OpenMPI  0.1.1
btl_sm.h
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2004-2007 The Trustees of Indiana University and Indiana
3  * University Research and Technology
4  * Corporation. All rights reserved.
5  * Copyright (c) 2004-2009 The University of Tennessee and The University
6  * of Tennessee Research Foundation. All rights
7  * reserved.
8  * Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
9  * University of Stuttgart. All rights reserved.
10  * Copyright (c) 2004-2005 The Regents of the University of California.
11  * All rights reserved.
12  * Copyright (c) 2006-2007 Voltaire. All rights reserved.
13  * Copyright (c) 2009-2010 Cisco Systems, Inc. All rights reserved.
14  * Copyright (c) 2010 Los Alamos National Security, LLC.
15  * All rights reserved.
16  * Copyright (c) 2010-2012 IBM Corporation. All rights reserved.
17  * $COPYRIGHT$
18  *
19  * Additional copyrights may follow
20  *
21  * $HEADER$
22  */
23 /**
24  * @file
25  */
26 #ifndef MCA_BTL_SM_H
27 #define MCA_BTL_SM_H
28 
29 #include "ompi_config.h"
30 #include <stddef.h>
31 #include <stdlib.h>
32 #include <string.h>
33 #ifdef HAVE_STDINT_H
34 #include <stdint.h>
35 #endif /* HAVE_STDINT_H */
36 #ifdef HAVE_SCHED_H
37 #include <sched.h>
38 #endif /* HAVE_SCHED_H */
39 #if OMPI_BTL_SM_HAVE_KNEM
40 #include "knem_io.h"
41 #endif /* OMPI_BTL_SM_HAVE_KNEM */
42 
43 #include "opal/util/bit_ops.h"
44 #include "opal/class/opal_free_list.h"
45 #include "ompi/mca/btl/btl.h"
46 #include "ompi/mca/common/sm/common_sm.h"
47 
48 BEGIN_C_DECLS
49 
50 /*
51  * Shared Memory FIFOs
52  *
53  * The FIFO is implemented as a circular queue with head and tail pointers
54  * (integer indices). For efficient wraparound indexing, the size of the
55  * queue is constrained to be a power of two and we "&" indices with a "mask".
56  *
57  * More than one process can write to the FIFO head. Therefore, there is a head
58  * lock. One cannot write until the head slot is empty, indicated by the special
59  * queue entry SM_FIFO_FREE.
60  *
61  * Only the receiver can read the FIFO tail. Therefore, the tail lock is
62  * required only in multithreaded applications. If a tail read returns the
63  * SM_FIFO_FREE value, that means the FIFO is empty. Once a non-FREE value
64  * has been read, the queue slot is *not* automatically reset to SM_FIFO_FREE.
65  * Rather, read tail slots are reset "lazily" (see "lazy_free" and "num_to_clear")
66  * to reduce the number of memory barriers and improve performance.
67  *
68  * Since the FIFO lives in shared memory that is mapped differently into
69  * each address space, the "queue" pointer is relative (each process must
70  * add its own offset) and the queue_recv pointer is meaningful only in the
71  * receiver's address space.
72  *
73  * Since multiple processes access different parts of the FIFO structure in
74  * different ways, we introduce padding to keep different parts on different
75  * cachelines.
