FD.io VPP  v17.07-30-g839fa73
Vector Packet Processing
threads.h
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2015 Cisco and/or its affiliates.
3  * Licensed under the Apache License, Version 2.0 (the "License");
4  * you may not use this file except in compliance with the License.
5  * You may obtain a copy of the License at:
6  *
7  * http://www.apache.org/licenses/LICENSE-2.0
8  *
9  * Unless required by applicable law or agreed to in writing, software
10  * distributed under the License is distributed on an "AS IS" BASIS,
11  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12  * See the License for the specific language governing permissions and
13  * limitations under the License.
14  */
15 #ifndef included_vlib_threads_h
16 #define included_vlib_threads_h
17 
18 #include <vlib/main.h>
19 #include <linux/sched.h>
20 
21 extern vlib_main_t **vlib_mains;
22 
23 void vlib_set_thread_name (char *name);
24 
25 /* arg is actually a vlib__thread_t * */
26 typedef void (vlib_thread_function_t) (void *arg);
27 
29 {
30  /* constructor generated list of thread registrations */
32 
33  /* config parameters */
34  char *name;
35  char *short_name;
42 
43  /* All threads of this type run on pthreads */
48 
49 /*
50  * Frames have their cpu / vlib_main_t index in the low-order N bits
51  * Make VLIB_MAX_CPUS a power-of-two, please...
52  */
53 
54 #ifndef VLIB_MAX_CPUS
55 #define VLIB_MAX_CPUS 256
56 #endif
57 
58 #if VLIB_MAX_CPUS > CLIB_MAX_MHEAPS
59 #error Please increase number of per-cpu mheaps
60 #endif
61 
62 #define VLIB_CPU_MASK (VLIB_MAX_CPUS - 1) /* 0x3f, max */
63 #define VLIB_OFFSET_MASK (~VLIB_CPU_MASK)
64 
65 #define VLIB_LOG2_THREAD_STACK_SIZE (21)
66 #define VLIB_THREAD_STACK_SIZE (1<<VLIB_LOG2_THREAD_STACK_SIZE)
67 
68 typedef enum
69 {
72 
73 typedef struct
74 {
75  CLIB_CACHE_LINE_ALIGN_MARK (cacheline0);
76  volatile u32 valid;
80 
81  /* 256 * 4 = 1024 bytes, even mult of cache line size */
82  u32 buffer_index[VLIB_FRAME_SIZE];
83 }
85 
86 typedef struct
87 {
88  /* First cache line */
89  CLIB_CACHE_LINE_ALIGN_MARK (cacheline0);
90  volatile u32 *wait_at_barrier;
92 
93  /* Second Cache Line */
94  CLIB_CACHE_LINE_ALIGN_MARK (cacheline1);
95  void *thread_mheap;
97  void (*thread_function) (void *);
105 
106  long lwp;
107  int lcore_id;
108  pthread_t thread_id;
110 
112 
113 typedef struct
114 {
115  /* enqueue side */
116  CLIB_CACHE_LINE_ALIGN_MARK (cacheline0);
117  volatile u64 tail;
122 
123  /* dequeue side */
124  CLIB_CACHE_LINE_ALIGN_MARK (cacheline1);
125  volatile u64 head;
131 
132  /* dequeue hint to enqueue side */
133  CLIB_CACHE_LINE_ALIGN_MARK (cacheline2);
134  volatile u64 head_hint;
135 
136  /* read-only, constant, shared */
137  CLIB_CACHE_LINE_ALIGN_MARK (cacheline3);
140 }
142 
143 typedef struct
144 {
147 
148  /* for frame queue tracing */
152 
153 /* Called early, in thread 0's context */
155 
156 int vlib_frame_queue_enqueue (vlib_main_t * vm, u32 node_runtime_index,
157  u32 frame_queue_index, vlib_frame_t * frame,
158  vlib_frame_queue_msg_type_t type);
159 
160 int
162 
164 
165 void vlib_create_worker_threads (vlib_main_t * vm, int n,
166  void (*thread_function) (void *));
167 
170 
171 /* Check for a barrier sync request every 30ms */
172 #define BARRIER_SYNC_DELAY (0.030000)
173 
174 #if CLIB_DEBUG > 0
175 /* long barrier timeout, for gdb... */
176 #define BARRIER_SYNC_TIMEOUT (600.1)
177 #else
178 #define BARRIER_SYNC_TIMEOUT (1.0)
179 #endif
180 
183 
186 {
187  return __os_thread_index;
188 }
189 
190 always_inline void
192 {
193  if (CLIB_DEBUG > 0)
194  {
195  if (vlib_get_thread_index ())
196  fformat (stderr, "%s: SMP unsafe warning...\n", __FUNCTION__);
197  }
198 }
199 
200 typedef enum
