- commits - mailman3.cels.anl.gov

r10732 - in mpich2/trunk/src/mpl: . include src
by gropp＠mcs.anl.gov 09 Dec '12

09 Dec '12

Author: gropp Date: 2012-12-09 13:23:05 -0600 (Sun, 09 Dec 2012) New Revision: 10732 Modified: mpich2/trunk/src/mpl/configure.ac mpich2/trunk/src/mpl/include/mpltrmem.h mpich2/trunk/src/mpl/src/mpltrmem.c Log: Enhanced trmem (tracing memory) code in preparation for changing to size_t from int for sizes Modified: mpich2/trunk/src/mpl/configure.ac =================================================================== --- mpich2/trunk/src/mpl/configure.ac 2012-12-09 17:13:56 UTC (rev 10731) +++ mpich2/trunk/src/mpl/configure.ac 2012-12-09 19:23:05 UTC (rev 10732) @@ -60,7 +60,7 @@ fi dnl Check if the necessary headers are available -AC_CHECK_HEADERS(stdio.h stdlib.h string.h stdarg.h ctype.h) +AC_CHECK_HEADERS(stdio.h stdlib.h string.h stdarg.h ctype.h search.h) # A C99 compliant compiler should have inttypes.h for fixed-size int types AC_CHECK_HEADERS(inttypes.h stdint.h) Modified: mpich2/trunk/src/mpl/include/mpltrmem.h =================================================================== --- mpich2/trunk/src/mpl/include/mpltrmem.h 2012-12-09 17:13:56 UTC (rev 10731) +++ mpich2/trunk/src/mpl/include/mpltrmem.h 2012-12-09 19:23:05 UTC (rev 10732) @@ -15,6 +15,7 @@ void *MPL_trmalloc(unsigned int, int, const char[]); void MPL_trfree(void *, int, const char[]); int MPL_trvalid(const char[]); +int MPL_trvalid2(const char[],int,const char[]); void MPL_trspace(int *, int *); void MPL_trid(int); void MPL_trlevel(int); Modified: mpich2/trunk/src/mpl/src/mpltrmem.c =================================================================== --- mpich2/trunk/src/mpl/src/mpltrmem.c 2012-12-09 17:13:56 UTC (rev 10731) +++ mpich2/trunk/src/mpl/src/mpltrmem.c 2012-12-09 19:23:05 UTC (rev 10732) @@ -84,7 +84,7 @@ #define ALREADY_FREED 0x0f0e0d9c typedef struct TRSPACE { - unsigned long size; + size_t size; int id; int lineno; int freed_lineno; @@ -106,9 +106,13 @@ * This package maintains some state about itself. These globals hold * this information. */ +#define TRHEAD_PRESENTINAL ((TRSPACE *)0xbacdef01) +#define TRHEAD_POSTSENTINAL ((TRSPACE *)0x10fedcba) static int world_rank = -1; -static volatile long allocated = 0, frags = 0; -static TRSPACE *volatile TRhead = 0; +static volatile size_t allocated = 0; +static volatile long frags = 0; +static TRSPACE *volatile TRhead[3] = + { TRHEAD_PRESENTINAL, 0, TRHEAD_POSTSENTINAL }; static volatile int TRid = 0; static volatile int TRidSet = 0; static volatile int TRlevel = 0; @@ -120,10 +124,10 @@ #define TR_FREE 0x2 /* Used to keep track of allocations */ -static volatile long TRMaxMem = 0; -static volatile long TRMaxMemId = 0; +static volatile size_t TRMaxMem = 0; +static volatile int TRMaxMemId = 0; /* Used to limit allocation */ -static volatile long TRMaxMemAllow = 0; +static volatile size_t TRMaxMemAllow = 0; /* * Printing of addresses. @@ -178,6 +182,11 @@ TRDefaultByte = 0; TRFreedByte = 0; } + s = getenv("MPL_TRMEM_TRACELEVEL"); + if (s && *s) { + int l = atoi(s); + TRlevel = l; + } } @@ -198,21 +207,19 @@ TRSPACE *head; char *new = NULL; unsigned long *nend; - unsigned int nsize; + size_t nsize; int l; if (TRdebugLevel > 0) { - char buf[256]; - MPL_snprintf(buf, 256, - "Invalid MALLOC arena detected at line %d in %s\n", lineno, fname); - if (MPL_trvalid(buf)) + if (MPL_trvalid2( "Invalid MALLOC arena detected at line %d in %s\n", + lineno, fname)) goto fn_exit; } nsize = a; if (nsize & TR_ALIGN_MASK) nsize += (TR_ALIGN_BYTES - (nsize & TR_ALIGN_MASK)); - if ((allocated + (long) nsize > TRMaxMemAllow) && TRMaxMemAllow) { + if ((allocated + nsize > TRMaxMemAllow) && TRMaxMemAllow) { /* Return a null when memory would be exhausted */ /* This is only called when additional debugging is enabled, * so the fact that this does not go through the regular error @@ -221,7 +228,7 @@ goto fn_exit; } - new = malloc((unsigned) (nsize + sizeof(TrSPACE) + sizeof(unsigned long))); + new = (char *)malloc(nsize + sizeof(TrSPACE) + sizeof(unsigned long)); if (!new) goto fn_exit; @@ -229,13 +236,17 @@ head = (TRSPACE *) new; new += sizeof(TrSPACE); - if (TRhead) { - MPL_VG_MAKE_MEM_DEFINED(&TRhead->prev, sizeof(TRhead->prev)); - TRhead->prev = head; - MPL_VG_MAKE_MEM_NOACCESS(&TRhead->prev, sizeof(TRhead->prev)); + if (TRhead[0] != TRHEAD_PRESENTINAL || TRhead[2] != TRHEAD_POSTSENTINAL) { + MPL_error_printf("TRhead corrupted - likely memory overwrite.\n"); + goto fn_exit; } - head->next = TRhead; - TRhead = head; + if (TRhead[1]) { + MPL_VG_MAKE_MEM_DEFINED(&TRhead[1]->prev, sizeof(TRhead[1]->prev)); + TRhead[1]->prev = head; + MPL_VG_MAKE_MEM_NOACCESS(&TRhead[1]->prev, sizeof(TRhead[1]->prev)); + } + head->next = TRhead[1]; + TRhead[1] = head; head->prev = 0; head->size = nsize; head->id = TRid; @@ -258,8 +269,8 @@ if (TRlevel & TR_MALLOC) { /* Note that %08p (what we'd like to use) isn't accepted by * all compilers */ - MPL_error_printf("[%d] Allocating %d bytes at %8p in %s:%d\n", - world_rank, a, new, fname, lineno); + MPL_error_printf("[%d] Allocating %ld(%ld) bytes at %8p in %s[%d]\n", + world_rank, (long)a, (long)nsize, new, fname, lineno); } /* Without these macros valgrind actually catches far fewer errors when @@ -284,55 +295,54 @@ void MPL_trfree(void *a_ptr, int line, const char file[]) { TRSPACE *head; - char *ahead; - char *a = (char *) a_ptr; unsigned long *nend; - int l, nset; + size_t nset; + int l; char hexstring[MAX_ADDRESS_CHARS]; /* Don't try to handle empty blocks */ - if (!a) + if (!a_ptr) return; if (TRdebugLevel > 0) { - if (MPL_trvalid("Invalid MALLOC arena detected by FREE")) + if (MPL_trvalid2("Invalid MALLOC arena detected by FREE at line %d in %s\n", line, file )) return; } - ahead = a; - a = a - sizeof(TrSPACE); - head = (TRSPACE *) a; + head = (TRSPACE *) ( ((char *)a_ptr) - sizeof(TrSPACE) ); /* We need to mark the memory as defined before performing our own error * checks or valgrind will flag the trfree function as erroneous. The real * free() at the end of this function will mark the whole block as NOACCESS * again. See the corresponding block in the trmalloc function for more * info. */ - MPL_VG_MAKE_MEM_DEFINED(a, sizeof(TrSPACE)); + MPL_VG_MAKE_MEM_DEFINED(head, sizeof(TrSPACE)); if (head->cookie != COOKIE_VALUE) { /* Damaged header */ /* Note that %08p (what we'd like to use) isn't accepted by * all compilers */ - MPL_error_printf("[%d] Block at address %8p is corrupted; cannot free;\n" + addrToHex( a_ptr, hexstring ); + MPL_error_printf("[%d] Block at address %s is corrupted; cannot free;\n" "may be block not allocated with MPL_trmalloc or MALLOC\n" - "called in %s at line %d\n", world_rank, a, file, line); + "called in %s at line %d\n", world_rank, hexstring, file, line); return; } - nend = (unsigned long *) (ahead + head->size); + nend = (unsigned long *) ((char *)a_ptr + head->size); /* Check that nend is properly aligned */ if ((sizeof(long) == 4 && ((long) nend & 0x3) != 0) || (sizeof(long) == 8 && ((long) nend & 0x7) != 0)) { + addrToHex( a_ptr, hexstring ); MPL_error_printf - ("[%d] Block at address %lx is corrupted (invalid address or header)\n" - "called in %s at line %d\n", world_rank, (long) a + sizeof(TrSPACE), file, line); + ("[%d] Block at address %s is corrupted (invalid address or header)\n" + "called in %s at line %d\n", world_rank, hexstring, file, line); return; } MPL_VG_MAKE_MEM_DEFINED(nend, sizeof(*nend)); if (*nend != COOKIE_VALUE) { if (*nend == ALREADY_FREED) { - addrToHex((char *) a + sizeof(TrSPACE), hexstring); + addrToHex(a_ptr, hexstring); if (TRidSet) { MPL_error_printf ("[%d] Block [id=%d(%lu)] at address %s was already freed\n", world_rank, @@ -352,7 +362,7 @@ } else { /* Damaged tail */ - addrToHex(a, hexstring); + addrToHex(a_ptr, hexstring); if (TRidSet) { MPL_error_printf ("[%d] Block [id=%d(%lu)] at address %s is corrupted (probably write past end)\n", @@ -364,8 +374,10 @@ world_rank, hexstring); } head->fname[TR_FNAME_LEN - 1] = 0; /* Just in case */ - MPL_error_printf("[%d] Block allocated in %s[%d]\n", + MPL_error_printf("[%d] Block being freed allocated in %s[%d]\n", world_rank, head->fname, head->lineno); + MPL_error_printf("[%d] Block cookie should be %lx but was %lx\n", + world_rank, (long)COOKIE_VALUE, *nend ); } } /* Mark the location freed */ @@ -383,7 +395,7 @@ MPL_VG_MAKE_MEM_NOACCESS(&head->prev->next, sizeof(head->prev->next)); } else { - TRhead = head->next; + TRhead[1] = head->next; } if (head->next) { @@ -393,8 +405,8 @@ } if (TRlevel & TR_FREE) { - addrToHex((char *) a + sizeof(TrSPACE), hexstring); - MPL_error_printf("[%d] Freeing %lu bytes at %s in %s:%d\n", + addrToHex(a_ptr, hexstring); + MPL_error_printf("[%d] Freeing %lu bytes at %s in %s[%d]\n", world_rank, head->size, hexstring, file, line); } @@ -403,16 +415,21 @@ * help catch access to already freed data */ /* FIXME why do we skip the first few ints? [goodell@] */ - nset = head->size - 2 * sizeof(int); - if (nset > 0) { + /* Answer lost in time. Probably because in some case, the + first few bytes provided useful information in tracking down + a problem. */ + if (head->size > 2*sizeof(int)) { + /* Now that nset is size_t, it might be defined as unsigned, + so we can't compare nset - 2*sizeof(int) against zero */ + nset = head->size - 2 * sizeof(int); /* If an upper layer (like the handle allocation code) ever used the * MPL_VG_MAKE_MEM_NOACCESS macro on part/all of the data we gave * them then our memset will elicit "invalid write" errors from * valgrind. Mark it as accessible but undefined here to prevent this. */ - MPL_VG_MAKE_MEM_UNDEFINED(ahead + 2 * sizeof(int), nset); - memset(ahead + 2 * sizeof(int), TRFreedByte, nset); + MPL_VG_MAKE_MEM_UNDEFINED((char *)a_ptr + 2 * sizeof(int), nset); + memset((char *)a_ptr + 2 * sizeof(int), TRFreedByte, nset); } - free(a); + free(head); } /*+C @@ -442,24 +459,38 @@ No output is generated if there are no problems detected. +*/ -int MPL_trvalid(const char str[]) +int MPL_trvalid( const char str[] ) { + return MPL_trvalid2( str, -1, (const char *)0 ); +} + +int MPL_trvalid2(const char str[], int line, const char file[] ) +{ TRSPACE *head; TRSPACE *next; - char *a; + char *a; unsigned long *nend; int errs = 0; char hexstring[MAX_ADDRESS_CHARS]; - head = TRhead; + if (TRhead[0] != TRHEAD_PRESENTINAL || TRhead[2] != TRHEAD_POSTSENTINAL) { + MPL_error_printf("TRhead corrupted - likely memory overwrite.\n"); + errs++; + goto fn_exit; + } + head = TRhead[1]; while (head) { /* mark defined before accessing head contents */ MPL_VG_MAKE_MEM_DEFINED(head, sizeof(*head)); if (head->cookie != COOKIE_VALUE) { - if (!errs) - MPL_error_printf("%s\n", str); + if (!errs) { + if (line > 0) + MPL_error_printf(str, line, file); + else + MPL_error_printf( "%s\n", str ); + } errs++; - addrToHex(head, hexstring); + addrToHex(head+1, hexstring); MPL_error_printf ("[%d] Block at address %s is corrupted (invalid cookie in head)\n", world_rank, hexstring); @@ -469,16 +500,24 @@ * SEGV or BUS */ goto fn_exit; } - /* FIXME why is this +1 here? */ - a = (char *) (((TrSPACE *) head) + 1); + /* Get the address of the first byte of the memory, which begins + just after the end of the header. We must use the full header + (TrSPACE) rather than the struct with the data (TRSPACE) because + the full header is padded to ensure correct byte alignment with + the data */ + a = (char *)( (TrSPACE *)head + 1 ); nend = (unsigned long *) (a + head->size); /* mark defined before accessing nend contents */ MPL_VG_MAKE_MEM_DEFINED(nend, sizeof(*nend)); if (nend[0] != COOKIE_VALUE) { - if (!errs) - MPL_error_printf("%s\n", str); + if (!errs) { + if (line > 0) + MPL_error_printf(str, line, file); + else + MPL_error_printf( "%s\n", str ); + } errs++; head->fname[TR_FNAME_LEN - 1] = 0; /* Just in case */ addrToHex(a, hexstring); @@ -494,6 +533,8 @@ } MPL_error_printf("[%d] Block allocated in %s[%d]\n", world_rank, head->fname, head->lineno); + MPL_error_printf("[%d] Block cookie should be %lx but was %lx\n", + world_rank, (long)COOKIE_VALUE, *nend ); } /* set both regions back to NOACCESS */ @@ -538,7 +579,11 @@ if (fp == 0) fp = stderr; - head = TRhead; + if (TRhead[0] != TRHEAD_PRESENTINAL || TRhead[2] != TRHEAD_POSTSENTINAL) { + MPL_error_printf("TRhead corrupted - likely memory overwrite.\n"); + return; + } + head = TRhead[1]; while (head) { MPL_VG_MAKE_MEM_DEFINED(head, sizeof(*head)); if (head->id >= minid) { @@ -585,7 +630,7 @@ } static volatile FILE *TRFP = 0; - /*ARGSUSED*/ static void PrintSum(TRINFO ** a, VISIT order, int level) +static void PrintSum(TRINFO ** a, VISIT order, int level) { if (order == postorder || order == leaf) fprintf(TRFP, "[%d]%s[%d] has %d\n", (*a)->id, (*a)->fname, (*a)->lineno, (*a)->size); @@ -611,7 +656,11 @@ if (fp == 0) fp = stderr; root = 0; - head = TRhead; + if (TRhead[0] != TRHEAD_PRESENTINAL || TRhead[2] != TRHEAD_POSTSENTINAL) { + MPL_error_printf("TRhead corrupted - likely memory overwrite.\n"); + return; + } + head = TRhead[1]; key = nspace; while (head) { if (head->id >= minid) { @@ -646,8 +695,8 @@ { if (fp == 0) fp = stderr; - fprintf(fp, "# [%d] The maximum space allocated was %ld bytes [%ld]\n", - world_rank, TRMaxMem, TRMaxMemId); + fprintf(fp, "# [%d] The maximum space allocated was %lu bytes [%d]\n", + world_rank, (unsigned long)TRMaxMem, TRMaxMemId); } #endif @@ -732,19 +781,17 @@ void *MPL_trrealloc(void *p, int size, int lineno, const char fname[]) { void *pnew; - char *pa; - int nsize; + size_t nsize; TRSPACE *head = 0; char hexstring[MAX_ADDRESS_CHARS]; /* We should really use the size of the old block... */ if (p) { - pa = (char *) p; - head = (TRSPACE *) (pa - sizeof(TrSPACE)); + head = (TRSPACE *) ((char *)p - sizeof(TrSPACE)); MPL_VG_MAKE_MEM_DEFINED(head, sizeof(*head)); /* mark defined before accessing contents */ if (head->cookie != COOKIE_VALUE) { /* Damaged header */ - addrToHex(pa, hexstring); + addrToHex(p, hexstring); MPL_error_printf("[%d] Block at address %s is corrupted; cannot realloc;\n" "may be block not allocated with MPL_trmalloc or MALLOC\n", world_rank, hexstring); @@ -766,13 +813,14 @@ if (p && pnew) { nsize = size; - if (head->size < (unsigned long) nsize) - nsize = (int) (head->size); + if (head->size < nsize) + nsize = head->size; memcpy(pnew, p, nsize); MPL_trfree(p, lineno, fname); } /* Re-mark the head as NOACCESS before returning. */ + /* FIXME: Note head is no longer valid after MPL_trfree above */ if (head) { MPL_VG_MAKE_MEM_NOACCESS(head, sizeof(*head)); } @@ -796,9 +844,9 @@ void *MPL_trstrdup(const char *str, int lineno, const char fname[]) { void *p; - unsigned len = (unsigned) strlen(str) + 1; + size_t len = strlen(str) + 1; - p = MPL_trmalloc(len, lineno, (char *) fname); + p = MPL_trmalloc((unsigned)len, lineno, (char *) fname); if (p) { memcpy(p, str, len); } @@ -891,13 +939,17 @@ TRSPACE *head; int cnt; - head = TRhead; + if (TRhead[0] != TRHEAD_PRESENTINAL || TRhead[2] != TRHEAD_POSTSENTINAL) { + MPL_error_printf("TRhead corrupted - likely memory overwrite.\n"); + return; + } + head = TRhead[1]; cnt = 0; while (head) { cnt++; head = head->next; } - TRhead = MPL_trIsort(TRhead, cnt); + TRhead[1] = MPL_trIsort(TRhead[1], cnt); } /* Takes sorted input and dumps as an aggregate */ @@ -909,8 +961,13 @@ if (fp == 0) fp = stderr; + if (TRhead[0] != TRHEAD_PRESENTINAL || TRhead[2] != TRHEAD_POSTSENTINAL) { + MPL_error_printf("TRhead corrupted - likely memory overwrite.\n"); + return; + } + MPL_trSortBlocks(); - head = TRhead; + head = TRhead[1]; cur = 0; while (head) { cur = head->next;

