1 #ifndef _LIB_RING_BUFFER_FRONTEND_API_H
2 #define _LIB_RING_BUFFER_FRONTEND_API_H
5 * lib/ringbuffer/frontend_api.h
7 * Ring Buffer Library Synchronization Header (buffer write API).
9 * Copyright (C) 2005-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
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.
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.
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
26 * Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
28 * See ring_buffer_frontend.c for more information on wait-free algorithms.
29 * See linux/ringbuffer/frontend.h for channel allocation and read-side API.
32 #include <wrapper/ringbuffer/frontend.h>
33 #include <wrapper/percpu-defs.h>
34 #include <linux/errno.h>
35 #include <linux/prefetch.h>
38 * lib_ring_buffer_get_cpu - Precedes ring buffer reserve/commit.
40 * Disables preemption (acts as a RCU read-side critical section) and keeps a
41 * ring buffer nesting count as supplementary safety net to ensure tracer client
42 * code will never trigger an endless recursion. Returns the processor ID on
43 * success, -EPERM on failure (nesting count too high).
45 * asm volatile and "memory" clobber prevent the compiler from moving
46 * instructions out of the ring buffer nesting count. This is required to ensure
47 * that probe side-effects which can cause recursion (e.g. unforeseen traps,
48 * divisions by 0, ...) are triggered within the incremented nesting count
52 int lib_ring_buffer_get_cpu(const struct lib_ring_buffer_config
*config
)
56 rcu_read_lock_sched_notrace();
57 cpu
= smp_processor_id();
58 nesting
= ++per_cpu(lib_ring_buffer_nesting
, cpu
);
61 if (unlikely(nesting
> 4)) {
63 per_cpu(lib_ring_buffer_nesting
, cpu
)--;
64 rcu_read_unlock_sched_notrace();
71 * lib_ring_buffer_put_cpu - Follows ring buffer reserve/commit.
74 void lib_ring_buffer_put_cpu(const struct lib_ring_buffer_config
*config
)
77 (*lttng_this_cpu_ptr(&lib_ring_buffer_nesting
))--;
78 rcu_read_unlock_sched_notrace();
82 * lib_ring_buffer_try_reserve is called by lib_ring_buffer_reserve(). It is not
83 * part of the API per se.
85 * returns 0 if reserve ok, or 1 if the slow path must be taken.
88 int lib_ring_buffer_try_reserve(const struct lib_ring_buffer_config
*config
,
89 struct lib_ring_buffer_ctx
*ctx
,
91 unsigned long *o_begin
, unsigned long *o_end
,
92 unsigned long *o_old
, size_t *before_hdr_pad
)
94 struct channel
*chan
= ctx
->chan
;
95 struct lib_ring_buffer
*buf
= ctx
->buf
;
96 *o_begin
= v_read(config
, &buf
->offset
);
99 ctx
->tsc
= lib_ring_buffer_clock_read(chan
);
100 if ((int64_t) ctx
->tsc
== -EIO
)
104 * Prefetch cacheline for read because we have to read the previous
105 * commit counter to increment it and commit seq value to compare it to
106 * the commit counter.
108 prefetch(&buf
->commit_hot
[subbuf_index(*o_begin
, chan
)]);
110 if (last_tsc_overflow(config
, buf
, ctx
->tsc
))
111 ctx
->rflags
|= RING_BUFFER_RFLAG_FULL_TSC
;
113 if (unlikely(subbuf_offset(*o_begin
, chan
) == 0))
116 ctx
->slot_size
= record_header_size(config
, chan
, *o_begin
,
117 before_hdr_pad
, ctx
, client_ctx
);
119 lib_ring_buffer_align(*o_begin
+ ctx
->slot_size
,
120 ctx
->largest_align
) + ctx
->data_size
;
121 if (unlikely((subbuf_offset(*o_begin
, chan
) + ctx
->slot_size
)
122 > chan
->backend
.subbuf_size
))
126 * Record fits in the current buffer and we are not on a switch
127 * boundary. It's safe to write.
129 *o_end
= *o_begin
+ ctx
->slot_size
;
131 if (unlikely((subbuf_offset(*o_end
, chan
)) == 0))
133 * The offset_end will fall at the very beginning of the next
142 * lib_ring_buffer_reserve - Reserve space in a ring buffer.