76  */
77 
78 #define SM_FIFO_FREE (void *) (-2)
79 /* We can't use opal_cache_line_size here because we need a
80  compile-time constant for padding the struct. We can't really have
81  a compile-time constant that is portable, either (e.g., compile on
82  one machine and run on another). So just use a big enough cache
83  line that should hopefully be good in most places. */
84 #define SM_CACHE_LINE_PAD 128
85 
86 struct sm_fifo_t {
87  /* This queue pointer is used only by the heads. */
88  volatile void **queue;
89  char pad0[SM_CACHE_LINE_PAD - sizeof(void **)];
90  /* This lock is used by the heads. */
91  opal_atomic_lock_t head_lock;
92  char pad1[SM_CACHE_LINE_PAD - sizeof(opal_atomic_lock_t)];
93  /* This index is used by the head holding the head lock. */
94  volatile int head;
95  char pad2[SM_CACHE_LINE_PAD - sizeof(int)];
96  /* This mask is used "read only" by all processes. */
97  unsigned int mask;
98  char pad3[SM_CACHE_LINE_PAD - sizeof(int)];
99  /* The following are used only by the tail. */
100  volatile void **queue_recv;
101  opal_atomic_lock_t tail_lock;
102  volatile int tail;
103  int num_to_clear;
104  int lazy_free;
105  char pad4[SM_CACHE_LINE_PAD - sizeof(void **) -
106  sizeof(opal_atomic_lock_t) -
107  sizeof(int) * 3];
108 };
109 typedef struct sm_fifo_t sm_fifo_t;
110 
111 /*
112  * Shared Memory resource managment
113  */
114 
115 #if OMPI_ENABLE_PROGRESS_THREADS == 1
116 #define DATA (char)0
117 #define DONE (char)1
118 #endif
119 
120 typedef struct mca_btl_sm_mem_node_t {
121  mca_mpool_base_module_t* sm_mpool; /**< shared memory pool */
123 
124 /**
125  * Shared Memory (SM) BTL module.
126  */
128  mca_btl_base_component_2_0_0_t super; /**< base BTL component */
129  int sm_free_list_num; /**< initial size of free lists */
130  int sm_free_list_max; /**< maximum size of free lists */
131  int sm_free_list_inc; /**< number of elements to alloc when growing free lists */
132  int32_t sm_max_procs; /**< upper limit on the number of processes using the shared memory pool */
133  int sm_extra_procs; /**< number of extra procs to allow */
134  char* sm_mpool_name; /**< name of shared memory pool module */
135  mca_mpool_base_module_t **sm_mpools; /**< shared memory pools (one for each memory node) */
136  mca_mpool_base_module_t *sm_mpool; /**< mpool on local node */
137  void* sm_mpool_base; /**< base address of shared memory pool */
138  size_t eager_limit; /**< first fragment size */
139  size_t max_frag_size; /**< maximum (second and beyone) fragment size */
140  opal_mutex_t sm_lock;
141  mca_common_sm_module_t *sm_seg; /**< description of shared memory segment */
142  volatile sm_fifo_t **shm_fifo; /**< pointer to fifo 2D array in shared memory */
143  char **shm_bases; /**< pointer to base pointers in shared memory */
144  uint16_t *shm_mem_nodes; /**< pointer to mem noded in shared memory */
145  sm_fifo_t **fifo; /**< cached copy of the pointer to the 2D
146  fifo array. The address in the shared
147  memory segment sm_ctl_header is a relative,
148  but this one, in process private memory, is
149  a real virtual address */
150  uint16_t *mem_nodes; /**< cached copy of mem nodes of each local rank */
151  size_t fifo_size; /**< number of FIFO queue entries */
152  size_t fifo_lazy_free; /**< number of reads before lazy fifo free is triggered */
153  int nfifos; /**< number of FIFOs per receiver */
154  int32_t num_smp_procs; /**< current number of smp procs on this host */
155  int32_t my_smp_rank; /**< My SMP process rank. Used for accessing
156  * SMP specfic data structures. */
157  ompi_free_list_t sm_frags_eager; /**< free list of sm first */
158  ompi_free_list_t sm_frags_max; /**< free list of sm second */
159  ompi_free_list_t sm_frags_user;
161  fragments that are
162  awaiting resources */
163  struct mca_btl_base_endpoint_t **sm_peers;
164 
165  opal_free_list_t pending_send_fl;
166  int num_outstanding_frags; /**< number of fragments sent but not yet returned to free list */
167  int num_pending_sends; /**< total number on all of my pending-send queues */
168  int mem_node;
169  int num_mem_nodes;
170 
171 #if OMPI_ENABLE_PROGRESS_THREADS == 1
172  char sm_fifo_path[PATH_MAX]; /**< path to fifo used to signal this process */
173  int sm_fifo_fd; /**< file descriptor corresponding to opened fifo */
174  opal_thread_t sm_fifo_thread;
175 #endif
176  struct mca_btl_sm_t **sm_btls;
177  struct mca_btl_sm_frag_t **table;
178  size_t sm_num_btls;
179  size_t sm_max_btls;
180 
181 #if OMPI_BTL_SM_HAVE_KNEM
182  /* Knem capabilities info */
183  struct knem_cmd_info knem_info;
184 #endif
185 
186  /** MCA: should we be using knem or not? neg=try but continue if
187  not available, 0=don't try, 1=try and fail if not available */
188  int use_knem;
189 
190  /** MCA: minimal message size (bytes) to offload on DMA engine
191  when using knem */
192  uint32_t knem_dma_min;
193 
194  /** MCA: how many simultaneous ongoing knem operations to
195  support */
197 
198  /** If we want DMA and DMA is supported, this will be loaded with
199  KNEM_FLAG_DMA. Otherwise, it'll be 0. */
201 
202  /** MCA: should we be using CMA or not?