1 0

r10728 - mpich2/trunk/test/mpi/f90/datatype
by gropp＠mcs.anl.gov 08 Dec '12

08 Dec '12

Author: gropp Date: 2012-12-08 08:44:28 -0600 (Sat, 08 Dec 2012) New Revision: 10728 Modified: mpich2/trunk/test/mpi/f90/datatype/ Log: Add missing ignore file Property changes on: mpich2/trunk/test/mpi/f90/datatype ___________________________________________________________________ Modified: svn:ignore - Makefile.in Makefile.sm Makefile testlist typenamef90.f90 packef90.f90 typecntsf90.f90 typem2f90.f90 typesnamef90.f90 typesubf90.f90 gaddressf90.f90 allctypesf90.f90 allctypes90.f90 hindex1f90.f90 hindexed_blockf90.f90 indtype *.gcda *.gcno Makefile.am-stamp Makefile.am + Makefile.in Makefile.sm Makefile testlist typenamef90.f90 typename3f90.f90 packef90.f90 typecntsf90.f90 typem2f90.f90 typesnamef90.f90 typesubf90.f90 gaddressf90.f90 allctypesf90.f90 allctypes90.f90 hindex1f90.f90 hindexed_blockf90.f90 indtype *.gcda *.gcno Makefile.am-stamp Makefile.am

