[mpich-commits] [mpich] MPICH primary repository branch, master, updated. v3.0.3-5-ga3c816a

This is an automated email from the git hooks/post-receive script. It was generated because a ref change was pushed to the repository containing the project "MPICH primary repository". The branch, master has been updated via a3c816ac90ad52ca1bdef10c6419c67f6c3c6373 (commit) via 9bb897880879d03a8eba22134d92edcc9fdd5fa0 (commit) via 317d58286e16f2307ecf66654fbb86c6d0459c8c (commit) via dcfe70c2b49a8fc64e03219d6525accd30544a95 (commit) from 25661856ca2226f05cb70eb9e10a3a83db76a104 (commit) Those revisions listed above that are new to this repository have not appeared on any other notification email; so we list those revisions in full, below. - Log ----------------------------------------------------------------- http://git.mpich.org/mpich.git/commitdiff/a3c816ac90ad52ca1bdef10c6419c67f6c... commit a3c816ac90ad52ca1bdef10c6419c67f6c3c6373 Author: Ralf Gunter <[email protected]> Date: Mon Apr 1 10:44:12 2013 -0500 Add per-communicator eager threshold support. Message transfers now respect the communicator-specific threshold. This change has not been carefully checked for impact on our shared-memory ping-pong latency. Reviewed-by: goodell diff --git a/src/mpid/ch3/channels/nemesis/src/ch3i_comm.c b/src/mpid/ch3/channels/nemesis/src/ch3i_comm.c index 60bbc2b..f25c9a7 100644 --- a/src/mpid/ch3/channels/nemesis/src/ch3i_comm.c +++ b/src/mpid/ch3/channels/nemesis/src/ch3i_comm.c @@ -27,6 +27,9 @@ int MPIDI_CH3I_comm_create(MPID_Comm *comm, void *param) MPIDI_FUNC_ENTER(MPID_STATE_MPIDI_CH3I_COMM_CREATE); + /* Use the VC's eager threshold by default. */ + comm->ch.eager_max_msg_sz = -1; + #ifndef ENABLED_SHM_COLLECTIVES goto fn_exit; #endif diff --git a/src/mpid/ch3/include/mpidimpl.h b/src/mpid/ch3/include/mpidimpl.h index 4efe3c2..a33c8c8 100644 --- a/src/mpid/ch3/include/mpidimpl.h +++ b/src/mpid/ch3/include/mpidimpl.h @@ -1946,5 +1946,12 @@ int MPIDI_CH3_ReqHandler_GetSendRespComplete( MPIDI_VC_t *, MPID_Request *, #endif /* MPICH_IS_THREADED */ -#endif /* !defined(MPICH_MPIDIMPL_H_INCLUDED) */ +#define MPIDI_CH3_GET_EAGER_THRESHOLD(eager_threshold_p, comm, vc) \ + do { \ + if ((comm)->ch.eager_max_msg_sz != -1) \ + *(eager_threshold_p) = (comm)->ch.eager_max_msg_sz; \ + else \ + *(eager_threshold_p) = (vc)->eager_max_msg_sz; \ + } while (0) +#endif /* !defined(MPICH_MPIDIMPL_H_INCLUDED) */ diff --git a/src/mpid/ch3/include/mpidpre.h b/src/mpid/ch3/include/mpidpre.h index fd8e846..032670a 100644 --- a/src/mpid/ch3/include/mpidpre.h +++ b/src/mpid/ch3/include/mpidpre.h @@ -175,6 +175,7 @@ typedef union { typedef struct MPIDI_CH3I_comm { + int eager_max_msg_sz; /* comm-wide eager/rendezvous message threshold */ int coll_active; /* TRUE iff this communicator is collectively active */ int anysource_enabled; /* TRUE iff this anysource recvs can be posted on this communicator */ struct MPID_nem_barrier_vars *barrier_vars; /* shared memory variables used in barrier */ diff --git a/src/mpid/ch3/src/mpid_init.c b/src/mpid/ch3/src/mpid_init.c index 92803b5..2a7b24e 100644 --- a/src/mpid/ch3/src/mpid_init.c +++ b/src/mpid/ch3/src/mpid_init.c @@ -58,6 +58,31 @@ static int finalize_failed_procs_group(void *param) return mpi_errno; } +#undef FUNCNAME +#define FUNCNAME MPIDI_CH3_Set_eager_threshold +#undef FCNAME +#define FCNAME MPIU_QUOTE(FUNCNAME) +int MPIDI_CH3_Set_eager_threshold(MPID_Comm *comm_ptr, MPID_Info *info, void *state) +{ + int mpi_errno = MPI_SUCCESS; + char *endptr; + MPID_MPI_STATE_DECL(MPID_STATE_MPIDI_CH3_SET_EAGER_THRESHOLD); + + MPID_MPI_FUNC_ENTER(MPID_STATE_MPIDI_CH3_SET_EAGER_THRESHOLD); + + comm_ptr->ch.eager_max_msg_sz = strtol(info->value, &endptr, 0); + + MPIU_ERR_CHKANDJUMP1(*endptr, mpi_errno, MPI_ERR_ARG, + "**infohintparse", "**infohintparse %s", + info->key); + + fn_exit: + MPID_MPI_FUNC_EXIT(MPID_STATE_MPIDI_CH3_SET_EAGER_THRESHOLD); + return mpi_errno; + fn_fail: + goto fn_exit; +} + #undef FUNCNAME #define FUNCNAME MPID_Init @@ -297,6 +322,11 @@ int MPID_Init(int *argc, char ***argv, int requested, int *provided, MPICH_THREAD_LEVEL : requested; } + mpi_errno = MPIR_Comm_register_hint("eager_rendezvous_threshold", + MPIDI_CH3_Set_eager_threshold, + NULL); + if (mpi_errno) MPIU_ERR_POP(mpi_errno); + fn_exit: MPIDI_FUNC_EXIT(MPID_STATE_MPID_INIT); return mpi_errno; diff --git a/src/mpid/ch3/src/mpid_isend.c b/src/mpid/ch3/src/mpid_isend.c index 0ac49ab..ae932a6 100644 --- a/src/mpid/ch3/src/mpid_isend.c +++ b/src/mpid/ch3/src/mpid_isend.c @@ -40,6 +40,7 @@ int MPID_Isend(const void * buf, int count, MPI_Datatype datatype, int rank, #if defined(MPID_USE_SEQUENCE_NUMBERS) MPID_Seqnum_t seqnum; #endif + int eager_threshold = -1; int mpi_errno = MPI_SUCCESS; MPIDI_STATE_DECL(MPID_STATE_MPID_ISEND); @@ -118,11 +119,12 @@ int MPID_Isend(const void * buf, int count, MPI_Datatype datatype, int rank, goto fn_exit; } - + + MPIDI_CH3_GET_EAGER_THRESHOLD(&eager_threshold, comm, vc); + /* FIXME: flow control: limit number of outstanding eager messsages containing data and need to be buffered by the receiver */ - - if (data_sz + sizeof(MPIDI_CH3_Pkt_eager_send_t) <= vc->eager_max_msg_sz) + if (data_sz + sizeof(MPIDI_CH3_Pkt_eager_send_t) <= eager_threshold) { if (dt_contig) { diff --git a/src/mpid/ch3/src/mpid_issend.c b/src/mpid/ch3/src/mpid_issend.c index 738876f..189907b 100644 --- a/src/mpid/ch3/src/mpid_issend.c +++ b/src/mpid/ch3/src/mpid_issend.c @@ -27,6 +27,7 @@ int MPID_Issend(const void * buf, int count, MPI_Datatype datatype, int rank, in #if defined(MPID_USE_SEQUENCE_NUMBERS) MPID_Seqnum_t seqnum; #endif + int eager_threshold = -1; int mpi_errno = MPI_SUCCESS; MPIDI_STATE_DECL(MPID_STATE_MPID_ISSEND); @@ -74,8 +75,10 @@ int MPID_Issend(const void * buf, int count, MPI_Datatype datatype, int rank, in context_offset ); goto fn_exit; } - - if (data_sz + sizeof(MPIDI_CH3_Pkt_eager_sync_send_t) <= vc->eager_max_msg_sz) + + MPIDI_CH3_GET_EAGER_THRESHOLD(&eager_threshold, comm, vc); + + if (data_sz + sizeof(MPIDI_CH3_Pkt_eager_sync_send_t) <= eager_threshold) { mpi_errno = MPIDI_CH3_EagerSyncNoncontigSend( &sreq, buf, count, datatype, data_sz, diff --git a/src/mpid/ch3/src/mpid_send.c b/src/mpid/ch3/src/mpid_send.c index e020b86..90dd507 100644 --- a/src/mpid/ch3/src/mpid_send.c +++ b/src/mpid/ch3/src/mpid_send.c @@ -28,6 +28,7 @@ int MPID_Send(const void * buf, int count, MPI_Datatype datatype, int rank, #if defined(MPID_USE_SEQUENCE_NUMBERS) MPID_Seqnum_t seqnum; #endif + int eager_threshold = -1; int mpi_errno = MPI_SUCCESS; MPIDI_STATE_DECL(MPID_STATE_MPID_SEND); @@ -112,7 +113,9 @@ int MPID_Send(const void * buf, int count, MPI_Datatype datatype, int rank, goto fn_exit; } - + + MPIDI_CH3_GET_EAGER_THRESHOLD(&eager_threshold, comm, vc); + /* FIXME: flow control: limit number of outstanding eager messsages containing data and need to be buffered by the receiver */ #ifdef USE_EAGER_SHORT @@ -125,16 +128,19 @@ int MPID_Send(const void * buf, int count, MPI_Datatype datatype, int rank, } else #endif - if (data_sz + sizeof(MPIDI_CH3_Pkt_eager_send_t) <= - vc->eager_max_msg_sz) { - if (dt_contig) { + + if (data_sz + sizeof(MPIDI_CH3_Pkt_eager_send_t) <= eager_threshold) + { + if (dt_contig) + { mpi_errno = MPIDI_CH3_EagerContigSend( &sreq, MPIDI_CH3_PKT_EAGER_SEND, (char *)buf + dt_true_lb, data_sz, rank, tag, comm, context_offset ); } - else { + else + { MPIDI_Request_create_sreq(sreq, mpi_errno, goto fn_exit); MPIDI_Request_set_type(sreq, MPIDI_REQUEST_TYPE_SEND); mpi_errno = MPIDI_CH3_EagerNoncontigSend( &sreq, @@ -144,7 +150,8 @@ int MPID_Send(const void * buf, int count, MPI_Datatype