ports/openmpi
ports/openmpi
1
1
Форкнуть 0
Код Релизы Активность
077b3df7cc
openmpi/ompi/mca/crcp/bkmrk/crcp_bkmrk_pml.c

6480 строки
248 KiB
C
Исходник Обычный вид История

Replace the old coordinated component ('coord') and replace it with a much more refined version ('bkmrk'). The new component fixes a number of problems with the old component. The core algorithm is the same, but by changing the data strucutres a bit we have improved performance and memory utilization. There are still a couple corner cases that still need some work. However, I did not want to delay bringing this into the trunk (and v1.3 branch) for too much longer. This commit was SVN r19537.
2008-09-10 22:29:17 +04:00
/*
* Copyright (c) 2004-2008 The Trustees of Indiana University.
* All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
/*
*
*/
#include "ompi_config.h"
#include <sys/types.h>
#ifdef HAVE_UNISTD_H
#include <unistd.h>
#endif /* HAVE_UNIST_H */
#include "opal/runtime/opal_cr.h"
#include "opal/event/event.h"
#include "opal/util/output.h"
#include "opal/util/argv.h"
#include "opal/util/opal_environ.h"
#include "opal/util/arch.h"
#include "opal/mca/mca.h"
#include "opal/mca/base/base.h"
#include "opal/mca/base/mca_base_param.h"
#include "orte/runtime/orte_globals.h"
#include "orte/util/name_fns.h"
#include "orte/mca/grpcomm/grpcomm.h"
#include "ompi/request/request.h"
#include "ompi/mca/dpm/dpm.h"
#include "ompi/mca/pml/pml.h"
#include "ompi/mca/pml/base/base.h"
#include "ompi/mca/pml/base/pml_base_request.h"
#include "ompi/mca/crcp/crcp.h"
#include "ompi/mca/crcp/base/base.h"
#include "ompi/class/ompi_free_list.h"
Commit to bring online OpenIB, MX, and shared memory support for Open MPI's checkpoint/restart functionality. Some tuning is still needed, but basic functionality is in place. There is still a problem with OpenIB and threads (external to C/R functionality). It has been reported in Ticket #1539 Additionally: * Fix a file cleanup bug in CRS Base. * Fix a possible deadlock in the TCP ft_event function * Add a mca_base_param_deregister() function to MCA base * Add whole process checkpoint timers * Add support for BTL: OpenIB, MX, Shared Memory * Add support Mpool: rdma, sm * Sundry bounds checking an cleanup in some scattered functions This commit was SVN r19756.
2008-10-16 19:09:00 +04:00
#include "ompi/runtime/ompi_cr.h"
Replace the old coordinated component ('coord') and replace it with a much more refined version ('bkmrk'). The new component fixes a number of problems with the old component. The core algorithm is the same, but by changing the data strucutres a bit we have improved performance and memory utilization. There are still a couple corner cases that still need some work. However, I did not want to delay bringing this into the trunk (and v1.3 branch) for too much longer. This commit was SVN r19537.
2008-09-10 22:29:17 +04:00
#include "crcp_bkmrk.h"
#include "crcp_bkmrk_pml.h"
/************************************
* Locally Global vars
************************************/
#define PROBE_ANY_SIZE ((size_t) 0)
#define PROBE_ANY_COUNT ((size_t) 0)
#define PERSIST_MARKER ((int) -1)
#define RECV_MATCH_RESP_DONE 0
#define RECV_MATCH_RESP_MORE 1
#define RECV_MATCH_RESP_ERROR 2
#define INVALID_INT -123456789
#define FIND_MSG_TRUE 0
#define FIND_MSG_FALSE 1
#define FIND_MSG_UNKNOWN 2
/* Pointers to the 'real' PML */
static mca_pml_base_component_t *wrapped_pml_component = NULL;
static mca_pml_base_module_t *wrapped_pml_module = NULL;
/* A unique ID for each message signature in the system */
static uint64_t message_seq_num = 1;
static uint64_t content_ref_seq_num = 1;
/* The current message being worked on */
static uint64_t current_msg_id = 0;
static ompi_crcp_bkmrk_pml_message_type_t current_msg_type = 0;
/* If we need to stall the C/R coordination until the current
* operation is complete */
static bool stall_for_completion;
/*
* State of the ft_event
*/
static int ft_event_state = OPAL_CRS_RUNNING;
/*
* List of known peers
*/
opal_list_t ompi_crcp_bkmrk_pml_peer_refs;
/*
* MPI_ANY_SOURCE recv lists
*/
opal_list_t unknown_recv_from_list;
opal_list_t unknown_persist_recv_list;
/*
* List of pending drain acks
*/
opal_list_t drained_msg_ack_list;
/*
* Free lists
*/
ompi_free_list_t coord_state_free_list;
ompi_free_list_t content_ref_free_list;
ompi_free_list_t peer_ref_free_list;
ompi_free_list_t traffic_msg_ref_free_list;
ompi_free_list_t drain_msg_ref_free_list;
ompi_free_list_t drain_ack_msg_ref_free_list;
/*
* Quiescence requests to wait on
*/
ompi_request_t ** quiesce_requests = NULL;
ompi_status_public_t ** quiesce_statuses = NULL;
int quiesce_request_count = 0;
/************************************
* Local Funcation Decls.
************************************/
static int ompi_crcp_bkmrk_pml_start_isend_init(ompi_request_t **request);
static int ompi_crcp_bkmrk_pml_start_irecv_init(ompi_request_t **request);
static int ompi_crcp_bkmrk_pml_start_drain_irecv_init(ompi_request_t **request, bool *found_drain);
static int ompi_crcp_bkmrk_request_complete_isend_init(struct ompi_request_t *request,
ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref,
int src, int tag, int tmp_ddt_size);
static int ompi_crcp_bkmrk_request_complete_isend(struct ompi_request_t *request,
ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref,
int src, int tag, int tmp_ddt_size);
static int ompi_crcp_bkmrk_request_complete_irecv_init(struct ompi_request_t *request,
ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref,
int src, int tag, int tmp_ddt_size);
static int ompi_crcp_bkmrk_request_complete_irecv(struct ompi_request_t *request,
ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref,
int src, int tag, int tmp_ddt_size);
/*
* Traffic Message: Append
*/
static int traffic_message_append(ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref,
opal_list_t * append_list,
ompi_crcp_bkmrk_pml_message_type_t msg_type,
size_t count,
ompi_datatype_t *datatype,
size_t ddt_size,
int tag,
int dest,
struct ompi_communicator_t* comm,
ompi_crcp_bkmrk_pml_traffic_message_ref_t **msg_ref);
/*
* Traffic Message: Start a persistent send/recv
*/
static int traffic_message_start(ompi_crcp_bkmrk_pml_traffic_message_ref_t *msg_ref,
ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref,
ompi_request_t **request,
opal_list_t * peer_list,
ompi_crcp_bkmrk_pml_message_content_ref_t ** content_ref);
/*
* Traffic Message: Move a message from one list to another
* - useful when moving messages from the unknown lists
*/
static int traffic_message_move(ompi_crcp_bkmrk_pml_traffic_message_ref_t *msg_ref,
ompi_crcp_bkmrk_pml_message_type_t msg_type,
ompi_crcp_bkmrk_pml_peer_ref_t *from_peer_ref,
opal_list_t * from_list,
ompi_crcp_bkmrk_pml_peer_ref_t *to_peer_ref,
opal_list_t * to_list,
Fix a problem with persistent sends found in MTT regression testing on Odin. This commit was SVN r19707.
2008-10-08 18:11:55 +04:00
ompi_crcp_bkmrk_pml_traffic_message_ref_t **new_msg_ref,
bool keep_active); /* If you have to create a new context, should it be initialized to active? */
Replace the old coordinated component ('coord') and replace it with a much more refined version ('bkmrk'). The new component fixes a number of problems with the old component. The core algorithm is the same, but by changing the data strucutres a bit we have improved performance and memory utilization. There are still a couple corner cases that still need some work. However, I did not want to delay bringing this into the trunk (and v1.3 branch) for too much longer. This commit was SVN r19537.
2008-09-10 22:29:17 +04:00
/*
* Traffic Message: Strip off the first matching request
*/
static int traffic_message_grab_content(ompi_crcp_bkmrk_pml_traffic_message_ref_t *msg_ref,
ompi_crcp_bkmrk_pml_message_content_ref_t ** content_ref,
bool remove,
bool already_drained);
/*
* Traffic Message: Find a persistent message, and mark it approprately
*/
static int traffic_message_find_mark_persistent(ompi_crcp_bkmrk_pml_traffic_message_ref_t *msg_ref,
ompi_request_t **request,
bool cur_active,
bool set_is_active,
ompi_crcp_bkmrk_pml_message_content_ref_t **content_ref);
/*
* Traffic Message: Find a message that matches the given signature
*/
static int traffic_message_find(opal_list_t * search_list,
size_t count, int tag, int peer, uint32_t comm_id,
size_t ddt_size,
ompi_crcp_bkmrk_pml_traffic_message_ref_t ** found_msg_ref,
int active);
/*
* Traffic Message: Determine if we have received a message matching this signature.
* Return a reference to the message on all matching lists.
*/
static int traffic_message_find_recv(ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref,
int rank, uint32_t comm_id, int tag,
size_t count, size_t datatype_size,
ompi_crcp_bkmrk_pml_traffic_message_ref_t ** posted_recv_msg_ref,
ompi_crcp_bkmrk_pml_traffic_message_ref_t ** posted_irecv_msg_ref,
ompi_crcp_bkmrk_pml_traffic_message_ref_t ** posted_precv_msg_ref,
ompi_crcp_bkmrk_pml_traffic_message_ref_t ** posted_unknown_recv_msg_ref,
ompi_crcp_bkmrk_pml_traffic_message_ref_t ** posted_unknown_precv_msg_ref);
/*
* Traffic Message: For all of the 'active' recvs, create a drain message
*/
static int traffic_message_create_drain_message(bool post_drain,
int max_post,
ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref,
ompi_crcp_bkmrk_pml_traffic_message_ref_t ** posted_msg_ref,
int *num_posted);
/*
* Drain Message: Append
*/
static int drain_message_append(ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref,
ompi_crcp_bkmrk_pml_message_type_t msg_type,
size_t count, size_t ddt_size,
int tag,int dest,
struct ompi_communicator_t* comm,
ompi_crcp_bkmrk_pml_drain_message_ref_t **msg_ref);
/*
* Drain Message: Remove
*/
static int drain_message_remove(ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref,
ompi_crcp_bkmrk_pml_drain_message_ref_t *msg_ref,
ompi_crcp_bkmrk_pml_message_content_ref_t *content_ref);
/*
* Drain Message: Check if this receive has been drained
*/
static int drain_message_check_recv(void **buf, size_t count,
ompi_datatype_t *datatype,
int *src, int *tag,
struct ompi_communicator_t* comm,
struct ompi_request_t **request,
ompi_status_public_t** status,
bool *found_drain);
/*
* Drain Message: Find a message matching the given signature on this peer list
*/
static int drain_message_find(opal_list_t * search_list,
size_t count, int tag, int peer,
uint32_t comm_id, size_t ddt_size,
ompi_crcp_bkmrk_pml_drain_message_ref_t ** found_msg_ref,
ompi_crcp_bkmrk_pml_message_content_ref_t ** content_ref);
/*
* Drain Message: Find a message matching the given signature on any list from any peer
*/
static int drain_message_find_any(size_t count, int tag, int peer,
struct ompi_communicator_t* comm, size_t ddt_size,
ompi_crcp_bkmrk_pml_drain_message_ref_t ** found_msg_ref,
ompi_crcp_bkmrk_pml_message_content_ref_t ** content_ref,
ompi_crcp_bkmrk_pml_peer_ref_t **peer_ref);
/*
* Drain Message: Grab a content reference, do not remove
*/
static int drain_message_grab_content(ompi_crcp_bkmrk_pml_drain_message_ref_t *drain_msg_ref,
ompi_crcp_bkmrk_pml_message_content_ref_t ** content_ref);
/*
* Drain Message: Copy this drain message to the signature provided, remove drain message
*/
static int drain_message_copy_remove(ompi_crcp_bkmrk_pml_drain_message_ref_t *msg_ref,
ompi_crcp_bkmrk_pml_message_content_ref_t * content_ref,
int *src, int *tag,
struct ompi_request_t **request,
ompi_status_public_t **status,
ompi_datatype_t *datatype, int count, void **buf,
ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref);
/*
* Drain Message: Copy this persistent drain message to the signature provided, remove drain message
*/
static int drain_message_copy_remove_persistent(ompi_crcp_bkmrk_pml_drain_message_ref_t *drain_msg_ref,
ompi_crcp_bkmrk_pml_message_content_ref_t *drain_content_ref,
ompi_crcp_bkmrk_pml_traffic_message_ref_t *traffic_msg_ref,
ompi_request_t *request,
ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref);
/*
* Peer List: Find the peer reference matching the ORTE process name
*/
static ompi_crcp_bkmrk_pml_peer_ref_t* find_peer(orte_process_name_t proc);
/*
* Peer List: Find the peer reference matching the index into the communicator
*/
static int find_peer_in_comm(struct ompi_communicator_t* comm, int proc_idx,
ompi_crcp_bkmrk_pml_peer_ref_t **peer_ref);
/*
* Coordinate Peers
* - Quiet channels
*/
static int ft_event_coordinate_peers(void);
/*
* Finalize the coordination of peers.
* - Mostly cleanup.
*/
static int ft_event_finalize_exchange(void);
/*
* Exchange the bookmarks
* - Staggered All-to-All
* LAM/MPI used a staggered all-to-all algoritm for bookmark exachange
* http://www.lam-mpi.org/papers/lacsi2003/
*/
static int ft_event_exchange_bookmarks(void);
/*
* Send Bookmarks to peer
*/
static int send_bookmarks(int peer_idx);
/*
* Recv Bookmarks from peer
*/
static int recv_bookmarks(int peer_idx);
/*
* Callback to receive the bookmarks from a peer
*/
static void recv_bookmarks_cbfunc(int status,
orte_process_name_t* sender,
opal_buffer_t *buffer,
orte_rml_tag_t tag,
void* cbdata);
static int total_recv_bookmarks = 0;
/*
* Now that we have all the bookmarks, check them to see if we need to
* drain any messages.
*/
static int ft_event_check_bookmarks(void);
/*
* Send message details to peer
* - matched with recv_msg_details()
*/
static int send_msg_details(ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref,
int total_sent, int total_matched);
/*
* Send a single message reference to a peer.
* found_match = true if peer found a message to drain.
*/
static int do_send_msg_detail(ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref,
ompi_crcp_bkmrk_pml_traffic_message_ref_t*msg_ref,
int *num_matches,
int *total_found,
bool *finished);
/*
* Recv message details from peer
* - matched with send_msg_details()
*/
static int recv_msg_details(ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref,
int total_recv, int total_matched);
/*
* Receive a single message reference from a peer.
*/
static int do_recv_msg_detail(ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref,
int *rank, uint32_t *comm_id, int *tag,
size_t *count, size_t *datatype_size,
int *p_num_sent);
/*
* Check the message reference to determine if:
* - We have received this message already, or
* - We need to post this message
*/
static int do_recv_msg_detail_check_drain(ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref,
int rank, uint32_t comm_id, int tag,
size_t count, size_t datatype_size,
int p_num_sent,
int *num_resolved);
/*
* Respond to peer regarding a received message detail
*/
static int do_recv_msg_detail_resp(ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref,
int resp,
int num_resolv,
int total_found);
/*
* Post the Drain Message Acks
* - These are sent once the receiver has finished receiving
* all of the messages it needed to drain off the wire.
*/
static int ft_event_post_drain_acks(void);
/*
* Callback to service drain message acks.
*/
static void drain_message_ack_cbfunc(int status,
orte_process_name_t* sender,
opal_buffer_t *buffer,
orte_rml_tag_t tag,
void* cbdata);
/*
* Post the Drain Messages
* - These are irecvs to be completed in any order.
*/
static int ft_event_post_drained(void);
static int ft_event_post_drain_message(ompi_crcp_bkmrk_pml_drain_message_ref_t *drain_msg_ref,
ompi_crcp_bkmrk_pml_message_content_ref_t *content_ref);
/*
* Wait for all drained messages and acks to complete.
* - Once this this finished then all channels associated
* with this process have been drained.
*/
static int ft_event_wait_quiesce(void);
/*
* Wait for all the posted drain messages to finish
*/
static int wait_quiesce_drained(void);
/*
* An optimized local version of waitall.
* - Remove some unnecessary logic
* - Remove logic to 'free' the request
*/
static int coord_request_wait_all( size_t count,
ompi_request_t ** requests,
ompi_status_public_t ** statuses);
/*
* An optimized local version of wait.
* - Remove some unnecessary logic
* - Remove logic to 'free' the request
* - Allow it to return if we need to stop waiting
*/
static int coord_request_wait( ompi_request_t * request,
ompi_status_public_t * status);
/*
* Wait for all the drain ACKs to be received
*/
static int wait_quiesce_drain_ack(void);
/************************************
* A few timing structures
*
* CRCP Timing Information
* -----------------------
* Pi Pj | Variable
* ---- -----+----------
* exchange_bookmark() | CRCP_TIMER_CKPT_EX_B
* -------------> | CRCP_TIMER_CKPT_EX_PEER_S
* <------------- | CRCP_TIMER_CKPT_EX_PEER_R
* -> wait_for_bk_done() |
* -- -- | CRCP_TIMER_CKPT_EX_WAIT
* check_bookmarks() |
* -- -- | CRCP_TIMER_CKPT_CHECK_B
* -> exchange_details (*) |
* -------------> | CRCP_TIMER_CKPT_CHECK_PEER_S
* <------------- | CRCP_TIMER_CKPT_CHECK_PEER_R
* post_drain[ack]() |
* -- -- | CRCP_TIMER_CKPT_POST_DRAIN
* wait_quiescence() |
* -- -- | CRCP_TIMER_CKPT_WAIT_QUI
* Finish checkpoint | -- Total Pre-Checkpoint
* -- -- | CRCP_TIMER_TOTAL_CKPT
* finalize_exchange() | -- Total Continue / Restart
* -- -- | CRCP_TIMER_TOTAL_CONT / _RST
*-----------------------------+
* (*) If needed.
*
* timing_enabled:
* < 0 : Off
* 1 : Summary only
* 2 : Per Peer messages + Barrier
* 3 : Messages from all procs
*
************************************/
#define CRCP_TIMER_TOTAL_CKPT 0
#define CRCP_TIMER_CKPT_EX_B 1
#define CRCP_TIMER_CKPT_EX_PEER_S 2
#define CRCP_TIMER_CKPT_EX_PEER_R 3
#define CRCP_TIMER_CKPT_EX_WAIT 4
#define CRCP_TIMER_CKPT_CHECK_B 5
#define CRCP_TIMER_CKPT_CHECK_PEER_S 6
#define CRCP_TIMER_CKPT_CHECK_PEER_R 7
#define CRCP_TIMER_CKPT_POST_DRAIN 8
#define CRCP_TIMER_CKPT_WAIT_QUI 9
#define CRCP_TIMER_TOTAL_CONT 10
#define CRCP_TIMER_TOTAL_RST 11
#define CRCP_TIMER_MAX 12
static double get_time(void);
static void start_time(int idx);
static void end_time(int idx);
static void display_indv_timer(int idx, int proc, int msgs);
static void display_indv_timer_core(int idx, int proc, int msgs, bool direct);
static void display_all_timers(int state);
static void clear_timers(void);
double timer_start[CRCP_TIMER_MAX];
double timer_end[CRCP_TIMER_MAX];
char * timer_label[CRCP_TIMER_MAX];
#define START_TIMER(idx) \
{ \
if(OPAL_UNLIKELY(timing_enabled > 0)) { \
start_time(idx); \
} \
}
#define END_TIMER(idx) \
{ \
if(OPAL_UNLIKELY(timing_enabled > 0)) { \
end_time(idx); \
} \
}
#define DISPLAY_INDV_TIMER(idx, proc, msg) \
{ \
if(OPAL_UNLIKELY(timing_enabled > 0)) { \
display_indv_timer(idx, proc, msg); \
} \
}
#define DISPLAY_ALL_TIMERS(var) \
{ \
if(OPAL_UNLIKELY(timing_enabled > 0)) { \
display_all_timers(var); \
} \
}
/************************************
* Additional Debuging dumps
************************************/
#if OMPI_ENABLE_DEBUG
static void traffic_message_dump_peer(ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref, char * msg, bool root_only);
static void traffic_message_dump_msg_list(opal_list_t *msg_list, bool is_drain);
static void traffic_message_dump_msg_indv(ompi_crcp_bkmrk_pml_traffic_message_ref_t * msg_ref, char * msg, bool vshort);
static void traffic_message_dump_msg_content_indv(ompi_crcp_bkmrk_pml_message_content_ref_t * content_ref);
static void traffic_message_dump_drain_msg_indv(ompi_crcp_bkmrk_pml_drain_message_ref_t * msg_ref, char * msg, bool vshort);
#define TRAFFIC_MSG_DUMP_PEER(lv, a) { \
if( lv <= mca_crcp_bkmrk_component.super.verbose ) { \
traffic_message_dump_peer a; \
} \
}
#define TRAFFIC_MSG_DUMP_MSG_LIST(lv, a) { \
if( lv <= mca_crcp_bkmrk_component.super.verbose ) { \
traffic_message_dump_msg_list a; \
} \
}
#define TRAFFIC_MSG_DUMP_MSG_INDV(lv, a) { \
if( lv <= mca_crcp_bkmrk_component.super.verbose ) { \
traffic_message_dump_msg_indv a; \
} \
}
#define TRAFFIC_MSG_DUMP_MSG_CONTENT_INDV(lv, a) { \
if( lv <= mca_crcp_bkmrk_component.super.verbose ) { \
traffic_message_dump_msg_content_indv a; \
} \
}
#define TRAFFIC_MSG_DUMP_DRAIN_MSG_INDV(lv, a) { \
if( lv <= mca_crcp_bkmrk_component.super.verbose ) { \
traffic_message_dump_drain_msg_indv a; \
} \
}
#else
#define TRAFFIC_MSG_DUMP_PEER(lv, a) ;
#define TRAFFIC_MSG_DUMP_MSG_LIST(lv, a) ;
#define TRAFFIC_MSG_DUMP_MSG_INDV(lv, a) ;
#define TRAFFIC_MSG_DUMP_MSG_CONTENT_INDV(lv, a) ;
#define TRAFFIC_MSG_DUMP_DRAIN_MSG_INDV(lv, a) ;
#endif
#define ERROR_SHOULD_NEVER_HAPPEN(msg) { \
opal_output(0, msg \
" ---------- This should never happen ---------- (%s:%d)", \
__FILE__, __LINE__); \
}
#define ERROR_SHOULD_NEVER_HAPPEN_ARG(msg, arg) { \
opal_output(0, msg \
" ---------- This should never happen ---------- (%s:%d)", \
arg, __FILE__, __LINE__); \
}
/************************************
* Declare/Define Object Structures
************************************/
/*
* Free List Maintenance
*/
#define HOKE_PEER_REF_ALLOC(peer_ref, rc) \
do { \
ompi_free_list_item_t* item; \
OMPI_FREE_LIST_WAIT(&peer_ref_free_list, item, rc); \
peer_ref = (ompi_crcp_bkmrk_pml_peer_ref_t*)item; \
} while(0);
#define HOKE_PEER_REF_RETURN(peer_ref) \
do { \
OMPI_FREE_LIST_RETURN(&peer_ref_free_list, \
(ompi_free_list_item_t*)peer_ref); \
} while(0);
#define HOKE_CONTENT_REF_ALLOC(content_ref, rc) \
do { \
ompi_free_list_item_t* item; \
OMPI_FREE_LIST_WAIT(&content_ref_free_list, item, rc); \
content_ref = (ompi_crcp_bkmrk_pml_message_content_ref_t*)item; \
content_ref->msg_id = content_ref_seq_num; \
content_ref_seq_num++;\
} while(0);
#define HOKE_CONTENT_REF_RETURN(content_ref) \
do { \
OMPI_FREE_LIST_RETURN(&content_ref_free_list, \
(ompi_free_list_item_t*)content_ref); \
} while(0);
#define HOKE_TRAFFIC_MSG_REF_ALLOC(msg_ref, rc) \
do { \
ompi_free_list_item_t* item; \
OMPI_FREE_LIST_WAIT(&traffic_msg_ref_free_list, item, rc); \
msg_ref = (ompi_crcp_bkmrk_pml_traffic_message_ref_t*)item; \
} while(0);
#define HOKE_TRAFFIC_MSG_REF_RETURN(msg_ref) \
do { \
OMPI_FREE_LIST_RETURN(&traffic_msg_ref_free_list, \
(ompi_free_list_item_t*)msg_ref); \
} while(0);
#define HOKE_DRAIN_MSG_REF_ALLOC(msg_ref, rc) \
do { \
ompi_free_list_item_t* item; \
OMPI_FREE_LIST_WAIT(&drain_msg_ref_free_list, item, rc); \
msg_ref = (ompi_crcp_bkmrk_pml_drain_message_ref_t*)item; \
} while(0);
#define HOKE_DRAIN_MSG_REF_RETURN(msg_ref) \
do { \
OMPI_FREE_LIST_RETURN(&drain_msg_ref_free_list, \
(ompi_free_list_item_t*)msg_ref); \
} while(0);
#define HOKE_DRAIN_ACK_MSG_REF_ALLOC(msg_ref, rc) \
do { \
ompi_free_list_item_t* item; \
OMPI_FREE_LIST_WAIT(&drain_ack_msg_ref_free_list, item, rc); \
msg_ref = (ompi_crcp_bkmrk_pml_drain_message_ack_ref_t*)item; \
} while(0);
#define HOKE_DRAIN_ACK_MSG_REF_RETURN(msg_ref) \
do { \
OMPI_FREE_LIST_RETURN(&drain_ack_msg_ref_free_list, \
(ompi_free_list_item_t*)msg_ref); \
} while(0);
/*
* Peer reference
*/
OBJ_CLASS_INSTANCE(ompi_crcp_bkmrk_pml_peer_ref_t,
opal_list_item_t,
ompi_crcp_bkmrk_pml_peer_ref_construct,
ompi_crcp_bkmrk_pml_peer_ref_destruct);
void ompi_crcp_bkmrk_pml_peer_ref_construct(ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref) {
peer_ref->proc_name.jobid = ORTE_JOBID_INVALID;
peer_ref->proc_name.vpid = ORTE_VPID_INVALID;
OBJ_CONSTRUCT(&peer_ref->send_list, opal_list_t);
OBJ_CONSTRUCT(&peer_ref->isend_list, opal_list_t);
OBJ_CONSTRUCT(&peer_ref->send_init_list, opal_list_t);
OBJ_CONSTRUCT(&peer_ref->recv_list, opal_list_t);
OBJ_CONSTRUCT(&peer_ref->irecv_list, opal_list_t);
OBJ_CONSTRUCT(&peer_ref->recv_init_list, opal_list_t);
OBJ_CONSTRUCT(&peer_ref->drained_list, opal_list_t);
peer_ref->total_msgs_sent = 0;
peer_ref->matched_msgs_sent = 0;
peer_ref->total_msgs_recvd = 0;
peer_ref->matched_msgs_recvd = 0;
peer_ref->total_drained_msgs = 0;
peer_ref->ack_required = false;
}
void ompi_crcp_bkmrk_pml_peer_ref_destruct( ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref) {
opal_list_item_t* item = NULL;
peer_ref->proc_name.jobid = ORTE_JOBID_INVALID;
peer_ref->proc_name.vpid = ORTE_VPID_INVALID;
while( NULL != (item = opal_list_remove_first(&peer_ref->send_list)) ) {
HOKE_TRAFFIC_MSG_REF_RETURN(item);
}
OBJ_DESTRUCT(&peer_ref->send_list);
while( NULL != (item = opal_list_remove_first(&peer_ref->isend_list)) ) {
HOKE_TRAFFIC_MSG_REF_RETURN(item);
}
OBJ_DESTRUCT(&peer_ref->isend_list);
while( NULL != (item = opal_list_remove_first(&peer_ref->send_init_list)) ) {
HOKE_TRAFFIC_MSG_REF_RETURN(item);
}
OBJ_DESTRUCT(&peer_ref->send_init_list);
while( NULL != (item = opal_list_remove_first(&peer_ref->recv_list)) ) {
HOKE_TRAFFIC_MSG_REF_RETURN(item);
}
OBJ_DESTRUCT(&peer_ref->recv_list);