1 0

r10724 - in mpich2/trunk: . src/binding/f77 src/mpi/datatype src/mpid/common/datatype test/mpi/datatype
by gropp＠mcs.anl.gov 07 Dec '12

07 Dec '12

Author: gropp Date: 2012-12-07 16:14:08 -0600 (Fri, 07 Dec 2012) New Revision: 10724 Modified: mpich2/trunk/configure.ac mpich2/trunk/src/binding/f77/buildiface mpich2/trunk/src/mpi/datatype/type_get_name.c mpich2/trunk/src/mpi/datatype/typeutil.c mpich2/trunk/src/mpid/common/datatype/mpid_datatype.h mpich2/trunk/test/mpi/datatype/typename.c Log: Correct implementation of MPI_COUNT, remove MPI_2COMPLEX and MPI_2DOUBLE_COMPLEX (invalid since MPI 1.1), and update generation of Fortran predefined MPI values Modified: mpich2/trunk/configure.ac =================================================================== --- mpich2/trunk/configure.ac 2012-12-07 22:13:22 UTC (rev 10723) +++ mpich2/trunk/configure.ac 2012-12-07 22:14:08 UTC (rev 10724) @@ -4260,10 +4260,23 @@ fi # Get the size for the bsendoverhead -AC_CHECK_SIZEOF(MPIR_Bsend_data_t,128,[ +AC_CHECK_SIZEOF(MPIR_Bsend_data_t,0,[ #define MPI_Datatype int +#ifdef HAVE_STDLIB_H +#include <stdlib.h> +#endif +#ifdef HAVE_STDINT_H +#include <stdint.h> +#endif #include "${master_top_srcdir}/src/include/mpibsend.h"] ) +if test "$ac_cv_sizeof_MPIR_Bsend_data_t" = "0" ; then + AC_MSG_ERROR([Unable to determine the size of MPI_BSEND_OVERHEAD"]) + # In the past, a default of 128 was used (still likely good enough), + # but the autoconf SIZEOF macro has been changed to ignore the second + # argument, so code that depended on the prior defined behavior now + # silently breaks. +fi BSEND_OVERHEAD=$ac_cv_sizeof_MPIR_Bsend_data_t export BSEND_OVERHEAD AC_SUBST(BSEND_OVERHEAD) @@ -5356,16 +5369,16 @@ # Find the size of OPA_ptr_t. This step needs to come after the OPA # configure above in order to get the size OPA_ptr_t evaluated for # this platform. -AC_CHECK_SIZEOF(OPA_ptr_t,-1,[ +AC_CHECK_SIZEOF(OPA_ptr_t,0,[ #include "${master_top_srcdir}/src/openpa/src/opa_primitives.h" pthread_mutex_t *OPA_emulation_lock; ]) if test "$enable_f77" != "yes" ; then # These are Fortran datatypes ONLY. Set to null if no Fortran compiler. + # Removed the invalid 2COMPLEX and 2DOUBLE_COMPLEX for name in CHARACTER INTEGER REAL LOGICAL COMPLEX DOUBLE_PRECISION \ - 2INTEGER 2REAL DOUBLE_COMPLEX 2COMPLEX 2DOUBLE_PRECISION \ - 2DOUBLE_COMPLEX ; do + 2INTEGER 2REAL DOUBLE_COMPLEX 2DOUBLE_PRECISION ; do fullname="MPI_$name" eval $fullname=MPI_DATATYPE_NULL done @@ -5384,8 +5397,8 @@ AC_SUBST(MPI_2REAL) AC_SUBST(MPI_DOUBLE_COMPLEX) AC_SUBST(MPI_2DOUBLE_PRECISION) -AC_SUBST(MPI_2COMPLEX) -AC_SUBST(MPI_2DOUBLE_COMPLEX) +dnl AC_SUBST(MPI_2COMPLEX) +dnl AC_SUBST(MPI_2DOUBLE_COMPLEX) AC_SUBST(MPI_FINT) # If ROMIO was successfully configured, then ROMIO will have exported the @@ -5535,7 +5548,7 @@ # 0x4c000045 is 1275068485 in decimal, add ($MPI_SIZEOF_COUNT * 256) and you get # the decimal equivalent of the hex number -MPI_F77_COUNT=`expr 1275068484 '+' '(' 256 '*' $MPI_SIZEOF_OFFSET ')'` +MPI_F77_COUNT=`expr 1275068485 '+' '(' 256 '*' $MPI_SIZEOF_OFFSET ')'` AC_SUBST([MPI_F77_COUNT]) # ---------------------------------------------------------------------------- Modified: mpich2/trunk/src/binding/f77/buildiface =================================================================== --- mpich2/trunk/src/binding/f77/buildiface 2012-12-07 22:13:22 UTC (rev 10723) +++ mpich2/trunk/src/binding/f77/buildiface 2012-12-07 22:14:08 UTC (rev 10724) @@ -3746,16 +3746,21 @@ foreach $key (IDENT,CONGRUENT,SIMILAR,UNEQUAL) { &print_mpif_int( "MPI_$key" ); } + # Window flavor and model + foreach $key (FLAVOR_CREATE,FLAVOR_ALLOCATE,FLAVOR_DYNAMIC,FLAVOR_SHARED, + SEPARATE, UNIFIED) { + &print_mpif_int( "MPI_WIN_$key" ); + } # Collective operations - foreach $key (MAX, MIN, SUM, PROD, LAND, BAND, LOR, BOR, LXOR, BXOR, MINLOC, MAXLOC, REPLACE ) { + foreach $key (MAX, MIN, SUM, PROD, LAND, BAND, LOR, BOR, LXOR, BXOR, MINLOC, MAXLOC, REPLACE, NO_OP ) { &print_mpif_int( "MPI_$key" ); } # Objects - foreach $key ('COMM_WORLD', 'COMM_SELF', 'GROUP_EMPTY', 'COMM_NULL', 'WIN_NULL', 'FILE_NULL', 'GROUP_NULL', 'OP_NULL', 'DATATYPE_NULL', 'REQUEST_NULL', 'ERRHANDLER_NULL', 'INFO_NULL', ) { + foreach $key ('COMM_WORLD', 'COMM_SELF', 'GROUP_EMPTY', 'COMM_NULL', 'WIN_NULL', 'FILE_NULL', 'GROUP_NULL', 'OP_NULL', 'DATATYPE_NULL', 'REQUEST_NULL', 'ERRHANDLER_NULL', 'INFO_NULL', 'INFO_ENV' ) { &print_mpif_int( "MPI_$key" ); } # Attributes - foreach $key (TAG_UB, HOST, IO, WTIME_IS_GLOBAL, UNIVERSE_SIZE, LASTUSEDCODE, APPNUM, WIN_BASE, WIN_SIZE, WIN_DISP_UNIT ) { + foreach $key (TAG_UB, HOST, IO, WTIME_IS_GLOBAL, UNIVERSE_SIZE, LASTUSEDCODE, APPNUM, WIN_BASE, WIN_SIZE, WIN_DISP_UNIT, WIN_CREATE_FLAVOR, WIN_MODEL ) { # Special cast: The Fortran versions of these attributes have # value 1 greater than the C versions $attrval = $mpidef{"MPI_$key"}; @@ -3768,7 +3773,7 @@ &print_mpif_int( "MPI_$key" ); } # String sizes - # See MPI-2 2.6.2 and 4.12.9; the constants for string lenghts are + # See MPI-2 2.6.2 and 4.12.9; the constants for string lengths are # defined as one less than the C/C++ version # Missing - max processor name! @@ -3805,10 +3810,14 @@ # # Datatypes # These are determined and set at configure time - foreach $key (COMPLEX, DOUBLE_COMPLEX, LOGICAL, REAL, DOUBLE_PRECISION, INTEGER, '2INTEGER', '2COMPLEX', '2DOUBLE_PRECISION', '2REAL', '2DOUBLE_COMPLEX', CHARACTER) { + foreach $key (COMPLEX, DOUBLE_COMPLEX, LOGICAL, REAL, DOUBLE_PRECISION, INTEGER, '2INTEGER', '2DOUBLE_PRECISION', '2REAL', CHARACTER) { print MPIFFD " INTEGER MPI_$key\n"; print MPIFFD " PARAMETER (MPI_$key=\@MPI_$key\@)\n"; } + # 2COMPLEX and 2DOUBLE_COMPLEX were defined in MPI 1 and removed in + # MPI 1.1. Don't define them for the Fortran interface. + #foreach $key ('2COMPLEX', '2DOUBLE_COMPLEX') { + #} # Value of MPI_BYTE from top level configure! $mpidef{"MPI_BYTE"} = hex "0x4c00010d"; foreach $key (BYTE, UB, LB, PACKED) { @@ -4918,9 +4927,12 @@ # Eat until we find the closing right brace $enum_line = $1; while (! ($enum_line =~ /}/)) { - $enum_line .= <MPIFD>; + my $newline = <MPIFD>; + $newline =~ s/\r*\n//; + $enum_line .= $newline; $linecount++; } + print "Handling enum $enum_line...\n" if $debug; # Now process for names and values while (($enum_line =~ /\s*(MPI[X]*_[A-Z_0-9]*)\s*=\s*([a-fx0-9]*)(.*)/)) { $mpidef{$1} = $2; Modified: mpich2/trunk/src/mpi/datatype/type_get_name.c =================================================================== --- mpich2/trunk/src/mpi/datatype/type_get_name.c 2012-12-07 22:13:22 UTC (rev 10723) +++ mpich2/trunk/src/mpi/datatype/type_get_name.c 2012-12-07 22:14:08 UTC (rev 10724) @@ -111,6 +111,7 @@ /* address/offset types */ type_name_entry(MPI_AINT), type_name_entry(MPI_OFFSET), + type_name_entry(MPI_COUNT), { 0, (char *) 0 }, /* Sentinel used to indicate the last element */ }; Modified: mpich2/trunk/src/mpi/datatype/typeutil.c =================================================================== --- mpich2/trunk/src/mpi/datatype/typeutil.c 2012-12-07 22:13:22 UTC (rev 10723) +++ mpich2/trunk/src/mpi/datatype/typeutil.c 2012-12-07 22:14:08 UTC (rev 10724) @@ -73,9 +73,10 @@ MPI_C_DOUBLE_COMPLEX, MPI_C_LONG_DOUBLE_COMPLEX, - /* address/offset types */ + /* address/offset/count types */ MPI_AINT, MPI_OFFSET, + MPI_COUNT, /* Fortran types */ MPI_COMPLEX, @@ -219,7 +220,6 @@ MPID_Datatype *dptr; MPI_Datatype d = MPI_DATATYPE_NULL; static int is_init = 0; - char error_msg[1024]; /* FIXME: This is actually an error, since this routine should only be called once */ @@ -246,14 +246,11 @@ if (dptr < MPID_Datatype_builtin || dptr > MPID_Datatype_builtin + MPID_DATATYPE_N_BUILTIN) { - MPIU_Snprintf(error_msg, 1024, - "%dth builtin datatype handle references invalid memory", - i); mpi_errno = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_FATAL, FCNAME, __LINE__, MPI_ERR_INTERN, - "**fail", "**fail %s", - error_msg); + "**typeinitbadmem", "**typeinitbadmem %d", + i); return mpi_errno; } /* --END ERROR HANDLING-- */ @@ -274,16 +271,11 @@ /* --BEGIN ERROR HANDLING-- */ if (d != -1 && i < sizeof(mpi_dtypes)/sizeof(*mpi_dtypes) && mpi_dtypes[i] != -1) { /* We did not hit the end-of-list */ - /*MPIU_Internal_error_printf( "Did not initialize all of the predefined datatypes (only did first %d)\n", i-1 );*/ - MPIU_Snprintf(error_msg, 1024, - "Did not initialize all of the predefined datatypes (only did first %d)\n", - i-1); - - mpi_errno = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_FATAL, - FCNAME, __LINE__, - MPI_ERR_INTERN, "**fail", - "**fail %s", error_msg); - return mpi_errno; + mpi_errno = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_FATAL, + FCNAME, __LINE__, + MPI_ERR_INTERN, "**typeinitfail", + "**typeinitfail %d", i-1); + return mpi_errno; } /* --END ERROR HANDLING-- */ is_init = 1; Modified: mpich2/trunk/src/mpid/common/datatype/mpid_datatype.h =================================================================== --- mpich2/trunk/src/mpid/common/datatype/mpid_datatype.h 2012-12-07 22:13:22 UTC (rev 10723) +++ mpich2/trunk/src/mpid/common/datatype/mpid_datatype.h 2012-12-07 22:14:08 UTC (rev 10724) @@ -427,7 +427,9 @@ * here. */ /* FIXME calculating this value this way is foolish, we should make this more * automatic and less error prone */ -#define MPID_DATATYPE_N_BUILTIN 68 +/* FIXME: Given that this is relatively static, an adequate alternative is + to provide a check that this value is valid. */ +#define MPID_DATATYPE_N_BUILTIN 69 extern MPID_Datatype MPID_Datatype_builtin[MPID_DATATYPE_N_BUILTIN + 1]; extern MPID_Datatype MPID_Datatype_direct[]; Modified: mpich2/trunk/test/mpi/datatype/typename.c =================================================================== --- mpich2/trunk/test/mpi/datatype/typename.c 2012-12-07 22:13:22 UTC (rev 10723) +++ mpich2/trunk/test/mpi/datatype/typename.c 2012-12-07 22:14:08 UTC (rev 10724) @@ -114,12 +114,12 @@ #if MTEST_HAVE_MIN_MPI_VERSION(2,2) /* added in MPI-2.2 */ { MPI_C_LONG_DOUBLE_COMPLEX, "MPI_C_LONG_DOUBLE_COMPLEX" }, + { MPI_AINT, "MPI_AINT" }, + { MPI_OFFSET, "MPI_OFFSET" }, #endif #if MTEST_HAVE_MIN_MPI_VERSION(3,0) /* added in MPI 3 */ { MPI_COUNT, "MPI_COUNT" }, - { MPI_AINT, "MPI_AINT" }, - { MPI_OFFSET, "MPI_OFFSET" }, #endif { 0, (char *)0 }, /* Sentinal used to indicate the last element */ };

