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 <linux/errno.h>
36 * lib_ring_buffer_get_cpu - Precedes ring buffer reserve/commit.
38 * Disables preemption (acts as a RCU read-side critical section) and keeps a
39 * ring buffer nesting count as supplementary safety net to ensure tracer client
40 * code will never trigger an endless recursion. Returns the processor ID on
41 * success, -EPERM on failure (nesting count too high).
43 * asm volatile and "memory" clobber prevent the compiler from moving
44 * instructions out of the ring buffer nesting count. This is required to ensure
45 * that probe side-effects which can cause recursion (e.g. unforeseen traps,
46 * divisions by 0, ...) are triggered within the incremented nesting count
50 int lib_ring_buffer_get_cpu(const struct lib_ring_buffer_config
*config
)
54 rcu_read_lock_sched_notrace();
55 cpu
= smp_processor_id();
56 nesting
= ++per_cpu(lib_ring_buffer_nesting
, cpu
);
59 if (unlikely(nesting
> 4)) {
61 per_cpu(lib_ring_buffer_nesting
, cpu
)--;
62 rcu_read_unlock_sched_notrace();
69 * lib_ring_buffer_put_cpu - Follows ring buffer reserve/commit.
72 void lib_ring_buffer_put_cpu(const struct lib_ring_buffer_config
*config
)
75 __get_cpu_var(lib_ring_buffer_nesting
)--;
76 rcu_read_unlock_sched_notrace();
80 * lib_ring_buffer_try_reserve is called by lib_ring_buffer_reserve(). It is not
81 * part of the API per se.
83 * returns 0 if reserve ok, or 1 if the slow path must be taken.
86 int lib_ring_buffer_try_reserve(const struct lib_ring_buffer_config
*config
,
87 struct lib_ring_buffer_ctx
*ctx
,
88 unsigned long *o_begin
, unsigned long *o_end
,
89 unsigned long *o_old
, size_t *before_hdr_pad
)
91 struct channel
*chan
= ctx
->chan
;
92 struct lib_ring_buffer
*buf
= ctx
->buf
;
93 *o_begin
= v_read(config
, &buf
->offset
);
96 ctx
->tsc
= lib_ring_buffer_clock_read(chan
);
97 if ((int64_t) ctx
->tsc
== -EIO
)
101 * Prefetch cacheline for read because we have to read the previous
102 * commit counter to increment it and commit seq value to compare it to
103 * the commit counter.
105 prefetch(&buf
->commit_hot
[subbuf_index(*o_begin
, chan
)]);
107 if (last_tsc_overflow(config
, buf
, ctx
->tsc
))
108 ctx
->rflags
|= RING_BUFFER_RFLAG_FULL_TSC
;
110 if (unlikely(subbuf_offset(*o_begin
, chan
) == 0))
113 ctx
->slot_size
= record_header_size(config
, chan
, *o_begin
,
114 before_hdr_pad
, ctx
);
116 lib_ring_buffer_align(*o_begin
+ ctx
->slot_size
,
117 ctx
->largest_align
) + ctx
->data_size
;
118 if (unlikely((subbuf_offset(*o_begin
, chan
) + ctx
->slot_size
)
119 > chan
->backend
.subbuf_size
))
123 * Record fits in the current buffer and we are not on a switch
124 * boundary. It's safe to write.
126 *o_end
= *o_begin
+ ctx
->slot_size
;
128 if (unlikely((subbuf_offset(*o_end
, chan
)) == 0))
130 * The offset_end will fall at the very beginning of the next
139 * lib_ring_buffer_reserve - Reserve space in a ring buffer.
140 * @config: ring buffer instance configuration.
141 * @ctx: ring buffer context. (input and output) Must be already initialized.
143 * Atomic wait-free slot reservation. The reserved space starts at the context
144 * "pre_offset". Its length is "slot_size". The associated time-stamp is "tsc".
148 * -EAGAIN if channel is disabled.
149 * -ENOSPC if event size is too large for packet.
150 * -ENOBUFS if there is currently not enough space in buffer for the event.
151 * -EIO if data cannot be written into the buffer for any other reason.
155 int lib_ring_buffer_reserve(const struct lib_ring_buffer_config
*config
,
156 struct lib_ring_buffer_ctx
*ctx
)
158 struct channel
*chan
= ctx
->chan
;
159 struct lib_ring_buffer
*buf
;
160 unsigned long o_begin
, o_end
, o_old
;
161 size_t before_hdr_pad
= 0;
163 if (atomic_read(&chan
->record_disabled
))
166 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
167 buf
= per_cpu_ptr(chan
->backend
.buf
, ctx
->cpu
);
169 buf
= chan
->backend
.buf
;
170 if (atomic_read(&buf
->record_disabled
))
175 * Perform retryable operations.
177 if (unlikely(lib_ring_buffer_try_reserve(config
, ctx
, &o_begin
,
178 &o_end
, &o_old
, &before_hdr_pad
)))
181 if (unlikely(v_cmpxchg(config
, &ctx
->buf
->offset
, o_old
, o_end
)
186 * Atomically update last_tsc. This update races against concurrent
187 * atomic updates, but the race will always cause supplementary full TSC
188 * record headers, never the opposite (missing a full TSC record header
189 * when it would be needed).
191 save_last_tsc(config
, ctx
->buf
, ctx
->tsc
);
194 * Push the reader if necessary
196 lib_ring_buffer_reserve_push_reader(ctx
->buf
, chan
, o_end
- 1);
199 * Clear noref flag for this subbuffer.
201 lib_ring_buffer_clear_noref(config
, &ctx
->buf
->backend
,
202 subbuf_index(o_end
- 1, chan
));
204 ctx
->pre_offset
= o_begin
;
205 ctx
->buf_offset
= o_begin
+ before_hdr_pad
;
208 return lib_ring_buffer_reserve_slow(ctx
);
212 * lib_ring_buffer_switch - Perform a sub-buffer switch for a per-cpu buffer.