201 {
205 
206 void vlib_worker_thread_fork_fixup (vlib_fork_fixup_t which);
207 
208 #define foreach_vlib_main(body) \
209 do { \
210  vlib_main_t ** __vlib_mains = 0, *this_vlib_main; \
211  int ii; \
212  \
213  for (ii = 0; ii < vec_len (vlib_mains); ii++) \
214  { \
215  this_vlib_main = vlib_mains[ii]; \
216  ASSERT (ii == 0 || \
217  this_vlib_main->parked_at_barrier == 1); \
218  if (this_vlib_main) \
219  vec_add1 (__vlib_mains, this_vlib_main); \
220  } \
221  \
222  for (ii = 0; ii < vec_len (__vlib_mains); ii++) \
223  { \
224  this_vlib_main = __vlib_mains[ii]; \
225  /* body uses this_vlib_main... */ \
226  (body); \
227  } \
228  vec_free (__vlib_mains); \
229 } while (0);
230 
231 #define foreach_sched_policy \
232  _(SCHED_OTHER, OTHER, "other") \
233  _(SCHED_BATCH, BATCH, "batch") \
234  _(SCHED_IDLE, IDLE, "idle") \
235  _(SCHED_FIFO, FIFO, "fifo") \
236  _(SCHED_RR, RR, "rr")
237 
238 typedef enum
239 {
240 #define _(v,f,s) SCHED_POLICY_##f = v,
242 #undef _
245 
246 typedef struct
247 {
248  clib_error_t *(*vlib_launch_thread_cb) (void *fp, vlib_worker_thread_t * w,
249  unsigned lcore_id);
250  clib_error_t *(*vlib_thread_set_lcore_cb) (u32 thread, u16 lcore);
251 } vlib_thread_callbacks_t;
252 
253 typedef struct
254 {
255  /* Link list of registrations, built by constructors */
257 
258  /* Vector of registrations, w/ non-data-structure clones at the top */
260 
262 
264 
265  /*
266  * Launch all threads as pthreads,
267  * not eal_rte_launch (strict affinity) threads
268  */
270 
271  /* Number of vlib_main / vnet_main clones */
273 
274  /* Number of thread stacks to create */
276 
277  /* Number of pthreads */
279 
280  /* Number of threads */
282 
283  /* Number of cores to skip, must match the core mask */
285 
286  /* Thread prefix name */
288 
289  /* main thread lcore */
291 
292  /* Bitmap of available CPU cores */
294 
295  /* Bitmap of available CPU sockets (NUMA nodes) */
297 
298  /* Worker handoff queues */
300 
301  /* worker thread initialization barrier */
303 
304  /* scheduling policy */
306 
307  /* scheduling policy priority */
309 
310  /* callbacks */
311  vlib_thread_callbacks_t cb;
314 
316 
317 #include <vlib/global_funcs.h>
318 
319 #define VLIB_REGISTER_THREAD(x,...) \
320  __VA_ARGS__ vlib_thread_registration_t x; \
321 static void __vlib_add_thread_registration_##x (void) \
322  __attribute__((__constructor__)) ; \
323 static void __vlib_add_thread_registration_##x (void) \
324 { \
325  vlib_thread_main_t * tm = &vlib_thread_main; \
326  x.next = tm->next; \
327  tm->next = &x; \
328 } \
329 __VA_ARGS__ vlib_thread_registration_t x
330 
333 {
334  return vlib_thread_main.n_vlib_mains - 1;
335 }
336 
339 {
340  return worker_index + 1;
341 }
342 
345 {
346  return thread_index - 1;
347 }
348 
351 {
352  return vlib_get_thread_index () - 1;
353 }
354 
355 static inline void
357 {
358  if (PREDICT_FALSE (*vlib_worker_threads->wait_at_barrier))
359  {
360  vlib_main_t *vm;
361  clib_smp_atomic_add (vlib_worker_threads->workers_at_barrier, 1);
362  if (CLIB_DEBUG > 0)
363  {
364  vm = vlib_get_main ();
365  vm->parked_at_barrier = 1;
366  }
367  while (*vlib_worker_threads->wait_at_barrier)
368  ;
369  if (CLIB_DEBUG > 0)
370  vm->parked_at_barrier = 0;
371  clib_smp_atomic_add (vlib_worker_threads->workers_at_barrier, -1);
372  }
373 }
374 
377 {
378  vlib_main_t *vm;
380  ASSERT (worker_index < tm->n_vlib_mains - 1);
381  vm = vlib_mains[worker_index + 1];
382  ASSERT (vm);
383  return vm;
384 }
385 
386 static inline void
388 {
390  hf->valid = 1;
391 }
392 
393 static inline vlib_frame_queue_elt_t *
394 vlib_get_frame_queue_elt (u32 frame_queue_index, u32 index)
395 {
396  vlib_frame_queue_t *fq;