143 * @config: ring buffer instance configuration.
144 * @ctx: ring buffer context. (input and output) Must be already initialized.
146 * Atomic wait-free slot reservation. The reserved space starts at the context
147 * "pre_offset". Its length is "slot_size". The associated time-stamp is "tsc".
151 * -EAGAIN if channel is disabled.
152 * -ENOSPC if event size is too large for packet.
153 * -ENOBUFS if there is currently not enough space in buffer for the event.
154 * -EIO if data cannot be written into the buffer for any other reason.
158 int lib_ring_buffer_reserve(const struct lib_ring_buffer_config
*config
,
159 struct lib_ring_buffer_ctx
*ctx
,
162 struct channel
*chan
= ctx
->chan
;
163 struct lib_ring_buffer
*buf
;
164 unsigned long o_begin
, o_end
, o_old
;
165 size_t before_hdr_pad
= 0;
167 if (unlikely(atomic_read(&chan
->record_disabled
)))
170 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
171 buf
= per_cpu_ptr(chan
->backend
.buf
, ctx
->cpu
);
173 buf
= chan
->backend
.buf
;
174 if (unlikely(atomic_read(&buf
->record_disabled
)))
179 * Perform retryable operations.
181 if (unlikely(lib_ring_buffer_try_reserve(config
, ctx
, client_ctx
, &o_begin
,
182 &o_end
, &o_old
, &before_hdr_pad
)))
185 if (unlikely(v_cmpxchg(config
, &ctx
->buf
->offset
, o_old
, o_end
)
190 * Atomically update last_tsc. This update races against concurrent
191 * atomic updates, but the race will always cause supplementary full TSC
192 * record headers, never the opposite (missing a full TSC record header
193 * when it would be needed).
195 save_last_tsc(config
, ctx
->buf
, ctx
->tsc
);
198 * Push the reader if necessary
200 lib_ring_buffer_reserve_push_reader(ctx
->buf
, chan
, o_end
- 1);
203 * Clear noref flag for this subbuffer.
205 lib_ring_buffer_clear_noref(config
, &ctx
->buf
->backend
,
206 subbuf_index(o_end
- 1, chan
));
208 ctx
->pre_offset
= o_begin
;
209 ctx
->buf_offset
= o_begin
+ before_hdr_pad
;
212 return lib_ring_buffer_reserve_slow(ctx
, client_ctx
);
216 * lib_ring_buffer_switch - Perform a sub-buffer switch for a per-cpu buffer.
217 * @config: ring buffer instance configuration.
219 * @mode: buffer switch mode (SWITCH_ACTIVE or SWITCH_FLUSH)
221 * This operation is completely reentrant : can be called while tracing is
222 * active with absolutely no lock held.
224 * Note, however, that as a v_cmpxchg is used for some atomic operations and
225 * requires to be executed locally for per-CPU buffers, this function must be
226 * called from the CPU which owns the buffer for a ACTIVE flush, with preemption
227 * disabled, for RING_BUFFER_SYNC_PER_CPU configuration.
230 void lib_ring_buffer_switch(const struct lib_ring_buffer_config
*config
,
231 struct lib_ring_buffer
*buf
, enum switch_mode mode
)
233 lib_ring_buffer_switch_slow(buf
, mode
);
236 /* See ring_buffer_frontend_api.h for lib_ring_buffer_reserve(). */
239 * lib_ring_buffer_commit - Commit an record.
240 * @config: ring buffer instance configuration.
241 * @ctx: ring buffer context. (input arguments only)
243 * Atomic unordered slot commit. Increments the commit count in the
244 * specified sub-buffer, and delivers it if necessary.
247 void lib_ring_buffer_commit(const struct lib_ring_buffer_config
*config
,
248 const struct lib_ring_buffer_ctx
*ctx
)
250 struct channel
*chan
= ctx
->chan
;
251 struct lib_ring_buffer
*buf
= ctx
->buf
;
252 unsigned long offset_end
= ctx
->buf_offset
;
253 unsigned long endidx
= subbuf_index(offset_end
- 1, chan
);
254 unsigned long commit_count
;
255 struct commit_counters_hot
*cc_hot
= &buf
->commit_hot
[endidx
];
258 * Must count record before incrementing the commit count.