while( NULL != (item = opal_list_remove_first(&peer_ref->irecv_list)) ) {
HOKE_TRAFFIC_MSG_REF_RETURN(item);
}
OBJ_DESTRUCT(&peer_ref->irecv_list);
while( NULL != (item = opal_list_remove_first(&peer_ref->recv_init_list)) ) {
HOKE_TRAFFIC_MSG_REF_RETURN(item);
}
OBJ_DESTRUCT(&peer_ref->recv_init_list);
while( NULL != (item = opal_list_remove_first(&peer_ref->drained_list)) ) {
HOKE_DRAIN_MSG_REF_RETURN(item);
}
OBJ_DESTRUCT(&peer_ref->drained_list);
peer_ref->total_msgs_sent = 0;
peer_ref->matched_msgs_sent = 0;
peer_ref->total_msgs_recvd = 0;
peer_ref->matched_msgs_recvd = 0;
peer_ref->total_drained_msgs = 0;
peer_ref->ack_required = false;
}
/*
* Message Content Structure
*/
OBJ_CLASS_INSTANCE(ompi_crcp_bkmrk_pml_message_content_ref_t,
opal_list_item_t,
ompi_crcp_bkmrk_pml_message_content_ref_construct,
ompi_crcp_bkmrk_pml_message_content_ref_destruct);
void ompi_crcp_bkmrk_pml_message_content_ref_construct(ompi_crcp_bkmrk_pml_message_content_ref_t *content_ref)
{
content_ref->buffer = NULL;
content_ref->request = NULL;
content_ref->active = false;
content_ref->done = false;
content_ref->active = false;
content_ref->already_posted = false;
content_ref->already_drained = false;
content_ref->msg_id = 0;
}
void ompi_crcp_bkmrk_pml_message_content_ref_destruct( ompi_crcp_bkmrk_pml_message_content_ref_t *content_ref)
{
if( NULL != content_ref->buffer ) {
free(content_ref->buffer);
}
content_ref->buffer = NULL;
if( NULL != content_ref->request ) {
OBJ_RELEASE(content_ref->request);
}
content_ref->request = NULL;
content_ref->active = false;
content_ref->done = false;
content_ref->active = false;
content_ref->already_posted = false;
content_ref->already_drained = false;
content_ref->msg_id = 0;
}
/*
* Traffic Message
*/
OBJ_CLASS_INSTANCE(ompi_crcp_bkmrk_pml_traffic_message_ref_t,
opal_list_item_t,
ompi_crcp_bkmrk_pml_traffic_message_ref_construct,
ompi_crcp_bkmrk_pml_traffic_message_ref_destruct);
void ompi_crcp_bkmrk_pml_traffic_message_ref_construct(ompi_crcp_bkmrk_pml_traffic_message_ref_t *msg_ref) {
msg_ref->msg_id = 0;
msg_ref->msg_type = COORD_MSG_TYPE_UNKNOWN;
msg_ref->count = 0;
msg_ref->ddt_size = 0;
msg_ref->tag = 0;
msg_ref->rank = 0;
msg_ref->comm = NULL;
OBJ_CONSTRUCT(&msg_ref->msg_contents, opal_list_t);
msg_ref->proc_name.jobid = ORTE_JOBID_INVALID;
msg_ref->proc_name.vpid = ORTE_VPID_INVALID;
msg_ref->matched = INVALID_INT;
msg_ref->done = INVALID_INT;
msg_ref->active = INVALID_INT;
msg_ref->posted = INVALID_INT;
msg_ref->active_drain = INVALID_INT;
}
void ompi_crcp_bkmrk_pml_traffic_message_ref_destruct( ompi_crcp_bkmrk_pml_traffic_message_ref_t *msg_ref) {
opal_list_item_t* item = NULL;
msg_ref->msg_id = 0;
msg_ref->msg_type = COORD_MSG_TYPE_UNKNOWN;
msg_ref->count = 0;
msg_ref->ddt_size = 0;
msg_ref->tag = 0;
msg_ref->rank = 0;
msg_ref->comm = NULL;
while( NULL != (item = opal_list_remove_first(&(msg_ref->msg_contents)) ) ) {
HOKE_CONTENT_REF_RETURN(item);
}
OBJ_DESTRUCT(&(msg_ref->msg_contents));
msg_ref->proc_name.jobid = ORTE_JOBID_INVALID;
msg_ref->proc_name.vpid = ORTE_VPID_INVALID;
msg_ref->matched = INVALID_INT;
msg_ref->done = INVALID_INT;
msg_ref->active = INVALID_INT;
msg_ref->posted = INVALID_INT;
msg_ref->active_drain = INVALID_INT;
}
/*
* Drain Message
*/
OBJ_CLASS_INSTANCE(ompi_crcp_bkmrk_pml_drain_message_ref_t,
opal_list_item_t,
ompi_crcp_bkmrk_pml_drain_message_ref_construct,
ompi_crcp_bkmrk_pml_drain_message_ref_destruct);
void ompi_crcp_bkmrk_pml_drain_message_ref_construct(ompi_crcp_bkmrk_pml_drain_message_ref_t *msg_ref) {
msg_ref->msg_id = 0;
msg_ref->msg_type = COORD_MSG_TYPE_UNKNOWN;
msg_ref->count = 0;
msg_ref->datatype = NULL;
msg_ref->ddt_size = 0;
msg_ref->tag = 0;
msg_ref->rank = 0;
msg_ref->comm = NULL;
OBJ_CONSTRUCT(&msg_ref->msg_contents, opal_list_t);
msg_ref->proc_name.jobid = ORTE_JOBID_INVALID;
msg_ref->proc_name.vpid = ORTE_VPID_INVALID;
msg_ref->done = INVALID_INT;
msg_ref->active = INVALID_INT;
msg_ref->already_posted = INVALID_INT;
}
void ompi_crcp_bkmrk_pml_drain_message_ref_destruct( ompi_crcp_bkmrk_pml_drain_message_ref_t *msg_ref) {
opal_list_item_t* item = NULL;
msg_ref->msg_id = 0;
msg_ref->msg_type = COORD_MSG_TYPE_UNKNOWN;
msg_ref->count = 0;
if( NULL != msg_ref->datatype ) {
OBJ_RELEASE(msg_ref->datatype);
msg_ref->datatype = NULL;
}
msg_ref->ddt_size = 0;
msg_ref->tag = 0;
msg_ref->rank = 0;
msg_ref->comm = NULL;
while( NULL != (item = opal_list_remove_first(&(msg_ref->msg_contents)) ) ) {
HOKE_CONTENT_REF_RETURN(item);
}
OBJ_DESTRUCT(&(msg_ref->msg_contents));
msg_ref->proc_name.jobid = ORTE_JOBID_INVALID;
msg_ref->proc_name.vpid = ORTE_VPID_INVALID;
msg_ref->done = INVALID_INT;
msg_ref->active = INVALID_INT;
msg_ref->already_posted = INVALID_INT;
}
/*
* Drain Ack Message
*/
OBJ_CLASS_INSTANCE(ompi_crcp_bkmrk_pml_drain_message_ack_ref_t,
opal_list_item_t,
ompi_crcp_bkmrk_pml_drain_message_ack_ref_construct,
ompi_crcp_bkmrk_pml_drain_message_ack_ref_destruct);
void ompi_crcp_bkmrk_pml_drain_message_ack_ref_construct(ompi_crcp_bkmrk_pml_drain_message_ack_ref_t *msg_ack_ref) {
msg_ack_ref->complete = false;
msg_ack_ref->peer.jobid = ORTE_JOBID_INVALID;
msg_ack_ref->peer.vpid = ORTE_VPID_INVALID;
}
void ompi_crcp_bkmrk_pml_drain_message_ack_ref_destruct( ompi_crcp_bkmrk_pml_drain_message_ack_ref_t *msg_ack_ref) {
msg_ack_ref->complete = false;
msg_ack_ref->peer.jobid = ORTE_JOBID_INVALID;
msg_ack_ref->peer.vpid = ORTE_VPID_INVALID;
}
/*
* PML state
*/
OBJ_CLASS_INSTANCE(ompi_crcp_bkmrk_pml_state_t,
ompi_crcp_base_pml_state_t,
NULL,
NULL
);
/************************************
* Some Macro shortcuts
************************************/
#define CRCP_COORD_STATE_ALLOC(state_ref, rc) \
do { \
ompi_free_list_item_t* item; \
OMPI_FREE_LIST_WAIT(&coord_state_free_list, item, rc); \
state_ref = (ompi_crcp_bkmrk_pml_state_t*)item; \
} while(0);
#define CRCP_COORD_STATE_RETURN(state_ref) \
do { \
OMPI_FREE_LIST_RETURN(&coord_state_free_list, \
(ompi_free_list_item_t*)state_ref); \
} while(0);
#define CREATE_COORD_STATE(coord_state, pml_state, v_peer_ref, v_msg_ref) \
{ \
CRCP_COORD_STATE_ALLOC(coord_state, ret); \
\
coord_state->prev_ptr = pml_state; \
coord_state->p_super.super = pml_state->super; \
coord_state->p_super.state = pml_state->state; \
coord_state->p_super.error_code = pml_state->error_code; \
coord_state->p_super.wrapped_pml_component = pml_state->wrapped_pml_component; \
coord_state->p_super.wrapped_pml_module = pml_state->wrapped_pml_module; \
\
coord_state->peer_ref = v_peer_ref; \
coord_state->msg_ref = v_msg_ref; \
}
#define EXTRACT_COORD_STATE(pml_state, v_coord_state, v_rtn_state, v_peer_ref, v_msg_ref) \
{ \
v_coord_state = (ompi_crcp_bkmrk_pml_state_t*)pml_state; \
v_rtn_state = v_coord_state->prev_ptr; \
v_peer_ref = v_coord_state->peer_ref; \
v_msg_ref = v_coord_state->msg_ref; \
}
#define CREATE_NEW_MSG(msg_ref, v_type, v_count, v_ddt_size, v_tag, v_rank, v_comm, p_jobid, p_vpid) \
{ \
HOKE_TRAFFIC_MSG_REF_ALLOC(msg_ref, ret); \
\
msg_ref->msg_id = message_seq_num; \
message_seq_num++; \
\
msg_ref->msg_type = v_type; \
\
msg_ref->count = v_count; \
\
msg_ref->ddt_size = v_ddt_size; \
\
msg_ref->tag = v_tag; \
msg_ref->rank = v_rank; \
msg_ref->comm = v_comm; \
\
msg_ref->proc_name.jobid = p_jobid; \
msg_ref->proc_name.vpid = p_vpid; \
\
msg_ref->matched = 0; \
msg_ref->done = 0; \
msg_ref->active = 0; \
msg_ref->posted = 0; \
msg_ref->active_drain = 0; \
}
#define CREATE_NEW_DRAIN_MSG(msg_ref, v_type, v_count, v_ddt_size, v_tag, v_rank, v_comm, p_jobid, p_vpid) \
{ \
HOKE_DRAIN_MSG_REF_ALLOC(msg_ref, ret); \
\
msg_ref->msg_id = message_seq_num; \
message_seq_num++; \
\
msg_ref->msg_type = v_type; \
\
msg_ref->count = v_count; \
\
msg_ref->datatype = NULL; \
msg_ref->ddt_size = ddt_size; \
\
msg_ref->tag = v_tag; \
msg_ref->rank = v_rank; \
msg_ref->comm = v_comm; \
\
msg_ref->proc_name.jobid = p_jobid; \
msg_ref->proc_name.vpid = p_vpid; \
}
#define PACK_BUFFER(buffer, var, count, type, error_msg) \
{ \
if (OMPI_SUCCESS != (ret = opal_dss.pack(buffer, &(var), count, type)) ) { \
opal_output(mca_crcp_bkmrk_component.super.output_handle, \
"%s (Return %d)", error_msg, ret); \
exit_status = ret; \
goto cleanup; \
} \
}
#define UNPACK_BUFFER(buffer, var, count, type, error_msg) \
{ \
orte_std_cntr_t n = count; \
if (ORTE_SUCCESS != (ret = opal_dss.unpack(buffer, &(var), &n, type)) ) { \
opal_output(mca_crcp_bkmrk_component.super.output_handle, \
"%s (Return %d)", error_msg, ret); \
exit_status = ret; \
goto cleanup; \
} \
}
/****************
* PML Wrapper Init/Finalize
****************/
int ompi_crcp_bkmrk_pml_init(void) {
message_seq_num = 1;
current_msg_id = 0;
current_msg_type = COORD_MSG_TYPE_UNKNOWN;
stall_for_completion = false;
ft_event_state = OPAL_CRS_RUNNING;
OBJ_CONSTRUCT(&ompi_crcp_bkmrk_pml_peer_refs, opal_list_t);
OBJ_CONSTRUCT(&unknown_recv_from_list, opal_list_t);
OBJ_CONSTRUCT(&unknown_persist_recv_list, opal_list_t);
OBJ_CONSTRUCT(&drained_msg_ack_list, opal_list_t);
/* Create free lists for
* - Coord State
* - Peer Refs
* - Traffic Message Refs
* - Drain Message Refs
* - Drain ACK Messsage Refs
* - Message Contents?
*/
OBJ_CONSTRUCT(&coord_state_free_list, ompi_free_list_t);
ompi_free_list_init_new( &coord_state_free_list,
sizeof(ompi_crcp_bkmrk_pml_state_t),
CACHE_LINE_SIZE,
OBJ_CLASS(ompi_crcp_bkmrk_pml_state_t),
0,CACHE_LINE_SIZE,
4, /* Initial number */
-1, /* Max = Unlimited */
4, /* Increment by */
NULL);
OBJ_CONSTRUCT(&content_ref_free_list, ompi_free_list_t);
ompi_free_list_init_new( &content_ref_free_list,
sizeof(ompi_crcp_bkmrk_pml_message_content_ref_t),
CACHE_LINE_SIZE,
OBJ_CLASS(ompi_crcp_bkmrk_pml_message_content_ref_t),
0,CACHE_LINE_SIZE,
80, /* Initial number */
-1, /* Max = Unlimited */
32, /* Increment by */
NULL);
OBJ_CONSTRUCT(&peer_ref_free_list, ompi_free_list_t);
ompi_free_list_init_new( &peer_ref_free_list,
sizeof(ompi_crcp_bkmrk_pml_peer_ref_t),
CACHE_LINE_SIZE,
OBJ_CLASS(ompi_crcp_bkmrk_pml_peer_ref_t),
0,CACHE_LINE_SIZE,
16, /* Initial number */
-1, /* Max = Unlimited */
16, /* Increment by */
NULL);
OBJ_CONSTRUCT(&traffic_msg_ref_free_list, ompi_free_list_t);
ompi_free_list_init_new( &traffic_msg_ref_free_list,
sizeof(ompi_crcp_bkmrk_pml_traffic_message_ref_t),
CACHE_LINE_SIZE,
OBJ_CLASS(ompi_crcp_bkmrk_pml_traffic_message_ref_t),
0,CACHE_LINE_SIZE,
32, /* Initial number */
-1, /* Max = Unlimited */
64, /* Increment by */
NULL);
OBJ_CONSTRUCT(&drain_msg_ref_free_list, ompi_free_list_t);
ompi_free_list_init_new( &drain_msg_ref_free_list,
sizeof(ompi_crcp_bkmrk_pml_drain_message_ref_t),
CACHE_LINE_SIZE,
OBJ_CLASS(ompi_crcp_bkmrk_pml_drain_message_ref_t),
0,CACHE_LINE_SIZE,
32, /* Initial number */
-1, /* Max = Unlimited */
64, /* Increment by */
NULL);
OBJ_CONSTRUCT(&drain_ack_msg_ref_free_list, ompi_free_list_t);
ompi_free_list_init_new( &drain_ack_msg_ref_free_list,
sizeof(ompi_crcp_bkmrk_pml_drain_message_ack_ref_t),
CACHE_LINE_SIZE,
OBJ_CLASS(ompi_crcp_bkmrk_pml_drain_message_ack_ref_t),
0,CACHE_LINE_SIZE,
16, /* Initial number */
-1, /* Max = Unlimited */
16, /* Increment by */
NULL);
clear_timers();
if( timing_enabled > 0 ) {
timer_label[CRCP_TIMER_TOTAL_CKPT] = strdup("Total Ckpt.");
timer_label[CRCP_TIMER_CKPT_EX_B] = strdup("Exchange Bookmarks");
timer_label[CRCP_TIMER_CKPT_EX_PEER_S] = strdup(" Ex.Bk. Send Peer");
timer_label[CRCP_TIMER_CKPT_EX_PEER_R] = strdup(" Ex.Bk. Recv Peer");
timer_label[CRCP_TIMER_CKPT_EX_WAIT] = strdup(" Ex.Bk. Wait");
timer_label[CRCP_TIMER_CKPT_CHECK_B] = strdup("Check Bookmarks");
timer_label[CRCP_TIMER_CKPT_CHECK_PEER_S] = strdup(" Ck.Bk. Send Peer");
timer_label[CRCP_TIMER_CKPT_CHECK_PEER_R] = strdup(" Ck.Bk. Recv Peer");
timer_label[CRCP_TIMER_CKPT_POST_DRAIN] = strdup("Post Drain Msgs.");
timer_label[CRCP_TIMER_CKPT_WAIT_QUI] = strdup("Wait for Quiescence");
timer_label[CRCP_TIMER_TOTAL_CONT] = strdup("Total Continue");
timer_label[CRCP_TIMER_TOTAL_RST] = strdup("Total Restart");
}
return OMPI_SUCCESS;
}
int ompi_crcp_bkmrk_pml_finalize(void) {
int i;
current_msg_id = 0;
current_msg_type = COORD_MSG_TYPE_UNKNOWN;
stall_for_completion = false;
ft_event_state = OPAL_CRS_RUNNING;
OBJ_DESTRUCT(&ompi_crcp_bkmrk_pml_peer_refs);
OBJ_DESTRUCT(&unknown_recv_from_list);
OBJ_DESTRUCT(&unknown_persist_recv_list);
OBJ_DESTRUCT(&drained_msg_ack_list);
/* Destroy All Free Lists */
OBJ_DESTRUCT(&peer_ref_free_list);
OBJ_DESTRUCT(&traffic_msg_ref_free_list);
OBJ_DESTRUCT(&drain_msg_ref_free_list);
OBJ_DESTRUCT(&drain_ack_msg_ref_free_list);
OBJ_DESTRUCT(&content_ref_free_list);
if( timing_enabled > 0 ) {
for(i = 0; i < CRCP_TIMER_MAX; ++i) {
free(timer_label[i]);
timer_label[i] = NULL;
}
}
return OMPI_SUCCESS;
}
/****************
* PML Wrapper
****************/
/**************** Enable *****************/
ompi_crcp_base_pml_state_t* ompi_crcp_bkmrk_pml_enable(
bool enable,
ompi_crcp_base_pml_state_t* pml_state )
{
OPAL_OUTPUT_VERBOSE((30, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: pml_enable()"));
pml_state->error_code = OMPI_SUCCESS;
return pml_state;
}
/**************** Progress *****************/
ompi_crcp_base_pml_state_t* ompi_crcp_bkmrk_pml_progress(
ompi_crcp_base_pml_state_t* pml_state)
{
OPAL_OUTPUT_VERBOSE((30, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: pml_progress()"));
pml_state->error_code = OMPI_SUCCESS;
return pml_state;
}
/**************** Probe *****************/
/* JJH - Code reuse: Combine iprobe and probe logic */
ompi_crcp_base_pml_state_t* ompi_crcp_bkmrk_pml_iprobe(
int dst, int tag,
struct ompi_communicator_t* comm,
int *matched,
ompi_status_public_t* status,
ompi_crcp_base_pml_state_t* pml_state )
{
ompi_crcp_bkmrk_pml_drain_message_ref_t *drain_msg_ref = NULL;
ompi_crcp_bkmrk_pml_message_content_ref_t *content_ref = NULL;
int exit_status = OMPI_SUCCESS;
int ret;
OPAL_OUTPUT_VERBOSE((30, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: pml_iprobe(%d, %d)", dst, tag));
/*
* Before PML Call
* - Determine if this can be satisfied from the drained list
* - Otherwise let the PML handle it
*/
if( OMPI_CRCP_PML_PRE == pml_state->state) {
/*
* Check to see if this message is in the drained message list
*/
if( OMPI_SUCCESS != (ret = drain_message_find_any(PROBE_ANY_COUNT, tag, dst,
comm, PROBE_ANY_SIZE,
&drain_msg_ref,
&content_ref,
NULL) ) ) {
ERROR_SHOULD_NEVER_HAPPEN("crcp:bkmrk: pml_iprobe(): Failed trying to find a drained message.");
exit_status = ret;
goto DONE;
}
/*
* If the message is a drained message
* - Copy of the status structure to pass back to the user
* - Mark the 'matched' flag as true
*/
if( NULL != drain_msg_ref ) {
OPAL_OUTPUT_VERBOSE((12, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: pml_iprobe(): Matched a drained message..."));
/* Copy the status information */
if( MPI_STATUS_IGNORE != status ) {
memcpy(status, &content_ref->status, sizeof(ompi_status_public_t));
}
/* Mark as complete */
*matched = 1;
/* This will identify to the wrapper that this message is complete */
pml_state->state = OMPI_CRCP_PML_DONE;
pml_state->error_code = OMPI_SUCCESS;
return pml_state;
}
/*
* Otherwise the message is not drained (common case), so let the PML deal with it
*/
else {
/* Mark as not complete */
*matched = 0;
}
}
DONE:
pml_state->error_code = exit_status;
return pml_state;
}
ompi_crcp_base_pml_state_t* ompi_crcp_bkmrk_pml_probe(
int dst, int tag,
struct ompi_communicator_t* comm,
ompi_status_public_t* status,
ompi_crcp_base_pml_state_t* pml_state )
{
ompi_crcp_bkmrk_pml_drain_message_ref_t *drain_msg_ref = NULL;
ompi_crcp_bkmrk_pml_message_content_ref_t *content_ref = NULL;
int exit_status = OMPI_SUCCESS;
int ret;
OPAL_OUTPUT_VERBOSE((30, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: pml_probe(%d, %d)", dst, tag));
/*
* Before PML Call
* - Determine if this can be satisfied from the drained list
* - Otherwise let the PML handle it
*/
if( OMPI_CRCP_PML_PRE == pml_state->state) {
/*
* Check to see if this message is in the drained message list
*/
if( OMPI_SUCCESS != (ret = drain_message_find_any(PROBE_ANY_COUNT, tag, dst,
comm, PROBE_ANY_SIZE,
&drain_msg_ref,
&content_ref,
NULL) ) ) {
ERROR_SHOULD_NEVER_HAPPEN("crcp:bkmrk: pml_probe(): Failed trying to find a drained message.");
exit_status = ret;
goto DONE;
}
/*
* If the message is a drained message
* - Copy of the status structure to pass back to the user
*/
if( NULL != drain_msg_ref ) {
OPAL_OUTPUT_VERBOSE((12, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: pml_iprobe(): Matched a drained message..."));
/* Copy the status information */
if( MPI_STATUS_IGNORE != status ) {
memcpy(status, &content_ref->status, sizeof(ompi_status_public_t));
}
/* This will identify to the wrapper that this message is complete */
pml_state->state = OMPI_CRCP_PML_DONE;
pml_state->error_code = OMPI_SUCCESS;
return pml_state;
}
}
DONE:
pml_state->error_code = exit_status;
return pml_state;
}
/**************** Dump *****************/
ompi_crcp_base_pml_state_t* ompi_crcp_bkmrk_pml_dump(
struct ompi_communicator_t* comm,
int verbose,
ompi_crcp_base_pml_state_t* pml_state )
{
OPAL_OUTPUT_VERBOSE((30, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: pml_dump()"));
pml_state->error_code = OMPI_SUCCESS;
return pml_state;
}
/**************** Communicator *****************/
ompi_crcp_base_pml_state_t* ompi_crcp_bkmrk_pml_add_comm(
struct ompi_communicator_t* comm,
ompi_crcp_base_pml_state_t* pml_state )
{
OPAL_OUTPUT_VERBOSE((30, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: pml_add_comm()"));
pml_state->error_code = OMPI_SUCCESS;
return pml_state;
}
ompi_crcp_base_pml_state_t* ompi_crcp_bkmrk_pml_del_comm(
struct ompi_communicator_t* comm,
ompi_crcp_base_pml_state_t* pml_state )
{
OPAL_OUTPUT_VERBOSE((30, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: pml_del_comm()"));
pml_state->error_code = OMPI_SUCCESS;
return pml_state;
}
/**************** Processes *****************/
ompi_crcp_base_pml_state_t* ompi_crcp_bkmrk_pml_add_procs(
struct ompi_proc_t **procs,
size_t nprocs,
ompi_crcp_base_pml_state_t* pml_state )
{
int ret;
ompi_crcp_bkmrk_pml_peer_ref_t *new_peer_ref;
size_t i;
if( OMPI_CRCP_PML_PRE != pml_state->state ){
goto DONE;
}
OPAL_OUTPUT_VERBOSE((30, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: pml_add_procs()"));
/*
* Save pointers to the wrapped PML
*/
wrapped_pml_component = pml_state->wrapped_pml_component;
wrapped_pml_module = pml_state->wrapped_pml_module;
/*
* Create a peer_ref for each peer added
*/
for( i = 0; i < nprocs; ++i) {
HOKE_PEER_REF_ALLOC(new_peer_ref, ret);
new_peer_ref->proc_name.jobid = procs[i]->proc_name.jobid;
new_peer_ref->proc_name.vpid = procs[i]->proc_name.vpid;
opal_list_append(&ompi_crcp_bkmrk_pml_peer_refs, &(new_peer_ref->super));
}
DONE:
pml_state->error_code = OMPI_SUCCESS;
return pml_state;
}
ompi_crcp_base_pml_state_t* ompi_crcp_bkmrk_pml_del_procs(
struct ompi_proc_t **procs,
size_t nprocs,
ompi_crcp_base_pml_state_t* pml_state )
{
opal_list_item_t *item = NULL;
ompi_crcp_bkmrk_pml_peer_ref_t *old_peer_ref;
int exit_status = OMPI_SUCCESS;
size_t i;
if( OMPI_CRCP_PML_PRE != pml_state->state ){
goto DONE;
}
OPAL_OUTPUT_VERBOSE((30, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: pml_del_procs()"));
for( i = 0; i < nprocs; ++i) {
item = (opal_list_item_t*)find_peer(procs[i]->proc_name);
if(NULL == item) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: del_procs: Unable to find peer %s\n",
ORTE_NAME_PRINT(&(procs[i]->proc_name)));
exit_status = OMPI_ERROR;
goto DONE;
}
/* Remove the found peer from the list */
opal_list_remove_item(&ompi_crcp_bkmrk_pml_peer_refs, item);
old_peer_ref = (ompi_crcp_bkmrk_pml_peer_ref_t*)item;
HOKE_PEER_REF_RETURN(old_peer_ref);
}
DONE:
pml_state->error_code = exit_status;
return pml_state;
}
/**************** Send *****************/
ompi_crcp_base_pml_state_t* ompi_crcp_bkmrk_pml_isend_init(
void *buf, size_t count,
ompi_datatype_t *datatype,
int dst, int tag,
mca_pml_base_send_mode_t mode,
struct ompi_communicator_t* comm,
struct ompi_request_t **request,
ompi_crcp_base_pml_state_t* pml_state )
{
ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref = NULL;
ompi_crcp_bkmrk_pml_traffic_message_ref_t *msg_ref = NULL;
ompi_crcp_bkmrk_pml_state_t *coord_state = NULL;
int exit_status = OMPI_SUCCESS;
int ret;
OPAL_OUTPUT_VERBOSE((30, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: pml_isend_init()"));
/*
* Before the PML gets the message:
* - Setup structure to track the message
*/
if( OMPI_CRCP_PML_PRE == pml_state->state ) {
/*
* Find the peer reference
*/
if( OMPI_SUCCESS != (ret = find_peer_in_comm(comm, dst, &peer_ref) ) ){
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: isend: Failed to find peer_ref\n");
exit_status = ret;
goto DONE;
}
/*
* Archive the message Message Object
*/
traffic_message_append(peer_ref, &(peer_ref->send_init_list),
COORD_MSG_TYPE_P_SEND,
count, datatype, 0, tag, dst, comm,
&msg_ref);
/* Save the pointers */
CREATE_COORD_STATE(coord_state, pml_state,
peer_ref, msg_ref);
coord_state->p_super.error_code = OMPI_SUCCESS;
return &coord_state->p_super;
}
/*
* After PML is done, update message reference
*/
else if( OMPI_CRCP_PML_POST == pml_state->state ) {
ompi_crcp_base_pml_state_t *rtn_state = NULL;
ompi_crcp_bkmrk_pml_message_content_ref_t *new_content = NULL;
EXTRACT_COORD_STATE(pml_state, coord_state, rtn_state,
peer_ref, msg_ref);
/*
* Update Message
*/
HOKE_CONTENT_REF_ALLOC(new_content, ret);
new_content->buffer = buf;
new_content->request = *request;
new_content->done = false;
new_content->active = false;
new_content->already_posted = true;
new_content->already_drained = false;
OBJ_RETAIN(*request);
opal_list_append(&(msg_ref->msg_contents), &(new_content->super) );
CRCP_COORD_STATE_RETURN(coord_state);
rtn_state->error_code = OMPI_SUCCESS;
return rtn_state;
}
DONE:
pml_state->error_code = exit_status;
return pml_state;
}
static int ompi_crcp_bkmrk_pml_start_isend_init(ompi_request_t **request)
{
int ret, exit_status = OMPI_SUCCESS;
ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref = NULL;
ompi_crcp_bkmrk_pml_traffic_message_ref_t *msg_ref = NULL;
ompi_crcp_bkmrk_pml_message_content_ref_t *content_ref = NULL;
mca_pml_base_request_t *breq = NULL;
size_t tmp_ddt_size = 0;
breq = (mca_pml_base_request_t *)(*request);
tmp_ddt_size = (breq->req_datatype)->size;
/*
* Find the peer reference
*/
if( OMPI_SUCCESS != (ret = find_peer_in_comm(breq->req_comm,
breq->req_peer,
&peer_ref) ) ){
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: req_start(): Failed to find peer_ref\n");
exit_status = ret;
goto DONE;
}
/* Check the send_init list */
if( OMPI_SUCCESS != (ret = traffic_message_find(&(peer_ref->send_init_list),
breq->req_count,
breq->req_tag,
breq->req_peer,
breq->req_comm->c_contextid,
tmp_ddt_size,
&msg_ref,
PERSIST_MARKER
) ) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: pml_start(): Unable to find the proper (send_init) message ref for this recv\n");
exit_status = ret;
goto DONE;
}
if( NULL == msg_ref ) {
ERROR_SHOULD_NEVER_HAPPEN("crcp:bkmrk: pml_start(): Could not find message ref");
exit_status = OMPI_ERROR;
goto DONE;
} else {
traffic_message_start(msg_ref,
peer_ref,
request,
&(peer_ref->send_init_list),
&content_ref);
if( !content_ref->already_drained ) {
/* Account for this inflight send */
peer_ref->total_msgs_sent += 1;
}
}
DONE:
return exit_status;
}
static int ompi_crcp_bkmrk_request_complete_isend_init(struct ompi_request_t *request,
ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref,
int src, int tag, int tmp_ddt_size)
{
int ret, exit_status = OMPI_SUCCESS;
mca_pml_base_request_t *breq = NULL;
ompi_crcp_bkmrk_pml_traffic_message_ref_t *msg_ref = NULL;
ompi_crcp_bkmrk_pml_message_content_ref_t *content_ref = NULL;
breq = (mca_pml_base_request_t *)request;
/* Check the isend_init list */
if( OMPI_SUCCESS != (ret = traffic_message_find(&(peer_ref->send_init_list),
breq->req_count,
tag, src,
breq->req_comm->c_contextid,
tmp_ddt_size,
&msg_ref,
FIND_MSG_TRUE
) ) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: req_complete: Unable to find the proper (send_init) message ref for this complete\n");
exit_status = ret;
goto DONE;
}
if( NULL == msg_ref ) {
/*
* It is possible that we did not 'find' the message because
* we could have previously marked it as done. Due to the
* logic in the Request Wait/Test routines we could
* receive multiple request complete calls for the
* same request.