203  0 = no, 1 = yes */
204  int use_cma;
205 };
207 OMPI_MODULE_DECLSPEC extern mca_btl_sm_component_t mca_btl_sm_component;
208 
209 /**
210  * SM BTL Interface
211  */
212 struct mca_btl_sm_t {
213  mca_btl_base_module_t super; /**< base BTL interface */
214  bool btl_inited; /**< flag indicating if btl has been inited */
216 
217 #if OMPI_BTL_SM_HAVE_KNEM
218 
219  /* File descriptor for knem */
220  int knem_fd;
221 
222  /* Array of knem status items for non-blocking knem requests */
223  knem_status_t *knem_status_array;
224 
225  /* Array of fragments currently being moved by knem non-blocking
226  operations */
227  struct mca_btl_sm_frag_t **knem_frag_array;
228 
229  /* First free/available location in knem_status_array */
230  int knem_status_first_avail;
231 
232  /* First currently-being used location in the knem_status_array */
233  int knem_status_first_used;
234 
235  /* Number of status items currently in use */
236  int knem_status_num_used;
237 #endif
238 };
239 typedef struct mca_btl_sm_t mca_btl_sm_t;
240 OMPI_MODULE_DECLSPEC extern mca_btl_sm_t mca_btl_sm;
241 
242 
243 
244 
245 
247 {
248  opal_free_list_item_t super;
249  void *data;
250 };
252 
253 /***
254  * FIFO support for sm BTL.
255  */
256 
257 /***
258  * One or more FIFO components may be a pointer that must be
259  * accessed by multiple processes. Since the shared region may
260  * be mmapped differently into each process's address space,
261  * these pointers will be relative to some base address. Here,
262  * we define macros to translate between relative addresses and
263  * virtual addresses.
264  */
265 #define VIRTUAL2RELATIVE(VADDR ) ((long)(VADDR) - (long)mca_btl_sm_component.shm_bases[mca_btl_sm_component.my_smp_rank])
266 #define RELATIVE2VIRTUAL(OFFSET) ((long)(OFFSET) + (long)mca_btl_sm_component.shm_bases[mca_btl_sm_component.my_smp_rank])
267 
268 static inline int sm_fifo_init(int fifo_size, mca_mpool_base_module_t *mpool,
269  sm_fifo_t *fifo, int lazy_free)
270 {
271  int i, qsize;
272 
273  /* figure out the queue size (a power of two that is at least 1) */
274  qsize = opal_next_poweroftwo_inclusive (fifo_size);
275 
276  /* allocate the queue in the receiver's address space */
277  fifo->queue_recv = (volatile void **)mpool->mpool_alloc(
278  mpool, sizeof(void *) * qsize, opal_cache_line_size, 0, NULL);
279  if(NULL == fifo->queue_recv) {
280  return OMPI_ERR_OUT_OF_RESOURCE;
281  }
282 
283  /* initialize the queue */
284  for ( i = 0; i < qsize; i++ )
285  fifo->queue_recv[i] = SM_FIFO_FREE;
286 
287  /* shift queue address to be relative */
288  fifo->queue = (volatile void **) VIRTUAL2RELATIVE(fifo->queue_recv);
289 
290  /* initialize the locks */
291  opal_atomic_init(&(fifo->head_lock), OPAL_ATOMIC_UNLOCKED);
292  opal_atomic_init(&(fifo->tail_lock), OPAL_ATOMIC_UNLOCKED);
293  opal_atomic_unlock(&(fifo->head_lock)); /* should be unnecessary */
294  opal_atomic_unlock(&(fifo->tail_lock)); /* should be unnecessary */
295 
296  /* other initializations */
297  fifo->head = 0;
298  fifo->mask = qsize - 1;
299  fifo->tail = 0;
300  fifo->num_to_clear = 0;
301  fifo->lazy_free = lazy_free;
302 
303  return OMPI_SUCCESS;
304 }
305 
306 