1 0

r10711 - mpich2/trunk/test/mpi
by gropp＠mcs.anl.gov 07 Dec '12

07 Dec '12

Author: gropp Date: 2012-12-07 10:18:51 -0600 (Fri, 07 Dec 2012) New Revision: 10711 Modified: mpich2/trunk/test/mpi/configure.ac Log: Added info about the -qfree=f90 option for xlf Modified: mpich2/trunk/test/mpi/configure.ac =================================================================== --- mpich2/trunk/test/mpi/configure.ac 2012-12-06 17:10:21 UTC (rev 10710) +++ mpich2/trunk/test/mpi/configure.ac 2012-12-07 16:18:51 UTC (rev 10711) @@ -984,6 +984,12 @@ # The Fortran90 tests rely on free-form input which needs to be tested # before any test that may modify FCFLAGS, e.g. the cray-pointer test. # The order of the tests is important in compiler like g95. + # Recent experience showed that the IBM xlf compiler, at least on + # some systems, requires -qfree=f90 instead of -qfree . At this + # writing (11/27/12), this Autoconf macro still uses -qfree and has + # no mechanism for extension. This test may fail in that case; if + # you encounter a problem, document it and submit it to the autoconf + # bug list, not the MPICH bug list. AC_FC_FREEFORM # See if the compiler supports the Cray-style pointers

1 0

r10710 - mpich2/trunk/test/mpi/datatype
by gropp＠mcs.anl.gov 06 Dec '12

06 Dec '12

Author: gropp Date: 2012-12-06 11:10:21 -0600 (Thu, 06 Dec 2012) New Revision: 10710 Modified: mpich2/trunk/test/mpi/datatype/typename.c Log: Added tests for MPI-3 predefined datatypes in C Modified: mpich2/trunk/test/mpi/datatype/typename.c =================================================================== --- mpich2/trunk/test/mpi/datatype/typename.c 2012-12-06 17:08:59 UTC (rev 10709) +++ mpich2/trunk/test/mpi/datatype/typename.c 2012-12-06 17:10:21 UTC (rev 10710) @@ -115,6 +115,12 @@ /* added in MPI-2.2 */ { MPI_C_LONG_DOUBLE_COMPLEX, "MPI_C_LONG_DOUBLE_COMPLEX" }, #endif +#if MTEST_HAVE_MIN_MPI_VERSION(3,0) + /* added in MPI 3 */ + { MPI_COUNT, "MPI_COUNT" }, + { MPI_AINT, "MPI_AINT" }, + { MPI_OFFSET, "MPI_OFFSET" }, +#endif { 0, (char *)0 }, /* Sentinal used to indicate the last element */ };

1 0

r10709 - mpich2/trunk/test/mpi/f77/datatype
by gropp＠mcs.anl.gov 06 Dec '12

06 Dec '12

Author: gropp Date: 2012-12-06 11:08:59 -0600 (Thu, 06 Dec 2012) New Revision: 10709 Added: mpich2/trunk/test/mpi/f77/datatype/typename3f.f Modified: mpich2/trunk/test/mpi/f77/datatype/Makefile.am mpich2/trunk/test/mpi/f77/datatype/testlist Log: Added tests for MPI-3 predefined datatypes in Fortran Modified: mpich2/trunk/test/mpi/f77/datatype/Makefile.am =================================================================== --- mpich2/trunk/test/mpi/f77/datatype/Makefile.am 2012-11-30 22:01:01 UTC (rev 10708) +++ mpich2/trunk/test/mpi/f77/datatype/Makefile.am 2012-12-06 17:08:59 UTC (rev 10709) @@ -13,7 +13,7 @@ AM_DEFAULT_SOURCE_EXT = .f noinst_PROGRAMS = typenamef typesnamef typecntsf typesubf typem2f gaddressf \ - packef allctypesf hindex1f hindexed_blockf + packef allctypesf hindex1f hindexed_blockf typename3f ## typeaints.h will be distributed because it's listed in AC_CONFIG_FILES/AC_OUTPUT Modified: mpich2/trunk/test/mpi/f77/datatype/testlist =================================================================== --- mpich2/trunk/test/mpi/f77/datatype/testlist 2012-11-30 22:01:01 UTC (rev 10708) +++ mpich2/trunk/test/mpi/f77/datatype/testlist 2012-12-06 17:08:59 UTC (rev 10709) @@ -1,4 +1,5 @@ typenamef 1 +typename3f 1 mpiversion=3.0 typesnamef 1 typecntsf 1 typem2f 1 Added: mpich2/trunk/test/mpi/f77/datatype/typename3f.f =================================================================== --- mpich2/trunk/test/mpi/f77/datatype/typename3f.f (rev 0) +++ mpich2/trunk/test/mpi/f77/datatype/typename3f.f 2012-12-06 17:08:59 UTC (rev 10709) @@ -0,0 +1,41 @@ +C -*- Mode: Fortran; -*- +C +C +C (C) 2012 by Argonne National Laboratory. +C See COPYRIGHT in top-level directory. +C + program main + implicit none + include 'mpif.h' + character*(MPI_MAX_OBJECT_NAME) name + integer namelen + integer ierr, errs + + errs = 0 + + call mtest_init( ierr ) +C +C Check each Fortran datatype, including the size-specific ones +C See the C version (typename.c) for the relevant MPI sections + + call MPI_Type_get_name( MPI_AINT, name, namelen, ierr ) + if (name(1:namelen) .ne. "MPI_AINT") then + errs = errs + 1 + print *, "Expected MPI_AINT but got "//name(1:namelen) + endif + + call MPI_Type_get_name( MPI_OFFSET, name, namelen, ierr ) + if (name(1:namelen) .ne. "MPI_OFFSET") then + errs = errs + 1 + print *, "Expected MPI_OFFSET but got "//name(1:namelen) + endif + + call MPI_Type_get_name( MPI_COUNT, name, namelen, ierr ) + if (name(1:namelen) .ne. "MPI_COUNT") then + errs = errs + 1 + print *, "Expected MPI_COUNT but got "//name(1:namelen) + endif + + call mtest_finalize( errs ) + call MPI_Finalize( ierr ) + end