*
* It is possible that we have 'completed' this message previously,
* so this case can occur during normal operation.
* This is caused by us checking for completeness twice in ompi_request_wait_all.
*/
OPAL_OUTPUT_VERBOSE((15, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: request_complete: No match found for this request :( %d, %d ): [%d/%d,%d]\n",
peer_ref->total_msgs_sent, peer_ref->total_msgs_recvd,
breq->req_peer, src, breq->req_comm->c_contextid));
exit_status = OMPI_SUCCESS;
goto DONE;
}
/* Mark request as inactive */
traffic_message_find_mark_persistent(msg_ref, &request,
true, /* Find currently active */
false, /* Mark as inactive */
&content_ref);
TRAFFIC_MSG_DUMP_PEER(15, (peer_ref, "-- Request Complete (Send_init) --", true));
if( !content_ref->already_drained ) {
msg_ref->done++;
msg_ref->active--;
} else {
msg_ref->active_drain--;
content_ref->already_drained = false;
}
OPAL_OUTPUT_VERBOSE((25, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: req_complete: Marked Message... ( %d, %d )\n",
peer_ref->total_msgs_sent, peer_ref->total_msgs_recvd));
DONE:
return exit_status;
}
ompi_crcp_base_pml_state_t* ompi_crcp_bkmrk_pml_isend(
void *buf, size_t count,
ompi_datatype_t *datatype,
int dst, int tag,
mca_pml_base_send_mode_t mode,
struct ompi_communicator_t* comm,
struct ompi_request_t **request,
ompi_crcp_base_pml_state_t* pml_state )
{
ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref = NULL;
ompi_crcp_bkmrk_pml_traffic_message_ref_t *msg_ref = NULL;
ompi_crcp_bkmrk_pml_state_t *coord_state = NULL;
int exit_status = OMPI_SUCCESS;
int ret;
OPAL_OUTPUT_VERBOSE((30, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: pml_isend()"));
/*
* Before the PML gets the message:
* - Setup structure to track the message
*/
if( OMPI_CRCP_PML_PRE == pml_state->state ) {
/*
* Find the peer reference
*/
if( OMPI_SUCCESS != (ret = find_peer_in_comm(comm, dst, &peer_ref) ) ){
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: isend: Failed to find peer_ref\n");
exit_status = ret;
goto DONE;
}
/*
* Archive the message Message Object
*/
traffic_message_append(peer_ref, &(peer_ref->isend_list),
COORD_MSG_TYPE_I_SEND,
count, datatype, 0, tag, dst, comm,
&msg_ref);
/* Bookkeeping */
peer_ref->total_msgs_sent += 1;
/* Save the pointers */
CREATE_COORD_STATE(coord_state, pml_state,
peer_ref, msg_ref);
coord_state->p_super.error_code = OMPI_SUCCESS;
return &coord_state->p_super;
}
/*
* After PML is done, update message reference
*/
else if( OMPI_CRCP_PML_POST == pml_state->state ) {
ompi_crcp_base_pml_state_t *rtn_state = NULL;
ompi_crcp_bkmrk_pml_message_content_ref_t *new_content = NULL;
EXTRACT_COORD_STATE(pml_state, coord_state, rtn_state,
peer_ref, msg_ref);
/*
* Update Message
*/
HOKE_CONTENT_REF_ALLOC(new_content, ret);
new_content->buffer = NULL; /* No Tracked */
new_content->request = *request;
new_content->done = false;
new_content->active = true;
new_content->already_posted = true;
new_content->already_drained = false;
OBJ_RETAIN(*request);
opal_list_append(&(msg_ref->msg_contents), &(new_content->super) );
TRAFFIC_MSG_DUMP_PEER(15, (peer_ref, "-- Append Message (isend) --", true));
CRCP_COORD_STATE_RETURN(coord_state);
rtn_state->error_code = OMPI_SUCCESS;
return rtn_state;
}
DONE:
pml_state->error_code = exit_status;
return pml_state;
}
static int ompi_crcp_bkmrk_request_complete_isend(struct ompi_request_t *request,
ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref,
int src, int tag, int tmp_ddt_size)
{
int ret, exit_status = OMPI_SUCCESS;
mca_pml_base_request_t *breq = NULL;
ompi_crcp_bkmrk_pml_traffic_message_ref_t *msg_ref = NULL;
ompi_crcp_bkmrk_pml_message_content_ref_t *content_ref = NULL;
breq = (mca_pml_base_request_t *)request;
/* Check the isend list */
if( OMPI_SUCCESS != (ret = traffic_message_find(&(peer_ref->isend_list),
breq->req_count,
tag, src,
breq->req_comm->c_contextid,
tmp_ddt_size,
&msg_ref,
FIND_MSG_TRUE
) ) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: req_complete: Unable to find the proper (isend) message ref for this complete\n");
exit_status = ret;
goto DONE;
}
if( NULL == msg_ref ) {
/*
* It is possible that we did not 'find' the message because
* we could have previously marked it as done. Due to the
* logic in the Request Wait/Test routines we could
* receive multiple request complete calls for the
* same request.
*
* It is possible that we have 'completed' this message previously,
* so this case can occur during normal operation.
* This is caused by us checking for completeness twice in ompi_request_wait_all.
*/
OPAL_OUTPUT_VERBOSE((15, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: request_complete: No match found for this request :( %d, %d ): [%d/%d,%d]\n",
peer_ref->total_msgs_sent, peer_ref->total_msgs_recvd,
breq->req_peer, src, breq->req_comm->c_contextid));
exit_status = OMPI_SUCCESS;
goto DONE;
}
OPAL_OUTPUT_VERBOSE((15, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: req_complete: Matched an iSend: total = %d",
peer_ref->total_msgs_sent));
/* Strip off an isend request */
traffic_message_grab_content(msg_ref, &content_ref, true, true); /* Remove, prefer already_drained */
if( !content_ref->already_drained ) {
msg_ref->done++;
msg_ref->active--;
} else {
msg_ref->active_drain--;
content_ref->already_drained = false;
}
HOKE_CONTENT_REF_RETURN(content_ref);
TRAFFIC_MSG_DUMP_PEER(15, (peer_ref, "-- Request Complete (iSend) --", true));
OPAL_OUTPUT_VERBOSE((25, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: req_complete: Marked Message... ( %d, %d )\n",
peer_ref->total_msgs_sent, peer_ref->total_msgs_recvd));
DONE:
return exit_status;
}
ompi_crcp_base_pml_state_t* ompi_crcp_bkmrk_pml_send(
void *buf, size_t count,
ompi_datatype_t *datatype,
int dst, int tag,
mca_pml_base_send_mode_t mode,
struct ompi_communicator_t* comm,
ompi_crcp_base_pml_state_t* pml_state )
{
ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref = NULL;
ompi_crcp_bkmrk_pml_traffic_message_ref_t *msg_ref = NULL;
ompi_crcp_bkmrk_pml_state_t *coord_state = NULL;
int exit_status = OMPI_SUCCESS;
int ret;
OPAL_OUTPUT_VERBOSE((30, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: pml_send()"));
/*
* Before the PML gets the message:
* - Setup structure to track the message
*/
if( OMPI_CRCP_PML_PRE == pml_state->state ) {
/*
* Find the peer reference
*/
if( OMPI_SUCCESS != (ret = find_peer_in_comm(comm, dst, &peer_ref) ) ){
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: send: Failed to find peer_ref\n");
exit_status = ret;
goto DONE;
}
/*
* Archive the message Message Object
*/
traffic_message_append(peer_ref, &(peer_ref->send_list),
COORD_MSG_TYPE_B_SEND,
count, datatype, 0, tag, dst, comm,
&msg_ref);
/* Bookkeeping */
peer_ref->total_msgs_sent += 1;
current_msg_id = msg_ref->msg_id;
current_msg_type = COORD_MSG_TYPE_B_SEND;
/* Save the pointers */
CREATE_COORD_STATE(coord_state, pml_state,
peer_ref, msg_ref);
coord_state->p_super.error_code = OMPI_SUCCESS;
return &coord_state->p_super;
}
/*
* After PML is done, update message reference
*/
else if( OMPI_CRCP_PML_POST == pml_state->state ) {
ompi_crcp_base_pml_state_t *rtn_state = NULL;
EXTRACT_COORD_STATE(pml_state, coord_state, rtn_state,
peer_ref, msg_ref);
/*
* Update Message
*/
msg_ref->done++;
msg_ref->active--;
current_msg_id = 0;
current_msg_type = COORD_MSG_TYPE_UNKNOWN;
TRAFFIC_MSG_DUMP_PEER(15, (peer_ref, "Send done", true));
CRCP_COORD_STATE_RETURN(coord_state);
rtn_state->error_code = OMPI_SUCCESS;
return rtn_state;
}
DONE:
pml_state->error_code = exit_status;
return pml_state;
}
/**************** Recv *****************/
ompi_crcp_base_pml_state_t* ompi_crcp_bkmrk_pml_irecv_init(
void *buf, size_t count,
ompi_datatype_t *datatype,
int src, int tag,
struct ompi_communicator_t* comm,
struct ompi_request_t **request,
ompi_crcp_base_pml_state_t* pml_state)
{
ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref = NULL;
ompi_crcp_bkmrk_pml_traffic_message_ref_t *msg_ref = NULL;
ompi_crcp_bkmrk_pml_state_t *coord_state = NULL;
int exit_status = OMPI_SUCCESS;
int ret;
OPAL_OUTPUT_VERBOSE((30, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: pml_irecv_init()"));
/*
* Before PML Call
* - Determine if this can be satisfied from the drained list
* - Otherwise create a new reference to it so we can track it
*/
if( OMPI_CRCP_PML_PRE == pml_state->state) {
/*
* A message will never be on the drained list for this function since
* it does not actually receive anything, just sets up the system.
* The receive for these reqeusts are done in the start() and wait()
* commands.
*/
/*
* Find the Peer
*/
if( MPI_ANY_SOURCE == src || src < 0) {
/*
* Archive the message Message Object
*/
traffic_message_append(NULL, &(unknown_persist_recv_list),
COORD_MSG_TYPE_P_RECV,
count, datatype, 0, tag, src, comm,
&msg_ref);
CREATE_COORD_STATE(coord_state, pml_state,
NULL, msg_ref);
}
else {
if( OMPI_SUCCESS != (ret = find_peer_in_comm(comm, src, &peer_ref) ) ){
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: recv: Failed to find peer_ref\n");
exit_status = ret;
goto DONE;
}
/*
* Archive the message Message Object
*/
traffic_message_append(peer_ref, &(peer_ref->recv_init_list),
COORD_MSG_TYPE_P_RECV,
count, datatype, 0, tag, src, comm,
&msg_ref);
CREATE_COORD_STATE(coord_state, pml_state,
peer_ref, msg_ref);
}
coord_state->p_super.error_code = OMPI_SUCCESS;
return &coord_state->p_super;
}
/*
* Post PML Call
* - bookkeeping...
*/
else if( OMPI_CRCP_PML_POST == pml_state->state) {
ompi_crcp_base_pml_state_t *rtn_state = NULL;
ompi_crcp_bkmrk_pml_message_content_ref_t *new_content = NULL;
EXTRACT_COORD_STATE(pml_state, coord_state, rtn_state,
peer_ref, msg_ref);
/*
* Do the update
*/
HOKE_CONTENT_REF_ALLOC(new_content, ret);
new_content->buffer = buf;
new_content->request = *request;
new_content->done = false;
new_content->active = false;
new_content->already_posted = true;
new_content->already_drained = false;
OBJ_RETAIN(*request);
opal_list_append(&(msg_ref->msg_contents), &(new_content->super) );
CRCP_COORD_STATE_RETURN(coord_state);
rtn_state->error_code = OMPI_SUCCESS;
return rtn_state;
}
DONE:
pml_state->error_code = exit_status;
return pml_state;
}
static int ompi_crcp_bkmrk_pml_start_drain_irecv_init(ompi_request_t **request, bool *found_drain)
{
int ret, exit_status = OMPI_SUCCESS;
ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref = NULL;
ompi_crcp_bkmrk_pml_traffic_message_ref_t *msg_ref = NULL;
ompi_crcp_bkmrk_pml_drain_message_ref_t *drain_msg_ref = NULL;
ompi_crcp_bkmrk_pml_message_content_ref_t *content_ref = NULL;
mca_pml_base_request_t *breq = NULL;
size_t tmp_ddt_size = 0;
*found_drain = false;
breq = (mca_pml_base_request_t *)(*request);
tmp_ddt_size = (breq->req_datatype)->size;
/*
* If peer rank is given then find the peer reference
*/
if( 0 <= breq->req_peer ) {
if( OMPI_SUCCESS != (ret = find_peer_in_comm(breq->req_comm,
breq->req_peer,
&peer_ref) ) ){
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: pml_start(): Failed to find peer_ref\n");
exit_status = ret;
goto DONE;
}
if( OMPI_SUCCESS != (ret = traffic_message_find(&(peer_ref->recv_init_list),
breq->req_count,
breq->req_tag,
breq->req_peer,
breq->req_comm->c_contextid,
tmp_ddt_size,
&msg_ref,
PERSIST_MARKER
) ) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: pml_start(): Unable to find the proper (recv) message ref for this recv\n");
exit_status = ret;
goto DONE;
}
}
/*
* Otherwise peer is not known
*/
else {
if( OMPI_SUCCESS != (ret = traffic_message_find(&(unknown_persist_recv_list),
breq->req_count,
breq->req_tag,
INVALID_INT,
breq->req_comm->c_contextid,
tmp_ddt_size,
&msg_ref,
PERSIST_MARKER
) ) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: pml_start(): Unable to find the proper (recv) message ref for this recv\n");
exit_status = ret;
goto DONE;
}
}
/*
* No message found :(
*/
if( NULL == msg_ref ) {
ERROR_SHOULD_NEVER_HAPPEN("crcp:bkmrk: pml_start(): Could not find message ref");
exit_status = OMPI_ERROR;
goto DONE;
}
/*
* See if this mesage was already drained.
*/
if( NULL != peer_ref ) {
if( OMPI_SUCCESS != (ret = drain_message_find(&(peer_ref->drained_list),
msg_ref->count, msg_ref->tag, msg_ref->rank,
msg_ref->comm->c_contextid, msg_ref->ddt_size,
&drain_msg_ref,
&content_ref) ) ) {
ERROR_SHOULD_NEVER_HAPPEN("crcp:bkmrk: pml_start(): Failed trying to find a drained message.");
exit_status = ret;
goto DONE;
}
} else {
if( OMPI_SUCCESS != (ret = drain_message_find_any(msg_ref->count, msg_ref->tag, msg_ref->rank,
msg_ref->comm, msg_ref->ddt_size,
&drain_msg_ref,
&content_ref,
&peer_ref) ) ) {
ERROR_SHOULD_NEVER_HAPPEN("crcp:bkmrk: pml_start(): Failed trying to find a drained message.");
exit_status = ret;
goto DONE;
}
}
/*
* Found a drained message!
*/
if( NULL != drain_msg_ref ) {
*found_drain = true;
OPAL_OUTPUT_VERBOSE((12, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: pml_start(): Matched a drained message..."));
if( OMPI_SUCCESS != (ret = drain_message_copy_remove_persistent(drain_msg_ref,
content_ref,
msg_ref,
*request,
peer_ref) ) ) {
opal_output( mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: pml_start(): Datatype copy failed (%d)",
ret);
}
peer_ref->total_drained_msgs -= 1;
}
DONE:
return exit_status;
}
static int ompi_crcp_bkmrk_pml_start_irecv_init(ompi_request_t **request)
{
int ret, exit_status = OMPI_SUCCESS;
ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref = NULL;
ompi_crcp_bkmrk_pml_traffic_message_ref_t *msg_ref = NULL;
mca_pml_base_request_t *breq = NULL;
size_t tmp_ddt_size = 0;
breq = (mca_pml_base_request_t *)(*request);
tmp_ddt_size = (breq->req_datatype)->size;
/*
* If peer rank is given then find the peer reference
*/
if( 0 <= breq->req_peer ) {
if( OMPI_SUCCESS != (ret = find_peer_in_comm(breq->req_comm,
breq->req_peer,
&peer_ref) ) ){
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: pml_start(): Failed to find peer_ref\n");
exit_status = ret;
goto DONE;
}
if( OMPI_SUCCESS != (ret = traffic_message_find(&(peer_ref->recv_init_list),
breq->req_count,
breq->req_tag,
breq->req_peer,
breq->req_comm->c_contextid,
tmp_ddt_size,
&msg_ref,
PERSIST_MARKER
) ) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: pml_start(): Unable to find the proper (recv) message ref for this recv\n");
exit_status = ret;
goto DONE;
}
if( NULL != msg_ref ) {
traffic_message_start(msg_ref,
peer_ref,
request,
&(peer_ref->recv_init_list),
NULL);
}
}
/*
* Else peer is not known
*/
else {
if( OMPI_SUCCESS != (ret = traffic_message_find(&(unknown_persist_recv_list),
breq->req_count,
breq->req_tag,
INVALID_INT,
breq->req_comm->c_contextid,
tmp_ddt_size,
&msg_ref,
PERSIST_MARKER
) ) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: pml_start(): Unable to find the proper (recv) message ref for this recv\n");
exit_status = ret;
goto DONE;
}
if( NULL != msg_ref ) {
traffic_message_start(msg_ref,
NULL,
request,
&(unknown_persist_recv_list),
NULL);
}
}
if( NULL == msg_ref ) {
ERROR_SHOULD_NEVER_HAPPEN("crcp:bkmrk: pml_start(): Could not find message ref");
exit_status = OMPI_ERROR;
goto DONE;
}
DONE:
return exit_status;
}
static int ompi_crcp_bkmrk_request_complete_irecv_init(struct ompi_request_t *request,
ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref,
int src, int tag, int tmp_ddt_size)
{
int ret, exit_status = OMPI_SUCCESS;
mca_pml_base_request_t *breq = NULL;
ompi_crcp_bkmrk_pml_traffic_message_ref_t *msg_ref = NULL, *new_msg_ref = NULL;
ompi_crcp_bkmrk_pml_message_content_ref_t *content_ref = NULL;
breq = (mca_pml_base_request_t *)request;
/*
* Check the irecv_init list
*/
if( OMPI_SUCCESS != (ret = traffic_message_find(&(peer_ref->recv_init_list),
breq->req_count,
tag, src,
breq->req_comm->c_contextid,
tmp_ddt_size,
&msg_ref,
FIND_MSG_TRUE
) ) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: req_complete: Unable to find the proper (recv_init) message ref for this complete\n");
exit_status = ret;
goto DONE;
}
/*
* If not found, check the unknown_irecv_list
*/
if( NULL == msg_ref ) {
if( OMPI_SUCCESS != (ret = traffic_message_find(&(unknown_persist_recv_list),
breq->req_count,
tag,
INVALID_INT,
breq->req_comm->c_contextid,
tmp_ddt_size,
&msg_ref,
FIND_MSG_TRUE
) ) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: requ_complete: Unable to find the proper (recv_init) message ref for this complete\n");
exit_status = ret;
goto DONE;
}
if( NULL != msg_ref ) {
traffic_message_move(msg_ref,
COORD_MSG_TYPE_P_RECV,
NULL, &(unknown_persist_recv_list),
peer_ref, &(peer_ref->recv_init_list),
Fix a problem with persistent sends found in MTT regression testing on Odin. This commit was SVN r19707.
2008-10-08 18:11:55 +04:00
&new_msg_ref,
true);
Replace the old coordinated component ('coord') and replace it with a much more refined version ('bkmrk'). The new component fixes a number of problems with the old component. The core algorithm is the same, but by changing the data strucutres a bit we have improved performance and memory utilization. There are still a couple corner cases that still need some work. However, I did not want to delay bringing this into the trunk (and v1.3 branch) for too much longer. This commit was SVN r19537.
2008-09-10 22:29:17 +04:00
msg_ref = new_msg_ref;
}
}
/*
* If still not found, then we must have completed this already
*/
if( NULL == msg_ref ) {
/*
* It is possible that we did not 'find' the message because
* we could have previously marked it as done. Due to the
* logic in the Request Wait/Test routines we could
* receive multiple request complete calls for the
* same request.
*
* It is possible that we have 'completed' this message previously,
* so this case can occur during normal operation.
* This is caused by us checking for completeness twice in ompi_request_wait_all.
*/
OPAL_OUTPUT_VERBOSE((15, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: request_complete: No match found for this request :( %d, %d ): [%d/%d,%d]\n",
peer_ref->total_msgs_sent, peer_ref->total_msgs_recvd,
breq->req_peer, src, breq->req_comm->c_contextid));
exit_status = OMPI_SUCCESS;
goto DONE;
}
/*
* Mark request as inactive
* Only increment the total count if this was not accounted for in the last checkpoint
*/
traffic_message_find_mark_persistent(msg_ref, &request,
true, /* Find currently active */
false, /* Mark as inactive */
&content_ref);
if( !content_ref->already_drained ) {
peer_ref->total_msgs_recvd += 1;
msg_ref->done++;
msg_ref->active--;
} else {
msg_ref->active_drain--;
content_ref->already_drained = false;
}
/* Do not return the content_ref, persistent sends re-use these */
if( NULL == new_msg_ref ) {
TRAFFIC_MSG_DUMP_PEER(15, (peer_ref, "-- Request Complete (Recv_Init) --", true));
} else {
TRAFFIC_MSG_DUMP_PEER(15, (peer_ref, "-- Request Complete (Recv_init - Unknown) --", true));
}
OPAL_OUTPUT_VERBOSE((25, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: req_complete: Marked Message... ( %d, %d )\n",
peer_ref->total_msgs_sent, peer_ref->total_msgs_recvd));
DONE:
return exit_status;
}
ompi_crcp_base_pml_state_t* ompi_crcp_bkmrk_pml_irecv(
void *buf, size_t count,
ompi_datatype_t *datatype,
int src, int tag,
struct ompi_communicator_t* comm,
struct ompi_request_t **request,
ompi_crcp_base_pml_state_t* pml_state )
{
int ret, exit_status = OMPI_SUCCESS;
ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref = NULL;
ompi_crcp_bkmrk_pml_traffic_message_ref_t *msg_ref = NULL;
ompi_crcp_bkmrk_pml_state_t *coord_state = NULL;
bool found_drain = false;
OPAL_OUTPUT_VERBOSE((30, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: pml_irecv()"));
/*
* Before PML Call
* - Determine if this can be satisfied from the drained list
* - Otherwise create a new reference to it so we can track it
*/
if( OMPI_CRCP_PML_PRE == pml_state->state) {
/*
* Check to see if this message is in the drained message list
*/
found_drain = false;
if( OMPI_SUCCESS != (ret = drain_message_check_recv(buf, count, datatype,
&src, &tag, comm, request, NULL,
&found_drain) ) ) {
ERROR_SHOULD_NEVER_HAPPEN("crcp:bkmrk: pml_recv(): Failed trying to find a drained message.");
exit_status = ret;
goto DONE;
}
if( found_drain ) {
/* Do *not* increment:
* peer_ref->total_msgs_recvd += 1;
* Because we accounted for this message during the last checkpoint.