datatype, int rank, comm, context_offset ); } } - else { + else + { MPIDI_Request_create_sreq(sreq, mpi_errno, goto fn_exit); MPIDI_Request_set_type(sreq, MPIDI_REQUEST_TYPE_SEND); mpi_errno = vc->rndvSend_fn( &sreq, buf, count, datatype, dt_contig, diff --git a/src/mpid/ch3/src/mpid_ssend.c b/src/mpid/ch3/src/mpid_ssend.c index 792cc60..6127d09 100644 --- a/src/mpid/ch3/src/mpid_ssend.c +++ b/src/mpid/ch3/src/mpid_ssend.c @@ -27,6 +27,7 @@ int MPID_Ssend(const void * buf, int count, MPI_Datatype datatype, int rank, int #if defined(MPID_USE_SEQUENCE_NUMBERS) MPID_Seqnum_t seqnum; #endif + int eager_threshold = -1; int mpi_errno = MPI_SUCCESS; MPIDI_STATE_DECL(MPID_STATE_MPID_SSEND); @@ -86,8 +87,10 @@ int MPID_Ssend(const void * buf, int count, MPI_Datatype datatype, int rank, int context_offset ); goto fn_exit; } - - if (data_sz + sizeof(MPIDI_CH3_Pkt_eager_sync_send_t) <= vc->eager_max_msg_sz) + + MPIDI_CH3_GET_EAGER_THRESHOLD(&eager_threshold, comm, vc); + + if (data_sz + sizeof(MPIDI_CH3_Pkt_eager_sync_send_t) <= eager_threshold) { mpi_errno = MPIDI_CH3_EagerSyncNoncontigSend( &sreq, buf, count, datatype, data_sz, http://git.mpich.org/mpich.git/commitdiff/9bb897880879d03a8eba22134d92edcc9f... commit 9bb897880879d03a8eba22134d92edcc9fdd5fa0 Author: Ralf Gunter <[email protected]> Date: Fri Mar 22 15:44:44 2013 -0500 Add info hint support to communicators. Communicators now accept info hints both at creation time (`MPI_Comm_dup_with_info`) as well as at any later point in their life cycle (`MPI_Comm_set_info`). Hints are stored in a hash table, and may be added by defining a function of type int fun(MPID_Comm *, MPID_Info *, void *) which applies the hint to the communicator, and then registered with MPICH via the function `MPIR_Comm_register_hint_fn`. The third (optional) argument can be used to pass extra data to the handler, such as local context from when it was registered with the runtime. Reviewed-by: goodell diff --git a/src/include/mpiimpl.h b/src/include/mpiimpl.h index 544ad44..507fdf0 100644 --- a/src/include/mpiimpl.h +++ b/src/include/mpiimpl.h @@ -1231,9 +1231,11 @@ typedef struct MPID_Comm { implementing the collective routines */ struct MPID_TopoOps *topo_fns; /* Pointer to a table of functions - implementting the topology routines - */ + implementting the topology routines */ int next_sched_tag; /* used by the NBC schedule code to allocate tags */ + + MPID_Info *info; /* Hints to the communicator */ + #ifdef MPID_HAS_HETERO int is_hetero; #endif @@ -1288,6 +1290,8 @@ static inline int MPIR_Comm_release(MPID_Comm * comm_ptr, int isDisconnect) */ int MPIR_Comm_release_always(MPID_Comm *comm_ptr, int isDisconnect); +/* applies the specified info chain to the specified communicator */ +int MPIR_Comm_apply_hints(MPID_Comm *comm_ptr, MPID_Info *info_ptr); /* Preallocated comm objects. There are 3: comm_world, comm_self, and a private (non-user accessible) dup of comm world that is provided @@ -4276,6 +4280,10 @@ int MPIR_Comm_init(MPID_Comm *); /* Miscellaneous */ void MPIU_SetTimeout( int ); +/* Communicator info hint functions */ +typedef int (*MPIR_Comm_hint_fn_t)(MPID_Comm *, MPID_Info *, void *); +int MPIR_Comm_register_hint(const char *hint_key, MPIR_Comm_hint_fn_t fn, void *state); + #if defined(HAVE_VSNPRINTF) && defined(NEEDS_VSNPRINTF_DECL) && !defined(vsnprintf) int vsnprintf(char *str, size_t size, const char *format, va_list ap); # endif @@ -4311,6 +4319,7 @@ void MPIR_Info_get_nkeys_impl(MPID_Info *info_ptr, int *nkeys); int MPIR_Info_get_nthkey_impl(MPID_Info *info, int n, char *key); void MPIR_Info_get_valuelen_impl(MPID_Info *info_ptr, const char *key, int *valuelen, int *flag); int MPIR_Info_set_impl(MPID_Info *info_ptr, const char *key, const char *value); +int MPIR_Info_dup_impl(MPID_Info *info_ptr, MPID_Info **new_info_ptr); int MPIR_Comm_delete_attr_impl(MPID_Comm *comm_ptr, MPID_Keyval *keyval_ptr); int MPIR_Comm_create_keyval_impl(MPI_Comm_copy_attr_function *comm_copy_attr_fn, MPI_Comm_delete_attr_function *comm_delete_attr_fn, diff --git a/src/mpi/comm/comm_set_info.c b/src/mpi/comm/comm_set_info.c index 37ebaaa..85be7b6 100644 --- a/src/mpi/comm/comm_set_info.c +++ b/src/mpi/comm/comm_set_info.c @@ -6,6 +6,7 @@ */ #include "mpiimpl.h" +#include "mpl_utlist.h" /* -- Begin Profiling Symbol Block for routine MPI_Comm_set_info */ #if defined(HAVE_PRAGMA_WEAK) @@ -30,8 +31,33 @@ int MPIR_Comm_set_info_impl(MPID_Comm * comm_ptr, MPID_Info * info_ptr) { int mpi_errno = MPI_SUCCESS; + MPID_Info *info; + MPID_Info *curr_info = NULL; + MPID_MPI_STATE_DECL(MPID_STATE_MPIR_COMM_SET_INFO_IMPL); - /* FIXME: We currently ignore the info passed */ + MPID_MPI_FUNC_ENTER(MPID_STATE_MPIR_COMM_SET_INFO_IMPL); + + mpi_errno = MPIR_Comm_apply_hints(comm_ptr, info_ptr); + if (mpi_errno != MPI_SUCCESS) + goto fn_fail; + + if (comm_ptr->info == NULL) { + /* Always have at least a blank info hint. */ + mpi_errno = MPIU_Info_alloc(&(comm_ptr->info)); + if (mpi_errno != MPI_SUCCESS) + goto fn_fail; + } + + /* MPIR_Info_set_impl will do an O(n) search to prevent duplicate keys, so + * this _FOREACH loop will cost O(m*n) time, where "m" is the number of keys + * in info_ptr and "n" is the number of keys in comm_ptr->info. */ + MPL_LL_FOREACH(info_ptr, curr_info) { + /* Have we hit the default, empty info hint? */ + if (curr_info->key == NULL) continue; + + mpi_errno = MPIR_Info_set_impl(comm_ptr->info, curr_info->key, curr_info->value); + if (mpi_errno) MPIU_ERR_POP(mpi_errno); + } fn_exit: return mpi_errno; diff --git a/src/mpi/comm/commutil.c b/src/mpi/comm/commutil.c index 70e2795..8dc7b55 100644 --- a/src/mpi/comm/commutil.c +++ b/src/mpi/comm/commutil.c @@ -6,6 +6,10 @@ #include "mpiimpl.h" #include "mpicomm.h" +#include "mpiinfo.h" /* MPIU_Info_free */ + +#include "mpl_utlist.h" +#include "mpiu_uthash.h" /* This is the utility file for comm that contains the basic comm items and storage management */ @@ -22,7 +26,14 @@ MPIU_Object_alloc_t MPID_Comm_mem = { 0, 0, 0, 0, MPID_COMM, MPID_COMM_PREALLOC}; /* Communicator creation functions */ -struct MPID_CommOps *MPID_Comm_fns = NULL; +struct MPID_CommOps *MPID_Comm_fns = NULL; +struct MPIR_Comm_hint_fn_elt { + char name[MPI_MAX_INFO_KEY]; + MPIR_Comm_hint_fn_t fn; + void *state; + UT_hash_handle hh; +}; +static struct MPIR_Comm_hint_fn_elt *MPID_hint_fns = NULL; /* utility function to pretty print a context ID for debugging purposes, see * mpiimpl.h for more info on the various fields */ @@ -92,6 +103,7 @@ int MPIR_Comm_init(MPID_Comm *comm_p) comm_p->coll_fns = NULL; comm_p->topo_fns = NULL; comm_p->name[0] = '\0'; + comm_p->info = NULL; comm_p->hierarchy_kind = MPID_HIERARCHY_FLAT; comm_p->node_comm = NULL; @@ -1555,6 +1567,13 @@ int MPIR_Comm_copy( MPID_Comm *comm_ptr, int size, MPID_Comm **outcomm_ptr ) /* Start with no attributes on this communicator */ newcomm_ptr->attributes = 0; + + /* Copy over the info hints from the original communicator. */ + mpi_errno = MPIR_Info_dup_impl(comm_ptr->info, &(newcomm_ptr->info)); + if (mpi_errno) MPIU_ERR_POP(mpi_errno); + mpi_errno = MPIR_Comm_apply_hints(newcomm_ptr, newcomm_ptr->info); + if (mpi_errno) MPIU_ERR_POP(mpi_errno); + *outcomm_ptr = newcomm_ptr; fn_fail: @@ -1684,7 +1703,12 @@ int MPIR_Comm_delete_internal(MPID_Comm * comm_ptr, int isDisconnect) destroyed */ mpi_errno = MPID_Dev_comm_destroy_hook(comm_ptr); if (mpi_errno) MPIU_ERR_POP(mpi_errno); - + + /* Free info hints */ + if (comm_ptr->info != NULL) { + MPIU_Info_free(comm_ptr->info); + } + /* release our reference to the collops structure, comes after the * destroy_hook to allow the device to manage