason.
138 int lib_ring_buffer_reserve(const struct lttng_ust_lib_ring_buffer_config
*config
,
139 struct lttng_ust_lib_ring_buffer_ctx
*ctx
,
142 struct channel
*chan
= ctx
->chan
;
143 struct lttng_ust_shm_handle
*handle
= ctx
->handle
;
144 struct lttng_ust_lib_ring_buffer
*buf
;
145 unsigned long o_begin
, o_end
, o_old
;
146 size_t before_hdr_pad
= 0;
148 if (caa_unlikely(uatomic_read(&chan
->record_disabled
)))
151 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
152 buf
= shmp(handle
, chan
->backend
.buf
[ctx
->cpu
].shmp
);
154 buf
= shmp(handle
, chan
->backend
.buf
[0].shmp
);
155 if (caa_unlikely(!buf
))
157 if (caa_unlikely(uatomic_read(&buf
->record_disabled
)))
162 * Perform retryable operations.
164 if (caa_unlikely(lib_ring_buffer_try_reserve(config
, ctx
, client_ctx
, &o_begin
,
165 &o_end
, &o_old
, &before_hdr_pad
)))
168 if (caa_unlikely(v_cmpxchg(config
, &ctx
->buf
->offset
, o_old
, o_end
)
173 * Atomically update last_tsc. This update races against concurrent
174 * atomic updates, but the race will always cause supplementary full TSC
175 * record headers, never the opposite (missing a full TSC record header
176 * when it would be needed).
178 save_last_tsc(config
, ctx
->buf
, ctx
->tsc
);
181 * Push the reader if necessary
183 lib_ring_buffer_reserve_push_reader(ctx
->buf
, chan
, o_end
- 1);
186 * Clear noref flag for this subbuffer.
188 lib_ring_buffer_clear_noref(config
, &ctx
->buf
->backend
,
189 subbuf_index(o_end
- 1, chan
), handle
);
191 ctx
->pre_offset
= o_begin
;
192 ctx
->buf_offset
= o_begin
+ before_hdr_pad
;
195 return lib_ring_buffer_reserve_slow(ctx
, client_ctx
);
199 * lib_ring_buffer_switch - Perform a sub-buffer switch for a per-cpu buffer.
200 * @config: ring buffer instance configuration.
202 * @mode: buffer switch mode (SWITCH_ACTIVE or SWITCH_FLUSH)
204 * This operation is completely reentrant : can be called while tracing is
205 * active with absolutely no lock held.
207 * Note, however, that as a v_cmpxchg is used for some atomic operations and
208 * requires to be executed locally for per-CPU buffers, this function must be
209 * called from the CPU which owns the buffer for a ACTIVE flush, with preemption
210 * disabled, for RING_BUFFER_SYNC_PER_CPU configuration.
213 void lib_ring_buffer_switch(const struct lttng_ust_lib_ring_buffer_config
*config
,
214 struct lttng_ust_lib_ring_buffer
*buf
, enum switch_mode mode
,
215 struct lttng_ust_shm_handle
*handle
)
217 lib_ring_buffer_switch_slow(buf
, mode
, handle
);
220 /* See ring_buffer_frontend_api.h for lib_ring_buffer_reserve(). */
223 * lib_ring_buffer_commit - Commit an record.
224 * @config: ring buffer instance configuration.
225 * @ctx: ring buffer context. (input arguments only)
227 * Atomic unordered slot commit. Increments the commit count in the
228 * specified sub-buffer, and delivers it if necessary.
231 void lib_ring_buffer_commit(const struct lttng_ust_lib_ring_buffer_config
*config
,
232 const struct lttng_ust_lib_ring_buffer_ctx
*ctx
)
234 struct channel
*chan
= ctx
->chan
;
235 struct lttng_ust_shm_handle
*handle
= ctx
->handle
;
236 struct lttng_ust_lib_ring_buffer
*buf
= ctx
->buf
;
237 unsigned long offset_end
= ctx
->buf_offset
;
238 unsigned long endidx
= subbuf_index(offset_end
- 1, chan
);
239 unsigned long commit_count
;
240 struct commit_counters_hot
*cc_hot
= shmp_index(handle
,
241 buf
->commit_hot
, endidx
);
243 if (caa_unlikely(!cc_hot
))
247 * Must count record before incrementing the commit count.