*/
/* This will identify to the wrapper that this message is complete */
pml_state->state = OMPI_CRCP_PML_DONE;
pml_state->error_code = OMPI_SUCCESS;
return pml_state;
}
/*
* Otherwise the message is not drained (common case)
*/
else {
/*
* Find the Peer
*/
if( MPI_ANY_SOURCE == src || src < 0) {
/*
* Archive the message Message Object
*/
traffic_message_append(NULL, &(unknown_recv_from_list),
COORD_MSG_TYPE_I_RECV,
count, datatype, 0, tag, src, comm,
&msg_ref);
CREATE_COORD_STATE(coord_state, pml_state,
NULL, msg_ref);
}
else {
if( OMPI_SUCCESS != (ret = find_peer_in_comm(comm, src, &peer_ref) ) ){
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: pml_irecv(): Failed to find peer_ref\n");
exit_status = ret;
goto DONE;
}
/*
* Archive the message Message Object
*/
traffic_message_append(peer_ref, &(peer_ref->irecv_list),
COORD_MSG_TYPE_I_RECV,
count, datatype, 0, tag, src, comm,
&msg_ref);
CREATE_COORD_STATE(coord_state, pml_state,
peer_ref, msg_ref);
}
coord_state->p_super.error_code = OMPI_SUCCESS;
return &coord_state->p_super;
}
}
/*
* Post PML Call
* - bookkeeping...
*/
else if( OMPI_CRCP_PML_POST == pml_state->state) {
ompi_crcp_base_pml_state_t *rtn_state = NULL;
ompi_crcp_bkmrk_pml_message_content_ref_t *new_content = NULL;
EXTRACT_COORD_STATE(pml_state, coord_state, rtn_state,
peer_ref, msg_ref);
/*
* Do the update
*/
HOKE_CONTENT_REF_ALLOC(new_content, ret);
new_content->buffer = NULL; /* No tracked */
new_content->request = *request;
new_content->done = false;
new_content->active = true;
new_content->already_posted = true;
new_content->already_drained = false;
OBJ_RETAIN(*request);
opal_list_append(&(msg_ref->msg_contents), &(new_content->super) );
TRAFFIC_MSG_DUMP_PEER(15, (peer_ref, "-- Append Message (irecv) --", true));
CRCP_COORD_STATE_RETURN(coord_state);
rtn_state->error_code = OMPI_SUCCESS;
return rtn_state;
}
DONE:
pml_state->error_code = exit_status;
return pml_state;
}
static int ompi_crcp_bkmrk_request_complete_irecv(struct ompi_request_t *request,
ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref,
int src, int tag, int tmp_ddt_size)
{
int ret, exit_status = OMPI_SUCCESS;
mca_pml_base_request_t *breq = NULL;
ompi_crcp_bkmrk_pml_traffic_message_ref_t *msg_ref = NULL, *new_msg_ref = NULL;
ompi_crcp_bkmrk_pml_message_content_ref_t *content_ref = NULL;
breq = (mca_pml_base_request_t *)request;
/*
* Check the irecv list
*/
if( OMPI_SUCCESS != (ret = traffic_message_find(&(peer_ref->irecv_list),
breq->req_count,
tag, src,
breq->req_comm->c_contextid,
tmp_ddt_size,
&msg_ref,
FIND_MSG_TRUE
) ) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: req_complete: Unable to find the proper (irecv) message ref for this complete\n");
exit_status = ret;
goto DONE;
}
/*
* If not found, try the unknown_irecv_list
*/
if( NULL == msg_ref ) {
if( OMPI_SUCCESS != (ret = traffic_message_find(&(unknown_recv_from_list),
breq->req_count,
tag,
INVALID_INT,
breq->req_comm->c_contextid,
tmp_ddt_size,
&msg_ref,
FIND_MSG_TRUE
) ) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: req_complete: Unable to find the proper (recv_init) message ref for this complete\n");
exit_status = ret;
goto DONE;
}
if( NULL != msg_ref ) {
traffic_message_move(msg_ref,
COORD_MSG_TYPE_I_RECV,
NULL, &(unknown_recv_from_list),
peer_ref, &(peer_ref->irecv_list),
Fix a problem with persistent sends found in MTT regression testing on Odin. This commit was SVN r19707.
2008-10-08 18:11:55 +04:00
&new_msg_ref,
true);
Replace the old coordinated component ('coord') and replace it with a much more refined version ('bkmrk'). The new component fixes a number of problems with the old component. The core algorithm is the same, but by changing the data strucutres a bit we have improved performance and memory utilization. There are still a couple corner cases that still need some work. However, I did not want to delay bringing this into the trunk (and v1.3 branch) for too much longer. This commit was SVN r19537.
2008-09-10 22:29:17 +04:00
msg_ref = new_msg_ref;
}
}
/*
* If still not found, then must have completed this twice
*/
if( NULL == msg_ref ) {
/*
* It is possible that we did not 'find' the message because
* we could have previously marked it as done. Due to the
* logic in the Request Wait/Test routines we could
* receive multiple request complete calls for the
* same request.
*
* It is possible that we have 'completed' this message previously,
* so this case can occur during normal operation.
* This is caused by us checking for completeness twice in ompi_request_wait_all.
*/
OPAL_OUTPUT_VERBOSE((15, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: request_complete: No match found for this request :( %d, %d ): [%d/%d,%d]\n",
peer_ref->total_msgs_sent, peer_ref->total_msgs_recvd,
breq->req_peer, src, breq->req_comm->c_contextid));
exit_status = OMPI_SUCCESS;
goto DONE;
}
/* Strip off an irecv request
* Only increment the total count if this was not accounted for in the last checkpoint
*/
traffic_message_grab_content(msg_ref, &content_ref, true, true); /* Remove, prefer already_drained */
if( !content_ref->already_drained ) {
peer_ref->total_msgs_recvd += 1;
msg_ref->done++;
msg_ref->active--;
} else {
msg_ref->active_drain--;
content_ref->already_drained = false;
}
HOKE_CONTENT_REF_RETURN(content_ref);
if( NULL == new_msg_ref ) {
TRAFFIC_MSG_DUMP_PEER(15, (peer_ref, "-- Request Complete (iRecv) --", true));
} else {
TRAFFIC_MSG_DUMP_PEER(15, (peer_ref, "-- Request Complete (iRecv - Unknown) --", true));
}
OPAL_OUTPUT_VERBOSE((15, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: req_complete: Matched an iRecv: total = %d",
peer_ref->total_msgs_recvd));
DONE:
return exit_status;
}
ompi_crcp_base_pml_state_t* ompi_crcp_bkmrk_pml_recv(
void *buf, size_t count,
ompi_datatype_t *datatype,
int src, int tag,
struct ompi_communicator_t* comm,
ompi_status_public_t* status,
ompi_crcp_base_pml_state_t* pml_state)
{
ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref = NULL;
ompi_crcp_bkmrk_pml_traffic_message_ref_t *msg_ref = NULL, *new_msg_ref = NULL;
ompi_crcp_bkmrk_pml_state_t *coord_state = NULL;
bool found_drain = false;
int exit_status = OMPI_SUCCESS;
int ret;
OPAL_OUTPUT_VERBOSE((30, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: pml_recv()"));
/*
* Before PML Call
* - Determine if this can be satisfied from the drained list
* - Otherwise create a new reference to it so we can track it
*/
if( OMPI_CRCP_PML_PRE == pml_state->state) {
/*
* Check to see if this message is in the drained message list
*/
found_drain = false;
if( OMPI_SUCCESS != (ret = drain_message_check_recv(buf, count, datatype,
&src, &tag, comm, NULL, &status,
&found_drain) ) ) {
ERROR_SHOULD_NEVER_HAPPEN("crcp:bkmrk: pml_recv(): Failed trying to find a drained message.");
exit_status = ret;
goto DONE;
}
if( found_drain ) {
/* Do *not* increment:
* peer_ref->total_msgs_recvd += 1;
* Because we accounted for this message during the last checkpoint.
*/
/* This will identify to the wrapper that this message is complete */
pml_state->state = OMPI_CRCP_PML_DONE;
pml_state->error_code = OMPI_SUCCESS;
return pml_state;
}
/*
* Otherwise the message is not drained (common case)
*/
else {
/*
* Find the Peer
*/
if( MPI_ANY_SOURCE == src || src < 0) {
traffic_message_append(NULL, &(unknown_recv_from_list),
COORD_MSG_TYPE_B_RECV,
count, datatype, 0, tag, src, comm,
&msg_ref);
CREATE_COORD_STATE(coord_state, pml_state,
NULL, msg_ref);
}
else {
if( OMPI_SUCCESS != (ret = find_peer_in_comm(comm, src, &peer_ref) ) ){
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: pml_recv(): Failed to find peer_ref\n");
exit_status = ret;
goto DONE;
}
traffic_message_append(peer_ref, &(peer_ref->recv_list),
COORD_MSG_TYPE_B_RECV,
count, datatype, 0, tag, src, comm,
&msg_ref);
CREATE_COORD_STATE(coord_state, pml_state,
peer_ref, msg_ref);
}
/* Bookkeeping */
current_msg_id = msg_ref->msg_id;
current_msg_type = COORD_MSG_TYPE_B_RECV;
coord_state->p_super.error_code = OMPI_SUCCESS;
return &coord_state->p_super;
}
}
/*
* Post PML Call
* - bookkeeping...
*/
else if( OMPI_CRCP_PML_POST == pml_state->state) {
ompi_crcp_base_pml_state_t *rtn_state = NULL;
EXTRACT_COORD_STATE(pml_state, coord_state, rtn_state,
peer_ref, msg_ref);
/*
* If MPI_ANY_SOUCE, then move the message from the unknown list
* to the list associated with the resolved process.
*/
if( NULL == peer_ref ) {
src = status->MPI_SOURCE;
if( OMPI_SUCCESS != (ret = find_peer_in_comm(comm, src, &peer_ref) ) ){
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: pml_recv(): Failed to resolve peer_ref (rank %d)\n",
src);
exit_status = ret;
goto DONE;
}
traffic_message_move(msg_ref,
COORD_MSG_TYPE_B_RECV,
NULL, &(unknown_recv_from_list),
peer_ref, &(peer_ref->recv_list),
Fix a problem with persistent sends found in MTT regression testing on Odin. This commit was SVN r19707.
2008-10-08 18:11:55 +04:00
&new_msg_ref,
false);
Replace the old coordinated component ('coord') and replace it with a much more refined version ('bkmrk'). The new component fixes a number of problems with the old component. The core algorithm is the same, but by changing the data strucutres a bit we have improved performance and memory utilization. There are still a couple corner cases that still need some work. However, I did not want to delay bringing this into the trunk (and v1.3 branch) for too much longer. This commit was SVN r19537.
2008-09-10 22:29:17 +04:00
new_msg_ref->done++;
new_msg_ref->active--;
} else {
/*
* Do the update
*/
msg_ref->done++;
msg_ref->active--;
}
peer_ref->total_msgs_recvd += 1;
current_msg_id = 0;
current_msg_type = COORD_MSG_TYPE_UNKNOWN;
TRAFFIC_MSG_DUMP_PEER(15, (peer_ref, "Recv Done", true));
CRCP_COORD_STATE_RETURN(coord_state);
rtn_state->error_code = OMPI_SUCCESS;
return rtn_state;
}
DONE:
pml_state->error_code = exit_status;
return pml_state;
}
/**************** Start *****************/
/* Start is connected to irecv_start or isend_start */
static ompi_request_type_t * coord_start_req_types = NULL;
ompi_crcp_base_pml_state_t* ompi_crcp_bkmrk_pml_start(
size_t count,
ompi_request_t** requests,
ompi_crcp_base_pml_state_t* pml_state )
{
int ret, exit_status = OMPI_SUCCESS;
mca_pml_base_request_t *breq = NULL;
size_t tmp_ddt_size = 0;
size_t iter_req;
bool found_drain = false;
OPAL_OUTPUT_VERBOSE((30, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: pml_start()"));
/*
* Handle all start() on send requests
*/
if( OMPI_CRCP_PML_POST == pml_state->state ) {
for(iter_req = 0; iter_req < count; iter_req++) {
breq = (mca_pml_base_request_t *)requests[iter_req];
if(breq->req_type == MCA_PML_REQUEST_SEND ) {
if( OMPI_SUCCESS != (ret = ompi_crcp_bkmrk_pml_start_isend_init(&(requests[iter_req]))) ) {
exit_status = ret;
goto DONE;
}
}
}
}
/*
* Handle all start() on recv requests
* - Pre: Check drain queue for a match
* - Post: Start the message, unless drained
*/
if( OMPI_CRCP_PML_PRE == pml_state->state ) {
/*
* Mark all saved requests as NOOP
*/
coord_start_req_types = (ompi_request_type_t *)malloc(sizeof(ompi_request_type_t) * count);
for(iter_req = 0; iter_req < count; iter_req++) {
coord_start_req_types[iter_req] = OMPI_REQUEST_NOOP;
}
for(iter_req = 0; iter_req < count; iter_req++) {
breq = (mca_pml_base_request_t *)requests[iter_req];
tmp_ddt_size = (breq->req_datatype)->size;
if( breq->req_type == MCA_PML_REQUEST_RECV ) {
found_drain = false;
if( OMPI_SUCCESS != (ret = ompi_crcp_bkmrk_pml_start_drain_irecv_init(&(requests[iter_req]), &found_drain)) ) {
exit_status = ret;
goto DONE;
}
if( found_drain ) {
coord_start_req_types[iter_req] = requests[iter_req]->req_type;
requests[iter_req]->req_type = OMPI_REQUEST_NOOP;
requests[iter_req]->req_complete = true;
}
}
}
goto DONE;
}
else if( OMPI_CRCP_PML_POST == pml_state->state) {
for(iter_req = 0; iter_req < count; iter_req++) {
breq = (mca_pml_base_request_t *)requests[iter_req];
tmp_ddt_size = (breq->req_datatype)->size;
if (breq->req_type == MCA_PML_REQUEST_RECV) {
/*
* If this was a drained message it will have it's type set to
* OMPI_REQUEST_NOOP so the PML does not try to start it again.
* So we need to replace it with the original type, but can
* skip starting it.
*/
if( NULL != coord_start_req_types ) {
if( OMPI_REQUEST_NOOP != coord_start_req_types[iter_req] ) {
requests[iter_req]->req_type = coord_start_req_types[iter_req];
continue;
}
}
if( OMPI_SUCCESS != (ret = ompi_crcp_bkmrk_pml_start_irecv_init(&(requests[iter_req]))) ) {
exit_status = ret;
goto DONE;
}
}
}
/*
* Clear out the temporary drain type structure.
*/
if( NULL != coord_start_req_types ) {
free(coord_start_req_types);
coord_start_req_types = NULL;
}
}
DONE:
pml_state->error_code = exit_status;
return pml_state;
}
/**************** Request Completed ********/
int ompi_crcp_bkmrk_request_complete(struct ompi_request_t *request)
{
int ret, exit_status = OMPI_SUCCESS;
ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref = NULL;
mca_pml_base_request_t *breq;
size_t tmp_ddt_size = 0;
int src, tag;
OPAL_OUTPUT_VERBOSE((30, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: pml_request_complete()"));
/*
* Extract & check the PML version of the request
*/
breq = (mca_pml_base_request_t *)request;
if( (breq->req_type != MCA_PML_REQUEST_SEND &&
breq->req_type != MCA_PML_REQUEST_RECV ) || /* JJH YYY -- req_state = OMPI_REQUEST_INACTIVE ??? */
request->req_type == OMPI_REQUEST_NOOP ||
request->req_type == OMPI_REQUEST_NULL) {
exit_status = OMPI_SUCCESS;
goto DONE;
}
/* Extract source/tag/ddt_size */
src = breq->req_peer;
tag = breq->req_tag;
tmp_ddt_size = (breq->req_datatype)->size;
/*
* Find the peer reference
*/
if( MPI_ANY_SOURCE == src ) {
if( OMPI_SUCCESS != (ret = find_peer_in_comm(breq->req_comm, request->req_status.MPI_SOURCE, &peer_ref) ) ){
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: req_complete(): Failed to find peer_ref\n");
exit_status = ret;
goto DONE;
}
} else {
if( OMPI_SUCCESS != (ret = find_peer_in_comm(breq->req_comm, src, &peer_ref) ) ){
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: req_complete(): Failed to find peer_ref\n");
exit_status = ret;
goto DONE;
}
}
/*******************************
* A send request is completing
******************************/
if(breq->req_type == MCA_PML_REQUEST_SEND ) {
/*
* ISEND Case:
*/
if( false == request->req_persistent ) {
if( OMPI_SUCCESS != (ret = ompi_crcp_bkmrk_request_complete_isend(request, peer_ref,
src, tag, tmp_ddt_size) ) ) {
exit_status = ret;
goto DONE;
}
}
/*
* SEND_INIT/START Case
*/
else {
if( OMPI_SUCCESS != (ret = ompi_crcp_bkmrk_request_complete_isend_init(request, peer_ref,
src, tag, tmp_ddt_size) ) ) {
exit_status = ret;
goto DONE;
}
}
}
/***********************************
* A receive request is completing
***********************************/
else if(breq->req_type == MCA_PML_REQUEST_RECV) {
/*
* IRECV Case:
*/
if( false == request->req_persistent ) {
if( OMPI_SUCCESS != (ret = ompi_crcp_bkmrk_request_complete_irecv(request, peer_ref,
src, tag, tmp_ddt_size) ) ) {
exit_status = ret;
goto DONE;
}
}
/*
* IRECV_INIT/START Case:
*/
else {
if( OMPI_SUCCESS != (ret = ompi_crcp_bkmrk_request_complete_irecv_init(request, peer_ref,
src, tag, tmp_ddt_size) ) ) {
exit_status = ret;
goto DONE;
}
}
}
DONE:
return exit_status;
}
/**************** FT Event *****************/
ompi_crcp_base_pml_state_t* ompi_crcp_bkmrk_pml_ft_event(
int state,
ompi_crcp_base_pml_state_t* pml_state)
{
static int step_to_return_to = 0;
Commit to bring online OpenIB, MX, and shared memory support for Open MPI's checkpoint/restart functionality. Some tuning is still needed, but basic functionality is in place. There is still a problem with OpenIB and threads (external to C/R functionality). It has been reported in Ticket #1539 Additionally: * Fix a file cleanup bug in CRS Base. * Fix a possible deadlock in the TCP ft_event function * Add a mca_base_param_deregister() function to MCA base * Add whole process checkpoint timers * Add support for BTL: OpenIB, MX, Shared Memory * Add support Mpool: rdma, sm * Sundry bounds checking an cleanup in some scattered functions This commit was SVN r19756.
2008-10-16 19:09:00 +04:00
static bool first_continue_pass = false;
Replace the old coordinated component ('coord') and replace it with a much more refined version ('bkmrk'). The new component fixes a number of problems with the old component. The core algorithm is the same, but by changing the data strucutres a bit we have improved performance and memory utilization. There are still a couple corner cases that still need some work. However, I did not want to delay bringing this into the trunk (and v1.3 branch) for too much longer. This commit was SVN r19537.
2008-09-10 22:29:17 +04:00
opal_list_item_t* item = NULL;
int exit_status = OMPI_SUCCESS;
int ret;
ft_event_state = state;
if( step_to_return_to == 1 ) {
goto STEP_1;
}
OPAL_OUTPUT_VERBOSE((30, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: pml_ft_event()"));
/**************************
* Prepare for a Checkpoint
**************************/
if(OPAL_CRS_CHECKPOINT == state) {
if( OMPI_CRCP_PML_PRE != pml_state->state){
goto DONE;
}
Commit to bring online OpenIB, MX, and shared memory support for Open MPI's checkpoint/restart functionality. Some tuning is still needed, but basic functionality is in place. There is still a problem with OpenIB and threads (external to C/R functionality). It has been reported in Ticket #1539 Additionally: * Fix a file cleanup bug in CRS Base. * Fix a possible deadlock in the TCP ft_event function * Add a mca_base_param_deregister() function to MCA base * Add whole process checkpoint timers * Add support for BTL: OpenIB, MX, Shared Memory * Add support Mpool: rdma, sm * Sundry bounds checking an cleanup in some scattered functions This commit was SVN r19756.
2008-10-16 19:09:00 +04:00
if( opal_cr_timing_barrier_enabled ) {
OPAL_CR_SET_TIMER(OPAL_CR_TIMER_CRCPBR0);
orte_grpcomm.barrier();
}
OPAL_CR_SET_TIMER(OPAL_CR_TIMER_CRCP0);
Replace the old coordinated component ('coord') and replace it with a much more refined version ('bkmrk'). The new component fixes a number of problems with the old component. The core algorithm is the same, but by changing the data strucutres a bit we have improved performance and memory utilization. There are still a couple corner cases that still need some work. However, I did not want to delay bringing this into the trunk (and v1.3 branch) for too much longer. This commit was SVN r19537.
2008-09-10 22:29:17 +04:00
START_TIMER(CRCP_TIMER_TOTAL_CKPT);
STEP_1:
step_to_return_to = 0;
/* Coordinate Peers:
* When we return from this function we know that all of our
* channels have been flushed.
*/
if( OMPI_SUCCESS != (ret = ft_event_coordinate_peers()) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s ft_event: Checkpoint Coordination Failed %d",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ret);
exit_status = ret;
goto DONE;
}
if( stall_for_completion ) {
stall_for_completion = false;
opal_cr_stall_check = true;
step_to_return_to = 1;
exit_status = OMPI_EXISTS;
goto DONE_STALL;
}
END_TIMER(CRCP_TIMER_TOTAL_CKPT);
DISPLAY_ALL_TIMERS(state);
clear_timers();
}
/*****************************
* Continue after a checkpoint
******************************/
else if(OPAL_CRS_CONTINUE == state) {
if( OMPI_CRCP_PML_POST != pml_state->state){
goto DONE;
}
Commit to bring online OpenIB, MX, and shared memory support for Open MPI's checkpoint/restart functionality. Some tuning is still needed, but basic functionality is in place. There is still a problem with OpenIB and threads (external to C/R functionality). It has been reported in Ticket #1539 Additionally: * Fix a file cleanup bug in CRS Base. * Fix a possible deadlock in the TCP ft_event function * Add a mca_base_param_deregister() function to MCA base * Add whole process checkpoint timers * Add support for BTL: OpenIB, MX, Shared Memory * Add support Mpool: rdma, sm * Sundry bounds checking an cleanup in some scattered functions This commit was SVN r19756.
2008-10-16 19:09:00 +04:00
first_continue_pass = !first_continue_pass;
/* Only finalize the Protocol after the PML has been rebuilt */
if( ompi_cr_continue_like_restart && first_continue_pass ) {
goto DONE;
}
Replace the old coordinated component ('coord') and replace it with a much more refined version ('bkmrk'). The new component fixes a number of problems with the old component. The core algorithm is the same, but by changing the data strucutres a bit we have improved performance and memory utilization. There are still a couple corner cases that still need some work. However, I did not want to delay bringing this into the trunk (and v1.3 branch) for too much longer. This commit was SVN r19537.
2008-09-10 22:29:17 +04:00
START_TIMER(CRCP_TIMER_TOTAL_CONT);
Commit to bring online OpenIB, MX, and shared memory support for Open MPI's checkpoint/restart functionality. Some tuning is still needed, but basic functionality is in place. There is still a problem with OpenIB and threads (external to C/R functionality). It has been reported in Ticket #1539 Additionally: * Fix a file cleanup bug in CRS Base. * Fix a possible deadlock in the TCP ft_event function * Add a mca_base_param_deregister() function to MCA base * Add whole process checkpoint timers * Add support for BTL: OpenIB, MX, Shared Memory * Add support Mpool: rdma, sm * Sundry bounds checking an cleanup in some scattered functions This commit was SVN r19756.
2008-10-16 19:09:00 +04:00
Replace the old coordinated component ('coord') and replace it with a much more refined version ('bkmrk'). The new component fixes a number of problems with the old component. The core algorithm is the same, but by changing the data strucutres a bit we have improved performance and memory utilization. There are still a couple corner cases that still need some work. However, I did not want to delay bringing this into the trunk (and v1.3 branch) for too much longer. This commit was SVN r19537.
2008-09-10 22:29:17 +04:00
/*
* Finish the coord protocol
*/
if( OMPI_SUCCESS != (ret = ft_event_finalize_exchange() ) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: pml_ft_event: Checkpoint Finalization Failed %d",
ret);
exit_status = ret;
goto DONE;
}
END_TIMER(CRCP_TIMER_TOTAL_CONT);
DISPLAY_ALL_TIMERS(state);
clear_timers();
Commit to bring online OpenIB, MX, and shared memory support for Open MPI's checkpoint/restart functionality. Some tuning is still needed, but basic functionality is in place. There is still a problem with OpenIB and threads (external to C/R functionality). It has been reported in Ticket #1539 Additionally: * Fix a file cleanup bug in CRS Base. * Fix a possible deadlock in the TCP ft_event function * Add a mca_base_param_deregister() function to MCA base * Add whole process checkpoint timers * Add support for BTL: OpenIB, MX, Shared Memory * Add support Mpool: rdma, sm * Sundry bounds checking an cleanup in some scattered functions This commit was SVN r19756.
2008-10-16 19:09:00 +04:00
if( opal_cr_timing_barrier_enabled ) {
OPAL_CR_SET_TIMER(OPAL_CR_TIMER_COREBR1);
orte_grpcomm.barrier();
}
OPAL_CR_SET_TIMER(OPAL_CR_TIMER_CORE2);
Replace the old coordinated component ('coord') and replace it with a much more refined version ('bkmrk'). The new component fixes a number of problems with the old component. The core algorithm is the same, but by changing the data strucutres a bit we have improved performance and memory utilization. There are still a couple corner cases that still need some work. However, I did not want to delay bringing this into the trunk (and v1.3 branch) for too much longer. This commit was SVN r19537.
2008-09-10 22:29:17 +04:00
}
/*****************************
* Restart from a checkpoint
*****************************/
else if(OPAL_CRS_RESTART == state) {
if( OMPI_CRCP_PML_POST != pml_state->state){
goto DONE;
}
START_TIMER(CRCP_TIMER_TOTAL_RST);
/*
* Refresh the jobids
*/
for(item = opal_list_get_first(&ompi_crcp_bkmrk_pml_peer_refs);
item != opal_list_get_end(&ompi_crcp_bkmrk_pml_peer_refs);
item = opal_list_get_next(item) ) {
ompi_crcp_bkmrk_pml_peer_ref_t *cur_peer_ref;
cur_peer_ref = (ompi_crcp_bkmrk_pml_peer_ref_t*)item;
/* JJH - Assuming only one global jobid at the moment */
cur_peer_ref->proc_name.jobid = ORTE_PROC_MY_NAME->jobid;
}
/*
* Finish the coord protocol
*/
if( OMPI_SUCCESS != (ret = ft_event_finalize_exchange() ) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: pml_ft_event: Checkpoint Finalization Failed %d",
ret);
exit_status = ret;
goto DONE;
}
END_TIMER(CRCP_TIMER_TOTAL_RST);
DISPLAY_ALL_TIMERS(state);
clear_timers();
}
/*****************************
* Terminating the process post checkpoint
*****************************/
else if(OPAL_CRS_TERM == state ) {
goto DONE;
}
/****************************
* Reached an error
****************************/
else {
goto DONE;
}
DONE:
step_to_return_to = 0;
ft_event_state = OPAL_CRS_RUNNING;
DONE_STALL:
pml_state->error_code = exit_status;
return pml_state;
}
/******************
* Local Utility functions
******************/
/************************************************
* Traffic Message Utility Functions
************************************************/
static int traffic_message_append(ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref,
opal_list_t * append_list,
ompi_crcp_bkmrk_pml_message_type_t msg_type,
size_t count,
ompi_datatype_t *datatype,
size_t in_ddt_size,
int tag,
int dest,
struct ompi_communicator_t* comm,
ompi_crcp_bkmrk_pml_traffic_message_ref_t **msg_ref)
{
int ret, exit_status = ORTE_SUCCESS;
size_t ddt_size = 0;
ddt_size = in_ddt_size;
if( NULL != datatype ) {
ompi_ddt_type_size(datatype,
&ddt_size);
} else {
ddt_size = 0;
}
/*
* Determine if message is currently in the list
* - If it is then increment the count.