213 * @config: ring buffer instance configuration.
215 * @mode: buffer switch mode (SWITCH_ACTIVE or SWITCH_FLUSH)
217 * This operation is completely reentrant : can be called while tracing is
218 * active with absolutely no lock held.
220 * Note, however, that as a v_cmpxchg is used for some atomic operations and
221 * requires to be executed locally for per-CPU buffers, this function must be
222 * called from the CPU which owns the buffer for a ACTIVE flush, with preemption
223 * disabled, for RING_BUFFER_SYNC_PER_CPU configuration.
226 void lib_ring_buffer_switch(const struct lib_ring_buffer_config
*config
,
227 struct lib_ring_buffer
*buf
, enum switch_mode mode
)
229 lib_ring_buffer_switch_slow(buf
, mode
);
232 /* See ring_buffer_frontend_api.h for lib_ring_buffer_reserve(). */
235 * lib_ring_buffer_commit - Commit an record.
236 * @config: ring buffer instance configuration.
237 * @ctx: ring buffer context. (input arguments only)
239 * Atomic unordered slot commit. Increments the commit count in the
240 * specified sub-buffer, and delivers it if necessary.
243 void lib_ring_buffer_commit(const struct lib_ring_buffer_config
*config
,
244 const struct lib_ring_buffer_ctx
*ctx
)
246 struct channel
*chan
= ctx
->chan
;
247 struct lib_ring_buffer
*buf
= ctx
->buf
;
248 unsigned long offset_end
= ctx
->buf_offset
;
249 unsigned long endidx
= subbuf_index(offset_end
- 1, chan
);
250 unsigned long commit_count
;
253 * Must count record before incrementing the commit count.
255 subbuffer_count_record(config
, &buf
->backend
, endidx
);
258 * Order all writes to buffer before the commit count update that will
259 * determine that the subbuffer is full.
261 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
263 * Must write slot data before incrementing commit count. This
264 * compiler barrier is upgraded into a smp_mb() by the IPI sent
271 v_add(config
, ctx
->slot_size
, &buf
->commit_hot
[endidx
].cc
);
274 * commit count read can race with concurrent OOO commit count updates.
275 * This is only needed for lib_ring_buffer_check_deliver (for
276 * non-polling delivery only) and for
277 * lib_ring_buffer_write_commit_counter. The race can only cause the
278 * counter to be read with the same value more than once, which could
280 * - Multiple delivery for the same sub-buffer (which is handled
281 * gracefully by the reader code) if the value is for a full
282 * sub-buffer. It's important that we can never miss a sub-buffer
283 * delivery. Re-reading the value after the v_add ensures this.
284 * - Reading a commit_count with a higher value that what was actually
285 * added to it for the lib_ring_buffer_write_commit_counter call
286 * (again caused by a concurrent committer). It does not matter,
287 * because this function is interested in the fact that the commit
288 * count reaches back the reserve offset for a specific sub-buffer,
289 * which is completely independent of the order.
291 commit_count
= v_read(config
, &buf
->commit_hot
[endidx
].cc
);
293 lib_ring_buffer_check_deliver(config
, buf
, chan
, offset_end
- 1,
294 commit_count
, endidx
);
296 * Update used size at each commit. It's needed only for extracting
297 * ring_buffer buffers from vmcore, after crash.
299 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, endidx
,
300 ctx
->buf_offset
, commit_count
,
305 * lib_ring_buffer_try_discard_reserve - Try discarding a record.
306 * @config: ring buffer instance configuration.
307 * @ctx: ring buffer context. (input arguments only)
309 * Only succeeds if no other record has been written after the record to
310 * discard. If discard fails, the record must be committed to the buffer.
312 * Returns 0 upon success, -EPERM if the record cannot be discarded.
315 int lib_ring_buffer_try_discard_reserve(const struct lib_ring_buffer_config
*config
,
316 const struct lib_ring_buffer_ctx
*ctx
)
318 struct lib_ring_buffer
*buf
= ctx
->buf
;
319 unsigned long end_offset
= ctx
->pre_offset
+ ctx
->slot_size
;
322 * We need to ensure that if the cmpxchg succeeds and discards the
323 * record, the next record will record a full TSC, because it cannot
324 * rely on the last_tsc associated with the discarded record to detect
325 * overflows. The only way to ensure this is to set the last_tsc to 0
326 * (assuming no 64-bit TSC overflow), which forces to write a 64-bit
327 * timestamp in the next record.
329 * Note: if discard fails, we must leave the TSC in the record header.
330 * It is needed to keep track of TSC overflows for the following
333 save_last_tsc(config
, buf
, 0ULL);
335 if (likely(v_cmpxchg(config
, &buf
->offset
, end_offset
, ctx
->pre_offset
)
343 void channel_record_disable(const struct lib_ring_buffer_config
*config
,
344 struct channel
*chan
)
346 atomic_inc(&chan
->record_disabled
);
350 void channel_record_enable(const struct lib_ring_buffer_config
*config
,
351 struct channel
*chan
)
353 atomic_dec(&chan
->record_disabled
);
357 void lib_ring_buffer_record_disable(const struct lib_ring_buffer_config
*config
,
358 struct lib_ring_buffer
*buf
)
360 atomic_inc(&buf
->record_disabled
);
364 void lib_ring_buffer_record_enable(const struct lib_ring_buffer_config
*config
,
365 struct lib_ring_buffer
*buf
)
367 atomic_dec(&buf
->record_disabled
);
370 #endif /* _LIB_RING_BUFFER_FRONTEND_API_H */