these vtables in a custom * fashion */ @@ -1801,3 +1825,98 @@ int MPIR_Comm_release_always(MPID_Comm *comm_ptr, int isDisconnect) goto fn_exit; } +/* Apply all known info hints in the specified info chain to the given + * communicator. */ +#undef FUNCNAME +#define FUNCNAME MPIR_Comm_apply_hints +#undef FCNAME +#define FCNAME MPIU_QUOTE(FUNCNAME) +int MPIR_Comm_apply_hints(MPID_Comm *comm_ptr, MPID_Info *info_ptr) +{ + int mpi_errno = MPI_SUCCESS; + MPID_Info *hint = NULL; + char hint_name[MPI_MAX_INFO_KEY] = { 0 }; + struct MPIR_Comm_hint_fn_elt *hint_fn = NULL; + MPID_MPI_STATE_DECL(MPID_STATE_MPIR_COMM_APPLY_HINTS); + + MPID_MPI_FUNC_ENTER(MPID_STATE_MPIR_COMM_APPLY_HINTS); + + MPL_LL_FOREACH(info_ptr, hint) { + /* Have we hit the default, empty info hint? */ + if (hint->key == NULL) continue; + + strncpy(hint_name, hint->key, MPI_MAX_INFO_KEY); + + HASH_FIND_STR(MPID_hint_fns, hint_name, hint_fn); + + /* Skip hints that MPICH doesn't recognize. */ + if (hint_fn) { + mpi_errno = hint_fn->fn(comm_ptr, hint, hint_fn->state); + if (mpi_errno) MPIU_ERR_POP(mpi_errno); + } + } + + fn_exit: + MPID_MPI_FUNC_EXIT(MPID_STATE_MPIR_COMM_APPLY_HINTS); + return mpi_errno; + fn_fail: + goto fn_exit; +} + +#undef FUNCNAME +#define FUNCNAME MPIR_Comm_free_hint_handles +#undef FCNAME +#define FCNAME MPIU_QUOTE(FUNCNAME) +static int MPIR_Comm_free_hint_handles(void *ignore) +{ + int mpi_errno = MPI_SUCCESS; + struct MPIR_Comm_hint_fn_elt *curr_hint = NULL, *tmp = NULL; + MPID_MPI_STATE_DECL(MPID_STATE_MPIR_COMM_FREE_HINT_HANDLES); + + MPID_MPI_FUNC_ENTER(MPID_STATE_MPIR_COMM_FREE_HINT_HANDLES); + + if (MPID_hint_fns) { + HASH_ITER(hh, MPID_hint_fns, curr_hint, tmp) { + HASH_DEL(MPID_hint_fns, curr_hint); + MPIU_Free(curr_hint); + } + } + + fn_exit: + MPID_MPI_FUNC_EXIT(MPID_STATE_MPIR_COMM_FREE_HINT_HANDLES); + return mpi_errno; + fn_fail: + goto fn_exit; +} + +/* The hint logic is stored in a uthash, with hint name as key and + * the function responsible for applying the hint as the value. */ +#undef FUNCNAME +#define FUNCNAME MPIR_Comm_register_hint +#undef FCNAME +#define FCNAME MPIU_QUOTE(FUNCNAME) +int MPIR_Comm_register_hint(const char *hint_key, MPIR_Comm_hint_fn_t fn, void *state) +{ + int mpi_errno = MPI_SUCCESS; + struct MPIR_Comm_hint_fn_elt *hint_elt = NULL; + MPID_MPI_STATE_DECL(MPID_STATE_MPIR_COMM_REGISTER_HINT); + + MPID_MPI_FUNC_ENTER(MPID_STATE_MPIR_COMM_REGISTER_HINT); + + if (MPID_hint_fns == NULL) { + MPIR_Add_finalize(MPIR_Comm_free_hint_handles, NULL, MPIR_FINALIZE_CALLBACK_PRIO - 1); + } + + hint_elt = MPIU_Malloc(sizeof(struct MPIR_Comm_hint_fn_elt)); + strncpy(hint_elt->name, hint_key, MPI_MAX_INFO_KEY); + hint_elt->state = state; + hint_elt->fn = fn; + + HASH_ADD_STR(MPID_hint_fns, name, hint_elt); + + fn_exit: + MPID_MPI_FUNC_EXIT(MPID_STATE_MPIR_COMM_REGISTER_HINT); + return mpi_errno; + fn_fail: + goto fn_exit; +} diff --git a/src/mpi/errhan/errnames.txt b/src/mpi/errhan/errnames.txt index cda2f7b..32bf69d 100644 --- a/src/mpi/errhan/errnames.txt +++ b/src/mpi/errhan/errnames.txt @@ -862,6 +862,9 @@ is too big (> MPIU_SHMW_GHND_SZ) **envvarparse:Unable to parse environment variable **envvarparse %s:Unable to parse environment variable (%s) +**infohintparse:Unable to parse info hint +**infohintparse %s:Unable to parse info hint (%s) + **signal:signal() failed **signal %s:signal() failed: %s **sigusr1:This version of MPICH requires the SIGUSR1 signal, but the application has already installed a handler diff --git a/src/mpi/info/info_dup.c b/src/mpi/info/info_dup.c index 63c3d5b..f838fd3 100644 --- a/src/mpi/info/info_dup.c +++ b/src/mpi/info/info_dup.c @@ -24,11 +24,54 @@ #ifndef MPICH_MPI_FROM_PMPI #undef MPI_Info_dup #define MPI_Info_dup PMPI_Info_dup + +#undef FUNCNAME +#define FUNCNAME MPIR_Info_dup_impl +#undef FCNAME +#define FCNAME MPIU_QUOTE(FUNCNAME) +int MPIR_Info_dup_impl(MPID_Info *info_ptr, MPID_Info **new_info_ptr) +{ + int mpi_errno = MPI_SUCCESS; + MPID_Info *curr_old, *curr_new; + + *new_info_ptr = NULL; + if (!info_ptr) goto fn_exit; + + /* Note that this routine allocates info elements one at a time. + In the multithreaded case, each allocation may need to acquire + and release the allocation lock. If that is ever a problem, we + may want to add an "allocate n elements" routine and execute this + it two steps: count and then allocate */ + /* FIXME : multithreaded */ + mpi_errno = MPIU_Info_alloc(&curr_new); + if (mpi_errno) MPIU_ERR_POP(mpi_errno); + *new_info_ptr = curr_new; + + curr_old = info_ptr->next; + while (curr_old) + { + mpi_errno = MPIU_Info_alloc(&curr_new->next); + if (mpi_errno) MPIU_ERR_POP(mpi_errno); + + curr_new = curr_new->next; + curr_new->key = MPIU_Strdup(curr_old->key); + curr_new->value = MPIU_Strdup(curr_old->value); + + curr_old = curr_old->next; + } + +fn_exit: + return mpi_errno; +fn_fail: + goto fn_exit; +} + #endif #undef FUNCNAME #define FUNCNAME MPI_Info_dup - +#undef FCNAME +#define FCNAME MPIU_QUOTE(FUNCNAME) /*@ MPI_Info_dup - Returns a duplicate of the info object @@ -48,16 +91,15 @@ Output Parameters: @*/ int MPI_Info_dup( MPI_Info info, MPI_Info *newinfo ) { - MPID_Info *info_ptr=0, *curr_old, *curr_new; - static const char FCNAME[] = "MPI_Info_dup"; + MPID_Info *info_ptr = 0, *new_info_ptr; int mpi_errno = MPI_SUCCESS; MPID_MPI_STATE_DECL(MPID_STATE_MPI_INFO_DUP); MPIR_ERRTEST_INITIALIZED_ORDIE(); - + MPIU_THREAD_CS_ENTER(ALLFUNC,); MPID_MPI_FUNC_ENTER(MPID_STATE_MPI_INFO_DUP); - + /* Validate parameters, especially handles needing to be converted */ # ifdef HAVE_ERROR_CHECKING { @@ -68,7 +110,7 @@ int MPI_Info_dup( MPI_Info info, MPI_Info *newinfo ) MPID_END_ERROR_CHECKS; } # endif /* HAVE_ERROR_CHECKING */ - + /* Convert MPI object handles to object pointers */ MPID_Info_get_ptr( info, info_ptr ); @@ -86,30 +128,12 @@ int MPI_Info_dup( MPI_Info info, MPI_Info *newinfo ) # endif /* HAVE_ERROR_CHECKING */ /* ... body of routine ... */ - - /* Note that this routine allocates info elements one at a time. - In the multithreaded case, each allocation may need to acquire - and release the allocation lock. If that is ever a problem, we - may want to add an "allocate n elements" routine and execute this - it two steps: count and then allocate */ - /* FIXME : multithreaded */ - mpi_errno = MPIU_Info_alloc(&curr_new); - if (mpi_errno) MPIU_ERR_POP(mpi_errno); - *newinfo = curr_new->handle; - curr_old = info_ptr->next; - while (curr_old) - { - mpi_errno = MPIU_Info_alloc(&curr_new->next); - if (mpi_errno) MPIU_ERR_POP(mpi_errno); + mpi_errno = MPIR_Info_dup_impl(info_ptr, &new_info_ptr); + if (mpi_errno != MPI_SUCCESS) goto fn_fail; - curr_new = curr_new->next; - curr_new->key = MPIU_Strdup(curr_old->key); - curr_new->value = MPIU_Strdup(curr_old->value); + *newinfo = new_info_ptr->handle; - curr_old = curr_old->next; - } - /* ... end of body of routine ... */ fn_exit: diff --git a/src/mpid/ch3/src/ch3u_comm.c b/src/mpid/ch3/src/ch3u_comm.c index 27176ea..23e182f 100644 --- a/src/mpid/ch3/src/ch3u_comm.c +++ b/src/mpid/ch3/src/ch3u_comm.c @@ -183,7 +183,7 @@ static int register_hook_finalize(void *param) MPL_LL_DELETE(destroy_hooks, elt); MPIU_Free(elt); } - + fn_exit: MPIDI_FUNC_EXIT(MPID_STATE_REGISTER_HOOK_FINALIZE); return mpi_errno; http://git.mpich.org/mpich.git/commitdiff/317d58286e16f2307ecf66654fbb86c6d0... commit 317d58286e16f2307ecf66654fbb86c6d0459c8c Author: Dave Goodell <[email protected]> Date: Mon Mar 25 15:43:01 2013 -0500 add `mpiu_uthash.h` With a couple of MPICH-specific tweaks. The `mx` netmod uses a non-namespaced "uthash.h" file based on an older version of the package. This is probably OK for