307 static inline int sm_fifo_write(void *value, sm_fifo_t *fifo)
308 {
309  volatile void **q = (volatile void **) RELATIVE2VIRTUAL(fifo->queue);
310 
311  /* if there is no free slot to write, report exhausted resource */
312  opal_atomic_rmb();
313  if ( SM_FIFO_FREE != q[fifo->head] )
314  return OMPI_ERR_OUT_OF_RESOURCE;
315 
316  /* otherwise, write to the slot and advance the head index */
317  q[fifo->head] = value;
318  opal_atomic_wmb();
319  fifo->head = (fifo->head + 1) & fifo->mask;
320  return OMPI_SUCCESS;
321 }
322 
323 
324 static inline void *sm_fifo_read(sm_fifo_t *fifo)
325 {
326  void *value;
327 
328  /* read the next queue entry */
329  value = (void *) fifo->queue_recv[fifo->tail];
330 
331  opal_atomic_rmb();
332 
333  /* if you read a non-empty slot, advance the tail pointer */
334  if ( SM_FIFO_FREE != value ) {
335 
336  fifo->tail = ( fifo->tail + 1 ) & fifo->mask;
337  fifo->num_to_clear += 1;
338 
339  /* check if it's time to free slots, which we do lazily */
340  if ( fifo->num_to_clear >= fifo->lazy_free ) {
341  int i = (fifo->tail - fifo->num_to_clear ) & fifo->mask;
342 
343  while ( fifo->num_to_clear > 0 ) {
344  fifo->queue_recv[i] = SM_FIFO_FREE;
345  i = (i+1) & fifo->mask;
346  fifo->num_to_clear -= 1;
347  }
348  opal_atomic_wmb();
349  }
350  }
351 
352  return value;
353 }
354 
355 /**
356  * shared memory component progress.
357  */
358 extern int mca_btl_sm_component_progress(void);
359 
360 
361 
362 /**
363  * Register a callback function that is called on error..
364  *
365  * @param btl (IN) BTL module
366  * @return Status indicating if cleanup was successful
367  */
368 
370  struct mca_btl_base_module_t* btl,
372 );
373 
374 /**
375  * Cleanup any resources held by the BTL.
376  *
377  * @param btl BTL instance.
378  * @return OMPI_SUCCESS or error status on failure.
379  */
380 
381 extern int mca_btl_sm_finalize(
382  struct mca_btl_base_module_t* btl
383 );
384 
385 
386 /**
387  * PML->BTL notification of change in the process list.
388  * PML->BTL Notification that a receive fragment has been matched.
389  * Called for message that is send from process with the virtual
390  * address of the shared memory segment being different than that of
391  * the receiver.
392  *
393  * @param btl (IN)
394  * @param proc (IN)
395  * @param peer (OUT)
396  * @return OMPI_SUCCESS or error status on failure.
397  *
398  */
399 
400 extern int mca_btl_sm_add_procs(
401  struct mca_btl_base_module_t* btl,
402  size_t nprocs,
403  struct ompi_proc_t **procs,
404  struct mca_btl_base_endpoint_t** peers,
405  struct opal_bitmap_t* reachability
406 );
407 
408 
409 /**
410  * PML->BTL notification of change in the process list.
411  *
412  * @param btl (IN) BTL instance
413  * @param proc (IN) Peer process
414  * @param peer (IN) Peer addressing information.
415  * @return Status indicating if cleanup was successful
416  *
417  */
418 extern int mca_btl_sm_del_procs(
419  struct mca_btl_base_module_t* btl,
420  size_t nprocs,
421  struct ompi_proc_t **procs,
422  struct mca_btl_base_endpoint_t **peers
423 );
424 
425 
426 /**
427  * Allocate a segment.
428  *
429  * @param btl (IN) BTL module
430  * @param size (IN) Request segment size.