* - ow add it to the list
*/
if( OMPI_SUCCESS != (ret = traffic_message_find(append_list,
count, tag, dest,
comm->c_contextid,
ddt_size,
msg_ref,
FIND_MSG_UNKNOWN /* Active? */
) ) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: traffic_message_append: Unable to find the proper message reference.\n");
return OMPI_ERROR;
}
if( NULL != *msg_ref ) {
if( msg_type == COORD_MSG_TYPE_P_SEND ||
msg_type == COORD_MSG_TYPE_P_RECV ) {
(*msg_ref)->posted++;
} else {
(*msg_ref)->active++;
}
} else {
if( NULL != peer_ref ) {
CREATE_NEW_MSG((*msg_ref), msg_type,
count, ddt_size, tag, dest, comm,
peer_ref->proc_name.jobid,
peer_ref->proc_name.vpid);
} else {
CREATE_NEW_MSG((*msg_ref), msg_type,
count, ddt_size, tag, dest, comm,
ORTE_JOBID_INVALID,
ORTE_VPID_INVALID);
}
if( msg_type == COORD_MSG_TYPE_P_SEND ||
msg_type == COORD_MSG_TYPE_P_RECV ) {
(*msg_ref)->matched = 0;
(*msg_ref)->done = 0;
(*msg_ref)->active = 0;
(*msg_ref)->posted = 1;
} else {
(*msg_ref)->matched = 0;
(*msg_ref)->done = 0;
(*msg_ref)->active = 1;
(*msg_ref)->posted = 0;
}
opal_list_append(append_list, &((*msg_ref)->super));
}
if( NULL != peer_ref ) {
if( msg_type == COORD_MSG_TYPE_B_SEND ) {
TRAFFIC_MSG_DUMP_PEER(15, (peer_ref, "-- Append Message (send) --", true));
}
else if( msg_type == COORD_MSG_TYPE_P_SEND ) {
TRAFFIC_MSG_DUMP_PEER(15, (peer_ref, "-- Append Message (send_init) --", true));
}
else if( msg_type == COORD_MSG_TYPE_B_RECV ) {
TRAFFIC_MSG_DUMP_PEER(15, (peer_ref, "-- Append Message (recv) --", true));
}
else if( msg_type == COORD_MSG_TYPE_P_RECV ) {
TRAFFIC_MSG_DUMP_PEER(15, (peer_ref, "-- Append Message (recv_init) --", true));
}
else if( msg_type == COORD_MSG_TYPE_I_SEND || msg_type == COORD_MSG_TYPE_I_RECV ) {
;
}
else {
TRAFFIC_MSG_DUMP_PEER(15, (peer_ref, "-- Append Message (Unknown) --", true));
}
}
return exit_status;
}
static int traffic_message_start(ompi_crcp_bkmrk_pml_traffic_message_ref_t *msg_ref,
ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref,
ompi_request_t **request,
opal_list_t * peer_list,
ompi_crcp_bkmrk_pml_message_content_ref_t ** content_ref)
{
/* This is only called by persistent calls.
* This will mark the current message as having one more active member.
* There is still only one posted message. */
msg_ref->active++;
traffic_message_find_mark_persistent(msg_ref, request,
false, /* Find currently not active */
true, /* Mark as active */
content_ref);
return OMPI_SUCCESS;
}
static int traffic_message_move(ompi_crcp_bkmrk_pml_traffic_message_ref_t *old_msg_ref,
ompi_crcp_bkmrk_pml_message_type_t msg_type,
ompi_crcp_bkmrk_pml_peer_ref_t *from_peer_ref,
opal_list_t * from_list,
ompi_crcp_bkmrk_pml_peer_ref_t *to_peer_ref,
opal_list_t * to_list,
Fix a problem with persistent sends found in MTT regression testing on Odin. This commit was SVN r19707.
2008-10-08 18:11:55 +04:00
ompi_crcp_bkmrk_pml_traffic_message_ref_t **new_msg_ref,
bool keep_active)
Replace the old coordinated component ('coord') and replace it with a much more refined version ('bkmrk'). The new component fixes a number of problems with the old component. The core algorithm is the same, but by changing the data strucutres a bit we have improved performance and memory utilization. There are still a couple corner cases that still need some work. However, I did not want to delay bringing this into the trunk (and v1.3 branch) for too much longer. This commit was SVN r19537.
2008-09-10 22:29:17 +04:00
{
int ret, exit_status = ORTE_SUCCESS;
ompi_crcp_bkmrk_pml_message_content_ref_t *new_content, *prev_content;
ompi_request_t *request = NULL;
bool loc_already_drained = false;
/* Append to the to_peer_ref */
if( COORD_MSG_TYPE_B_RECV != msg_type ) {
traffic_message_grab_content(old_msg_ref, &prev_content, true, true); /* Remove, prefer already_drained */
request = prev_content->request;
loc_already_drained = prev_content->already_drained;
prev_content->request = NULL;
HOKE_CONTENT_REF_RETURN(prev_content);
}
ret = traffic_message_append(to_peer_ref, to_list,
old_msg_ref->msg_type,
old_msg_ref->count,
NULL,
old_msg_ref->ddt_size,
old_msg_ref->tag,
old_msg_ref->rank,
old_msg_ref->comm,
new_msg_ref);
if( loc_already_drained ) {
old_msg_ref->active_drain--;
(*new_msg_ref)->active--; /* Undo the action from _append() */
(*new_msg_ref)->active_drain++;
} else {
/* 'remove' from from_peer_ref */
old_msg_ref->active--;
}
if( COORD_MSG_TYPE_B_RECV != msg_type && NULL == request ) {
ERROR_SHOULD_NEVER_HAPPEN("Error: Must match a non-blocking send, and there is no matching request.");
}
if( NULL != request ) {
HOKE_CONTENT_REF_ALLOC(new_content, ret);
new_content->buffer = NULL;
new_content->request = request;
new_content->done = false;
Fix a problem with persistent sends found in MTT regression testing on Odin. This commit was SVN r19707.
2008-10-08 18:11:55 +04:00
new_content->active = keep_active;
Replace the old coordinated component ('coord') and replace it with a much more refined version ('bkmrk'). The new component fixes a number of problems with the old component. The core algorithm is the same, but by changing the data strucutres a bit we have improved performance and memory utilization. There are still a couple corner cases that still need some work. However, I did not want to delay bringing this into the trunk (and v1.3 branch) for too much longer. This commit was SVN r19537.
2008-09-10 22:29:17 +04:00
new_content->already_posted = true;
new_content->already_drained = loc_already_drained;
OBJ_RETAIN(request);
opal_list_append(&((*new_msg_ref)->msg_contents), &(new_content->super) );
}
if( NULL == from_peer_ref && NULL != to_peer_ref ) {
(*new_msg_ref)->proc_name.jobid = to_peer_ref->proc_name.jobid;
(*new_msg_ref)->proc_name.vpid = to_peer_ref->proc_name.vpid;
}
return exit_status;
}
static int traffic_message_find_mark_persistent(ompi_crcp_bkmrk_pml_traffic_message_ref_t *msg_ref,
ompi_request_t **request,
bool cur_active,
bool set_is_active,
ompi_crcp_bkmrk_pml_message_content_ref_t **c_ref)
{
mca_pml_base_request_t * breq = NULL;
opal_list_item_t* item = NULL;
breq = (mca_pml_base_request_t *)(*request);
for(item = opal_list_get_first(&(msg_ref->msg_contents));
item != opal_list_get_end( &(msg_ref->msg_contents));
item = opal_list_get_next(item) ) {
ompi_crcp_bkmrk_pml_message_content_ref_t *content_ref = NULL;
mca_pml_base_request_t * loc_breq = NULL;
content_ref = (ompi_crcp_bkmrk_pml_message_content_ref_t*)item;
loc_breq = (mca_pml_base_request_t *)(content_ref->request);
if( content_ref->active != cur_active ) {
continue;
}
else if( loc_breq->req_sequence == breq->req_sequence ) {
OPAL_OUTPUT_VERBOSE((25, mca_crcp_bkmrk_component.super.output_handle,
"%s %8s Request [%d] (%s) %d : %d",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
(set_is_active ? "Start" : (NULL != c_ref ? "Drain" : "Complete")),
(int)msg_ref->msg_id,
(content_ref->active ? "T" : "F"),
(int)loc_breq->req_sequence,
(int)breq->req_sequence));
content_ref->active = set_is_active;
if( NULL != c_ref ) {
*c_ref = content_ref;
}
break;
}
}
return OMPI_SUCCESS;
}
static int traffic_message_grab_content(ompi_crcp_bkmrk_pml_traffic_message_ref_t *msg_ref,
ompi_crcp_bkmrk_pml_message_content_ref_t ** content_ref,
bool remove,
bool already_drained)
{
ompi_crcp_bkmrk_pml_message_content_ref_t *new_content = NULL;
ompi_crcp_bkmrk_pml_message_content_ref_t *loc_content_ref = NULL;
opal_list_item_t* item = NULL;
/*
* If there is no request list, return NULL
*/
if( 0 >= opal_list_get_size(&msg_ref->msg_contents) ) {
return OMPI_SUCCESS;
}
/*
* Otherwise look though the list, and grab something 'already_drained' if
* possible, otherwise just get the first element.
*/
if( already_drained ) {
item = opal_list_get_first(&(msg_ref->msg_contents));
new_content = (ompi_crcp_bkmrk_pml_message_content_ref_t*)item;
}
for(item = opal_list_get_first(&(msg_ref->msg_contents));
item != opal_list_get_end( &(msg_ref->msg_contents));
item = opal_list_get_next(item) ) {
loc_content_ref = (ompi_crcp_bkmrk_pml_message_content_ref_t*)item;
if( !already_drained ) {
TRAFFIC_MSG_DUMP_MSG_CONTENT_INDV(10, (loc_content_ref));
}
if( loc_content_ref->already_drained == already_drained ) {
new_content = (ompi_crcp_bkmrk_pml_message_content_ref_t*)item;
break;
}
}
if( remove ) {
opal_list_remove_item(&msg_ref->msg_contents, &(new_content->super));
}
if( NULL != content_ref ) {
*content_ref = new_content;
} else if( remove && NULL != new_content ) {
HOKE_CONTENT_REF_RETURN(new_content);
}
return OMPI_SUCCESS;
}
static int traffic_message_create_drain_message(bool post_drain,
int max_post,
ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref,
ompi_crcp_bkmrk_pml_traffic_message_ref_t ** posted_msg_ref,
int *num_posted)
{
ompi_crcp_bkmrk_pml_drain_message_ref_t *drain_msg_ref = NULL;
ompi_crcp_bkmrk_pml_message_content_ref_t *new_content = NULL, *prev_content = NULL;
int m_iter, m_total, ret;
*num_posted = 0;
/*
* Nothing to do here
*/
if( NULL == (*posted_msg_ref) || max_post <= 0) {
return OMPI_SUCCESS;
}
/*
* For each active message or if not active message then max_post, create a drain message
*/
m_total = max_post;
if( !post_drain && max_post > (*posted_msg_ref)->active ) {
m_total = (*posted_msg_ref)->active;
}
OPAL_OUTPUT_VERBOSE((10, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s <-- %s "
" --> Create Drain Msg: %s %4d = min(%4d / %4d)",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer_ref->proc_name)),
(post_drain ? "Posting" : "Not Posting"),
m_total, (*posted_msg_ref)->active, max_post ));
TRAFFIC_MSG_DUMP_MSG_INDV(10, ((*posted_msg_ref), "Drain", true));
/*
* Get a drained message reference for this signature.
*/
drain_message_append(peer_ref,
COORD_MSG_TYPE_I_RECV,
(*posted_msg_ref)->count,
(*posted_msg_ref)->ddt_size,
(*posted_msg_ref)->tag,
(*posted_msg_ref)->rank,
(*posted_msg_ref)->comm,
&drain_msg_ref);
/*
* Create a new message content for each message to be drained.
*/
for(m_iter = 0; m_iter < m_total; ++m_iter) {
new_content = NULL;
OPAL_OUTPUT_VERBOSE((10, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s <-- %s "
" \t--> Find Content: %s (%4d of %4d)",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer_ref->proc_name)),
(post_drain ? "Posting" : "Not Posting"),
m_iter, m_total));
/* Grab a request if there are any
* - if we are posting, then we created a dummy message which will not
* have any contents, so this is still valid.
* - if we are not posting, and this is an iRecv, then we *must* find a content!
*/
traffic_message_grab_content((*posted_msg_ref), &prev_content, false, false); /* Do not remove, No already drained */
if( NULL != prev_content ) {
prev_content->already_drained = true;
}
/* YYY JJH YYY - Is this needed? */
if( !post_drain && (*posted_msg_ref)->msg_type != COORD_MSG_TYPE_B_RECV ) {
assert( NULL != prev_content );
}
(*posted_msg_ref)->active--;
(*posted_msg_ref)->active_drain++;
/* Create a new content for the drained message */
HOKE_CONTENT_REF_ALLOC(new_content, ret);
new_content->buffer = NULL;
if( NULL == prev_content ) {
new_content->request = NULL;
} else {
new_content->request = prev_content->request;
if( NULL != new_content->request ) {
OBJ_RETAIN(new_content->request);
}
}
opal_list_append(&(drain_msg_ref->msg_contents), &(new_content->super) );
if( !post_drain ) {
new_content->done = false;
new_content->active = true;
new_content->already_posted = true;
new_content->already_drained = true;
drain_msg_ref->active++;
drain_msg_ref->already_posted++;
} else {
new_content->done = false;
new_content->active = false;
new_content->already_posted = false;
new_content->already_drained = true;
/*
* Put the true count here so we can properly match the drain.
* The post_drained() will properly handle the packed datatype
* by changing the count to (count * ddt_size).
*/
ompi_ddt_duplicate(&ompi_mpi_packed, &(drain_msg_ref->datatype));
/* Create a buffer of the necessary type/size */
if(drain_msg_ref->count > 0 ) {
new_content->buffer = (void *) malloc(drain_msg_ref->count * drain_msg_ref->ddt_size);
} else {
new_content->buffer = (void *) malloc(1 * drain_msg_ref->ddt_size);
}
/* JJH - Performance Optimization? - Post drained messages right away? */
}
(*num_posted)++;
}
peer_ref->total_drained_msgs += *num_posted;
OPAL_OUTPUT_VERBOSE((15, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s <-- %s "
"Added %d messages to the drained list (size = %d)",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer_ref->proc_name)),
(*num_posted),
(int)opal_list_get_size(&(peer_ref->drained_list)) ));
return OMPI_SUCCESS;
}
static int traffic_message_find_recv(ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref,
int rank, uint32_t comm_id, int tag,
size_t count, size_t datatype_size,
ompi_crcp_bkmrk_pml_traffic_message_ref_t ** posted_recv_msg_ref,
ompi_crcp_bkmrk_pml_traffic_message_ref_t ** posted_irecv_msg_ref,
ompi_crcp_bkmrk_pml_traffic_message_ref_t ** posted_precv_msg_ref,
ompi_crcp_bkmrk_pml_traffic_message_ref_t ** posted_unknown_recv_msg_ref,
ompi_crcp_bkmrk_pml_traffic_message_ref_t ** posted_unknown_precv_msg_ref)
{
int ret;
*posted_recv_msg_ref = NULL;
*posted_irecv_msg_ref = NULL;
*posted_precv_msg_ref = NULL;
*posted_unknown_recv_msg_ref = NULL;
*posted_unknown_precv_msg_ref = NULL;
/*
* Check the recv_list
*/
if( OMPI_SUCCESS != (ret = traffic_message_find(&(peer_ref->recv_list),
count, tag, INVALID_INT,
comm_id, datatype_size,
posted_recv_msg_ref,
FIND_MSG_UNKNOWN) ) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: traffic_message_find_recv: Unable to find the proper message reference.\n");
return OMPI_ERROR;
}
/*
* Check the irecv_list
*/
if( OMPI_SUCCESS != (ret = traffic_message_find(&(peer_ref->irecv_list),
count, tag, INVALID_INT,
comm_id, datatype_size,
posted_irecv_msg_ref,
FIND_MSG_UNKNOWN) ) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: traffic_message_find_recv: Unable to find the proper message reference.\n");
return OMPI_ERROR;
}
/*
* Check the recv_init_list
*/
if( OMPI_SUCCESS != (ret = traffic_message_find(&(peer_ref->recv_init_list),
count, tag, INVALID_INT,
comm_id, datatype_size,
posted_precv_msg_ref,
FIND_MSG_UNKNOWN) ) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: traffic_message_find_recv: Unable to find the proper message reference.\n");
return OMPI_ERROR;
}
/*
* Check the unknown list of non-persistant
*/
if( OMPI_SUCCESS != (ret = traffic_message_find(&(unknown_recv_from_list),
count, tag, INVALID_INT,
comm_id, datatype_size,
posted_unknown_recv_msg_ref,
FIND_MSG_UNKNOWN) ) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: traffic_message_find_recv: Unable to find the proper message reference.\n");
return OMPI_ERROR;
}
/*
* Check the unknown list of persistant
*/
if( OMPI_SUCCESS != (ret = traffic_message_find(&(unknown_persist_recv_list),
count, tag, INVALID_INT,
comm_id, datatype_size,
posted_unknown_precv_msg_ref,
FIND_MSG_UNKNOWN) ) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: traffic_message_find_recv: Unable to find the proper message reference.\n");
return OMPI_ERROR;
}
/*
* JJH -- Should we check the drained list?
* If we checkpoint again before dimishing the drained list, then
* the peer could be requesting that a drained send complete...
*/
return OMPI_SUCCESS;
}
static int traffic_message_find(opal_list_t * search_list,
size_t count, int tag, int peer,
uint32_t comm_id, size_t ddt_size,
ompi_crcp_bkmrk_pml_traffic_message_ref_t ** found_msg_ref,
int active )
{
opal_list_item_t* item = NULL;
*found_msg_ref = NULL;
#if OMPI_ENABLE_DEBUG == 1
/*
* Dummy checks:
*/
if( NULL == search_list) {
opal_output(0, "WARNING (Debug): Search_list NULL! (%s:%d)", __FILE__, __LINE__);
return OMPI_ERROR;
}
#endif
/*
* Check the search list
*/
for(item = opal_list_get_last(search_list);
item != opal_list_get_begin(search_list);
item = opal_list_get_prev(item) ) {
ompi_crcp_bkmrk_pml_traffic_message_ref_t * msg_ref;
msg_ref = (ompi_crcp_bkmrk_pml_traffic_message_ref_t*)item;
if( active != FIND_MSG_UNKNOWN ) {
if( active == PERSIST_MARKER ) {
if( 0 >= msg_ref->posted ) {
continue;
}
}
else if( (active == FIND_MSG_TRUE && 0 >= (msg_ref->active + msg_ref->active_drain) ) ||
(active == FIND_MSG_FALSE && 0 <= (msg_ref->active + msg_ref->active_drain) ) ) {
continue;
}
}
if(msg_ref->count == count &&
(NULL != msg_ref->comm && msg_ref->comm->c_contextid == comm_id) &&
(msg_ref->tag == MPI_ANY_TAG || msg_ref->tag == tag) &&
(peer == INVALID_INT || msg_ref->rank == peer) &&
msg_ref->ddt_size == ddt_size) {
OPAL_OUTPUT_VERBOSE((30, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: traffic_message_find: Found Message -- Comm list (%d, %d)\n",
tag, peer));
*found_msg_ref = msg_ref;
return OMPI_SUCCESS;
}
}
return OMPI_SUCCESS;
}
/************************************************
* Drain Message Utility Functions
************************************************/
static int drain_message_append(ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref,
ompi_crcp_bkmrk_pml_message_type_t msg_type,
size_t count, size_t ddt_size,
int tag,int dest,
struct ompi_communicator_t* comm,
ompi_crcp_bkmrk_pml_drain_message_ref_t **msg_ref)
{
int ret, exit_status = ORTE_SUCCESS;
ompi_crcp_bkmrk_pml_message_content_ref_t *content_ref = NULL;
/*
* Determine if message is currently in the list
* - If it is then increment the count.
* - ow add it to the list
*/
if( OMPI_SUCCESS != (ret = drain_message_find(&(peer_ref->drained_list),
count, tag, dest,
comm->c_contextid,
ddt_size,
msg_ref,
&content_ref) ) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: traffic_message_append: Unable to find the proper message reference.\n");
return OMPI_ERROR;
}
if( NULL == *msg_ref ) {
CREATE_NEW_DRAIN_MSG((*msg_ref), msg_type,
count, NULL, tag, dest, comm,
peer_ref->proc_name.jobid,
peer_ref->proc_name.vpid);
(*msg_ref)->done = 0;
(*msg_ref)->active = 0;
(*msg_ref)->already_posted = 0;
opal_list_append(&(peer_ref->drained_list), &((*msg_ref)->super));
}
/* If message does exist then the calling function needs to handle the msg_contents and counts */
return exit_status;
}
static int drain_message_remove(ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref,
ompi_crcp_bkmrk_pml_drain_message_ref_t *msg_ref,
ompi_crcp_bkmrk_pml_message_content_ref_t *content_ref)
{
/*
* Remove the message content from the list attached to the message
*/
opal_list_remove_item(&(msg_ref->msg_contents), &(content_ref->super));
HOKE_CONTENT_REF_RETURN(content_ref);
/*
* If there are no more drained messages of this signature,
* then remove the signature from the peers drained list.
*/
if( 0 >= opal_list_get_size(&(msg_ref->msg_contents) ) ) {
TRAFFIC_MSG_DUMP_DRAIN_MSG_INDV(10, (msg_ref, "D*remove", true));
opal_list_remove_item(&(peer_ref->drained_list), &(msg_ref->super));
HOKE_DRAIN_MSG_REF_RETURN(msg_ref);
} else {
TRAFFIC_MSG_DUMP_DRAIN_MSG_INDV(10, (msg_ref, "Dremove", true));
}
return OMPI_SUCCESS;
}
static int drain_message_check_recv(void **buf, size_t count,
ompi_datatype_t *datatype,
int *src, int *tag,
struct ompi_communicator_t* comm,
struct ompi_request_t **request,
ompi_status_public_t** status,
bool *found_drain)
{
int ret, exit_status = OMPI_SUCCESS;
ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref = NULL;
ompi_crcp_bkmrk_pml_drain_message_ref_t *drain_msg_ref = NULL;
ompi_crcp_bkmrk_pml_message_content_ref_t *content_ref = NULL;
*found_drain = false;
/*
* Check to see if this message is in the drained message list
*/
if( OMPI_SUCCESS != (ret = drain_message_find_any(count, *tag, *src,
comm, datatype->size,
&drain_msg_ref,
&content_ref,
&peer_ref) ) ) {
ERROR_SHOULD_NEVER_HAPPEN("crcp:bkmrk: drain_check_recv(): Failed trying to find a drained message.");
exit_status = ret;
goto DONE;
}
/*
* If the message is a drained message
* - Complete it right now
* - We do not need to increment any counters here since we already have
* when we originally drained the message.
*/
if( NULL != drain_msg_ref ) {
OPAL_OUTPUT_VERBOSE((12, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: drain_check_recv(): Matched a drained message..."));
if( OMPI_SUCCESS != (ret = drain_message_copy_remove(drain_msg_ref,
content_ref,
src, tag, request, status,
datatype, count, buf,
peer_ref) ) ) {
opal_output( mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: drain_check_recv(): Datatype copy failed (%d)",
ret);
exit_status = ret;
goto DONE;
}
peer_ref->total_drained_msgs -= 1;
*found_drain = true;
}
DONE:
return exit_status;
}
static int drain_message_find_any(size_t count, int tag, int peer,
struct ompi_communicator_t* comm, size_t ddt_size,
ompi_crcp_bkmrk_pml_drain_message_ref_t ** found_msg_ref,
ompi_crcp_bkmrk_pml_message_content_ref_t ** content_ref,
ompi_crcp_bkmrk_pml_peer_ref_t **peer_ref)
{
ompi_crcp_bkmrk_pml_peer_ref_t *cur_peer_ref = NULL;
opal_list_item_t* item = NULL;
*found_msg_ref = NULL;
for(item = opal_list_get_first(&ompi_crcp_bkmrk_pml_peer_refs);
item != opal_list_get_end(&ompi_crcp_bkmrk_pml_peer_refs);
item = opal_list_get_next(item) ) {
cur_peer_ref = (ompi_crcp_bkmrk_pml_peer_ref_t*)item;
/*
* If we ware not MPI_ANY_SOURCE, then extract the process name from the
* communicator, and search only the peer that matches.
*/
if( MPI_ANY_SOURCE != peer && peer >= 0) {
Make sure to check the bounds of the communicator array before referencing it. Fixes some of the FT related Intel bugs seen in Odin testing last night. This commit was SVN r19701.
2008-10-07 23:15:12 +04:00
/* Check to see if peer could possibly be in this communicator */
if( comm->c_local_group->grp_proc_count <= peer ) {
continue;
}
Replace the old coordinated component ('coord') and replace it with a much more refined version ('bkmrk'). The new component fixes a number of problems with the old component. The core algorithm is the same, but by changing the data strucutres a bit we have improved performance and memory utilization. There are still a couple corner cases that still need some work. However, I did not want to delay bringing this into the trunk (and v1.3 branch) for too much longer. This commit was SVN r19537.
2008-09-10 22:29:17 +04:00
if( OPAL_EQUAL != orte_util_compare_name_fields(ORTE_NS_CMP_ALL,
&(cur_peer_ref->proc_name),
&(comm->c_local_group->grp_proc_pointers[peer]->proc_name)) ) {
continue;
}
}
drain_message_find(&(cur_peer_ref->drained_list),
count, tag, peer,
comm->c_contextid, ddt_size,
found_msg_ref,
content_ref);
if( NULL != *found_msg_ref) {
if( NULL != peer_ref ) {
*peer_ref = cur_peer_ref;
}
return OMPI_SUCCESS;
}
}
return OMPI_SUCCESS;
}
static int drain_message_find(opal_list_t * search_list,
size_t count, int tag, int peer,
uint32_t comm_id, size_t ddt_size,
ompi_crcp_bkmrk_pml_drain_message_ref_t ** found_msg_ref,
ompi_crcp_bkmrk_pml_message_content_ref_t ** content_ref)
{
ompi_crcp_bkmrk_pml_drain_message_ref_t * drain_msg = NULL;
opal_list_item_t* item = NULL;
*found_msg_ref = NULL;
*content_ref = NULL;
/* Dummy Check:
* If the list is empty...