now, but we run the risk of symbol conflict and the update is nontrivial since the API changed between the two versions. No reviewer. diff --git a/src/include/Makefile.mk b/src/include/Makefile.mk index 5a5f453..4165748 100644 --- a/src/include/Makefile.mk +++ b/src/include/Makefile.mk @@ -56,6 +56,7 @@ noinst_HEADERS += \ src/include/mpiu_ex.h \ src/include/mpiu_thread.h \ src/include/mpiu_utarray.h \ + src/include/mpiu_uthash.h \ src/include/mpiutil.h \ src/include/nopackage.h \ src/include/oputil.h \ diff --git a/src/include/mpiu_uthash.h b/src/include/mpiu_uthash.h new file mode 100644 index 0000000..f455310 --- /dev/null +++ b/src/include/mpiu_uthash.h @@ -0,0 +1,951 @@ +/* MPICH-local modifications: + * + * 1) Rename header from "uthash.h" to "mpiu_uthash.h" to avoid accidentally + * pulling in a system-installed version of the header. + * + * 2) Use MPIU_Malloc/MPIU_Free instead of malloc/free + * + * The actual macros themselves have not been namespaced, so if a plain + * "uthash.h" is included anywhere in the tree then there could be a conflict. + */ + +/* +Copyright (c) 2003-2013, Troy D. Hanson http://troydhanson.github.com/uthash/ +All rights reserved. + +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions are met: + + * Redistributions of source code must retain the above copyright + notice, this list of conditions and the following disclaimer. + +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS +IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED +TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A +PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER +OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, +EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, +PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR +PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF +LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING +NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS +SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. +*/ + +#ifndef UTHASH_H +#define UTHASH_H + +#include <string.h> /* memcmp,strlen */ +#include <stddef.h> /* ptrdiff_t */ +#include <stdlib.h> /* exit() */ + +/* These macros use decltype or the earlier __typeof GNU extension. + As decltype is only available in newer compilers (VS2010 or gcc 4.3+ + when compiling c++ source) this code uses whatever method is needed + or, for VS2008 where neither is available, uses casting workarounds. */ +#ifdef _MSC_VER /* MS compiler */ +#if _MSC_VER >= 1600 && defined(__cplusplus) /* VS2010 or newer in C++ mode */ +#define DECLTYPE(x) (decltype(x)) +#else /* VS2008 or older (or VS2010 in C mode) */ +#define NO_DECLTYPE +#define DECLTYPE(x) +#endif +#else /* GNU, Sun and other compilers */ +#define DECLTYPE(x) (__typeof(x)) +#endif + +#ifdef NO_DECLTYPE +#define DECLTYPE_ASSIGN(dst,src) \ +do { \ + char **_da_dst = (char**)(&(dst)); \ + *_da_dst = (char*)(src); \ +} while(0) +#else +#define DECLTYPE_ASSIGN(dst,src) \ +do { \ + (dst) = DECLTYPE(dst)(src); \ +} while(0) +#endif + +/* a number of the hash function use uint32_t which isn't defined on win32 */ +#ifdef _MSC_VER +typedef unsigned int uint32_t; +typedef unsigned char uint8_t; +#else +#include <inttypes.h> /* uint32_t */ +#endif + +#define UTHASH_VERSION 1.9.8 + +#ifndef uthash_fatal +#define uthash_fatal(msg) exit(-1) /* fatal error (out of memory,etc) */ +#endif +#ifndef uthash_malloc +#define uthash_malloc(sz) MPIU_Malloc(sz) /* malloc fcn */ +#endif +#ifndef uthash_free +#define uthash_free(ptr,sz) MPIU_Free(ptr) /* free fcn */ +#endif + +#ifndef uthash_noexpand_fyi +#define uthash_noexpand_fyi(tbl) /* can be defined to log noexpand */ +#endif +#ifndef uthash_expand_fyi +#define uthash_expand_fyi(tbl) /* can be defined to log expands */ +#endif + +/* initial number of buckets */ +#define HASH_INITIAL_NUM_BUCKETS 32 /* initial number of buckets */ +#define HASH_INITIAL_NUM_BUCKETS_LOG2 5 /* lg2 of initial number of buckets */ +#define HASH_BKT_CAPACITY_THRESH 10 /* expand when bucket count reaches */ + +/* calculate the element whose hash handle address is hhe */ +#define ELMT_FROM_HH(tbl,hhp) ((void*)(((char*)(hhp)) - ((tbl)->hho))) + +#define HASH_FIND(hh,head,keyptr,keylen,out) \ +do { \ + unsigned _hf_bkt,_hf_hashv; \ + out=NULL; \ + if (head) { \ + HASH_FCN(keyptr,keylen, (head)->hh.tbl->num_buckets, _hf_hashv, _hf_bkt); \ + if (HASH_BLOOM_TEST((head)->hh.tbl, _hf_hashv)) { \ + HASH_FIND_IN_BKT((head)->hh.tbl, hh, (head)->hh.tbl->buckets[ _hf_bkt ], \ + keyptr,keylen,out); \ + } \ + } \ +} while (0) + +#ifdef HASH_BLOOM +#define HASH_BLOOM_BITLEN (1ULL << HASH_BLOOM) +#define HASH_BLOOM_BYTELEN (HASH_BLOOM_BITLEN/8) + ((HASH_BLOOM_BITLEN%8) ? 1:0) +#define HASH_BLOOM_MAKE(tbl) \ +do { \ + (tbl)->bloom_nbits = HASH_BLOOM; \ + (tbl)->bloom_bv = (uint8_t*)uthash_malloc(HASH_BLOOM_BYTELEN); \ + if (!((tbl)->bloom_bv)) { uthash_fatal( "out of memory"); } \ + memset((tbl)->bloom_bv, 0, HASH_BLOOM_BYTELEN); \ + (tbl)->bloom_sig = HASH_BLOOM_SIGNATURE; \ +} while (0) + +#define HASH_BLOOM_FREE(tbl) \ +do { \ + uthash_free((tbl)->bloom_bv, HASH_BLOOM_BYTELEN); \ +} while (0) + +#define HASH_BLOOM_BITSET(bv,idx) (bv[(idx)/8] |= (1U << ((idx)%8))) +#define HASH_BLOOM_BITTEST(bv,idx) (bv[(idx)/8] & (1U << ((idx)%8))) + +#define HASH_BLOOM_ADD(tbl,hashv) \ + HASH_BLOOM_BITSET((tbl)->bloom_bv, (hashv & (uint32_t)((1ULL << (tbl)->bloom_nbits) - 1))) + +#define HASH_BLOOM_TEST(tbl,hashv) \ + HASH_BLOOM_BITTEST((tbl)->bloom_bv, (hashv & (uint32_t)((1ULL << (tbl)->bloom_nbits) - 1))) + +#else +#define HASH_BLOOM_MAKE(tbl) +#define HASH_BLOOM_FREE(tbl) +#define HASH_BLOOM_ADD(tbl,hashv) +#define HASH_BLOOM_TEST(tbl,hashv) (1) +#define HASH_BLOOM_BYTELEN 0 +#endif + +#define HASH_MAKE_TABLE(hh,head) \ +do { \ + (head)->hh.tbl = (UT_hash_table*)uthash_malloc( \ + sizeof(UT_hash_table)); \ + if (!((head)->hh.tbl)) { uthash_fatal( "out of memory"); } \ + memset((head)->hh.tbl, 0, sizeof(UT_hash_table)); \ + (head)->hh.tbl->tail = &((head)->hh); \ + (head)->hh.tbl->num_buckets = HASH_INITIAL_NUM_BUCKETS; \ + (head)->hh.tbl->log2_num_buckets = HASH_INITIAL_NUM_BUCKETS_LOG2; \ + (head)->hh.tbl->hho = (char*)(&(head)->hh) - (char*)(head); \ + (head)->hh.tbl->buckets = (UT_hash_bucket*)uthash_malloc( \ + HASH_INITIAL_NUM_BUCKETS*sizeof(struct UT_hash_bucket)); \ + if (! (head)->hh.tbl->buckets) { uthash_fatal( "out of memory"); } \ + memset((head)->hh.tbl->buckets, 0, \ + HASH_INITIAL_NUM_BUCKETS*sizeof(struct UT_hash_bucket)); \ + HASH_BLOOM_MAKE((head)->hh.tbl); \ + (head)->hh.tbl->signature = HASH_SIGNATURE; \ +} while(0) + +#define HASH_ADD(hh,head,fieldname,keylen_in,add) \ + HASH_ADD_KEYPTR(hh,head,&((add)->fieldname),keylen_in,add) + +#define HASH_REPLACE(hh,head,fieldname,keylen_in,add,replaced) \ +do { \ + replaced=NULL; \ + HASH_FIND(hh,head,&((add)->fieldname),keylen_in,replaced); \ + if (replaced!=NULL) { \ + HASH_DELETE(hh,head,replaced); \ + }; \ + HASH_ADD(hh,head,fieldname,keylen_in,add); \ +} while(0) + +#define HASH_ADD_KEYPTR(hh,head,keyptr,keylen_in,add) \ +do { \ + unsigned _ha_bkt; \ + (add)->hh.next = NULL; \ + (add)->hh.key = (char*)keyptr; \ + (add)->hh.keylen = (unsigned)keylen_in; \ + if (!