400  vec_elt_at_index (tm->frame_queue_mains, frame_queue_index);
401  u64 new_tail;
402 
403  fq = fqm->vlib_frame_queues[index];
404  ASSERT (fq);
405 
406  new_tail = __sync_add_and_fetch (&fq->tail, 1);
407 
408  /* Wait until a ring slot is available */
409  while (new_tail >= fq->head_hint + fq->nelts)
411 
412  elt = fq->elts + (new_tail & (fq->nelts - 1));
413 
414  /* this would be very bad... */
415  while (elt->valid)
416  ;
417 
419  elt->last_n_vectors = elt->n_vectors = 0;
420 
421  return elt;
422 }
423 
424 static inline vlib_frame_queue_t *
426  u32 index,
427  u32 queue_hi_thresh,
429  handoff_queue_by_worker_index)
430 {
431  vlib_frame_queue_t *fq;
434  vec_elt_at_index (tm->frame_queue_mains, frame_queue_index);
435 
436  fq = handoff_queue_by_worker_index[index];
437  if (fq != (vlib_frame_queue_t *) (~0))
438  return fq;
439 
440  fq = fqm->vlib_frame_queues[index];
441  ASSERT (fq);
442 
443  if (PREDICT_FALSE (fq->tail >= (fq->head_hint + queue_hi_thresh)))
444  {
445  /* a valid entry in the array will indicate the queue has reached
446  * the specified threshold and is congested
447  */
448  handoff_queue_by_worker_index[index] = fq;
449  fq->enqueue_full_events++;
450  return fq;
451  }
452 
453  return NULL;
454 }
455 
456 static inline vlib_frame_queue_elt_t *
458  u32 vlib_worker_index,
460  handoff_queue_elt_by_worker_index)
461 {
463 
464  if (handoff_queue_elt_by_worker_index[vlib_worker_index])
465  return handoff_queue_elt_by_worker_index[vlib_worker_index];
466 
467  elt = vlib_get_frame_queue_elt (frame_queue_index, vlib_worker_index);
468 
469  handoff_queue_elt_by_worker_index[vlib_worker_index] = elt;
470 
471  return elt;
472 }
473 
474 u8 *vlib_thread_stack_init (uword thread_index);
475 
476 int vlib_thread_cb_register (struct vlib_main_t *vm,
477  vlib_thread_callbacks_t * cb);
478 
479 #endif /* included_vlib_threads_h */
480 
481 /*
482  * fd.io coding-style-patch-verification: ON
483  *
484  * Local Variables:
485  * eval: (c-set-style "gnu")
486  * End:
487  */
#define CLIB_CACHE_LINE_ALIGN_MARK(mark)
Definition: cache.h:68
void vlib_worker_thread_init(vlib_worker_thread_t *w)
Definition: threads.c:417
int vlib_frame_queue_enqueue(vlib_main_t *vm, u32 node_runtime_index, u32 frame_queue_index, vlib_frame_t *frame, vlib_frame_queue_msg_type_t type)
#define NULL
Definition: clib.h:55
void vlib_set_thread_name(char *name)
Definition: threads.c:51
void * thread_function_arg
Definition: threads.h:98
frame_queue_trace_t * frame_queue_traces
Definition: threads.h:149
elog_track_t elog_track
Definition: threads.h:100
vlib_worker_thread_t * vlib_worker_threads
Definition: threads.c:35
struct vlib_thread_registration_ * next
Definition: threads.h:31
int vlib_thread_cb_register(struct vlib_main_t *vm, vlib_thread_callbacks_t *cb)
Definition: threads.c:1417
volatile u32 valid
Definition: threads.h:76
static vlib_frame_queue_t * is_vlib_frame_queue_congested(u32 frame_queue_index, u32 index, u32 queue_hi_thresh, vlib_frame_queue_t **handoff_queue_by_worker_index)
Definition: threads.h:425
static void vlib_smp_unsafe_warning(void)
Definition: threads.h:191
volatile int parked_at_barrier
Definition: main.h:179
static void vlib_worker_thread_barrier_check(void)
Definition: threads.h:356
vlib_thread_registration_t * next
Definition: threads.h:256
static u32 vlib_get_worker_index(u32 thread_index)
Definition: threads.h:344
#define static_always_inline
Definition: clib.h:85
#define clib_smp_atomic_add(addr, increment)
Definition: smp.h:46
#define always_inline
Definition: clib.h:84
vlib_frame_queue_msg_type_t
Definition: threads.h:68
#define vec_elt_at_index(v, i)
Get vector value at index i checking that i is in bounds.