*/
if( 0 >= opal_list_get_size(search_list) ) {
return OMPI_SUCCESS;
}
for(item = opal_list_get_first(search_list);
item != opal_list_get_end(search_list);
item = opal_list_get_next(item) ) {
drain_msg = (ompi_crcp_bkmrk_pml_drain_message_ref_t*)item;
OPAL_OUTPUT_VERBOSE((15, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: find_drain_msg(): Compare [%d, %d, %d, %d] to [%d, %d, %d, %d]",
(int)ddt_size, (int)count, tag, peer,
(int)drain_msg->ddt_size, (int)drain_msg->count, (int)drain_msg->tag, (int)drain_msg->rank));
/* Check the communicator for a match */
if( NULL != drain_msg->comm ) {
if( drain_msg->comm->c_contextid != comm_id ) {
continue;
}
}
/* If a specific tag was requested, then make sure this messages matches */
if( MPI_ANY_TAG != tag &&
drain_msg->tag != tag) {
continue;
}
/* If a specific rank was requested, then make sure this messages matches */
if( INVALID_INT != peer ) {
if( MPI_ANY_SOURCE != peer &&
drain_msg->rank != peer) {
continue;
}
}
/* Check the datatype size, if specified for a match */
if( ddt_size != PROBE_ANY_SIZE &&
count != PROBE_ANY_COUNT) {
/* Check the datatype size and count to make sure it matches */
if((drain_msg->count ) != count ||
(drain_msg->ddt_size) != ddt_size) {
continue;
}
}
/* If we get here then the message matches */
*found_msg_ref = drain_msg;
break;
}
/*
* Find a content to return
*/
if( NULL != *found_msg_ref ) {
drain_message_grab_content((*found_msg_ref), content_ref );
/* If there are no contents that match, then there are no drained messages that match. */
if( NULL == *content_ref ) {
*found_msg_ref = NULL;
}
}
return OMPI_SUCCESS;
}
static int drain_message_grab_content(ompi_crcp_bkmrk_pml_drain_message_ref_t *drain_msg_ref,
ompi_crcp_bkmrk_pml_message_content_ref_t ** content_ref)
{
ompi_crcp_bkmrk_pml_message_content_ref_t *loc_content_ref = NULL;
opal_list_item_t* item = NULL;
*content_ref = NULL;
for(item = opal_list_get_first(&(drain_msg_ref->msg_contents));
item != opal_list_get_end(&(drain_msg_ref->msg_contents));
item = opal_list_get_next(item) ) {
loc_content_ref = (ompi_crcp_bkmrk_pml_message_content_ref_t*)item;
/* If the buffer is invalid then this is not a valid message or
* has not been completed draining just yet */
if(NULL != loc_content_ref->buffer) {
*content_ref = loc_content_ref;
break;
}
}
return OMPI_SUCCESS;
}
static int drain_message_copy_remove_persistent(ompi_crcp_bkmrk_pml_drain_message_ref_t *drain_msg_ref,
ompi_crcp_bkmrk_pml_message_content_ref_t *drain_content_ref,
ompi_crcp_bkmrk_pml_traffic_message_ref_t *traffic_msg_ref,
ompi_request_t *request,
ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref)
{
int ret, exit_status = OMPI_SUCCESS;
ompi_crcp_bkmrk_pml_message_content_ref_t *content_ref = NULL;
/*
* Find the request in the list that has been posted, but not started
*/
traffic_message_find_mark_persistent(traffic_msg_ref, &request,
false, /* Find currently not active */
false, /* Keep inactive */
&content_ref);
/* These two requests should be exactly the same, so this is redundant, but here for completeness */
content_ref->request = request;
memcpy(&(content_ref->status), &drain_content_ref->status, sizeof(ompi_status_public_t));
if( 0 != (ret = ompi_ddt_copy_content_same_ddt(drain_msg_ref->datatype,
drain_msg_ref->count,
content_ref->buffer,
drain_content_ref->buffer) ) ) {
opal_output( mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: drain_message_copy_remove_p(): Datatype copy failed (%d)",
ret);
exit_status = ret;
}
/* Remove the message from the list */
drain_content_ref->request = NULL;
drain_message_remove(peer_ref, drain_msg_ref, drain_content_ref);
return exit_status;
}
static int drain_message_copy_remove(ompi_crcp_bkmrk_pml_drain_message_ref_t *drain_msg_ref,
ompi_crcp_bkmrk_pml_message_content_ref_t * drain_content_ref,
int *src, int *tag,
struct ompi_request_t **request,
ompi_status_public_t **status,
ompi_datatype_t *datatype, int count, void **buf,
ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref)
{
int ret, exit_status = OMPI_SUCCESS;
if( NULL != src ) {
*src = drain_msg_ref->rank;
}
if( NULL != tag ) {
*tag = drain_msg_ref->tag;
}
if( NULL != request ) {
*request = drain_content_ref->request;
OBJ_RETAIN(*request);
}
if( NULL != status && MPI_STATUS_IGNORE != *status ) {
memcpy(*status, &drain_content_ref->status, sizeof(ompi_status_public_t));
}
if( 0 != (ret = ompi_ddt_copy_content_same_ddt(datatype, count,
(void*)buf, drain_content_ref->buffer) ) ) {
opal_output( mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: drain_message_copy_remove(): Datatype copy failed (%d)",
ret);
exit_status = ret;
}
/* Remove the message from the list */
drain_content_ref->request = NULL;
drain_message_remove(peer_ref, drain_msg_ref, drain_content_ref);
return exit_status;
}
/************************************************
* Peer List Utility Functions
************************************************/
static ompi_crcp_bkmrk_pml_peer_ref_t * find_peer(orte_process_name_t proc)
{
opal_list_item_t* item = NULL;
for(item = opal_list_get_first(&ompi_crcp_bkmrk_pml_peer_refs);
item != opal_list_get_end(&ompi_crcp_bkmrk_pml_peer_refs);
item = opal_list_get_next(item) ) {
ompi_crcp_bkmrk_pml_peer_ref_t *cur_peer_ref;
cur_peer_ref = (ompi_crcp_bkmrk_pml_peer_ref_t*)item;
if( OPAL_EQUAL == orte_util_compare_name_fields(ORTE_NS_CMP_ALL,
&(cur_peer_ref->proc_name),
&proc) ) {
return cur_peer_ref;
}
}
return NULL;
}
static int find_peer_in_comm(struct ompi_communicator_t* comm, int proc_idx,
ompi_crcp_bkmrk_pml_peer_ref_t **peer_ref)
{
*peer_ref = find_peer(comm->c_remote_group->grp_proc_pointers[proc_idx]->proc_name);
if( NULL == *peer_ref) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: find_peer_in_comm(): Failed to find peer_ref - peer_ref is NULL\n");
return OMPI_ERROR;
}
return OMPI_SUCCESS;
}
/************************************************
* FT Event Utility Functions
************************************************/
static int ft_event_coordinate_peers(void)
{
static int step_to_return_to = 0;
int exit_status = OMPI_SUCCESS;
int ret;
if( step_to_return_to == 1 ) {
goto STEP_1;
}
/*
* Exchange Bookmarks with peers
*/
START_TIMER(CRCP_TIMER_CKPT_EX_B);
if( OMPI_SUCCESS != (ret = ft_event_exchange_bookmarks() ) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: ft_event_coordinate_peers: Bookmark Exchange Failed %d",
ret);
exit_status = ret;
goto DONE;
}
END_TIMER(CRCP_TIMER_CKPT_EX_B);
/*
* Check exchanged bookmarks
*/
START_TIMER(CRCP_TIMER_CKPT_CHECK_B);
if( OMPI_SUCCESS != (ret = ft_event_check_bookmarks() ) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: ft_event_coordinate_peers: Bookmark Check Failed %d",
ret);
exit_status = ret;
goto DONE;
}
END_TIMER(CRCP_TIMER_CKPT_CHECK_B);
/*
* Post Drain Acks and Msgs
*/
START_TIMER(CRCP_TIMER_CKPT_POST_DRAIN);
if( OMPI_SUCCESS != (ret = ft_event_post_drain_acks() ) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: ft_event_coordinate_peers: Bookmark Post Drain ACKS Failed %d",
ret);
exit_status = ret;
goto DONE;
}
if( OMPI_SUCCESS != (ret = ft_event_post_drained() ) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: ft_event_coordinate_peers: Bookmark Post Drain Msgs Failed %d",
ret);
exit_status = ret;
goto DONE;
}
END_TIMER(CRCP_TIMER_CKPT_POST_DRAIN);
DISPLAY_INDV_TIMER(CRCP_TIMER_CKPT_POST_DRAIN, -1, 0);
/*
* Check if we need to stall for completion of tasks
*/
/*
* If we know that we are in the middle of a blocking send then we
* need to stall the coordination algorithm while we wait for this to
* complete.
*/
if( 0 < current_msg_id &&
current_msg_type == COORD_MSG_TYPE_B_SEND) {
stall_for_completion = true;
}
START_TIMER(CRCP_TIMER_CKPT_WAIT_QUI);
if( stall_for_completion ) {
OPAL_OUTPUT_VERBOSE((10, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s **** STALLING %s in PID %d ***",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
(current_msg_type == COORD_MSG_TYPE_B_SEND ? "Send" : "Recv"),
getpid() ));
step_to_return_to = 1;
exit_status = OMPI_SUCCESS;
goto DONE;
}
STEP_1:
step_to_return_to = 0;
/*
* Wait for any messages that needed resolved.
* - Outstanding Receives (to drain wire) -- Receiver
* - Outstanding Irecvs (for drain ack) -- Sender
*/
if( OMPI_SUCCESS != (ret = ft_event_wait_quiesce() ) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: ft_event_coordinate_peers: Wait Quiesce Failed %d",
ret);
exit_status = ret;
goto DONE;
}
END_TIMER(CRCP_TIMER_CKPT_WAIT_QUI);
OPAL_OUTPUT_VERBOSE((5, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s Coordination Finished...\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
/*
* Now that all our peer channels are marked as drained
* continue with the checkpoint.
* Note: This does not guarentee that all of the peers
* are at this same position, but that our
* checkpoint will be consistent with all of the
* peers once they finish the protocol.
*/
DONE:
return exit_status;
}
static int ft_event_finalize_exchange(void)
{
int exit_status = OMPI_SUCCESS;
opal_list_item_t* item = NULL, *rm_item = NULL;
ompi_crcp_bkmrk_pml_traffic_message_ref_t * msg_ref;
ompi_crcp_bkmrk_pml_message_content_ref_t *content_ref = NULL;
opal_list_item_t* cont_item = NULL;
/*
* Clear bookmark totals
*/
for(item = opal_list_get_first(&ompi_crcp_bkmrk_pml_peer_refs);
item != opal_list_get_end(&ompi_crcp_bkmrk_pml_peer_refs);
item = opal_list_get_next(item) ) {
ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref;
peer_ref = (ompi_crcp_bkmrk_pml_peer_ref_t*)item;
if( OPAL_EQUAL != orte_util_compare_name_fields(ORTE_NS_CMP_ALL,
(ORTE_PROC_MY_NAME),
&(peer_ref->proc_name)) ) {
TRAFFIC_MSG_DUMP_PEER(10, (peer_ref, "finalize_exchange", false));
}
peer_ref->total_msgs_sent = 0;
peer_ref->total_msgs_recvd = 0;
peer_ref->matched_msgs_sent = 0;
peer_ref->matched_msgs_recvd = 0;
peer_ref->ack_required = false;
/* Clear send_list */
for(rm_item = opal_list_get_last(&peer_ref->send_list);
rm_item != opal_list_get_begin(&peer_ref->send_list);
rm_item = opal_list_get_prev(rm_item) ) {
msg_ref = (ompi_crcp_bkmrk_pml_traffic_message_ref_t*)rm_item;
msg_ref->matched = 0;
msg_ref->done = 0;
msg_ref->active_drain += msg_ref->active;
msg_ref->active = 0;
for(cont_item = opal_list_get_first(&(msg_ref->msg_contents));
cont_item != opal_list_get_end( &(msg_ref->msg_contents));
cont_item = opal_list_get_next(cont_item) ) {
content_ref = (ompi_crcp_bkmrk_pml_message_content_ref_t*)cont_item;
if( content_ref->active ) {
content_ref->already_drained = true;
}
}
}
/* Clear isend_list */
for(rm_item = opal_list_get_last(&peer_ref->isend_list);
rm_item != opal_list_get_begin(&peer_ref->isend_list);
rm_item = opal_list_get_prev(rm_item) ) {
msg_ref = (ompi_crcp_bkmrk_pml_traffic_message_ref_t*)rm_item;
msg_ref->matched = 0;
msg_ref->done = 0;
msg_ref->active_drain += msg_ref->active;
msg_ref->active = 0;
for(cont_item = opal_list_get_first(&(msg_ref->msg_contents));
cont_item != opal_list_get_end( &(msg_ref->msg_contents));
cont_item = opal_list_get_next(cont_item) ) {
content_ref = (ompi_crcp_bkmrk_pml_message_content_ref_t*)cont_item;
if( content_ref->active ) {
content_ref->already_drained = true;
}
}
}
/* Clear send_init_list */
for(rm_item = opal_list_get_last(&peer_ref->send_list);
rm_item != opal_list_get_begin(&peer_ref->send_list);
rm_item = opal_list_get_prev(rm_item) ) {
msg_ref = (ompi_crcp_bkmrk_pml_traffic_message_ref_t*)rm_item;
msg_ref->matched = 0;
msg_ref->done = 0;
msg_ref->active_drain += msg_ref->active;
msg_ref->active = 0;
for(cont_item = opal_list_get_first(&(msg_ref->msg_contents));
cont_item != opal_list_get_end( &(msg_ref->msg_contents));
cont_item = opal_list_get_next(cont_item) ) {
content_ref = (ompi_crcp_bkmrk_pml_message_content_ref_t*)cont_item;
if( content_ref->active ) {
content_ref->already_drained = true;
}
}
}
/* Clear recv_list */
for(rm_item = opal_list_get_last(&peer_ref->recv_list);
rm_item != opal_list_get_begin(&peer_ref->recv_list);
rm_item = opal_list_get_prev(rm_item) ) {
msg_ref = (ompi_crcp_bkmrk_pml_traffic_message_ref_t*)rm_item;
msg_ref->matched = 0;
msg_ref->done = 0;
}
/* Clear irecv_list */
for(rm_item = opal_list_get_last(&peer_ref->irecv_list);
rm_item != opal_list_get_begin(&peer_ref->irecv_list);
rm_item = opal_list_get_prev(rm_item) ) {
msg_ref = (ompi_crcp_bkmrk_pml_traffic_message_ref_t*)rm_item;
msg_ref->matched = 0;
msg_ref->done = 0;
}
/* Clear recv_init_list */
for(rm_item = opal_list_get_last(&peer_ref->recv_list);
rm_item != opal_list_get_begin(&peer_ref->recv_list);
rm_item = opal_list_get_prev(rm_item) ) {
msg_ref = (ompi_crcp_bkmrk_pml_traffic_message_ref_t*)rm_item;
msg_ref->matched = 0;
msg_ref->done = 0;
}
}
return exit_status;
}
static int ft_event_exchange_bookmarks(void)
{
int peer_idx = 0;
int my_idx = ORTE_PROC_MY_NAME->vpid;
int iter = 0;
int num_peers = 0;
num_peers = opal_list_get_size(&ompi_crcp_bkmrk_pml_peer_refs);
for( peer_idx = (num_peers - my_idx - 1), iter = 0;
iter < num_peers;
peer_idx = (peer_idx + 1) % num_peers, ++iter)
{
if(my_idx > peer_idx) {
/* Send our bookmark status */
send_bookmarks(peer_idx);
/* Recv peer bookmark status */
recv_bookmarks(peer_idx);
}
else if(my_idx < peer_idx) {
/* Recv peer bookmark status */
recv_bookmarks(peer_idx);
/* Send our bookmark status */
send_bookmarks(peer_idx);
}
}
/* Wait for all bookmarks to arrive */
START_TIMER(CRCP_TIMER_CKPT_EX_WAIT);
while( total_recv_bookmarks > 0 ) {
opal_progress();
opal_event_loop(OPAL_EVLOOP_NONBLOCK);
}
total_recv_bookmarks = 0;
END_TIMER(CRCP_TIMER_CKPT_EX_WAIT);
return OMPI_SUCCESS;
}
static int ft_event_check_bookmarks(void)
{
opal_list_item_t* item = NULL;
int ret;
int p_n_to_p_m = 0;
int p_n_from_p_m = 0;
if( 10 <= mca_crcp_bkmrk_component.super.verbose ) {
sleep(ORTE_PROC_MY_NAME->vpid);
OPAL_OUTPUT_VERBOSE((10, mca_crcp_bkmrk_component.super.output_handle,
"---------------------------------------------"));
OPAL_OUTPUT_VERBOSE((10, mca_crcp_bkmrk_component.super.output_handle,
"Process %s Match Table",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
OPAL_OUTPUT_VERBOSE((10, mca_crcp_bkmrk_component.super.output_handle,
"%s %5s | %7s | %7s | %7s | %7s |",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
"Vpid", "T_Send", "M_Recv", "M_Send", "T_Recv"));
for(item = opal_list_get_first(&ompi_crcp_bkmrk_pml_peer_refs);
item != opal_list_get_end(&ompi_crcp_bkmrk_pml_peer_refs);
item = opal_list_get_next(item) ) {
ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref;
int t_send, m_send;
int t_recv, m_recv;
peer_ref = (ompi_crcp_bkmrk_pml_peer_ref_t*)item;
t_send = peer_ref->total_msgs_sent;
m_send = peer_ref->matched_msgs_sent;
t_recv = peer_ref->total_msgs_recvd;
m_recv = peer_ref->matched_msgs_recvd;
OPAL_OUTPUT_VERBOSE((10, mca_crcp_bkmrk_component.super.output_handle,
"%s %5d | %7d | %7d | %7d | %7d |",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
peer_ref->proc_name.vpid,
t_send, m_recv, m_send, t_recv));
}
OPAL_OUTPUT_VERBOSE((10, mca_crcp_bkmrk_component.super.output_handle,
"---------------------------------------------"));
}
/*
* For each peer:
* - Check bookmarks
* - if mis-matched then post outstanding recvs.
*/
for(item = opal_list_get_first(&ompi_crcp_bkmrk_pml_peer_refs);
item != opal_list_get_end(&ompi_crcp_bkmrk_pml_peer_refs);
item = opal_list_get_next(item) ) {
ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref;
peer_ref = (ompi_crcp_bkmrk_pml_peer_ref_t*)item;
if( OPAL_EQUAL == orte_util_compare_name_fields(ORTE_NS_CMP_ALL,
(ORTE_PROC_MY_NAME),
&(peer_ref->proc_name)) ) {
continue;
}
TRAFFIC_MSG_DUMP_PEER(15, (peer_ref, "-- Bookmark Details --", false));
/* Lowest Rank sends first */
if( ORTE_PROC_MY_NAME->vpid < peer_ref->proc_name.vpid ) {
/********************
* Check P_n --> P_m
* Has the peer received all the messages that I have put on the wire?
********************/
p_n_to_p_m = peer_ref->total_msgs_sent;
p_n_from_p_m = peer_ref->matched_msgs_recvd;
/* T_Send >= M_Recv */
if( p_n_to_p_m < p_n_from_p_m ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s --> %s "
"Total Sent (%4d) = Matched Recv. (%4d) => Diff (%4d). "
" WARNING: Peer received more than was sent. :(\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer_ref->proc_name)),
p_n_to_p_m,
p_n_from_p_m,
(p_n_to_p_m - p_n_from_p_m)
);
}
/* I've send more than my peer has received,
* so need to coordinate with peer. */
if( p_n_to_p_m > p_n_from_p_m) {
OPAL_OUTPUT_VERBOSE((10, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s --> %s "
"Total Sent (%4d) = Matched Recv. (%4d). Peer needs %4d.\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer_ref->proc_name)),
p_n_to_p_m,
p_n_from_p_m,
(p_n_to_p_m - p_n_from_p_m)
));
/*
* Tell the peer what the outstanding messages looked like.
* Since we can't tell which ones they are, we need to send the
* information for all of the messages since the last checkpoint
*/
if( OMPI_SUCCESS != (ret = send_msg_details(peer_ref, p_n_to_p_m, p_n_from_p_m) ) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: check_bookmarks: Unable to send message details to peer %s: Return %d\n",
ORTE_NAME_PRINT(&peer_ref->proc_name),
ret);
return ret;
}
}
/********************
* Check P_n <-- P_m
* Have I received all the messages that my peer has put on the wire?
********************/
p_n_to_p_m = peer_ref->matched_msgs_sent;
p_n_from_p_m = peer_ref->total_msgs_recvd;
/* M_Send >= T_Recv */
if( p_n_to_p_m < p_n_from_p_m ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s --> %s "
"Matched Sent (%4d) = Total Recv. (%4d) => Diff (%4d). "
" WARNING: I received more than the peer sent. :(\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer_ref->proc_name)),
p_n_to_p_m,
p_n_from_p_m,
(p_n_to_p_m - p_n_from_p_m)
);
}
/* I've recv'ed less than my peer has sent,
* so need to coordinate with peer. */
if( p_n_to_p_m > p_n_from_p_m) {
OPAL_OUTPUT_VERBOSE((10, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s <-- %s "
"Matched Sent (%4d) = Total Recv. (%4d). I need %4d.\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer_ref->proc_name)),
p_n_to_p_m,
p_n_from_p_m,
(p_n_to_p_m - p_n_from_p_m)
));
/*
* Receive from peer the datatypes of the outstanding sends
* As we figure out what they are post Irecv's for them into a drained buffer list.
*/
if( OMPI_SUCCESS != (ret = recv_msg_details(peer_ref, p_n_to_p_m, p_n_from_p_m) ) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: check_bookmarks: Unable to recv message details from peer %s: Return %d\n",
ORTE_NAME_PRINT(&peer_ref->proc_name),
ret);
return ret;
}
}
}
/* Highest rank recvs first */
else {
/********************
* Check P_n <-- P_m
* Have I received all the messages that my peer has put on the wire?
********************/
p_n_to_p_m = peer_ref->matched_msgs_sent;
p_n_from_p_m = peer_ref->total_msgs_recvd;
/* M_Send >= T_Recv */
if( p_n_to_p_m < p_n_from_p_m ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s --> %s "
"Matched Sent (%4d) = Total Recv. (%4d) => Diff (%4d). "
" WARNING: I received more than the peer sent. :(\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer_ref->proc_name)),
p_n_to_p_m,
p_n_from_p_m,
(p_n_to_p_m - p_n_from_p_m)
);
}
/* I've recv'ed less than my peer has sent,
* so need to coordinate with peer. */
if( p_n_to_p_m > p_n_from_p_m) {
OPAL_OUTPUT_VERBOSE((10, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s <-- %s "
"Matched Sent (%4d) = Total Recv. (%4d). I need %4d.\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer_ref->proc_name)),
p_n_to_p_m,
p_n_from_p_m,
(p_n_to_p_m - p_n_from_p_m)
));
/*
* Receive from peer the datatypes of the outstanding sends
* As we figure out what they are post Irecv's for them into a drained buffer list.
*/
if( OMPI_SUCCESS != (ret = recv_msg_details(peer_ref, p_n_to_p_m, p_n_from_p_m) ) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: check_bookmarks: Unable to recv message details from peer %s: Return %d\n",
ORTE_NAME_PRINT(&peer_ref->proc_name),
ret);
return ret;
}
}
/********************
* Check P_n --> P_m
* Has the peer received all the messages that I have put on the wire?
********************/
p_n_to_p_m = peer_ref->total_msgs_sent;
p_n_from_p_m = peer_ref->matched_msgs_recvd;
/* T_Send >= M_Recv */
if( p_n_to_p_m < p_n_from_p_m ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s --> %s "
"Total Sent (%4d) = Matched Recv. (%4d) => Diff (%4d). "
" WARNING: Peer received more than was sent. :(\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer_ref->proc_name)),
p_n_to_p_m,
p_n_from_p_m,
(p_n_to_p_m - p_n_from_p_m)
);
}
/* I've send more than my peer has received,
* so need to coordinate with peer. */
if( p_n_to_p_m > p_n_from_p_m) {
OPAL_OUTPUT_VERBOSE((10, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s --> %s "
"Total Sent (%4d) = Matched Recv. (%4d). Peer needs %4d.\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer_ref->proc_name)),
p_n_to_p_m,
p_n_from_p_m,
(p_n_to_p_m - p_n_from_p_m)
));
/*
* Tell the peer what the outstanding messages looked like.
* Since we can't tell which ones they are, we need to send the
* information for all of the messages since the last checkpoint
*/
if( OMPI_SUCCESS != (ret = send_msg_details(peer_ref, p_n_to_p_m, p_n_from_p_m) ) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: check_bookmarks: Unable to send message details to peer %s: Return %d\n",
ORTE_NAME_PRINT(&peer_ref->proc_name),
ret);
return ret;
}
}
}
}
return OMPI_SUCCESS;
}
static int ft_event_post_drain_acks(void)
{
ompi_crcp_bkmrk_pml_drain_message_ack_ref_t * drain_msg_ack = NULL;
opal_list_item_t* item = NULL;
size_t req_size;
int ret;
req_size = opal_list_get_size(&drained_msg_ack_list);
if(req_size <= 0) {
return OMPI_SUCCESS;
}
OPAL_OUTPUT_VERBOSE((10, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s Wait on %d Drain ACK Messages.\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
(int)req_size));
/*
* We have loaded our peer with the message information
* Now wait for the ack from them
*/
for(item = opal_list_get_first(&drained_msg_ack_list);
item != opal_list_get_end(&drained_msg_ack_list);
item = opal_list_get_next(item) ) {
drain_msg_ack = (ompi_crcp_bkmrk_pml_drain_message_ack_ref_t*)item;
/* Post the receive */
if( OMPI_SUCCESS != (ret = orte_rml.recv_buffer_nb( &drain_msg_ack->peer,
OMPI_CRCP_COORD_BOOKMARK_TAG,
0,
drain_message_ack_cbfunc,
NULL) ) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s <-- %s: Failed to post a RML receive to the peer\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(drain_msg_ack->peer)));
return ret;
}
}
return OMPI_SUCCESS;
}
static void drain_message_ack_cbfunc(int status,
orte_process_name_t* sender,
opal_buffer_t *buffer,
orte_rml_tag_t tag,
void* cbdata)
{
int ret, exit_status = OMPI_SUCCESS;
opal_list_item_t* item = NULL;
size_t ckpt_status;
/*
* Unpack the buffer
*/
UNPACK_BUFFER(buffer, ckpt_status, 1, OPAL_SIZE, "");
/*
* Update the outstanding message queue
*/
for(item = opal_list_get_first(&drained_msg_ack_list);
item != opal_list_get_end(&drained_msg_ack_list);
item = opal_list_get_next(item) ) {
ompi_crcp_bkmrk_pml_drain_message_ack_ref_t * drain_msg_ack;
drain_msg_ack = (ompi_crcp_bkmrk_pml_drain_message_ack_ref_t*)item;
/* If this ACK has not completed yet */
if(!drain_msg_ack->complete) {
/* If it is the correct peer */
if( OPAL_EQUAL == orte_util_compare_name_fields(ORTE_NS_CMP_ALL,
&(drain_msg_ack->peer),
sender) ) {
/* We found it! */
drain_msg_ack->complete = true;
OPAL_OUTPUT_VERBOSE((5, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s --> %s Received ACK of FLUSH from peer\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(sender) ));
return;
}
}
}
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s --> %s ERROR: Unable to match ACK to peer\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(sender) );
cleanup:
return;
}
static int ft_event_post_drained(void)
{
int ret, exit_status = OMPI_SUCCESS;
ompi_crcp_bkmrk_pml_peer_ref_t *cur_peer_ref = NULL;
ompi_crcp_bkmrk_pml_drain_message_ref_t * drain_msg_ref = NULL;
ompi_crcp_bkmrk_pml_message_content_ref_t *content_ref = NULL;
opal_list_item_t* item = NULL, *d_item = NULL, *c_item = NULL;
int i, total_number_to_drain = 0, peer_total = 0;
/* First Pass just to get a count */
for(item = opal_list_get_first(&ompi_crcp_bkmrk_pml_peer_refs);
item != opal_list_get_end(&ompi_crcp_bkmrk_pml_peer_refs);
item = opal_list_get_next(item) ) {
cur_peer_ref = (ompi_crcp_bkmrk_pml_peer_ref_t*)item;
if( OPAL_EQUAL == orte_util_compare_name_fields(ORTE_NS_CMP_ALL,
(ORTE_PROC_MY_NAME),
&(cur_peer_ref->proc_name)) ) {
continue;
}
for(d_item = opal_list_get_first(&(cur_peer_ref->drained_list));
d_item != opal_list_get_end(&(cur_peer_ref->drained_list));
d_item = opal_list_get_next(d_item) ) {
drain_msg_ref = (ompi_crcp_bkmrk_pml_drain_message_ref_t*)d_item;
for(c_item = opal_list_get_first(&(drain_msg_ref->msg_contents));
c_item != opal_list_get_end(&(drain_msg_ref->msg_contents));
c_item = opal_list_get_next(c_item) ) {
content_ref = (ompi_crcp_bkmrk_pml_message_content_ref_t*)c_item;
if( !content_ref->done ) {
++total_number_to_drain;
}
}
}
}
/*
* Check to make sure there is something to post
*/
if( 0 >= total_number_to_drain ) {
return OMPI_SUCCESS;
}
/* Allocate memory */
if( NULL != quiesce_requests ) {
free(quiesce_requests);
quiesce_requests = NULL;
}
quiesce_requests = (ompi_request_t **)malloc( (total_number_to_drain) * sizeof(ompi_request_t *));
if( NULL == quiesce_requests){
exit_status = OMPI_ERROR;
goto cleanup;
}
if( NULL != quiesce_statuses ) {
free(quiesce_statuses);
quiesce_statuses = NULL;
}
quiesce_statuses = (ompi_status_public_t **)malloc( (total_number_to_drain) * sizeof(ompi_status_public_t *));
if( NULL == quiesce_statuses){
exit_status = OMPI_ERROR;
goto cleanup;
}
/* Initalize to invalid values */
for(i = 0; i < (total_number_to_drain); ++i) {
quiesce_requests[i] = &ompi_request_null;
quiesce_statuses[i] = &ompi_status_empty;
}
quiesce_request_count = 0;
/* Second pass to post */
for(item = opal_list_get_first(&ompi_crcp_bkmrk_pml_peer_refs);
item != opal_list_get_end(&ompi_crcp_bkmrk_pml_peer_refs);
item = opal_list_get_next(item) ) {
cur_peer_ref = (ompi_crcp_bkmrk_pml_peer_ref_t*)item;
peer_total = 0;
if( OPAL_EQUAL == orte_util_compare_name_fields(ORTE_NS_CMP_ALL,
(ORTE_PROC_MY_NAME),
&(cur_peer_ref->proc_name)) ) {
continue;
}
for(d_item = opal_list_get_first(&(cur_peer_ref->drained_list));
d_item != opal_list_get_end(&(cur_peer_ref->drained_list));
d_item = opal_list_get_next(d_item) ) {
drain_msg_ref = (ompi_crcp_bkmrk_pml_drain_message_ref_t*)d_item;
for(c_item = opal_list_get_first(&(drain_msg_ref->msg_contents));
c_item != opal_list_get_end(&(drain_msg_ref->msg_contents));
c_item = opal_list_get_next(c_item) ) {
content_ref = (ompi_crcp_bkmrk_pml_message_content_ref_t*)c_item;
if( content_ref->done ) {
continue;
}
if( OMPI_SUCCESS != (ret = ft_event_post_drain_message(drain_msg_ref, content_ref) ) ) {
exit_status = ret;
goto cleanup;
}
cur_peer_ref->ack_required = true;
/* Wait on all drain requests, newly posted or not */
if( NULL != content_ref->request) {
quiesce_requests[quiesce_request_count] = content_ref->request;
quiesce_statuses[quiesce_request_count] = &content_ref->status;
quiesce_request_count++;
peer_total++;
}
/* If a NULL request, and already_posted then this is an indicator that we need to stall */
else if( content_ref->already_posted ) {
stall_for_completion = true;
}
else {
ERROR_SHOULD_NEVER_HAPPEN("crcp:bkmrk: ft_event_post_drained(): Found a drain message with a NULL request.");
}
}
}
if( peer_total > 0 || stall_for_completion ) {
OPAL_OUTPUT_VERBOSE((10, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s <-- %s Will be draining %4d messages from this peer. Total %4d %s\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(cur_peer_ref->proc_name)),
peer_total,
quiesce_request_count,
(stall_for_completion ? "(And Stalling)" : "") ));
}
}
cleanup:
return exit_status;
}
static int ft_event_post_drain_message(ompi_crcp_bkmrk_pml_drain_message_ref_t *drain_msg_ref,
ompi_crcp_bkmrk_pml_message_content_ref_t *content_ref)
{
int ret;
/*
* This message has already been posted and drained in a previous
* checkpoint, do not post it again.