(head)) { \ + head = (add); \ + (head)->hh.prev = NULL; \ + HASH_MAKE_TABLE(hh,head); \ + } else { \ + (head)->hh.tbl->tail->next = (add); \ + (add)->hh.prev = ELMT_FROM_HH((head)->hh.tbl, (head)->hh.tbl->tail); \ + (head)->hh.tbl->tail = &((add)->hh); \ + } \ + (head)->hh.tbl->num_items++; \ + (add)->hh.tbl = (head)->hh.tbl; \ + HASH_FCN(keyptr,keylen_in, (head)->hh.tbl->num_buckets, \ + (add)->hh.hashv, _ha_bkt); \ + HASH_ADD_TO_BKT((head)->hh.tbl->buckets[_ha_bkt],&(add)->hh); \ + HASH_BLOOM_ADD((head)->hh.tbl,(add)->hh.hashv); \ + HASH_EMIT_KEY(hh,head,keyptr,keylen_in); \ + HASH_FSCK(hh,head); \ +} while(0) + +#define HASH_TO_BKT( hashv, num_bkts, bkt ) \ +do { \ + bkt = ((hashv) & ((num_bkts) - 1)); \ +} while(0) + +/* delete "delptr" from the hash table. + * "the usual" patch-up process for the app-order doubly-linked-list. + * The use of _hd_hh_del below deserves special explanation. + * These used to be expressed using (delptr) but that led to a bug + * if someone used the same symbol for the head and deletee, like + * HASH_DELETE(hh,users,users); + * We want that to work, but by changing the head (users) below + * we were forfeiting our ability to further refer to the deletee (users) + * in the patch-up process. Solution: use scratch space to + * copy the deletee pointer, then the latter references are via that + * scratch pointer rather than through the repointed (users) symbol. + */ +#define HASH_DELETE(hh,head,delptr) \ +do { \ + unsigned _hd_bkt; \ + struct UT_hash_handle *_hd_hh_del; \ + if ( ((delptr)->hh.prev == NULL) && ((delptr)->hh.next == NULL) ) { \ + uthash_free((head)->hh.tbl->buckets, \ + (head)->hh.tbl->num_buckets*sizeof(struct UT_hash_bucket) ); \ + HASH_BLOOM_FREE((head)->hh.tbl); \ + uthash_free((head)->hh.tbl, sizeof(UT_hash_table)); \ + head = NULL; \ + } else { \ + _hd_hh_del = &((delptr)->hh); \ + if ((delptr) == ELMT_FROM_HH((head)->hh.tbl,(head)->hh.tbl->tail)) { \ + (head)->hh.tbl->tail = \ + (UT_hash_handle*)((ptrdiff_t)((delptr)->hh.prev) + \ + (head)->hh.tbl->hho); \ + } \ + if ((delptr)->hh.prev) { \ + ((UT_hash_handle*)((ptrdiff_t)((delptr)->hh.prev) + \ + (head)->hh.tbl->hho))->next = (delptr)->hh.next; \ + } else { \ + DECLTYPE_ASSIGN(head,(delptr)->hh.next); \ + } \ + if (_hd_hh_del->next) { \ + ((UT_hash_handle*)((ptrdiff_t)_hd_hh_del->next + \ + (head)->hh.tbl->hho))->prev = \ + _hd_hh_del->prev; \ + } \ + HASH_TO_BKT( _hd_hh_del->hashv, (head)->hh.tbl->num_buckets, _hd_bkt); \ + HASH_DEL_IN_BKT(hh,(head)->hh.tbl->buckets[_hd_bkt], _hd_hh_del); \ + (head)->hh.tbl->num_items--; \ + } \ + HASH_FSCK(hh,head); \ +} while (0) + + +/* convenience forms of HASH_FIND/HASH_ADD/HASH_DEL */ +#define HASH_FIND_STR(head,findstr,out) \ + HASH_FIND(hh,head,findstr,strlen(findstr),out) +#define HASH_ADD_STR(head,strfield,add) \ + HASH_ADD(hh,head,strfield,strlen(add->strfield),add) +#define HASH_REPLACE_STR(head,strfield,add,replaced) \ + HASH_REPLACE(hh,head,strfield,strlen(add->strfield),add,replaced) +#define HASH_FIND_INT(head,findint,out) \ + HASH_FIND(hh,head,findint,sizeof(int),out) +#define HASH_ADD_INT(head,intfield,add) \ + HASH_ADD(hh,head,intfield,sizeof(int),add) +#define HASH_REPLACE_INT(head,intfield,add,replaced) \ + HASH_REPLACE(hh,head,intfield,sizeof(int),add,replaced) +#define HASH_FIND_PTR(head,findptr,out) \ + HASH_FIND(hh,head,findptr,sizeof(void *),out) +#define HASH_ADD_PTR(head,ptrfield,add) \ + HASH_ADD(hh,head,ptrfield,sizeof(void *),add) +#define HASH_REPLACE_PTR(head,ptrfield,add) \ + HASH_REPLACE(hh,head,ptrfield,sizeof(void *),add,replaced) +#define HASH_DEL(head,delptr) \ + HASH_DELETE(hh,head,delptr) + +/* HASH_FSCK checks hash integrity on every add/delete when HASH_DEBUG is defined. + * This is for uthash developer only; it compiles away if HASH_DEBUG isn't defined. + */ +#ifdef HASH_DEBUG +#define HASH_OOPS(...) do { fprintf(stderr,__VA_ARGS__); exit(-1); } while (0) +#define HASH_FSCK(hh,head) \ +do { \ + unsigned _bkt_i; \ + unsigned _count, _bkt_count; \ + char *_prev; \ + struct UT_hash_handle *_thh; \ + if (head) { \ + _count = 0; \ + for( _bkt_i = 0; _bkt_i < (head)->hh.tbl->num_buckets; _bkt_i++) { \ + _bkt_count = 0; \ + _thh = (head)->hh.tbl->buckets[_bkt_i].hh_head; \ + _prev = NULL; \ + while (_thh) { \ + if (_prev != (char*)(_thh->hh_prev)) { \ + HASH_OOPS("invalid hh_prev %p, actual %p\n", \ + _thh->hh_prev, _prev ); \ + } \ + _bkt_count++; \ + _prev = (char*)(_thh); \ + _thh = _thh->hh_next; \ + } \ + _count += _bkt_count; \ + if ((head)->hh.tbl->buckets[_bkt_i].count != _bkt_count) { \ + HASH_OOPS("invalid bucket count %d, actual %d\n", \ + (head)->hh.tbl->buckets[_bkt_i].count, _bkt_count); \ + } \ + } \ + if (_count != (head)->hh.tbl->num_items) { \ + HASH_OOPS("invalid hh item count %d, actual %d\n", \ + (head)->hh.tbl->num_items, _count ); \ + } \ + /* traverse hh in app order; check next/prev integrity, count */ \ + _count = 0; \ + _prev = NULL; \ + _thh = &(head)->hh; \ + while (_thh) { \ + _count++; \ + if (_prev !=(char*)(_thh->prev)) { \ + HASH_OOPS("invalid prev %p, actual %p\n", \ + _thh->prev, _prev ); \ + } \ + _prev = (char*)ELMT_FROM_HH((head)->hh.tbl, _thh); \ + _thh = ( _thh->next ? (UT_hash_handle*)((char*)(_thh->next) + \ + (head)->hh.tbl->hho) : NULL ); \ + } \ + if (_count != (head)->hh.tbl->num_items) { \ + HASH_OOPS("invalid app item count %d, actual %d\n", \ + (head)->hh.tbl->num_items, _count ); \ + } \ + } \ +} while (0) +#else +#define HASH_FSCK(hh,head) +#endif + +/* When compiled with -DHASH_EMIT_KEYS, length-prefixed keys are emitted to + * the descriptor to which this macro is defined for tuning the hash function. + * The app can #include <unistd.h> to get the prototype for write(2). */ +#ifdef HASH_EMIT_KEYS +#define HASH_EMIT_KEY(hh,head,keyptr,fieldlen) \ +do { \ + unsigned _klen = fieldlen; \ + write(HASH_EMIT_KEYS, &_klen, sizeof(_klen)); \ + write(HASH_EMIT_KEYS, keyptr, fieldlen); \ +} while (0) +#else +#define HASH_EMIT_KEY(hh,head,keyptr,fieldlen) +#endif + +/* default to Jenkin's hash unless overridden e.g. DHASH_FUNCTION=HASH_SAX */ +#ifdef HASH_FUNCTION +#define HASH_FCN HASH_FUNCTION +#else +#define HASH_FCN HASH_JEN +#endif + +/* The Bernstein hash function, used in Perl prior to v5.6 */ +#define HASH_BER(key,keylen,num_bkts,hashv,bkt) \ +do { \ + unsigned _hb_keylen=keylen; \ + char *_hb_key=(char*)(key); \ + (hashv) = 0; \ + while (_hb_keylen--) { (hashv) = ((hashv) * 33) + *_hb_key++; } \ + bkt = (hashv) & (num_bkts-1); \ +} while (0) + + +/* SAX/FNV/OAT/JEN hash functions are macro variants of those listed at + * http://eternallyconfuzzled.com/tuts/algorithms/jsw_tut_hashing.aspx */ +#define HASH_SAX(key,keylen,num_bkts,hashv,bkt) \ +do { \ + unsigned _sx_i; \ + char *_hs_key=(char*)(key); \ + hashv = 0; \ + for(_sx_i=0; _sx_i < keylen; _sx_i++) \ + hashv ^= (hashv << 5) + (hashv >> 2) + _hs_key[_sx_i]; \ + bkt = hashv & (num_bkts-1); \ +} while (0) + +#define HASH_FNV(key,keylen,num_bkts,hashv,bkt) \ +do { \ + unsigned _fn_i; \ + char *_hf_key=(char*)(key); \ + hashv = 2166136261UL; \ + for(_fn_i=0; _fn_i < keylen; _fn_i++) \ + hashv = (hashv * 16777619) ^ _hf_key[_fn_i]; \ + bkt = hashv & (num_bkts-1); \ +} while(0) + +#define HASH_OAT(key,keylen,num_bkts,hashv,bkt) \ +do { \ + unsigned _ho_i; \ + char *_ho_key=(char*)(key); \ + hashv = 0; \ + for(_ho_i=0; _ho_i < keylen; _ho_i++) { \ + hashv += _ho_key[_ho_i]; \ + hashv += (hashv << 10); \ + hashv ^= (hashv >> 6); \ + } \ + hashv += (hashv << 3); \ + hashv ^= (hashv >> 11); \ + hashv += (hashv << 15); \ + bkt = hashv & (num_bkts-1); \ +} while(0) + +#define HASH_JEN_MIX(a,b,c) \ +do { \ + a -= b; a -= c; a ^= ( c >> 13 ); \ + b -= c; b -= a; b ^= ( a << 8 ); \ + c -= a; c -= b; c ^= ( b >> 13 ); \ + a -= b; a -= c; a ^= ( c >> 12 ); \ + b -= c; b -= a; b ^= ( a << 16 ); \ + c -= a; c -= b; c ^= ( b >> 5 ); \ + a -= b; a -= c; a ^= ( c >> 3 ); \ + b -= c; b -= a; b ^= ( a << 10 ); \ + c -= a; c -= b; c ^= ( b >> 15 ); \ +} while (0) + +#define HASH_JEN(key,keylen,num_bkts,hashv,bkt) \ +do { \ + unsigned _hj_i,_hj_j,_hj_k; \ + unsigned char *_hj_key=(unsigned char*)(key); \ + hashv = 0xfeedbeef; \ + _hj_i = _hj_j = 0x9e3779b9; \ + _hj_k = (unsigned)keylen; \ + while (_hj_k >= 12) { \ + _hj_i += (_hj_key[0] + ( (unsigned)_hj_key[1] << 8 ) \ + + ( (unsigned)_hj_key[2] << 16 ) \ + + ( (unsigned)_hj_key[3] << 24 ) ); \ + _hj_j += (_hj_key[4] + ( (unsigned)_hj_key[5] << 8 ) \ + + ( (unsigned)_hj_key[6] << 16 ) \ + + ( (unsigned)_hj_key[7] << 24 ) ); \ + hashv += (_hj_key[8] + ( (unsigned)_hj_key[9] << 8 ) \ + + ( (unsigned)_hj_key[10] << 16 ) \ + + ( (unsigned)_hj_key[11] << 24 ) ); \ + \ + HASH_JEN_MIX(_hj_i, _hj_j, hashv); \ + \ + _hj_key += 12; \ + _hj_k -= 12; \ + } \ + hashv += keylen; \ + switch ( _hj_k ) { \ + case 11: hashv += ( (unsigned)_hj_key[10] << 24 ); \ + case 10: hashv += ( (unsigned)_hj_key[9] << 16 ); \ + case 9: hashv += ( (unsigned)_hj_key[8] << 8 ); \ + case 8: _hj_j += ( (unsigned)_hj_key[7] << 24 ); \ + case 7: _hj_j += ( (unsigned)_hj_key[6] << 16 ); \ + case 6: _hj_j += ( (unsigned)_hj_key[5] << 8 ); \ + case 5: _hj_j += _hj_key[4]; \ + case 4: _hj_i += ( (unsigned)_hj_key[3] << 24 ); \ + case 3: _hj_i += ( (unsigned)_hj_key[2] << 16 ); \ + case 2: _hj_i += ( (unsigned)_hj_key[1] << 8 ); \ + case 1: _hj_i += _hj_key[0]; \ + } \ + HASH_JEN_MIX(_hj_i, _hj_j, hashv); \ + bkt = hashv & (num_bkts-1); \ +} while(0) + +/* The Paul Hsieh hash function */ +#undef get16bits +#if (defined(__GNUC__) && defined(__i386__)) || defined(__WATCOMC__) \ + || defined(_MSC_VER) || defined (__BORLANDC__) || defined (__TURBOC__) +#define get16bits(d) (*((const uint16_t *) (d))) +#endif + +#if !defined (get16bits) +#define get16bits(d) ((((uint32_t)(((const uint8_t *)(d))[1])) << 8) \ + +(uint32_t)(((const uint8_t *)(d))[0]) ) +#endif +#define HASH_SFH(key,keylen,num_bkts,hashv,bkt) \ +do { \ + unsigned char *_sfh_key=(unsigned char*)(key); \ + uint32_t _sfh_tmp, _sfh_len = keylen; \ + \ + int _sfh_rem = _sfh_len & 3; \ + _sfh_len >>= 2; \ + hashv = 0xcafebabe; \ + \ + /* Main loop */ \ + for (;_sfh_len > 0; _sfh_len--) { \ + hashv += get16bits (_sfh_key); \ + _sfh_tmp = (uint32_t)(get16bits (_sfh_key+2)) << 11 ^ hashv; \ + hashv = (hashv << 16) ^ _sfh_tmp; \ + _sfh_key += 2*sizeof (uint16_t); \ + hashv += hashv >> 11; \ + } \ + \ + /* Handle end cases */ \ + switch (_sfh_rem) { \ + case 3: hashv += get16bits (_sfh_key); \ + hashv ^= hashv << 16; \ + hashv ^= (uint32_t)(_sfh_key[sizeof (uint16_t)] << 18); \ + hashv += hashv >> 11; \ + break; \ + case 2: hashv += get16bits (_sfh_key); \ + hashv ^= hashv << 11; \ + hashv += hashv >> 17; \ + break; \ + case 1: hashv += *_sfh_key; \ + hashv ^= hashv << 10; \ + hashv += hashv >> 1; \ + } \ + \ + /* Force "avalanching" of final 127 bits */ \ + hashv ^= hashv << 3; \ + hashv += hashv >> 5; \ + hashv ^= hashv << 4; \ + hashv += hashv >> 17; \ + hashv ^= hashv << 25; \ + hashv += hashv >> 6; \ + bkt = hashv & (num_bkts-1); \ +} while(0) + +#ifdef HASH_USING_NO_STRICT_ALIASING +/* The MurmurHash exploits some CPU's (x86,x86_64) tolerance for unaligned reads. + * For other types of CPU's (e.g. Sparc) an unaligned read causes a bus error. + * MurmurHash uses the faster approach only on CPU's where we know it's safe. + * + * Note the preprocessor built-in defines can be emitted using: + * + * gcc -m64 -dM -E - < /dev/null (on gcc) + * cc -## a.c (where a.c is a simple test file) (Sun Studio) + */ +#if (defined(__i386__) || defined(__x86_64__) || defined(_M_IX86)) +#define MUR_GETBLOCK(p,i) p[i] +#else /* non intel */ +#define MUR_PLUS0_ALIGNED(p) (((unsigned long)p & 0x3) == 0) +#define MUR_PLUS1_ALIGNED(p) (((unsigned long)p & 0x3) == 1) +#define MUR_PLUS2_ALIGNED(p) (((unsigned long)p & 0x3) == 2) +#define MUR_PLUS3_ALIGNED(p) (((unsigned long)p & 0x3) == 3) +#define WP(p) ((uint32_t*)((unsigned long)(p) & ~3UL)) +#if (defined(__BIG_ENDIAN__) || defined(SPARC) || defined(__ppc__) || defined(__ppc64__)) +#define MUR_THREE_ONE(p) ((((*WP(p))&0x00ffffff) << 8) | (((*(WP(p)+1))&0xff000000) >> 24)) +#define MUR_TWO_TWO(p) ((((*WP(p))&0x0000ffff) <<16) | (((*(WP(p)+1))&0xffff0000) >> 16)) +#define MUR_ONE_THREE(p) ((((*WP(p))&0x000000ff) <<24) | (((*(WP(p)+1))&0xffffff00) >> 8)) +#else /* assume little endian non-intel */ +#define MUR_THREE_ONE(p) ((((*WP(p))&0xffffff00) >> 8) | (((*(WP(p)+1))&0x000000ff) << 24)) +#define MUR_TWO_TWO(p) ((((*WP(p))&0xffff0000) >>16) | (((*(WP(p)+1))&0x0000ffff) << 16)) +#define MUR_ONE_THREE(p) ((((*WP(p))&0xff000000) >>24) | (((*(WP(p)+1))&0x00ffffff) << 8)) +#endif +#define MUR_GETBLOCK(p,i) (MUR_PLUS0_ALIGNED(p) ? ((p)[i]) : \ + (MUR_PLUS1_ALIGNED(p) ? MUR_THREE_ONE(p) : \ + (MUR_PLUS2_ALIGNED(p) ? MUR_TWO_TWO(p) : \ + MUR_ONE_THREE(p)))) +#endif +#define MUR_ROTL32(x,r) (((x) << (r)) | ((x) >> (32 - (r)))) +#define MUR_FMIX(_h) \ +do { \ + _h ^= _h >> 16; \ + _h *= 0x85ebca6b; \ + _h ^= _h >> 13; \ + _h *= 0xc2b2ae35l; \ + _h ^= _h >> 16; \ +} while(0) + +#define HASH_MUR(key,keylen,num_bkts,hashv,bkt) \ +do { \ + const uint8_t *_mur_data = (const uint8_t*)(key); \ + const int _mur_nblocks = (keylen) / 4; \ + uint32_t _mur_h1 = 0xf88D5353; \ + uint32_t _mur_c1 = 0xcc9e2d51; \ + uint32_t _mur_c2 = 0x1b873593; \ + uint32_t _mur_k1 = 0; \ + const uint8_t *_mur_tail; \ + const uint32_t *_mur_blocks = (const uint32_t*)(_mur_data+_mur_nblocks*4); \ + int _mur_i; \ + for(_mur_i = -_mur_nblocks; _mur_i; _mur_i++) { \ + _mur_k1 = MUR_GETBLOCK(_mur_blocks,_mur_i); \ + _mur_k1 *= _mur_c1; \ + _mur_k1 = MUR_ROTL32(_mur_k1,15); \ + _mur_k1 *= _mur_c2; \ + \ + _mur_h1 ^= _mur_k1; \ + _mur_h1 = MUR_ROTL32(_mur_h1,13); \ + _mur_h1 = _mur_h1*5+0xe6546b64; \ + } \ + _mur_tail = (const uint8_t*)(_mur_data + _mur_nblocks*4); \ + _mur_k1=0; \ + switch((keylen) & 3) { \ + case 3: _mur_k1 ^= _mur_tail[2] << 16; \ + case 2: _mur_k1 ^= _mur_tail[1] << 8; \ + case 1: _mur_k1 ^= _mur_tail[0]; \ + _mur_k1 *= _mur_c1; \ + _mur_k1 = MUR_ROTL32(_mur_k1,15); \ + _mur_k1 *= _mur_c2; \ + _mur_h1 ^= _mur_k1; \ + } \ + _mur_h1 ^= (keylen); \ + MUR_FMIX(_mur_h1); \ + hashv = _mur_h1; \ + bkt = hashv & (num_bkts-1); \ +} while(0) +#endif /* HASH_USING_NO_STRICT_ALIASING */ + +/* key comparison function; return 0 if keys equal */ +#define HASH_KEYCMP(a,b,len) memcmp(a,b,len) + +/* iterate over items in a known bucket to find desired item */ +#define HASH_FIND_IN_BKT(tbl,hh,head,keyptr,keylen_in,out) \ +do { \ + if (head.hh_head) DECLTYPE_ASSIGN(out,ELMT_FROM_HH(tbl,head.hh_head)); \ + else out=NULL; \ + while (out) { \ + if ((out)->hh.keylen == keylen_in) { \ + if ((HASH_KEYCMP((out)->hh.key,keyptr,keylen_in)) == 0) break; \ + } \ + if ((out)->hh.hh_next) DECLTYPE_ASSIGN(out,ELMT_FROM_HH(tbl,(out)->hh.hh_next)); \ + else out = NULL; \ + } \ +} while(0) + +/* add an item to a bucket */ +#define HASH_ADD_TO_BKT(head,addhh) \ +do { \ + head.count++; \ + (addhh)->hh_next = head.hh_head; \ + (addhh)->hh_prev = NULL; \ + if (head.hh_head) { (head).hh_head->hh_prev = (addhh); } \ + (head).hh_head=addhh; \ + if (head.count >= ((head.expand_mult+1) * HASH_BKT_CAPACITY_THRESH) \ + && (addhh)->tbl->noexpand != 1) { \ + HASH_EXPAND_BUCKETS((addhh)->tbl); \ + } \ +} while(0) + +/* remove an item from a given bucket */ +#define HASH_DEL_IN_BKT(hh,head,hh_del) \ + (head).count--; \ + if ((head).hh_head == hh_del) { \ + (head).hh_head = hh_del->hh_next; \ + } \ + if (hh_del->hh_prev) { \ + hh_del->hh_prev->hh_next = hh_del->hh_next; \ + } \ + if (hh_del->hh_next) { \ + hh_del->hh_next->hh_prev = hh_del->hh_prev; \ + } + +/* Bucket expansion has the effect of doubling the number of buckets + * and redistributing the items into the new buckets. Ideally the + * items will