249 subbuffer_count_record(config
, ctx
, &buf
->backend
, endidx
, handle
);
252 * Order all writes to buffer before the commit count update that will
253 * determine that the subbuffer is full.
257 v_add(config
, ctx
->slot_size
, &cc_hot
->cc
);
260 * commit count read can race with concurrent OOO commit count updates.
261 * This is only needed for lib_ring_buffer_check_deliver (for
262 * non-polling delivery only) and for
263 * lib_ring_buffer_write_commit_counter. The race can only cause the
264 * counter to be read with the same value more than once, which could
266 * - Multiple delivery for the same sub-buffer (which is handled
267 * gracefully by the reader code) if the value is for a full
268 * sub-buffer. It's important that we can never miss a sub-buffer
269 * delivery. Re-reading the value after the v_add ensures this.
270 * - Reading a commit_count with a higher value that what was actually
271 * added to it for the lib_ring_buffer_write_commit_counter call
272 * (again caused by a concurrent committer). It does not matter,
273 * because this function is interested in the fact that the commit
274 * count reaches back the reserve offset for a specific sub-buffer,
275 * which is completely independent of the order.
277 commit_count
= v_read(config
, &cc_hot
->cc
);
279 lib_ring_buffer_check_deliver(config
, buf
, chan
, offset_end
- 1,
280 commit_count
, endidx
, handle
, ctx
->tsc
);
282 * Update used size at each commit. It's needed only for extracting
283 * ring_buffer buffers from vmcore, after crash.
285 lib_ring_buffer_write_commit_counter(config
, buf
, chan
,
286 offset_end
, commit_count
, handle
, cc_hot
);
290 * lib_ring_buffer_try_discard_reserve - Try discarding a record.
291 * @config: ring buffer instance configuration.
292 * @ctx: ring buffer context. (input arguments only)
294 * Only succeeds if no other record has been written after the record to
295 * discard. If discard fails, the record must be committed to the buffer.
297 * Returns 0 upon success, -EPERM if the record cannot be discarded.
300 int lib_ring_buffer_try_discard_reserve(const struct lttng_ust_lib_ring_buffer_config
*config
,
301 const struct lttng_ust_lib_ring_buffer_ctx
*ctx
)
303 struct lttng_ust_lib_ring_buffer
*buf
= ctx
->buf
;
304 unsigned long end_offset
= ctx
->pre_offset
+ ctx
->slot_size
;
307 * We need to ensure that if the cmpxchg succeeds and discards the
308 * record, the next record will record a full TSC, because it cannot
309 * rely on the last_tsc associated with the discarded record to detect
310 * overflows. The only way to ensure this is to set the last_tsc to 0
311 * (assuming no 64-bit TSC overflow), which forces to write a 64-bit
312 * timestamp in the next record.
314 * Note: if discard fails, we must leave the TSC in the record header.
315 * It is needed to keep track of TSC overflows for the following
318 save_last_tsc(config
, buf
, 0ULL);
320 if (caa_likely(v_cmpxchg(config
, &buf
->offset
, end_offset
, ctx
->pre_offset
)
328 void channel_record_disable(const struct lttng_ust_lib_ring_buffer_config
*config
,
329 struct channel
*chan
)
331 uatomic_inc(&chan
->record_disabled
);
335 void channel_record_enable(const struct lttng_ust_lib_ring_buffer_config
*config
,
336 struct channel
*chan
)
338 uatomic_dec(&chan
->record_disabled
);
342 void lib_ring_buffer_record_disable(const struct lttng_ust_lib_ring_buffer_config
*config
,
343 struct lttng_ust_lib_ring_buffer
*buf
)
345 uatomic_inc(&buf
->record_disabled
);
349 void lib_ring_buffer_record_enable(const struct lttng_ust_lib_ring_buffer_config
*config
,
350 struct lttng_ust_lib_ring_buffer
*buf
)
352 uatomic_dec(&buf
->record_disabled
);
355 #endif /* _LTTNG_RING_BUFFER_FRONTEND_API_H */