*/
if( content_ref->done ) {
return OMPI_SUCCESS;
}
/* Do not repost those that are already posted, and
* we have requests for
*/
if( content_ref->already_posted ) {
OPAL_OUTPUT_VERBOSE((15, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s <-- %s Found a message that we do not need to post.\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(drain_msg_ref->proc_name)) ));
return OMPI_SUCCESS;
}
/* Match counts in traffic_message_create_drain_message() */
content_ref->active = true;
drain_msg_ref->active++;
/*
* Post a receive to drain this message
*/
OPAL_OUTPUT_VERBOSE((20, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s <-- %s Posting a message to be drained from rank %d.\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(drain_msg_ref->proc_name)),
drain_msg_ref->rank));
if( OMPI_SUCCESS != (ret = wrapped_pml_module->pml_irecv(content_ref->buffer,
(drain_msg_ref->count * drain_msg_ref->ddt_size),
drain_msg_ref->datatype,
drain_msg_ref->rank,
drain_msg_ref->tag,
drain_msg_ref->comm,
&(content_ref->request) ) ) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s <-- %s Failed to post the Draining PML iRecv\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(drain_msg_ref->proc_name)) );
return ret;
}
return OMPI_SUCCESS;
}
static int ft_event_wait_quiesce(void)
{
int exit_status = OMPI_SUCCESS;
int ret;
/*********************************************
* Wait for all draining receives to complete
**********************************************/
if( OMPI_SUCCESS != (ret = wait_quiesce_drained() ) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: wait_quiesce: %s Failed to quiesce drained messages\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME) );
exit_status = ret;
goto cleanup;
}
/*******************************************************************
* If we are waiting for All Clear messages from peers wait on them.
*******************************************************************/
if( OMPI_SUCCESS != (ret = wait_quiesce_drain_ack() ) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: wait_quiesce: %s Failed to recv all drain ACKs\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME) );
exit_status = ret;
goto cleanup;
}
cleanup:
return exit_status;
}
static int wait_quiesce_drained(void)
{
int ret, exit_status = OMPI_SUCCESS;
ompi_crcp_bkmrk_pml_peer_ref_t *cur_peer_ref = NULL;
ompi_crcp_bkmrk_pml_drain_message_ref_t * drain_msg_ref = NULL;
ompi_crcp_bkmrk_pml_message_content_ref_t *content_ref = NULL;
opal_list_item_t* item = NULL, *d_item = NULL, *d_next = NULL, *c_item = NULL, *c_next = NULL;
bool prev_stall = false;
/* Can we shortcut this? */
OPAL_OUTPUT_VERBOSE((5, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s Waiting on %d messages to drain\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
(int)quiesce_request_count));
/*
* Wait on all of the message to complete in any order
* Created in ft_event_post_drained()
*/
prev_stall = opal_cr_stall_check;
opal_cr_stall_check = true;
if( OMPI_SUCCESS != (ret = coord_request_wait_all(quiesce_request_count,
quiesce_requests,
quiesce_statuses) ) ) {
exit_status = ret;
goto cleanup;
}
opal_cr_stall_check = prev_stall;
/*
* Send ACKs to all peers
*
* Remove only the already posted members of the drained list.
* All other elements need to be left in the list since we need
* to match them as new receives come in.
*/
for(item = opal_list_get_first(&ompi_crcp_bkmrk_pml_peer_refs);
item != opal_list_get_end(&ompi_crcp_bkmrk_pml_peer_refs);
item = opal_list_get_next(item) ) {
cur_peer_ref = (ompi_crcp_bkmrk_pml_peer_ref_t*)item;
if( OPAL_EQUAL == orte_util_compare_name_fields(ORTE_NS_CMP_ALL,
(ORTE_PROC_MY_NAME),
&(cur_peer_ref->proc_name)) ) {
continue;
}
/*
* Send ACK to peer if wanted
*/
if( cur_peer_ref->ack_required ) {
opal_buffer_t *buffer = NULL;
size_t response = 1;
OPAL_OUTPUT_VERBOSE((5, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s --> %s Send ACKs to Peer\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(cur_peer_ref->proc_name)) ));
/* Send All Clear to Peer */
if (NULL == (buffer = OBJ_NEW(opal_buffer_t))) {
exit_status = OMPI_ERROR;
goto cleanup;
}
PACK_BUFFER(buffer, response, 1, OPAL_SIZE, "");
/* JJH - Performance Optimization? - Why not post all isends, then wait? */
if ( 0 > ( ret = orte_rml.send_buffer(&(cur_peer_ref->proc_name), buffer, OMPI_CRCP_COORD_BOOKMARK_TAG, 0)) ) {
exit_status = ret;
goto cleanup;
}
if( NULL != buffer) {
OBJ_RELEASE(buffer);
buffer = NULL;
}
}
cur_peer_ref->ack_required = false;
/*
* Remove already_posted drained items
*/
for(d_item = opal_list_get_first(&(cur_peer_ref->drained_list));
d_item != opal_list_get_end(&(cur_peer_ref->drained_list));
d_item = d_next ) {
drain_msg_ref = (ompi_crcp_bkmrk_pml_drain_message_ref_t*)d_item;
d_next = opal_list_get_next(d_item);
for(c_item = opal_list_get_first(&(drain_msg_ref->msg_contents));
c_item != opal_list_get_end(&(drain_msg_ref->msg_contents));
c_item = c_next ) {
content_ref = (ompi_crcp_bkmrk_pml_message_content_ref_t*)c_item;
c_next = opal_list_get_next(c_item);
/*
* This message has already been posted and drained in a previous
* checkpoint, do not do anything to it.
*/
if( content_ref->done ) {
continue;
}
if( content_ref->already_posted ) {
drain_message_remove(cur_peer_ref, drain_msg_ref, content_ref);
/* Match counts in traffic_message_create_drain_message() */
drain_msg_ref->active--;
drain_msg_ref->already_posted--;
} else {
content_ref->done = true;
content_ref->active = false;
/* Match counts in traffic_message_create_drain_message() */
drain_msg_ref->done++;
drain_msg_ref->active--;
}
}
}
}
cleanup:
if( NULL != quiesce_requests ) {
free(quiesce_requests);
quiesce_requests = NULL;
}
if( NULL != quiesce_statuses ) {
free(quiesce_statuses);
quiesce_statuses = NULL;
}
quiesce_request_count = 0;
return exit_status;
}
static int coord_request_wait_all( size_t count,
ompi_request_t ** requests,
ompi_status_public_t ** statuses )
{
int exit_status = OMPI_SUCCESS;
ompi_status_public_t * status;
ompi_request_t *req;
size_t i;
/*
* Just wait on each request in order
*/
for( i = 0; i < count; ++i) {
req = requests[i];
status = statuses[i];
coord_request_wait(req, status);
OPAL_OUTPUT_VERBOSE((15, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s Request Wait: Done with idx %d of %d\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
(int)i, (int)count));
}
return exit_status;
}
static int coord_request_wait( ompi_request_t * req,
ompi_status_public_t * status)
{
ompi_request_wait_completion(req);
if( MPI_STATUS_IGNORE != status ) {
status->MPI_TAG = req->req_status.MPI_TAG;
status->MPI_SOURCE = req->req_status.MPI_SOURCE;
status->_count = req->req_status._count;
status->_cancelled = req->req_status._cancelled;
}
return OMPI_SUCCESS;
}
static int wait_quiesce_drain_ack(void)
{
opal_list_item_t* item = NULL;
opal_list_item_t* next = NULL;
ompi_crcp_bkmrk_pml_drain_message_ack_ref_t * drain_msg_ack;
int num_outstanding;
/* YYY JJH YYY Should we wait explicitly on the send requests pending first? */
num_outstanding = opal_list_get_size(&drained_msg_ack_list);
if(num_outstanding <= 0) {
/* Nothing to do */
return OMPI_SUCCESS;
}
OPAL_OUTPUT_VERBOSE((10, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s Waiting on %d Drain ACK messages\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
num_outstanding));
while(0 < num_outstanding) {
for(item = opal_list_get_first(&drained_msg_ack_list);
item != opal_list_get_end(&drained_msg_ack_list);
item = next) {
drain_msg_ack = (ompi_crcp_bkmrk_pml_drain_message_ack_ref_t*)item;
next = opal_list_get_next(item);
if(drain_msg_ack->complete) {
num_outstanding--;
opal_list_remove_item(&drained_msg_ack_list, &(drain_msg_ack->super) );
HOKE_DRAIN_ACK_MSG_REF_RETURN(item);
break;
}
}
opal_progress();
opal_event_loop(OPAL_EVLOOP_NONBLOCK);
}
/* Clear the ack queue if it isn't already clear (it should already be) */
while (NULL != (item = opal_list_remove_first(&drained_msg_ack_list) ) ) {
HOKE_DRAIN_ACK_MSG_REF_RETURN(item);
}
return OMPI_SUCCESS;
}
/* Paired with recv_bookmarks */
static int send_bookmarks(int peer_idx)
{
ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref;
orte_process_name_t peer_name;
opal_buffer_t *buffer = NULL;
int exit_status = OMPI_SUCCESS;
int ret;
START_TIMER(CRCP_TIMER_CKPT_EX_PEER_S);
/*
* Find the peer structure for this peer
*/
peer_name.jobid = ORTE_PROC_MY_NAME->jobid;
peer_name.vpid = peer_idx;
if( NULL == (peer_ref = find_peer(peer_name))) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: send_bookmarks: Error: Could not find peer indexed %d\n",
peer_idx);
exit_status = OMPI_ERROR;
goto cleanup;
}
OPAL_OUTPUT_VERBOSE((15, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s --> %s Sending bookmark (S[%6d] R[%6d])\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&peer_name),
peer_ref->total_msgs_sent,
peer_ref->total_msgs_recvd));
/*
* Send the bookmarks to peer
*/
if (NULL == (buffer = OBJ_NEW(opal_buffer_t))) {
exit_status = OMPI_ERROR;
goto cleanup;
}
PACK_BUFFER(buffer, (peer_ref->total_msgs_sent), 1, OPAL_UINT32,
"crcp:bkmrk: send_bookmarks: Unable to pack total_msgs_sent");
PACK_BUFFER(buffer, (peer_ref->total_msgs_recvd), 1, OPAL_UINT32,
"crcp:bkmrk: send_bookmarks: Unable to pack total_msgs_recvd");
if ( 0 > ( ret = orte_rml.send_buffer(&peer_name, buffer, OMPI_CRCP_COORD_BOOKMARK_TAG, 0)) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: send_bookmarks: Failed to send bookmark to peer %s: Return %d\n",
ORTE_NAME_PRINT(&peer_name),
ret);
exit_status = ret;
goto cleanup;
}
cleanup:
if(NULL != buffer) {
OBJ_RELEASE(buffer);
}
END_TIMER(CRCP_TIMER_CKPT_EX_PEER_S);
DISPLAY_INDV_TIMER(CRCP_TIMER_CKPT_EX_PEER_S, peer_idx, 1);
return exit_status;
}
/* Paired with send_bookmarks */
static int recv_bookmarks(int peer_idx)
{
orte_process_name_t peer_name;
int exit_status = OMPI_SUCCESS;
int ret;
START_TIMER(CRCP_TIMER_CKPT_EX_PEER_R);
peer_name.jobid = ORTE_PROC_MY_NAME->jobid;
peer_name.vpid = peer_idx;
if ( 0 > (ret = orte_rml.recv_buffer_nb(&peer_name,
OMPI_CRCP_COORD_BOOKMARK_TAG,
0,
recv_bookmarks_cbfunc,
NULL) ) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: recv_bookmarks: Failed to post receive bookmark from peer %s: Return %d\n",
ORTE_NAME_PRINT(&peer_name),
ret);
exit_status = ret;
goto cleanup;
}
++total_recv_bookmarks;
cleanup:
END_TIMER(CRCP_TIMER_CKPT_EX_PEER_R);
/* JJH Doesn't make much sense to print this. The real bottleneck is always the send_bookmarks()
* DISPLAY_INDV_TIMER(CRCP_TIMER_CKPT_EX_PEER_R, peer_idx, 1);
*/
return exit_status;
}
static void recv_bookmarks_cbfunc(int status,
orte_process_name_t* sender,
opal_buffer_t *buffer,
orte_rml_tag_t tag,
void* cbdata)
{
ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref;
int exit_status = OMPI_SUCCESS;
int ret, tmp_int;
orte_vpid_t peer_idx;
peer_idx = sender->vpid;
/*
* Find the peer structure for this peer
*/
if( NULL == (peer_ref = find_peer(*sender))) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: recv_bookmarks: Could not find peer indexed %d\n",
peer_idx);
exit_status = OMPI_ERROR;
goto cleanup;
}
UNPACK_BUFFER(buffer, tmp_int, 1, OPAL_UINT32,
"crcp:bkmrk: recv_bookmarks: Unable to unpack total_msgs_sent");
peer_ref->matched_msgs_sent = tmp_int;
UNPACK_BUFFER(buffer, tmp_int, 1, OPAL_UINT32,
"crcp:bkmrk: recv_bookmarks: Unable to unpack total_msgs_recvd");
peer_ref->matched_msgs_recvd = tmp_int;
OPAL_OUTPUT_VERBOSE((15, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s <-- %s Received bookmark (S[%6d] R[%6d]) vs. (S[%6d] R[%6d])\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(sender),
peer_ref->matched_msgs_sent,
peer_ref->matched_msgs_recvd,
peer_ref->total_msgs_sent,
peer_ref->total_msgs_recvd));
cleanup:
--total_recv_bookmarks;
return;
}
static int send_msg_details(ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref,
int total_sent, int total_matched)
{
int ret, exit_status = OMPI_SUCCESS;
ompi_crcp_bkmrk_pml_drain_message_ack_ref_t * d_msg_ack = NULL;
opal_list_t *search_list = NULL;
opal_list_item_t* msg_item = NULL;
bool finished;
int pass_num = 1;
int need, found;
int total_details_sent = 0;
int num_matches = 0;
int p_total_found = 0;
need = total_sent - total_matched;
found = 0;
finished = false;
assert(need > 0);
START_TIMER(CRCP_TIMER_CKPT_CHECK_PEER_S);
/*
* Check the 'send_list' for this peer
*/
search_list = &(peer_ref->send_list);
pass_num = 1;
SEARCH_AGAIN:
for(msg_item = opal_list_get_last(search_list);
msg_item != opal_list_get_begin(search_list);
msg_item = opal_list_get_prev(msg_item) ) {
ompi_crcp_bkmrk_pml_traffic_message_ref_t * msg_ref;
msg_ref = (ompi_crcp_bkmrk_pml_traffic_message_ref_t*)msg_item;
num_matches = 0;
OPAL_OUTPUT_VERBOSE((10, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: send_msg_details: Stage 1: [M/A/D/AD] [%3d/%3d/%3d/%3d] (%s)",
msg_ref->matched, msg_ref->active, msg_ref->done, msg_ref->active_drain,
(msg_ref->msg_type == COORD_MSG_TYPE_B_SEND ? " Send" :
(msg_ref->msg_type == COORD_MSG_TYPE_I_SEND ? "iSend" : "pSend"))
));
/* If this message has not seen any traffic, then skip it */
if( 0 >= (msg_ref->active + msg_ref->done) ) {
continue;
}
/* YYY JJH YYY Keep this as a sanity check? if( msg_ref->matched >= (msg_ref->active + msg_ref->done) ) { continue; } */
if(OMPI_SUCCESS != (ret = do_send_msg_detail(peer_ref, msg_ref, &num_matches, &p_total_found, &finished)) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: send_msg_details: %s --> %s Failed to send message details to peer. Return %d\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer_ref->proc_name)),
ret);
}
OPAL_OUTPUT_VERBOSE((10, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: send_msg_details: Stage 2: [M/A/D/AD] [%3d/%3d/%3d/%3d] (%s) [%3d, %3d, %s] [%3d, %3d]",
msg_ref->matched, msg_ref->active, msg_ref->done, msg_ref->active_drain,
(msg_ref->msg_type == COORD_MSG_TYPE_B_SEND ? " Send" :
(msg_ref->msg_type == COORD_MSG_TYPE_I_SEND ? "iSend" : "pSend")),
num_matches, p_total_found, (finished ? "T" : "F"),
total_details_sent, found
));
total_details_sent += num_matches;
if(0 < num_matches ) {
found += num_matches;
}
if(finished) {
goto ALL_SENT;
}
}
/*
* We tried the 'send_list' and need more,
* so match off the 'isend_list'
*/
if( 1 == pass_num ) {
search_list = &(peer_ref->isend_list);
pass_num = 2;
goto SEARCH_AGAIN;
}
/*
* We tried the 'send_list' and 'isend_list' and need more,
* so match off the 'send_init_list'
*/
if( 2 == pass_num ) {
search_list = &(peer_ref->send_init_list);
pass_num = 3;
goto SEARCH_AGAIN;
}
ALL_SENT:
if( need > found ) {
OPAL_OUTPUT_VERBOSE((10, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: send_msg_details: ERROR: ****** Need (%d) vs Found (%d)",
need, found));
}
assert(need <= found);
/* Prepare to post a Recv for the ACK All Clear signal from the peer
* which is sent when they have finished receiving all of the
* inflight messages into a local buffer
*/
HOKE_DRAIN_ACK_MSG_REF_ALLOC(d_msg_ack, ret);
d_msg_ack->peer.jobid = peer_ref->proc_name.jobid;
d_msg_ack->peer.vpid = peer_ref->proc_name.vpid;
d_msg_ack->complete = false;
opal_list_append(&drained_msg_ack_list, &(d_msg_ack->super));
OPAL_OUTPUT_VERBOSE((10, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s <-> %s Message Inflight! Will wait on ACK from this peer.\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer_ref->proc_name))));
END_TIMER(CRCP_TIMER_CKPT_CHECK_PEER_S);
DISPLAY_INDV_TIMER(CRCP_TIMER_CKPT_CHECK_PEER_S, peer_ref->proc_name.vpid, total_details_sent);
return exit_status;
}
static int do_send_msg_detail(ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref,
ompi_crcp_bkmrk_pml_traffic_message_ref_t*msg_ref,
int *num_matches,
int *total_found,
bool *finished)
{
int ret, exit_status = OMPI_SUCCESS;
opal_buffer_t *buffer = NULL;
int32_t recv_response = RECV_MATCH_RESP_ERROR;
int32_t num_resolv = -1;
int32_t p_total_found = -1;
int comm_my_rank = -1;
int total_sent;
*num_matches = 0;
*total_found = 0;;
*finished = false;
if( NULL != buffer) {
OBJ_RELEASE(buffer);
buffer = NULL;
}
if (NULL == (buffer = OBJ_NEW(opal_buffer_t))) {
exit_status = OMPI_ERROR;
goto cleanup;
}
/*
* Send:
* - Communicator Context ID
* - My Rank in Communicator
*/
comm_my_rank = ompi_comm_rank(msg_ref->comm);
PACK_BUFFER(buffer, msg_ref->comm->c_contextid, 1, OPAL_UINT32,
"crcp:bkmrk: send_msg_details: Unable to pack communicator ID");
PACK_BUFFER(buffer, comm_my_rank, 1, OPAL_INT,
"crcp:bkmrk: send_msg_details: Unable to pack comm rank ID");
/*
* Send:
* - Message tag
* - Message count
* - Message Datatype size
*/
PACK_BUFFER(buffer, msg_ref->tag, 1, OPAL_INT,
"crcp:bkmrk: send_msg_details: Unable to pack tag");
PACK_BUFFER(buffer, msg_ref->count, 1, OPAL_SIZE,
"crcp:bkmrk: send_msg_details: Unable to pack count");
PACK_BUFFER(buffer, msg_ref->ddt_size, 1, OPAL_SIZE,
"crcp:bkmrk: send_msg_details: Unable to pack datatype size");
/*
* Send:
* - Message done
* - Message active
*/
total_sent = msg_ref->done + msg_ref->active;
PACK_BUFFER(buffer, total_sent, 1, OPAL_INT,
"crcp:bkmrk: send_msg_details: Unable to pack done+active count");
/*
* Do the send...
*/
if ( 0 > ( ret = orte_rml.send_buffer(&peer_ref->proc_name, buffer,
OMPI_CRCP_COORD_BOOKMARK_TAG, 0)) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: do_send_msg_detail: Unable to send message details to peer %s: Return %d\n",
ORTE_NAME_PRINT(&peer_ref->proc_name),
ret);
exit_status = OMPI_ERROR;
goto cleanup;
}
if( NULL != buffer) {
OBJ_RELEASE(buffer);
buffer = NULL;
}
/*
* Check return value from peer to see if we found a match.
*/
if (NULL == (buffer = OBJ_NEW(opal_buffer_t))) {
exit_status = ORTE_ERROR;
goto cleanup;
}
/*
* Recv the ACK msg
*/
if ( 0 > (ret = orte_rml.recv_buffer(&peer_ref->proc_name, buffer,
OMPI_CRCP_COORD_BOOKMARK_TAG, 0) ) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: do_send_msg_detail: %s --> %s Failed to receive ACK buffer from peer. Return %d\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer_ref->proc_name)),
ret);
exit_status = ret;
goto cleanup;
}
UNPACK_BUFFER(buffer, recv_response, 1, OPAL_UINT32,
"crcp:bkmrk: send_msg_details: Failed to unpack the ACK from peer buffer.");
UNPACK_BUFFER(buffer, num_resolv, 1, OPAL_UINT32,
"crcp:bkmrk: send_msg_details: Failed to unpack the num_resolv from peer buffer.");
UNPACK_BUFFER(buffer, p_total_found, 1, OPAL_UINT32,
"crcp:bkmrk: send_msg_details: Failed to unpack the total_found from peer buffer.");
/* Mark message as matched */
msg_ref->matched += num_resolv;
*num_matches = num_resolv;
*total_found = p_total_found;
/*
*
*/
if( RECV_MATCH_RESP_DONE == recv_response ) {
*finished = true;
}
else if( RECV_MATCH_RESP_MORE == recv_response ) {
*finished = false;
}
OPAL_OUTPUT_VERBOSE((15, mca_crcp_bkmrk_component.super.output_handle,
"**************************\n"));
OPAL_OUTPUT_VERBOSE((15, mca_crcp_bkmrk_component.super.output_handle,
"send_msg_details: %d, %d = %s [%d / %d]\n",
*num_matches, *total_found,
(*finished ? "Done" : "Continue..."),
msg_ref->done, msg_ref->active));
TRAFFIC_MSG_DUMP_MSG_INDV(15, (msg_ref, "", false));
OPAL_OUTPUT_VERBOSE((15, mca_crcp_bkmrk_component.super.output_handle,
"**************************\n"));
cleanup:
return exit_status;
}
/*
* Recv message details from peer
* - matched with send_msg_details()
*/
static int recv_msg_details(ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref,
int total_recv, int total_matched)
{
int need, found;
int response;
int exit_status = OMPI_SUCCESS;
int ret;
int total_details_recv = 0;
need = total_recv - total_matched;
found = 0;
assert( need > 0);
START_TIMER(CRCP_TIMER_CKPT_CHECK_PEER_R);
/*
* While we are still looking for messages to drain
*/
while(need > found) {
uint32_t p_comm_id;
size_t p_count;
size_t p_datatype_size;
int p_rank;
int p_tag;
int p_num_sent;
int num_resolved = 0;
/*
* Receive the message details from peer
*/
if( OMPI_SUCCESS != (ret = do_recv_msg_detail(peer_ref,
&p_rank, &p_comm_id,
&p_tag, &p_count,
&p_datatype_size,
&p_num_sent)) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: recv_msg_details: %s <-- %s "
"Failed to receive message detail from peer. Return %d\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer_ref->proc_name)),
ret);
exit_status = ret;
goto cleanup;
}
/*
* Determine if we have matched this message or not.
* Also take approprate action.