distribute more or less evenly into the new buckets + * (the extent to which this is true is a measure of the quality of + * the hash function as it applies to the key domain). + * + * With the items distributed into more buckets, the chain length + * (item count) in each bucket is reduced. Thus by expanding buckets + * the hash keeps a bound on the chain length. This bounded chain + * length is the essence of how a hash provides constant time lookup. + * + * The calculation of tbl->ideal_chain_maxlen below deserves some + * explanation. First, keep in mind that we're calculating the ideal + * maximum chain length based on the *new* (doubled) bucket count. + * In fractions this is just n/b (n=number of items,b=new num buckets). + * Since the ideal chain length is an integer, we want to calculate + * ceil(n/b). We don't depend on floating point arithmetic in this + * hash, so to calculate ceil(n/b) with integers we could write + * + * ceil(n/b) = (n/b) + ((n%b)?1:0) + * + * and in fact a previous version of this hash did just that. + * But now we have improved things a bit by recognizing that b is + * always a power of two. We keep its base 2 log handy (call it lb), + * so now we can write this with a bit shift and logical AND: + * + * ceil(n/b) = (n>>lb) + ( (n & (b-1)) ? 1:0) + * + */ +#define HASH_EXPAND_BUCKETS(tbl) \ +do { \ + unsigned _he_bkt; \ + unsigned _he_bkt_i; \ + struct UT_hash_handle *_he_thh, *_he_hh_nxt; \ + UT_hash_bucket *_he_new_buckets, *_he_newbkt; \ + _he_new_buckets = (UT_hash_bucket*)uthash_malloc( \ + 2 * tbl->num_buckets * sizeof(struct UT_hash_bucket)); \ + if (!_he_new_buckets) { uthash_fatal( "out of memory"); } \ + memset(_he_new_buckets, 0, \ + 2 * tbl->num_buckets * sizeof(struct UT_hash_bucket)); \ + tbl->ideal_chain_maxlen = \ + (tbl->num_items >> (tbl->log2_num_buckets+1)) + \ + ((tbl->num_items & ((tbl->num_buckets*2)-1)) ? 1 : 0); \ + tbl->nonideal_items = 0; \ + for(_he_bkt_i = 0; _he_bkt_i < tbl->num_buckets; _he_bkt_i++) \ + { \ + _he_thh = tbl->buckets[ _he_bkt_i ].hh_head; \ + while (_he_thh) { \ + _he_hh_nxt = _he_thh->hh_next; \ + HASH_TO_BKT( _he_thh->hashv, tbl->num_buckets*2, _he_bkt); \ + _he_newbkt = &(_he_new_buckets[ _he_bkt ]); \ + if (++(_he_newbkt->count) > tbl->ideal_chain_maxlen) { \ + tbl->nonideal_items++; \ + _he_newbkt->expand_mult = _he_newbkt->count / \ + tbl->ideal_chain_maxlen; \ + } \ + _he_thh->hh_prev = NULL; \ + _he_thh->hh_next = _he_newbkt->hh_head; \ + if (_he_newbkt->hh_head) _he_newbkt->hh_head->hh_prev = \ + _he_thh; \ + _he_newbkt->hh_head = _he_thh; \ + _he_thh = _he_hh_nxt; \ + } \ + } \ + uthash_free( tbl->buckets, tbl->num_buckets*sizeof(struct UT_hash_bucket) ); \ + tbl->num_buckets *= 2; \ + tbl->log2_num_buckets++; \ + tbl->buckets = _he_new_buckets; \ + tbl->ineff_expands = (tbl->nonideal_items > (tbl->num_items >> 1)) ? \ + (tbl->ineff_expands+1) : 0; \ + if (tbl->ineff_expands > 1) { \ + tbl->noexpand=1; \ + uthash_noexpand_fyi(tbl); \ + } \ + uthash_expand_fyi(tbl); \ +} while(0) + + +/* This is an adaptation of Simon Tatham's O(n log(n)) mergesort */ +/* Note that HASH_SORT assumes the hash handle name to be hh. + * HASH_SRT was added to allow the hash handle name to be passed in. */ +#define HASH_SORT(head,cmpfcn) HASH_SRT(hh,head,cmpfcn) +#define HASH_SRT(hh,head,cmpfcn) \ +do { \ + unsigned _hs_i; \ + unsigned _hs_looping,_hs_nmerges,_hs_insize,_hs_psize,_hs_qsize; \ + struct UT_hash_handle *_hs_p, *_hs_q, *_hs_e, *_hs_list, *_hs_tail; \ + if (head) { \ + _hs_insize = 1; \ + _hs_looping = 1; \ + _hs_list = &((head)->hh); \ + while (_hs_looping) { \ + _hs_p = _hs_list; \ + _hs_list = NULL; \ + _hs_tail = NULL; \ + _hs_nmerges = 0; \ + while (_hs_p) { \ + _hs_nmerges++; \ + _hs_q = _hs_p; \ + _hs_psize = 0; \ + for ( _hs_i = 0; _hs_i < _hs_insize; _hs_i++ ) { \ + _hs_psize++; \ + _hs_q = (UT_hash_handle*)((_hs_q->next) ? \ + ((void*)((char*)(_hs_q->next) + \ + (head)->hh.tbl->hho)) : NULL); \ + if (! (_hs_q) ) break; \ + } \ + _hs_qsize = _hs_insize; \ + while ((_hs_psize > 0) || ((_hs_qsize > 0) && _hs_q )) { \ + if (_hs_psize == 0) { \ + _hs_e = _hs_q; \ + _hs_q = (UT_hash_handle*)((_hs_q->next) ? \ + ((void*)((char*)(_hs_q->next) + \ + (head)->hh.tbl->hho)) : NULL); \ + _hs_qsize--; \ + } else if ( (_hs_qsize == 0) || !(_hs_q) ) { \ + _hs_e = _hs_p; \ + _hs_p = (UT_hash_handle*)((_hs_p->next) ? \ + ((void*)((char*)(_hs_p->next) + \ + (head)->hh.tbl->hho)) : NULL); \ + _hs_psize--; \ + } else if (( \ + cmpfcn(DECLTYPE(head)(ELMT_FROM_HH((head)->hh.tbl,_hs_p)), \ + DECLTYPE(head)(ELMT_FROM_HH((head)->hh.tbl,_hs_q))) \ + ) <= 0) { \ + _hs_e = _hs_p; \ + _hs_p = (UT_hash_handle*)((_hs_p->next) ? \ + ((void*)((char*)(_hs_p->next) + \ + (head)->hh.tbl->hho)) : NULL); \ + _hs_psize--; \ + } else { \ + _hs_e = _hs_q; \ + _hs_q = (UT_hash_handle*)((_hs_q->next) ? \ + ((void*)((char*)(_hs_q->next) + \ + (head)->hh.tbl->hho)) : NULL); \ + _hs_qsize--; \ + } \ + if ( _hs_tail ) { \ + _hs_tail->next = ((_hs_e) ? \ + ELMT_FROM_HH((head)->hh.tbl,_hs_e) : NULL); \ + } else { \ + _hs_list = _hs_e; \ + } \ + _hs_e->prev = ((_hs_tail) ? \ + ELMT_FROM_HH((head)->hh.tbl,_hs_tail) : NULL); \ + _hs_tail = _hs_e; \ + } \ + _hs_p = _hs_q; \ + } \ + _hs_tail->next = NULL; \ + if ( _hs_nmerges <= 1 ) { \ + _hs_looping=0; \ + (head)->hh.tbl->tail = _hs_tail; \ + DECLTYPE_ASSIGN(head,ELMT_FROM_HH((head)->hh.tbl, _hs_list)); \ + } \ + _hs_insize *= 2; \ + } \ + HASH_FSCK(hh,head); \ + } \ +} while (0) + +/* This function selects items from one hash into another hash. + * The end result is that the selected items have dual presence + * in both hashes. There is no copy of the items made; rather + * they are added into the new hash through a secondary hash + * hash handle that must be present in the structure. */ +#define HASH_SELECT(hh_dst, dst, hh_src, src, cond) \ +do { \ + unsigned _src_bkt, _dst_bkt; \ + void *_last_elt=NULL, *_elt; \ + UT_hash_handle *_src_hh, *_dst_hh, *_last_elt_hh=NULL; \ + ptrdiff_t _dst_hho = ((char*)(&(dst)->hh_dst) - (char*)(dst)); \ + if (src) { \ + for(_src_bkt=0; _src_bkt < (src)->hh_src.tbl->num_buckets; _src_bkt++) { \ + for(_src_hh = (src)->hh_src.tbl->buckets[_src_bkt].hh_head; \ + _src_hh; \ + _src_hh = _src_hh->hh_next) { \ + _elt = ELMT_FROM_HH((src)->hh_src.tbl, _src_hh); \ + if (cond(_elt)) { \ + _dst_hh = (UT_hash_handle*)(((char*)_elt) + _dst_hho); \ + _dst_hh->key = _src_hh->key; \ + _dst_hh->keylen = _src_hh->keylen; \ + _dst_hh->hashv = _src_hh->hashv; \ + _dst_hh->prev = _last_elt; \ + _dst_hh->next = NULL; \ + if (_last_elt_hh) { _last_elt_hh->next = _elt; } \ + if (!dst) { \ + DECLTYPE_ASSIGN(dst,_elt); \ + HASH_MAKE_TABLE(hh_dst,dst); \ + } else { \ + _dst_hh->tbl = (dst)->hh_dst.tbl; \ + } \ + HASH_TO_BKT(_dst_hh->hashv, _dst_hh->tbl->num_buckets, _dst_bkt); \ + HASH_ADD_TO_BKT(_dst_hh->tbl->buckets[_dst_bkt],_dst_hh); \ + (dst)->hh_dst.tbl->num_items++; \ + _last_elt = _elt; \ + _last_elt_hh = _dst_hh; \ + } \ + } \ + } \ + } \ + HASH_FSCK(hh_dst,dst); \ +} while (0) + +#define HASH_CLEAR(hh,head) \ +do { \ + if (head) { \ + uthash_free((head)->hh.tbl->buckets, \ + (head)->hh.tbl->num_buckets*sizeof(struct UT_hash_bucket)); \ + HASH_BLOOM_FREE((head)->hh.tbl); \ + uthash_free((head)->hh.tbl, sizeof(UT_hash_table)); \ + (head)=NULL; \ + } \ +} while(0) + +#define HASH_OVERHEAD(hh,head) \ + (size_t)((((head)->hh.tbl->num_items * sizeof(UT_hash_handle)) + \ + ((head)->hh.tbl->num_buckets * sizeof(UT_hash_bucket)) + \ + (sizeof(UT_hash_table)) + \ + (HASH_BLOOM_BYTELEN))) + +#ifdef NO_DECLTYPE +#define HASH_ITER(hh,head,el,tmp) \ +for((el)=(head), (*(char**)(&(tmp)))=(char*)((head)?