1 0

r10701 - in mpich2/trunk/src/pm/hydra: tools/demux ui/mpich
by balaji＠mcs.anl.gov 30 Nov '12

30 Nov '12

Author: balaji Date: 2012-11-30 02:50:40 -0600 (Fri, 30 Nov 2012) New Revision: 10701 Modified: mpich2/trunk/src/pm/hydra/tools/demux/demux.c mpich2/trunk/src/pm/hydra/ui/mpich/mpiexec.c Log: Check for STDIN validity before any other sockets are opened. This is to cover the case where an external script closed STDIN before calling mpiexec. No reviewer. Modified: mpich2/trunk/src/pm/hydra/tools/demux/demux.c =================================================================== --- mpich2/trunk/src/pm/hydra/tools/demux/demux.c 2012-11-30 08:50:33 UTC (rev 10700) +++ mpich2/trunk/src/pm/hydra/tools/demux/demux.c 2012-11-30 08:50:40 UTC (rev 10701) @@ -12,6 +12,7 @@ struct HYDT_dmxu_fns HYDT_dmxu_fns = { 0 }; static int got_sigttin = 0; +static int stdin_valid; #if defined(SIGTTIN) static void signal_cb(int sig) @@ -63,6 +64,9 @@ "cannot find an appropriate demux engine\n"); } + status = HYDT_dmxu_fns.stdin_valid(&stdin_valid); + HYDU_ERR_POP(status, "error checking for stdin validity\n"); + fn_exit: HYDU_FUNC_EXIT(); return status; @@ -277,5 +281,7 @@ HYD_status HYDT_dmx_stdin_valid(int *out) { - return HYDT_dmxu_fns.stdin_valid(out); + *out = stdin_valid; + + return HYD_SUCCESS; } Modified: mpich2/trunk/src/pm/hydra/ui/mpich/mpiexec.c =================================================================== --- mpich2/trunk/src/pm/hydra/ui/mpich/mpiexec.c 2012-11-30 08:50:33 UTC (rev 10700) +++ mpich2/trunk/src/pm/hydra/ui/mpich/mpiexec.c 2012-11-30 08:50:40 UTC (rev 10701) @@ -144,18 +144,13 @@ status = HYDU_set_common_signals(signal_cb); HYDU_ERR_POP(status, "unable to set signal\n"); - if (pipe(HYD_server_info.cmd_pipe) < 0) - HYDU_ERR_SETANDJUMP(status, HYD_INTERNAL_ERROR, "pipe error\n"); - status = HYDT_ftb_init(); HYDU_ERR_POP(status, "unable to initialize FTB\n"); - /* Get user preferences */ status = HYD_uii_mpx_get_parameters(argv); HYDU_ERR_POP(status, "error parsing parameters\n"); - /* Now we initialize engines that require us to know user * preferences */ #if HAVE_ALARM @@ -163,6 +158,10 @@ alarm(HYD_ui_mpich_info.ckpoint_int); #endif /* HAVE_ALARM */ + /* The demux engine should be initialized before any sockets are + * created, since it checks for STDIN's validity. If STDIN was + * closed and we opened a socket that got the same fd as STDIN, + * this test will not be possible. */ status = HYDT_dmx_init(&HYD_server_info.user_global.demux); HYDU_ERR_POP(status, "unable to initialize the demux engine\n"); @@ -318,6 +317,10 @@ HYD_server_info.stdout_cb = HYD_uiu_stdout_cb; HYD_server_info.stderr_cb = HYD_uiu_stderr_cb; + /* Create a pipe connection to wake up the process manager */ + if (pipe(HYD_server_info.cmd_pipe) < 0) + HYDU_ERR_SETANDJUMP(status, HYD_INTERNAL_ERROR, "pipe error\n"); + /* Launch the processes */ status = HYD_pmci_launch_procs(); HYDU_ERR_POP(status, "process manager returned error launching processes\n");

1 0

r10700 - mpich2/trunk/src/pm/hydra/tools/demux
by balaji＠mcs.anl.gov 30 Nov '12

30 Nov '12

Author: balaji Date: 2012-11-30 02:50:33 -0600 (Fri, 30 Nov 2012) New Revision: 10700 Modified: mpich2/trunk/src/pm/hydra/tools/demux/demux.c Log: Whether mpiexec is attached to a terminal does not make any difference for our stdin strategy. Remove that test. No reviewer. Modified: mpich2/trunk/src/pm/hydra/tools/demux/demux.c =================================================================== --- mpich2/trunk/src/pm/hydra/tools/demux/demux.c 2012-11-30 07:49:00 UTC (rev 10699) +++ mpich2/trunk/src/pm/hydra/tools/demux/demux.c 2012-11-30 08:50:33 UTC (rev 10700) @@ -13,7 +13,7 @@ static int got_sigttin = 0; -#if defined(SIGTTIN) && defined(HAVE_ISATTY) +#if defined(SIGTTIN) static void signal_cb(int sig) { HYDU_FUNC_ENTER(); @@ -25,7 +25,7 @@ HYDU_FUNC_EXIT(); return; } -#endif /* SIGTTIN and HAVE_ISATTY */ +#endif /* SIGTTIN */ HYD_status HYDT_dmx_init(char **demux) { @@ -228,25 +228,44 @@ * that, we catch SIGTTIN and ignore it. But that causes the * read() call to return an error (with errno == EINTR) when we * are not attached to the terminal. */ -#if defined(SIGTTIN) && defined(HAVE_ISATTY) - if (isatty(STDIN_FILENO)) { - status = HYDU_set_signal(SIGTTIN, signal_cb); - HYDU_ERR_POP(status, "unable to set SIGTTIN\n"); - } -#endif /* SIGTTIN and HAVE_ISATTY */ + /* + * We need to allow for the following cases: + * + * 1. mpiexec -n 2 ./foo --> type something on the terminal + * Attached to terminal, and can read from stdin + * + * 2. mpiexec -n 2 ./foo & + * Attached to terminal, but cannot read from stdin + * + * 3. mpiexec -n 2 ./foo < bar + * Not attached to terminal, but can read from stdin + * + * 4. mpiexec -n 2 ./foo < /dev/null + * Not attached to terminal, and can read from stdin + * + * 5. mpiexec -n 2 ./foo < bar & + * Not attached to terminal, but can read from stdin + * + * 6. mpiexec -n 2 ./foo < /dev/null & + * Not attached to terminal, and can read from stdin + */ + +#if defined(SIGTTIN) + status = HYDU_set_signal(SIGTTIN, signal_cb); + HYDU_ERR_POP(status, "unable to set SIGTTIN\n"); +#endif /* SIGTTIN */ + ret = read(STDIN_FILENO, NULL, 0); - if (ret < 0 && errno == EINTR && got_sigttin) + if (ret == 0) + *out = 1; + else *out = 0; - else - *out = 1; -#if defined(SIGTTIN) && defined(HAVE_ISATTY) - if (isatty(STDIN_FILENO)) { - status = HYDU_set_signal(SIGTTIN, SIG_IGN); - HYDU_ERR_POP(status, "unable to set SIGTTIN\n"); - } -#endif /* SIGTTIN and HAVE_ISATTY */ +#if defined(SIGTTIN) + status = HYDU_set_signal(SIGTTIN, SIG_IGN); + HYDU_ERR_POP(status, "unable to set SIGTTIN\n"); +#endif /* SIGTTIN */ fn_exit: HYDU_FUNC_EXIT();