260 subbuffer_count_record(config
, &buf
->backend
, endidx
);
263 * Order all writes to buffer before the commit count update that will
264 * determine that the subbuffer is full.
266 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
268 * Must write slot data before incrementing commit count. This
269 * compiler barrier is upgraded into a smp_mb() by the IPI sent
276 v_add(config
, ctx
->slot_size
, &cc_hot
->cc
);
279 * commit count read can race with concurrent OOO commit count updates.
280 * This is only needed for lib_ring_buffer_check_deliver (for
281 * non-polling delivery only) and for
282 * lib_ring_buffer_write_commit_counter. The race can only cause the
283 * counter to be read with the same value more than once, which could
285 * - Multiple delivery for the same sub-buffer (which is handled
286 * gracefully by the reader code) if the value is for a full
287 * sub-buffer. It's important that we can never miss a sub-buffer
288 * delivery. Re-reading the value after the v_add ensures this.
289 * - Reading a commit_count with a higher value that what was actually
290 * added to it for the lib_ring_buffer_write_commit_counter call
291 * (again caused by a concurrent committer). It does not matter,
292 * because this function is interested in the fact that the commit
293 * count reaches back the reserve offset for a specific sub-buffer,
294 * which is completely independent of the order.
296 commit_count
= v_read(config
, &cc_hot
->cc
);
298 lib_ring_buffer_check_deliver(config
, buf
, chan
, offset_end
- 1,
299 commit_count
, endidx
, ctx
->tsc
);
301 * Update used size at each commit. It's needed only for extracting
302 * ring_buffer buffers from vmcore, after crash.
304 lib_ring_buffer_write_commit_counter(config
, buf
, chan
,
305 offset_end
, commit_count
, cc_hot
);
309 * lib_ring_buffer_try_discard_reserve - Try discarding a record.
310 * @config: ring buffer instance configuration.
311 * @ctx: ring buffer context. (input arguments only)
313 * Only succeeds if no other record has been written after the record to
314 * discard. If discard fails, the record must be committed to the buffer.
316 * Returns 0 upon success, -EPERM if the record cannot be discarded.
319 int lib_ring_buffer_try_discard_reserve(const struct lib_ring_buffer_config
*config
,
320 const struct lib_ring_buffer_ctx
*ctx
)
322 struct lib_ring_buffer
*buf
= ctx
->buf
;
323 unsigned long end_offset
= ctx
->pre_offset
+ ctx
->slot_size
;
326 * We need to ensure that if the cmpxchg succeeds and discards the
327 * record, the next record will record a full TSC, because it cannot
328 * rely on the last_tsc associated with the discarded record to detect
329 * overflows. The only way to ensure this is to set the last_tsc to 0
330 * (assuming no 64-bit TSC overflow), which forces to write a 64-bit
331 * timestamp in the next record.
333 * Note: if discard fails, we must leave the TSC in the record header.
334 * It is needed to keep track of TSC overflows for the following
337 save_last_tsc(config
, buf
, 0ULL);
339 if (likely(v_cmpxchg(config
, &buf
->offset
, end_offset
, ctx
->pre_offset
)
347 void channel_record_disable(const struct lib_ring_buffer_config
*config
,
348 struct channel
*chan
)
350 atomic_inc(&chan
->record_disabled
);
354 void channel_record_enable(const struct lib_ring_buffer_config
*config
,
355 struct channel
*chan
)
357 atomic_dec(&chan
->record_disabled
);
361 void lib_ring_buffer_record_disable(const struct lib_ring_buffer_config
*config
,
362 struct lib_ring_buffer
*buf
)
364 atomic_inc(&buf
->record_disabled
);
368 void lib_ring_buffer_record_enable(const struct lib_ring_buffer_config
*config
,
369 struct lib_ring_buffer
*buf
)
371 atomic_dec(&buf
->record_disabled
);
374 #endif /* _LIB_RING_BUFFER_FRONTEND_API_H */