*/
num_resolved = 0;
if( OMPI_SUCCESS != (ret = do_recv_msg_detail_check_drain(peer_ref,
p_rank, p_comm_id,
p_tag, p_count,
p_datatype_size,
p_num_sent,
&num_resolved)) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: recv_msg_details: %s <-- %s "
"Failed to check message detail from peer. Return %d\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer_ref->proc_name)),
ret);
exit_status = ret;
goto cleanup;
}
found += num_resolved;
total_details_recv += num_resolved;
OPAL_OUTPUT_VERBOSE((10, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s <-- %s Recv Detail: Stage --: [%3d / %3d] [%3d, %3d, %s]",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer_ref->proc_name)),
need, found,
num_resolved, total_details_recv,
( need <= found ? "T" : "F") ));
/* If we do not need any more, respond DONE */
if( need <= found ) {
response = RECV_MATCH_RESP_DONE; /* All done */
}
/* Otherwise respond need more */
else {
response = RECV_MATCH_RESP_MORE;
}
if(OMPI_SUCCESS != (ret = do_recv_msg_detail_resp(peer_ref, response, num_resolved, found))) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: recv_msg_details: %s <-- %s Failed to respond to peer. Return %d\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer_ref->proc_name)),
ret);
exit_status = ret;
goto cleanup;
}
}
cleanup:
END_TIMER(CRCP_TIMER_CKPT_CHECK_PEER_R);
DISPLAY_INDV_TIMER(CRCP_TIMER_CKPT_CHECK_PEER_R, peer_ref->proc_name.vpid, total_details_recv);
return exit_status;
}
static int do_recv_msg_detail(ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref,
int *rank, uint32_t *comm_id, int *tag,
size_t *count, size_t *datatype_size,
int *p_num_sent)
{
opal_buffer_t * buffer = NULL;
int exit_status = OMPI_SUCCESS;
int ret;
if (NULL == (buffer = OBJ_NEW(opal_buffer_t))) {
exit_status = ORTE_ERROR;
goto cleanup;
}
/*
* Recv the msg
*/
if ( 0 > (ret = orte_rml.recv_buffer(&peer_ref->proc_name, buffer, OMPI_CRCP_COORD_BOOKMARK_TAG, 0) ) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: do_recv_msg_detail: %s <-- %s Failed to receive buffer from peer. Return %d\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer_ref->proc_name)),
ret);
exit_status = ret;
goto cleanup;
}
/* Pull out the communicator ID */
UNPACK_BUFFER(buffer, (*comm_id), 1, OPAL_UINT32,
"crcp:bkmrk: recv_msg_details: Failed to unpack the communicator ID");
UNPACK_BUFFER(buffer, (*rank), 1, OPAL_INT,
"crcp:bkmrk: recv_msg_details: Failed to unpack the communicator rank ID");
/* Pull out the message details */
UNPACK_BUFFER(buffer, (*tag), 1, OPAL_INT,
"crcp:bkmrk: recv_msg_details: Failed to unpack the tag");
UNPACK_BUFFER(buffer, (*count), 1, OPAL_SIZE,
"crcp:bkmrk: recv_msg_details: Failed to unpack the count");
UNPACK_BUFFER(buffer, (*datatype_size), 1, OPAL_SIZE,
"crcp:bkmrk: recv_msg_details: Failed to unpack the datatype size");
/* Pull out the counts */
UNPACK_BUFFER(buffer, (*p_num_sent), 1, OPAL_INT,
"crcp:bkmrk: recv_msg_details: Failed to unpack the sent count");
cleanup:
if( NULL != buffer) {
OBJ_RELEASE(buffer);
buffer = NULL;
}
return exit_status;
}
static int do_recv_msg_detail_check_drain(ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref,
int rank, uint32_t comm_id, int tag,
size_t count, size_t datatype_size,
int p_num_sent,
int *num_resolved)
{
int ret, exit_status = OMPI_SUCCESS;
ompi_crcp_bkmrk_pml_traffic_message_ref_t *posted_tmp_msg_ref = NULL;
ompi_crcp_bkmrk_pml_traffic_message_ref_t *posted_recv_msg_ref = NULL;
ompi_crcp_bkmrk_pml_traffic_message_ref_t *posted_irecv_msg_ref = NULL;
ompi_crcp_bkmrk_pml_traffic_message_ref_t *posted_precv_msg_ref = NULL;
ompi_crcp_bkmrk_pml_traffic_message_ref_t *posted_unknown_recv_msg_ref = NULL;
ompi_crcp_bkmrk_pml_traffic_message_ref_t *posted_unknown_precv_msg_ref = NULL;
/* Number of messages left not-matched */
int num_left_unresolved = 0;
/* Number of active messages need to be drained */
int num_still_active = 0;
/* Number of drain messages posted */
int num_posted = 0;
*num_resolved = 0;
num_left_unresolved = p_num_sent;
OPAL_OUTPUT_VERBOSE((10, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s <-- %s "
"Stage 0: Ck.Drain: [TR %3d/MS %3d] sent %4d, unres %4d, res %4d",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer_ref->proc_name)),
peer_ref->total_msgs_recvd,
peer_ref->matched_msgs_sent,
p_num_sent,
num_left_unresolved,
*num_resolved));
TRAFFIC_MSG_DUMP_PEER(15, (peer_ref, "Recv Check...", true));
/*
* Find all references to this message signature.
*/
ret = traffic_message_find_recv(peer_ref, /* Peer to resolve with */
rank, comm_id, tag, count, datatype_size, /* Message signature */
&posted_recv_msg_ref, /* One of 5 lists where this signature could match */
&posted_irecv_msg_ref,
&posted_precv_msg_ref,
&posted_unknown_recv_msg_ref,
&posted_unknown_precv_msg_ref);
if( OMPI_SUCCESS != ret) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: recv_msg_detail_check: %s -- %s "
"Failed to determine if we have received this message. Return %d\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer_ref->proc_name)),
ret);
exit_status = ret;
goto cleanup;
}
/*
* Peer sent 'p_num_sent'.
* For each msg_ref from recv lists:
* Mark all as 'matched'
* Subtract recv->{'active' + 'done'} from 'p_num_sent'
* If recv->active
* need to make sure to drain these and possibly stall
* If 'p_num_sent' > 0
* Post outstanding messages in drain queue
*/
/*
* First pass: Count all 'done'
*/
if( NULL != posted_recv_msg_ref ) {
posted_recv_msg_ref->matched += posted_recv_msg_ref->done;
num_left_unresolved -= posted_recv_msg_ref->done;
TRAFFIC_MSG_DUMP_MSG_INDV(11, (posted_recv_msg_ref, "Ck. Recv", true));
}
if( NULL != posted_irecv_msg_ref ) {
posted_irecv_msg_ref->matched += posted_irecv_msg_ref->done;
num_left_unresolved -= posted_irecv_msg_ref->done;
TRAFFIC_MSG_DUMP_MSG_INDV(11, (posted_irecv_msg_ref, "Ck. iRecv", true));
}
if( NULL != posted_precv_msg_ref ) {
posted_precv_msg_ref->matched += posted_precv_msg_ref->done;
num_left_unresolved -= posted_precv_msg_ref->done;
TRAFFIC_MSG_DUMP_MSG_INDV(11, (posted_precv_msg_ref, "Ck. pRecv", true));
}
if( NULL != posted_unknown_recv_msg_ref ) {
posted_unknown_recv_msg_ref->matched += posted_unknown_recv_msg_ref->done;
num_left_unresolved -= posted_unknown_recv_msg_ref->done;
TRAFFIC_MSG_DUMP_MSG_INDV(11, (posted_unknown_recv_msg_ref, "Ck. uRecv", true));
}
if( NULL != posted_unknown_precv_msg_ref ) {
posted_unknown_precv_msg_ref->matched += posted_unknown_precv_msg_ref->done;
num_left_unresolved -= posted_unknown_precv_msg_ref->done;
TRAFFIC_MSG_DUMP_MSG_INDV(11, (posted_unknown_precv_msg_ref, "Ck. upRecv", true));
}
OPAL_OUTPUT_VERBOSE((10, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s <-- %s "
"Stage 1: Ck.Drain: [TR %3d/MS %3d] sent %4d, unres %4d, res %4d",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer_ref->proc_name)),
peer_ref->total_msgs_recvd,
peer_ref->matched_msgs_sent,
p_num_sent,
num_left_unresolved,
*num_resolved));
/* Short cut if we have completed everything necessary
* This should never happen since we are here because there is a message
* that was sent that has not been started.
*/
if( num_left_unresolved <= 0 ) {
goto cleanup;
}
/*
* Next pass: Count all 'active'
* - if active > unresolved then match all unresolved, and jump to end
* - if active < unresolved then match all active, and continue looking
*/
if( NULL != posted_recv_msg_ref ) {
if( posted_recv_msg_ref->active > num_left_unresolved ) {
posted_recv_msg_ref->matched += num_left_unresolved;
num_still_active += num_left_unresolved;
num_left_unresolved = 0;
} else {
posted_recv_msg_ref->matched += posted_recv_msg_ref->active;
num_still_active += posted_recv_msg_ref->active;
num_left_unresolved -= posted_recv_msg_ref->active;
}
}
if( num_left_unresolved > 0 && NULL != posted_irecv_msg_ref ) {
if( posted_irecv_msg_ref->active > num_left_unresolved ) {
posted_irecv_msg_ref->matched += num_left_unresolved;
num_still_active += num_left_unresolved;
num_left_unresolved = 0;
} else {
posted_irecv_msg_ref->matched += posted_irecv_msg_ref->active;
num_still_active += posted_irecv_msg_ref->active;
num_left_unresolved -= posted_irecv_msg_ref->active;
}
}
if( num_left_unresolved > 0 && NULL != posted_precv_msg_ref ) {
if( posted_precv_msg_ref->active > num_left_unresolved ) {
posted_precv_msg_ref->matched += num_left_unresolved;
num_still_active += num_left_unresolved;
num_left_unresolved = 0;
} else {
posted_precv_msg_ref->matched += posted_precv_msg_ref->active;
num_still_active += posted_precv_msg_ref->active;
num_left_unresolved -= posted_precv_msg_ref->active;
}
}
if( num_left_unresolved > 0 && NULL != posted_unknown_recv_msg_ref ) {
if( posted_unknown_recv_msg_ref->active > num_left_unresolved ) {
posted_unknown_recv_msg_ref->matched += num_left_unresolved;
num_still_active += num_left_unresolved;
num_left_unresolved = 0;
} else {
posted_unknown_recv_msg_ref->matched += posted_unknown_recv_msg_ref->active;
num_still_active += posted_unknown_recv_msg_ref->active;
num_left_unresolved -= posted_unknown_recv_msg_ref->active;
}
}
if( num_left_unresolved > 0 && NULL != posted_unknown_precv_msg_ref ) {
if( posted_unknown_precv_msg_ref->active > num_left_unresolved ) {
posted_unknown_precv_msg_ref->matched += num_left_unresolved;
num_still_active += num_left_unresolved;
num_left_unresolved = 0;
} else {
posted_unknown_precv_msg_ref->matched += posted_unknown_precv_msg_ref->active;
num_still_active += posted_unknown_precv_msg_ref->active;
num_left_unresolved -= posted_unknown_precv_msg_ref->active;
}
}
/*
* If we happen to have more active Recvs than the peer has posted sends, then
* we need to reset the number still active to reflect that only a subset
* of the active sends should be drained.
*/
OPAL_OUTPUT_VERBOSE((10, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s <-- %s "
"Stage 2: Ck.Drain: [TR %3d/MS %3d] sent %4d, unres %4d, res %4d, active %4d",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer_ref->proc_name)),
peer_ref->total_msgs_recvd,
peer_ref->matched_msgs_sent,
p_num_sent,
num_left_unresolved,
*num_resolved,
num_still_active
));
/*
* Check the math at this point, and make sure we did not mess up above.
*/
if(num_left_unresolved < 0 ) {
ERROR_SHOULD_NEVER_HAPPEN_ARG("crcp:bkmrk: Ck.Drain: Unresolved (%3d) < 0", num_left_unresolved);
exit_status = OMPI_ERROR;
goto cleanup;
}
/*
* Fast Track: If there are no outstanding messages to post, and nothing 'active'
* If all the matched messages were found 'done' (none were 'active')
* -> Nothing to do.
*/
if( num_left_unresolved <= 0 &&
num_still_active <= 0) {
goto cleanup;
}
/*
* Stage 3: Resolve 'active' messages by posting a drain message for each
* -> then we need to make sure to wait for them to complete before the checkpoint
* -> Create a drain message
* -> Point the 'request' at it
* -> Make sure not to post this message to be drained, but just wait on the request.
*/
if( num_still_active > 0 ) {
/*
* If this is the current blocking recv, then we need to stall for it to
* complete properly.
* - Only applies to Blocking Recv.
*/
if( NULL != posted_recv_msg_ref ) {
/* Is this the signature of the current blocking recv? */
if (current_msg_id == posted_recv_msg_ref->msg_id &&
COORD_MSG_TYPE_B_RECV == posted_recv_msg_ref->msg_type) {
OPAL_OUTPUT_VERBOSE((10, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s <-- %s "
"Recv Check: Found a message that is 'active'! Prepare to STALL.\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer_ref->proc_name)) ));
stall_for_completion = true;
}
else {
OPAL_OUTPUT_VERBOSE((10, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s <-- %s "
"Recv Check: Found a message that is 'active', but is not the current recv! "
"No stall required [%3d, %3d, %3d, %3d].\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer_ref->proc_name)),
(int)current_msg_id,
(int)current_msg_type,
(int)posted_recv_msg_ref->msg_id,
(int)posted_recv_msg_ref->msg_type));
}
}
/*
* Construct a message for draining for each active message.
* This message will *not* be posted for draining since it is already
* posted in the system. We will simply wait for it to complete.
* - Only applies to messages that are not Blocking Recv
*/
traffic_message_create_drain_message(false, num_still_active,
peer_ref,
&posted_recv_msg_ref,
&num_posted);
num_still_active -= num_posted;
*num_resolved += num_posted;
peer_ref->total_msgs_recvd += num_posted;
traffic_message_create_drain_message(false, num_still_active,
peer_ref,
&posted_irecv_msg_ref,
&num_posted);
num_still_active -= num_posted;
*num_resolved += num_posted;
peer_ref->total_msgs_recvd += num_posted;
traffic_message_create_drain_message(false, num_still_active,
peer_ref,
&posted_precv_msg_ref,
&num_posted);
num_still_active -= num_posted;
*num_resolved += num_posted;
peer_ref->total_msgs_recvd += num_posted;
traffic_message_create_drain_message(false, num_still_active,
peer_ref,
&posted_unknown_recv_msg_ref,
&num_posted);
num_still_active -= num_posted;
*num_resolved += num_posted;
peer_ref->total_msgs_recvd += num_posted;
traffic_message_create_drain_message(false, num_still_active,
peer_ref,
&posted_unknown_precv_msg_ref,
&num_posted);
num_still_active -= num_posted;
*num_resolved += num_posted;
peer_ref->total_msgs_recvd += num_posted;
}
OPAL_OUTPUT_VERBOSE((10, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s <-- %s "
"Stage 3: Ck.Drain: [TR %3d/MS %3d] sent %4d, unres %4d, res %4d, active %4d",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer_ref->proc_name)),
peer_ref->total_msgs_recvd,
peer_ref->matched_msgs_sent,
p_num_sent,
num_left_unresolved,
*num_resolved,
num_still_active
));
/*
* Post all unresolved messages to the drain queue
* - Create a new message to drain
* - Notify peer of resolution of N messages
*/
if( num_left_unresolved > 0 ) {
/* Create a stamp for the drained message */
CREATE_NEW_MSG(posted_tmp_msg_ref, COORD_MSG_TYPE_I_RECV,
count, datatype_size, tag, rank,
ompi_comm_lookup(comm_id),
peer_ref->proc_name.jobid,
peer_ref->proc_name.vpid);
traffic_message_create_drain_message(true, num_left_unresolved,
peer_ref,
&posted_tmp_msg_ref,
&num_posted);
num_left_unresolved -= num_posted;
*num_resolved += num_posted;
peer_ref->total_msgs_recvd += num_posted;
HOKE_TRAFFIC_MSG_REF_RETURN(posted_tmp_msg_ref);
}
OPAL_OUTPUT_VERBOSE((10, mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: %s <-- %s "
"Stage 4: Ck.Drain: [TR %3d/MS %3d] sent %4d, unres %4d, res %4d, active %4d",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer_ref->proc_name)),
peer_ref->total_msgs_recvd,
peer_ref->matched_msgs_sent,
p_num_sent,
num_left_unresolved,
*num_resolved,
num_still_active
));
/* YYY JJH YYY Should we check for no-action? */
cleanup:
return exit_status;
}
static int do_recv_msg_detail_resp(ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref,
int resp, int num_resolv, int total_found)
{
opal_buffer_t * buffer = NULL;
int exit_status = OMPI_SUCCESS;
int ret;
if (NULL == (buffer = OBJ_NEW(opal_buffer_t))) {
exit_status = OMPI_ERROR;
goto cleanup;
}
PACK_BUFFER(buffer, resp, 1, OPAL_UINT32,
"crcp:bkmrk: recv_msg_details: Unable to ask peer for more messages");
PACK_BUFFER(buffer, num_resolv, 1, OPAL_UINT32,
"crcp:bkmrk: recv_msg_details: Unable to ask peer for more messages");
PACK_BUFFER(buffer, total_found, 1, OPAL_UINT32,
"crcp:bkmrk: recv_msg_details: Unable to ask peer for more messages");
if ( 0 > ( ret = orte_rml.send_buffer(&peer_ref->proc_name, buffer, OMPI_CRCP_COORD_BOOKMARK_TAG, 0)) ) {
opal_output(mca_crcp_bkmrk_component.super.output_handle,
"crcp:bkmrk: recv_msg_detail_resp: Unable to send message detail response to peer %s: Return %d\n",
ORTE_NAME_PRINT(&peer_ref->proc_name),
ret);
exit_status = OMPI_ERROR;
goto cleanup;
}
cleanup:
if( NULL != buffer) {
OBJ_RELEASE(buffer);
buffer = NULL;
}
return exit_status;
}
/************************************************
* Timer Utility Functions
************************************************/
static void start_time(int idx) {
if(idx < CRCP_TIMER_MAX ) {
timer_start[idx] = get_time();
}
}
static void end_time(int idx) {
if(idx < CRCP_TIMER_MAX ) {
timer_end[idx] = get_time();
}
}
static double get_time() {
double wtime;
#if OPAL_TIMER_USEC_NATIVE
wtime = (double)opal_timer_base_get_usec() / 1000000.0;
#else
struct timeval tv;
gettimeofday(&tv, NULL);
wtime = tv.tv_sec;
wtime += (double)tv.tv_usec / 1000000.0;
#endif
return wtime;
}
static void clear_timers(void) {
int i;
for(i = 0; i < CRCP_TIMER_MAX; ++i) {
timer_start[i] = 0.0;
timer_end[i] = 0.0;
}
}
static void display_all_timers(int state) {
bool report_ready = false;
double barrier_start, barrier_stop;
int i;
if( 0 != ORTE_PROC_MY_NAME->vpid ) {
if( 2 > timing_enabled ) {
return;
}
else if( 2 == timing_enabled ) {
orte_grpcomm.barrier();
return;
}
}
for( i = 0; i < CRCP_TIMER_MAX; ++i) {
if(timer_end[i] > 0.001) {
report_ready = true;
}
}
if( !report_ready ) {
return;
}
opal_output(0, "crcp:bkmrk: timing(%20s): ******************** Begin: [State = %12s]\n", "Summary", opal_crs_base_state_str(state));
for( i = 0; i < CRCP_TIMER_MAX; ++i) {
display_indv_timer_core(i, 0, 0, false);
}
if( timing_enabled >= 2) {
barrier_start = get_time();
orte_grpcomm.barrier();
barrier_stop = get_time();
opal_output(0,
"crcp:bkmrk: timing(%20s): %20s = %10.2f s\n",
"",
"Group Barrier",
(barrier_stop - barrier_start));
}
opal_output(0, "crcp:bkmrk: timing(%20s): ******************** End: [State = %12s]\n", "Summary", opal_crs_base_state_str(state));
}
static void display_indv_timer(int idx, int proc, int msgs) {
display_indv_timer_core(idx, proc, msgs, true);
}
static void display_indv_timer_core(int idx, int proc, int msgs, bool direct) {
double diff = timer_end[idx] - timer_start[idx];
char * str = NULL;
if( 0 != ORTE_PROC_MY_NAME->vpid && timing_enabled < 3 ) {
return;
}
/* Only display the timer if an end value was set */
if(timer_end[idx] <= 0.001) {
return;
}
switch(idx) {
case CRCP_TIMER_CKPT_EX_PEER_S:
case CRCP_TIMER_CKPT_EX_PEER_R:
case CRCP_TIMER_CKPT_CHECK_PEER_S:
case CRCP_TIMER_CKPT_CHECK_PEER_R:
/* These timers do not mean anything in the aggregate, so only display
* them when directly asked for */
if( direct && timing_enabled >= 2) {
asprintf(&str, "Proc %2d, Msg %5d", proc, msgs);
} else {
return;
}
break;
default:
str = strdup("");
break;
}
opal_output(0,
"crcp:bkmrk: timing(%20s): %20s = %10.2f s\n",
str,
timer_label[idx],
diff);
free(str);
str = NULL;
}
/**************** Message Dump functionality ********************/
#if OMPI_ENABLE_DEBUG
static void traffic_message_dump_msg_content_indv(ompi_crcp_bkmrk_pml_message_content_ref_t * content_ref)
{
OPAL_OUTPUT_VERBOSE((10, mca_crcp_bkmrk_component.super.output_handle,
"\t\t(%3d) Content: [A/D/P/Dr] [%s / %s / %s /%s]",
(int)content_ref->msg_id,
(content_ref->active ? "T" : "F"),
(content_ref->done ? "T" : "F"),
(content_ref->already_posted ? "T" : "F"),
(content_ref->already_drained ? "T" : "F")));
}
static void traffic_message_dump_msg_indv(ompi_crcp_bkmrk_pml_traffic_message_ref_t * msg_ref, char * msg, bool vshort)
{
ompi_crcp_bkmrk_pml_message_content_ref_t *content_ref = NULL;
opal_list_item_t* cont_item = NULL;
char * type_name = NULL;
switch(msg_ref->msg_type) {
case COORD_MSG_TYPE_B_SEND:
type_name = strdup(" Send");
break;
case COORD_MSG_TYPE_I_SEND:
type_name = strdup("iSend");
break;
case COORD_MSG_TYPE_P_SEND:
type_name = strdup("pSend");
break;
case COORD_MSG_TYPE_B_RECV:
type_name = strdup(" Recv");
break;
case COORD_MSG_TYPE_I_RECV:
type_name = strdup("iRecv");
break;
case COORD_MSG_TYPE_P_RECV:
type_name = strdup("pRecv");
break;
default:
type_name = strdup("Unknown");
break;
}
if( !vshort ) {
opal_output(0, "\t%s %10s (%3d): [m %3d/d %3d/a %3d/ad %3d/p %3d] Contents %2d ... count %6d, tag %6d, rank %3d",
type_name,
msg,
(int)msg_ref->msg_id,
msg_ref->matched,
msg_ref->done,
msg_ref->active,
msg_ref->active_drain,
msg_ref->posted,
(int)opal_list_get_size(&msg_ref->msg_contents),
(int)msg_ref->count,
msg_ref->tag,
msg_ref->rank);
} else {
opal_output(0, "\t%s %10s (%3d): [m %3d/d %3d/a %3d/ad %3d/p %3d] Contents %2d ... count %6d",
type_name,
msg,
(int)msg_ref->msg_id,
msg_ref->matched,
msg_ref->done,
msg_ref->active,
msg_ref->active_drain,
msg_ref->posted,
(int)opal_list_get_size(&msg_ref->msg_contents),
(int)msg_ref->count);
}
free(type_name);
for(cont_item = opal_list_get_first(&(msg_ref->msg_contents));
cont_item != opal_list_get_end( &(msg_ref->msg_contents));
cont_item = opal_list_get_next(cont_item) ) {
content_ref = (ompi_crcp_bkmrk_pml_message_content_ref_t*)cont_item;
traffic_message_dump_msg_content_indv(content_ref);
}
}
static void traffic_message_dump_drain_msg_indv(ompi_crcp_bkmrk_pml_drain_message_ref_t * msg_ref, char * msg, bool vshort)
{
ompi_crcp_bkmrk_pml_message_content_ref_t *content_ref = NULL;
opal_list_item_t* cont_item = NULL;
char * type_name = NULL;
switch(msg_ref->msg_type) {
case COORD_MSG_TYPE_B_SEND:
type_name = strdup(" Send");
break;
case COORD_MSG_TYPE_I_SEND:
type_name = strdup("iSend");
break;
case COORD_MSG_TYPE_P_SEND:
type_name = strdup("pSend");
break;
case COORD_MSG_TYPE_B_RECV:
type_name = strdup(" Recv");
break;
case COORD_MSG_TYPE_I_RECV:
type_name = strdup("iRecv");
break;
case COORD_MSG_TYPE_P_RECV:
type_name = strdup("pRecv");
break;
default:
type_name = strdup("Unknown");
break;
}
if( !vshort ) {
opal_output(0, "\t%s %10s (%3d): [d %3d/a %3d] Contents %2d ... count %6d, tag %6d, rank %3d",
type_name,
msg,
(int)msg_ref->msg_id,
msg_ref->done,
msg_ref->active,
(int)opal_list_get_size(&msg_ref->msg_contents),
(int)msg_ref->count,
msg_ref->tag,
msg_ref->rank);
} else {
opal_output(0, "\t%s %10s (%3d): [d %3d/a %3d] Contents %2d ... count %6d",
type_name,
msg,
(int)msg_ref->msg_id,
msg_ref->done,
msg_ref->active,
(int)opal_list_get_size(&msg_ref->msg_contents),
(int)msg_ref->count);
}
free(type_name);
for(cont_item = opal_list_get_first(&(msg_ref->msg_contents));
cont_item != opal_list_get_end( &(msg_ref->msg_contents));
cont_item = opal_list_get_next(cont_item) ) {
content_ref = (ompi_crcp_bkmrk_pml_message_content_ref_t*)cont_item;
traffic_message_dump_msg_content_indv(content_ref);
}
}
static void traffic_message_dump_msg_list(opal_list_t *msg_list, bool is_drain)
{
opal_list_item_t* item = NULL;
ompi_crcp_bkmrk_pml_traffic_message_ref_t * msg_ref = NULL;
ompi_crcp_bkmrk_pml_drain_message_ref_t * drain_msg_ref = NULL;
for(item = opal_list_get_last(msg_list);
item != opal_list_get_begin(msg_list);
item = opal_list_get_prev(item) ) {
if( !is_drain ) {
msg_ref = (ompi_crcp_bkmrk_pml_traffic_message_ref_t*)item;
traffic_message_dump_msg_indv(msg_ref, "", false);
} else {
drain_msg_ref = (ompi_crcp_bkmrk_pml_drain_message_ref_t*)item;
traffic_message_dump_drain_msg_indv(drain_msg_ref, "Drain", false);
}
}
}
static void traffic_message_dump_peer(ompi_crcp_bkmrk_pml_peer_ref_t *peer_ref, char * msg, bool root_only)
{
if( root_only && orte_process_info.my_name.vpid != 0 ) {
return;
} else {
sleep(orte_process_info.my_name.vpid * 2);
}
opal_output(0, "------------- %s ---------------------------------", msg);
opal_output(0, "%s <-> %s Totals Sent [ %3d / %3d ] Recv [ %3d / %3d ]",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer_ref->proc_name)),
peer_ref->total_msgs_sent,
peer_ref->matched_msgs_sent,
peer_ref->total_msgs_recvd,
peer_ref->matched_msgs_recvd);
opal_output(0, "\n");
traffic_message_dump_msg_list(&(peer_ref->send_list), false);
traffic_message_dump_msg_list(&(peer_ref->isend_list), false);
traffic_message_dump_msg_list(&(peer_ref->send_init_list), false);
traffic_message_dump_msg_list(&(peer_ref->recv_list), false);
traffic_message_dump_msg_list(&(peer_ref->irecv_list), false);
traffic_message_dump_msg_list(&(peer_ref->recv_init_list), false);
traffic_message_dump_msg_list(&(peer_ref->drained_list), true);
opal_output(0, "--------------------------------------------------");
usleep(250000);
}
#endif
Копировать постоянную ссылку