431  */
433  struct mca_btl_base_module_t* btl,
434  struct mca_btl_base_endpoint_t* endpoint,
435  uint8_t order,
436  size_t size,
437  uint32_t flags
438 );
439 
440 /**
441  * Return a segment allocated by this BTL.
442  *
443  * @param btl (IN) BTL module
444  * @param segment (IN) Allocated segment.
445  */
446 extern int mca_btl_sm_free(
447  struct mca_btl_base_module_t* btl,
449 );
450 
451 
452 /**
453  * Pack data
454  *
455  * @param btl (IN) BTL module
456  * @param peer (IN) BTL peer addressing
457  */
459  struct mca_btl_base_module_t* btl,
460  struct mca_btl_base_endpoint_t* endpoint,
461  mca_mpool_base_registration_t* registration,
462  struct opal_convertor_t* convertor,
463  uint8_t order,
464  size_t reserve,
465  size_t* size,
466  uint32_t flags
467 );
468 
469 
470 /**
471  * Initiate an inlined send to the peer or return a descriptor.
472  *
473  * @param btl (IN) BTL module
474  * @param peer (IN) BTL peer addressing
475  */
476 extern int mca_btl_sm_sendi( struct mca_btl_base_module_t* btl,
477  struct mca_btl_base_endpoint_t* endpoint,
478  struct opal_convertor_t* convertor,
479  void* header,
480  size_t header_size,
481  size_t payload_size,
482  uint8_t order,
483  uint32_t flags,
484  mca_btl_base_tag_t tag,
485  mca_btl_base_descriptor_t** descriptor );
486 
487 /**
488  * Initiate a send to the peer.
489  *
490  * @param btl (IN) BTL module
491  * @param peer (IN) BTL peer addressing
492  */
493 extern int mca_btl_sm_send(
494  struct mca_btl_base_module_t* btl,
495  struct mca_btl_base_endpoint_t* endpoint,
496  struct mca_btl_base_descriptor_t* descriptor,
497  mca_btl_base_tag_t tag
498 );
499 
500 #if OMPI_BTL_SM_HAVE_KNEM || OMPI_BTL_SM_HAVE_CMA
501 /*
502  * Synchronous knem/cma get
503  */
504 extern int mca_btl_sm_get_sync(
505  struct mca_btl_base_module_t* btl,
506  struct mca_btl_base_endpoint_t* endpoint,
507  struct mca_btl_base_descriptor_t* des );
508 
509 extern struct mca_btl_base_descriptor_t* mca_btl_sm_prepare_dst(
510  struct mca_btl_base_module_t* btl,
511  struct mca_btl_base_endpoint_t* endpoint,
512  struct mca_mpool_base_registration_t* registration,
513  struct opal_convertor_t* convertor,
514  uint8_t order,
515  size_t reserve,
516  size_t* size,
517  uint32_t flags);
518 #endif
519 
520 #if OMPI_BTL_SM_HAVE_KNEM
521 /*
522  * Asynchronous knem get
523  */
524 extern int mca_btl_sm_get_async(
525  struct mca_btl_base_module_t* btl,
526  struct mca_btl_base_endpoint_t* endpoint,
527  struct mca_btl_base_descriptor_t* des );
528 
529 #endif
530 
531 /**
532  * Fault Tolerance Event Notification Function
533  * @param state Checkpoint Stae
534  * @return OMPI_SUCCESS or failure status
535  */
536 int mca_btl_sm_ft_event(int state);
537 
538 #if OMPI_ENABLE_PROGRESS_THREADS == 1
539 void mca_btl_sm_component_event_thread(opal_object_t*);
540 #endif
541 
542 #if OMPI_ENABLE_PROGRESS_THREADS == 1
543 #define MCA_BTL_SM_SIGNAL_PEER(peer) \
544 { \
545  unsigned char cmd = DATA; \
546  if(write(peer->fifo_fd, &cmd, sizeof(cmd)) != sizeof(cmd)) { \
547  opal_output(0, "mca_btl_sm_send: write fifo failed: errno=%d\n", errno); \
548  } \
549 }
550 #else
551 #define MCA_BTL_SM_SIGNAL_PEER(peer)
552 #endif
553 
554 END_C_DECLS
555 
556 #endif
557 
Shared Memory (SM) BTL module.