uword * cpu_core_bitmap
Definition: threads.h:293
unsigned long u64
Definition: types.h:89
int vlib_frame_queue_dequeue(vlib_main_t *vm, vlib_frame_queue_main_t *fqm)
Definition: threads.c:1228
vlib_frame_queue_elt_t * elts
Definition: threads.h:138
static u32 vlib_get_current_worker_index()
Definition: threads.h:350
vlib_fork_fixup_t
Definition: threads.h:200
int extern_thread_mgmt
Definition: threads.h:312
sched_policy_t
Definition: threads.h:238
void * thread_mheap
Definition: threads.h:95
volatile u64 head
Definition: threads.h:125
u8 * vlib_thread_stack_init(uword thread_index)
Definition: main.c:521
vlib_thread_callbacks_t cb
Definition: threads.h:311
void vlib_worker_thread_barrier_release(vlib_main_t *vm)
Definition: threads.c:1198
#define PREDICT_FALSE(x)
Definition: clib.h:97
static vlib_frame_queue_elt_t * vlib_get_worker_handoff_queue_elt(u32 frame_queue_index, u32 vlib_worker_index, vlib_frame_queue_elt_t **handoff_queue_elt_by_worker_index)
Definition: threads.h:457
#define VLIB_FRAME_SIZE
Definition: node.h:329
long i64
Definition: types.h:82
word fformat(FILE *f, char *fmt,...)
Definition: format.c:453
void vlib_worker_thread_node_runtime_update(void)
Definition: threads.c:781
void vlib_worker_thread_fork_fixup(vlib_fork_fixup_t which)
Definition: threads.c:1143
volatile u64 tail
Definition: threads.h:117
vlib_frame_queue_t ** vlib_frame_queues
Definition: threads.h:146
static_always_inline uword vlib_get_thread_index(void)
Definition: threads.h:185
frame_queue_nelt_counter_t * frame_queue_histogram
Definition: threads.h:150
vlib_worker_thread_t * worker_threads
Definition: threads.h:263
static vlib_frame_queue_elt_t * vlib_get_frame_queue_elt(u32 frame_queue_index, u32 index)
Definition: threads.h:394
volatile u32 * wait_at_barrier
Definition: threads.h:90
#define ASSERT(truth)
unsigned int u32
Definition: types.h:88
vlib_frame_queue_main_t * frame_queue_mains
Definition: threads.h:299
struct vlib_thread_registration_ vlib_thread_registration_t
uword * thread_registrations_by_name
Definition: threads.h:261
u32 vlib_frame_queue_main_init(u32 node_index, u32 frame_queue_nelts)
Definition: threads.c:1390
static vlib_main_t * vlib_get_main(void)
Definition: global_funcs.h:23
clib_error_t * vlib_thread_init(vlib_main_t *vm)
Definition: threads.c:106
u64 uword
Definition: types.h:112
u32 enqueue_full_events
Definition: threads.h:121
unsigned short u16
Definition: types.h:57
unsigned char u8
Definition: types.h:56
vlib_main_t ** vlib_mains
void vlib_create_worker_threads(vlib_main_t *vm, int n, void(*thread_function)(void *))
volatile u64 head_hint
Definition: threads.h:134
static void vlib_put_frame_queue_elt(vlib_frame_queue_elt_t *hf)
Definition: threads.h:387
void vlib_worker_thread_barrier_sync(vlib_main_t *vm)
Definition: threads.c:1166
volatile u32 * workers_at_barrier
Definition: threads.h:91
static u32 vlib_num_workers()
Definition: threads.h:332
#define CLIB_MEMORY_BARRIER()
Definition: clib.h:101
uword * cpu_socket_bitmap
Definition: threads.h:296
#define foreach_sched_policy
Definition: threads.h:231
void( vlib_thread_function_t)(void *arg)
Definition: threads.h:26
vlib_thread_registration_t ** registrations
Definition: threads.h:259
vlib_thread_main_t vlib_thread_main
Definition: threads.c:36
pthread_t thread_id
Definition: threads.h:108
vlib_thread_registration_t * registration
Definition: threads.h:102
volatile u32 worker_thread_release
Definition: threads.h:302
static vlib_main_t * vlib_get_worker_vlib_main(u32 worker_index)
Definition: threads.h:376
static u32 vlib_get_worker_thread_index(u32 worker_index)
Definition: threads.h:338