(head)->hh.next:NULL); \ + el; (el)=(tmp),(*(char**)(&(tmp)))=(char*)((tmp)?(tmp)->hh.next:NULL)) +#else +#define HASH_ITER(hh,head,el,tmp) \ +for((el)=(head),(tmp)=DECLTYPE(el)((head)?(head)->hh.next:NULL); \ + el; (el)=(tmp),(tmp)=DECLTYPE(el)((tmp)?(tmp)->hh.next:NULL)) +#endif + +/* obtain a count of items in the hash */ +#define HASH_COUNT(head) HASH_CNT(hh,head) +#define HASH_CNT(hh,head) ((head)?((head)->hh.tbl->num_items):0) + +typedef struct UT_hash_bucket { + struct UT_hash_handle *hh_head; + unsigned count; + + /* expand_mult is normally set to 0. In this situation, the max chain length + * threshold is enforced at its default value, HASH_BKT_CAPACITY_THRESH. (If + * the bucket's chain exceeds this length, bucket expansion is triggered). + * However, setting expand_mult to a non-zero value delays bucket expansion + * (that would be triggered by additions to this particular bucket) + * until its chain length reaches a *multiple* of HASH_BKT_CAPACITY_THRESH. + * (The multiplier is simply expand_mult+1). The whole idea of this + * multiplier is to reduce bucket expansions, since they are expensive, in + * situations where we know that a particular bucket tends to be overused. + * It is better to let its chain length grow to a longer yet-still-bounded + * value, than to do an O(n) bucket expansion too often. + */ + unsigned expand_mult; + +} UT_hash_bucket; + +/* random signature used only to find hash tables in external analysis */ +#define HASH_SIGNATURE 0xa0111fe1 +#define HASH_BLOOM_SIGNATURE 0xb12220f2 + +typedef struct UT_hash_table { + UT_hash_bucket *buckets; + unsigned num_buckets, log2_num_buckets; + unsigned num_items; + struct UT_hash_handle *tail; /* tail hh in app order, for fast append */ + ptrdiff_t hho; /* hash handle offset (byte pos of hash handle in element */ + + /* in an ideal situation (all buckets used equally), no bucket would have + * more than ceil(#items/#buckets) items. that's the ideal chain length. */ + unsigned ideal_chain_maxlen; + + /* nonideal_items is the number of items in the hash whose chain position + * exceeds the ideal chain maxlen. these items pay the penalty for an uneven + * hash distribution; reaching them in a chain traversal takes >ideal steps */ + unsigned nonideal_items; + + /* ineffective expands occur when a bucket doubling was performed, but + * afterward, more than half the items in the hash had nonideal chain + * positions. If this happens on two consecutive expansions we inhibit any + * further expansion, as it's not helping; this happens when the hash + * function isn't a good fit for the key domain. When expansion is inhibited + * the hash will still work, albeit no longer in constant time. */ + unsigned ineff_expands, noexpand; + + uint32_t signature; /* used only to find hash tables in external analysis */ +#ifdef HASH_BLOOM + uint32_t bloom_sig; /* used only to test bloom exists in external analysis */ + uint8_t *bloom_bv; + char bloom_nbits; +#endif + +} UT_hash_table; + +typedef struct UT_hash_handle { + struct UT_hash_table *tbl; + void *prev; /* prev element in app order */ + void *next; /* next element in app order */ + struct UT_hash_handle *hh_prev; /* previous hh in bucket order */ + struct UT_hash_handle *hh_next; /* next hh in bucket order */ + void *key; /* ptr to enclosing struct's key */ + unsigned keylen; /* enclosing struct's key len */ + unsigned hashv; /* result of hash-fcn(key) */ +} UT_hash_handle; + +#endif /* UTHASH_H */ http://git.mpich.org/mpich.git/commitdiff/dcfe70c2b49a8fc64e03219d6525accd30... commit dcfe70c2b49a8fc64e03219d6525accd30544a95 Author: Ralf Gunter <[email protected]> Date: Fri Mar 22 15:33:45 2013 -0500 VCs now use eager threshold from environment var. Previously only the Nemesis shared memory eager threshold could be changed via an environment variable. Now the default (if not overridden by the netmod) eager/rendezvous threshold can also be set via an environment variable. Reviewed-by: goodell@ diff --git a/src/mpid/ch3/channels/nemesis/include/mpidi_ch3_post.h b/src/mpid/ch3/channels/nemesis/include/mpidi_ch3_post.h index 3dd3883..13a9c79 100644 --- a/src/mpid/ch3/channels/nemesis/include/mpidi_ch3_post.h +++ b/src/mpid/ch3/channels/nemesis/include/mpidi_ch3_post.h @@ -7,9 +7,6 @@ #if !defined(MPICH_MPIDI_CH3_POST_H_INCLUDED) #define MPICH_MPIDI_CH3_POST_H_INCLUDED -/* #define MPIDI_CH3_EAGER_MAX_MSG_SIZE (1500 - sizeof(MPIDI_CH3_Pkt_t)) */ -#define MPIDI_CH3_EAGER_MAX_MSG_SIZE (128*1024) - #define MPIDI_CH3_Progress_start(progress_state_) \ (progress_state_)->ch.completion_count = OPA_load_int(&MPIDI_CH3I_progress_completion_count); #define MPIDI_CH3_Progress_end(progress_state_) diff --git a/src/mpid/ch3/src/mpid_vc.c b/src/mpid/ch3/src/mpid_vc.c index 0076e3b..0d4afa6 100644 --- a/src/mpid/ch3/src/mpid_vc.c +++ b/src/mpid/ch3/src/mpid_vc.c @@ -721,8 +721,8 @@ int MPIDI_VC_Init( MPIDI_VC_t *vc, MPIDI_PG_t *pg, int rank ) MPIDI_VC_Init_seqnum_recv(vc); vc->rndvSend_fn = MPIDI_CH3_RndvSend; vc->rndvRecv_fn = MPIDI_CH3_RecvRndv; - vc->eager_max_msg_sz = MPIDI_CH3_EAGER_MAX_MSG_SIZE; vc->ready_eager_max_msg_sz = -1; /* no limit */; + vc->eager_max_msg_sz = MPIR_PARAM_CH3_EAGER_MAX_MSG_SIZE; vc->sendNoncontig_fn = MPIDI_CH3_SendNoncontig_iov; #ifdef ENABLE_COMM_OVERRIDES diff --git a/src/util/param/params.yml b/src/util/param/params.yml index ecd9a82..5d5fe3f 100644 --- a/src/util/param/params.yml +++ b/src/util/param/params.yml @@ -38,6 +38,8 @@ categories: description : parameters that control fault tolerance behavior - name : threads description : multi-threading parameters + - name : ch3 + description : parameters that control behavior of ch3 - name : nemesis description : "parameters that control behavior of the ch3:nemesis channel" - name : sockets @@ -369,8 +371,9 @@ parameters: default : -1 description : >- This parameter controls the message size at which Nemesis - switches from eager to rendezvous mode. If this parameter - is set to -1, then Nemesis will choose an appropriate value. + switches from eager to rendezvous mode for shared memory. + If this parameter is set to -1, then Nemesis will choose + an appropriate value. - category : nemesis name : SHM_READY_EAGER_MAX_SZ @@ -394,6 +397,15 @@ parameters: # Note the netmod must implement this functionality by not # setting the comm_ops table in any VC. + - category : ch3 + name : CH3_EAGER_MAX_MSG_SIZE + type : int + default : 131072 + description : >- + This parameter controls the message size at which CH3 switches + from eager to rendezvous mode. + + ############################################################## # debugging parameters - category : debugger ----------------------------------------------------------------------- Summary of changes: src/include/Makefile.mk | 1 + src/include/mpiimpl.h | 13 +- src/include/mpiu_uthash.h | 951 ++++++++++++++++++++ src/mpi/comm/comm_set_info.c | 28 +- src/mpi/comm/commutil.c | 123 +++- src/mpi/errhan/errnames.txt | 3 + src/mpi/info/info_dup.c | 78 ++- .../ch3/channels/nemesis/include/mpidi_ch3_post.h | 3 - src/mpid/ch3/channels/nemesis/src/ch3i_comm.c | 3 + src/mpid/ch3/include/mpidimpl.h | 9 +- src/mpid/ch3/include/mpidpre.h | 1 + src/mpid/ch3/src/ch3u_comm.c | 2 +- src/mpid/ch3/src/mpid_init.c | 30 + src/mpid/ch3/src/mpid_isend.c | 8 +- src/mpid/ch3/src/mpid_issend.c | 7 +- src/mpid/ch3/src/mpid_send.c | 19 +- src/mpid/ch3/src/mpid_ssend.c | 7 +- src/mpid/ch3/src/mpid_vc.c | 2 +- src/util/param/params.yml | 16 +- 19 files changed, 1251 insertions(+), 53 deletions(-) create mode 100644 src/include/mpiu_uthash.h hooks/post-receive -- MPICH primary repository