Definition: btl_sm.h:127
char ** shm_bases
pointer to base pointers in shared memory
Definition: btl_sm.h:143
size_t max_frag_size
maximum (second and beyone) fragment size
Definition: btl_sm.h:139
int32_t my_smp_rank
My SMP process rank.
Definition: btl_sm.h:155
int32_t sm_max_procs
upper limit on the number of processes using the shared memory pool
Definition: btl_sm.h:132
int knem_max_simultaneous
MCA: how many simultaneous ongoing knem operations to support.
Definition: btl_sm.h:196
bool btl_inited
flag indicating if btl has been inited
Definition: btl_sm.h:214
A descriptor that holds the parameters to a send/put/get operation along w/ a callback routine that i...
Definition: btl.h:275
Definition: btl_sm.h:120
uint16_t * shm_mem_nodes
pointer to mem noded in shared memory
Definition: btl_sm.h:144
int sm_free_list_max
maximum size of free lists
Definition: btl_sm.h:130
int mca_btl_sm_ft_event(int state)
Fault Tolerance Event Notification Function.
Definition: btl_sm.c:1158
void opal_atomic_rmb(void)
Read memory barrier.
Definition: opal_bitmap.h:53
Definition: common_sm.h:60
int sm_free_list_inc
number of elements to alloc when growing free lists
Definition: btl_sm.h:131
void(* mca_btl_base_module_error_cb_fn_t)(struct mca_btl_base_module_t *btl, int32_t flags, struct ompi_proc_t *errproc, char *btlinfo)
Callback function that is called asynchronously on receipt of an error from the transport layer...
Definition: btl.h:538
Definition: mutex_unix.h:53
sm_fifo_t ** fifo
cached copy of the pointer to the 2D fifo array.
Definition: btl_sm.h:145
mca_btl_base_module_t super
base BTL interface
Definition: btl_sm.h:213
struct mca_btl_base_descriptor_t * mca_btl_sm_prepare_src(struct mca_btl_base_module_t *btl, struct mca_btl_base_endpoint_t *endpoint, mca_mpool_base_registration_t *registration, struct opal_convertor_t *convertor, uint8_t order, size_t reserve, size_t *size, uint32_t flags)
Pack data.
Definition: btl_sm.c:678
int sm_free_list_num
initial size of free lists
Definition: btl_sm.h:129
Definition: mpool.h:44
Remote Open MPI process structure.
Definition: proc.h:56
volatile sm_fifo_t ** shm_fifo
pointer to fifo 2D array in shared memory
Definition: btl_sm.h:142
Volatile lock object (with optional padding).
Definition: atomic.h:102
ompi_free_list_t sm_frags_max
free list of sm second
Definition: btl_sm.h:158
static void opal_atomic_unlock(opal_atomic_lock_t *lock)
Release a lock.
int mca_btl_sm_sendi(struct mca_btl_base_module_t *btl, struct mca_btl_base_endpoint_t *endpoint, struct opal_convertor_t *convertor, void *header, size_t header_size, size_t payload_size, uint8_t order, uint32_t flags, mca_btl_base_tag_t tag, mca_btl_base_descriptor_t **descriptor)
Initiate an inlined send to the peer or return a descriptor.
Definition: btl_sm.c:815
void * sm_mpool_base
base address of shared memory pool
Definition: btl_sm.h:137
int mca_btl_sm_register_error_cb(struct mca_btl_base_module_t *btl, mca_btl_base_module_error_cb_fn_t cbfunc)
Register a callback function that is called on error.
Definition: btl_sm.c:619
char * sm_mpool_name
name of shared memory pool module
Definition: btl_sm.h:134
Definition: opal_free_list.h:47
mca_btl_base_component_2_0_0_t super
base BTL component
Definition: btl_sm.h:128
size_t fifo_lazy_free
number of reads before lazy fifo free is triggered
Definition: btl_sm.h:152
size_t fifo_size
number of FIFO queue entries
Definition: btl_sm.h:151
int use_knem
MCA: should we be using knem or not? neg=try but continue if not available, 0=don't try...
Definition: btl_sm.h:188
mca_common_sm_module_t * sm_seg
description of shared memory segment
Definition: btl_sm.h:141
size_t eager_limit
first fragment size
Definition: btl_sm.h:138
Definition: btl_sm.h:246
Definition: opal_free_list.h:31
int mca_btl_sm_add_procs(struct mca_btl_base_module_t *btl, size_t nprocs, struct ompi_proc_t **procs, struct mca_btl_base_endpoint_t **peers, struct opal_bitmap_t *reachability)
PML->BTL notification of change in the process list.
Definition: btl_sm.c:436
ompi_free_list_t sm_frags_eager
free list of sm first
Definition: btl_sm.h:157
Byte Transfer Layer (BTL)
SM BTL Interface.
Definition: btl_sm.h:212
int mca_btl_sm_component_progress(void)
shared memory component progress.
Definition: btl_sm_component.c:608
int mca_btl_sm_send(struct mca_btl_base_module_t *btl, struct mca_btl_base_endpoint_t *endpoint, struct mca_btl_base_descriptor_t *descriptor, mca_btl_base_tag_t tag)
Initiate a send to the peer.
Definition: btl_sm.c:900
shared memory send fragment derived type.
Definition: btl_sm_frag.h:51
Definition: ompi_free_list.h:39
int mca_btl_sm_del_procs(struct mca_btl_base_module_t *btl, size_t nprocs, struct ompi_proc_t **procs, struct mca_btl_base_endpoint_t **peers)
PML->BTL notification of change in the process list.
Definition: btl_sm.c:587
int knem_dma_flag
If we want DMA and DMA is supported, this will be loaded with KNEM_FLAG_DMA.
Definition: btl_sm.h:200
State of ELAN endpoint connection.
Definition: btl_elan_endpoint.h:33
BTL component descriptor.
Definition: btl.h:411
Base object.
Definition: opal_object.h:182
Definition: opal_convertor.h:90
int num_pending_sends
total number on all of my pending-send queues
Definition: btl_sm.h:167
uint8_t order
order value, this is only valid in the local completion callback and may be used in subsequent calls ...
Definition: btl.h:292
uint16_t * mem_nodes
cached copy of mem nodes of each local rank
Definition: btl_sm.h:150
void opal_atomic_wmb(void)
Write memory barrier.
mca_btl_base_descriptor_t * mca_btl_sm_alloc(struct mca_btl_base_module_t *btl, struct mca_btl_base_endpoint_t *endpoint, uint8_t order, size_t size, uint32_t flags)
Allocate a segment.
Definition: btl_sm.c:634
static void opal_atomic_init(opal_atomic_lock_t *lock, int32_t value)
Initialize a lock to value.
int mca_btl_sm_free(struct mca_btl_base_module_t *btl, mca_btl_base_descriptor_t *segment)
Return a segment allocated by this BTL.
Definition: btl_sm.c:662
int nfifos
number of FIFOs per receiver
Definition: btl_sm.h:153
int num_outstanding_frags
number of fragments sent but not yet returned to free list
Definition: btl_sm.h:166
Definition: threads.h:46
ompi_free_list_t sm_first_frags_to_progress
list of first fragments that are awaiting resources
Definition: btl_sm.h:160
mca_mpool_base_module_alloc_fn_t mpool_alloc
allocate function
Definition: mpool.h:177
Definition: btl_sm.h:86
int sm_extra_procs
number of extra procs to allow
Definition: btl_sm.h:133
BTL module interface functions and attributes.
Definition: btl.h:786
uint32_t knem_dma_min
MCA: minimal message size (bytes) to offload on DMA engine when using knem.
Definition: btl_sm.h:192
int mca_btl_sm_finalize(struct mca_btl_base_module_t *btl)
Cleanup any resources held by the BTL.
Definition: btl_sm.c:610
mca_mpool_base_module_t * sm_mpool
shared memory pool
Definition: btl_sm.h:121
int32_t num_smp_procs
current number of smp procs on this host
Definition: btl_sm.h:154
mca_mpool_base_module_t ** sm_mpools
shared memory pools (one for each memory node)
Definition: btl_sm.h:135
int use_cma
MCA: should we be using CMA or not? 0 = no, 1 = yes.
Definition: btl_sm.h:204
mca_mpool_base_module_t * sm_mpool
mpool on local node
Definition: btl_sm.h:136
mpool module descriptor.
Definition: mpool.h:174