1 0

r10699 - mpich2/trunk/src/pm/hydra/tools/demux
by balaji＠mcs.anl.gov 30 Nov '12

30 Nov '12

Author: balaji Date: 2012-11-30 01:49:00 -0600 (Fri, 30 Nov 2012) New Revision: 10699 Modified: mpich2/trunk/src/pm/hydra/tools/demux/demux.c Log: Improve the check for a valid stdin. No reviewer. Modified: mpich2/trunk/src/pm/hydra/tools/demux/demux.c =================================================================== --- mpich2/trunk/src/pm/hydra/tools/demux/demux.c 2012-11-30 07:48:57 UTC (rev 10698) +++ mpich2/trunk/src/pm/hydra/tools/demux/demux.c 2012-11-30 07:49:00 UTC (rev 10699) @@ -235,19 +235,7 @@ } #endif /* SIGTTIN and HAVE_ISATTY */ - /* FIXME: The below read() check for STDIN validity is somehow - * interfering in the case where the application is run in the - * background, but expects some STDIN. The correct behavior is to - * close the STDIN socket for the application in that - * case. However, this seems to hang. I'm still investigating this - * case. In the meanwhile, I have commented out this check. PLEASE - * DO NOT DELETE. All other cases dealing with STDIN seem to be - * working correctly. */ -#if 0 ret = read(STDIN_FILENO, NULL, 0); -#else - ret = 0; -#endif if (ret < 0 && errno == EINTR && got_sigttin) *out = 0; else

1 0

r10698 - in mpich2/trunk/src/pm/hydra: include pm/pmiserv utils/sock
by balaji＠mcs.anl.gov 30 Nov '12

30 Nov '12

Author: balaji Date: 2012-11-30 01:48:57 -0600 (Fri, 30 Nov 2012) New Revision: 10698 Modified: mpich2/trunk/src/pm/hydra/include/hydra.h mpich2/trunk/src/pm/hydra/pm/pmiserv/pmip_cb.c mpich2/trunk/src/pm/hydra/utils/sock/sock.c Log: Our stdin handling logic is not correct. I don't see a good way to fix this, unless we add some form of flow-control logic between the proxy and mpiexec. For the time being, at least throw an error when we can't keep up and are giving up, instead of hanging. No reviewer. Modified: mpich2/trunk/src/pm/hydra/include/hydra.h =================================================================== --- mpich2/trunk/src/pm/hydra/include/hydra.h 2012-11-30 07:48:54 UTC (rev 10697) +++ mpich2/trunk/src/pm/hydra/include/hydra.h 2012-11-30 07:48:57 UTC (rev 10698) @@ -535,6 +535,7 @@ enum HYDU_sock_comm_flag flag); HYD_status HYDU_sock_write(int fd, const void *buf, int maxlen, int *sent, int *closed, enum HYDU_sock_comm_flag flag); +HYD_status HYDU_sock_set_nonblock(int fd); HYD_status HYDU_sock_forward_stdio(int in, int out, int *closed); HYD_status HYDU_sock_get_iface_ip(char *iface, char **ip); HYD_status HYDU_sock_is_local(char *host, int *is_local); Modified: mpich2/trunk/src/pm/hydra/pm/pmiserv/pmip_cb.c =================================================================== --- mpich2/trunk/src/pm/hydra/pm/pmiserv/pmip_cb.c 2012-11-30 07:48:54 UTC (rev 10697) +++ mpich2/trunk/src/pm/hydra/pm/pmiserv/pmip_cb.c 2012-11-30 07:48:57 UTC (rev 10698) @@ -711,6 +711,9 @@ process_id); HYDU_ERR_POP(status, "create process returned error\n"); + status = HYDU_sock_set_nonblock(HYD_pmcd_pmip.downstream.in); + HYDU_ERR_POP(status, "unable to set stdin socket to non-blocking\n"); + HYDU_free_strlist(client_args); if (pmi_fds[1] != HYD_FD_UNSET) { @@ -920,9 +923,6 @@ int count; if (hdr.buflen) { - if (HYD_pmcd_pmip.downstream.in == HYD_FD_CLOSED) - goto fn_exit; - HYDU_MALLOC(buf, char *, hdr.buflen, status); HYDU_ERR_POP(status, "unable to allocate memory\n"); @@ -931,10 +931,20 @@ HYDU_ERR_POP(status, "unable to read from control socket\n"); HYDU_ASSERT(!closed, status); + if (HYD_pmcd_pmip.downstream.in == HYD_FD_CLOSED) { + HYDU_FREE(buf); + goto fn_exit; + } + status = HYDU_sock_write(HYD_pmcd_pmip.downstream.in, buf, hdr.buflen, &count, - &closed, HYDU_SOCK_COMM_MSGWAIT); + &closed, HYDU_SOCK_COMM_NONE); HYDU_ERR_POP(status, "unable to write to downstream stdin\n"); + HYDU_ERR_CHKANDJUMP(status, count != hdr.buflen, HYD_INTERNAL_ERROR, + "process reading stdin too slowly; can't keep up\n"); + + HYDU_ASSERT(count == hdr.buflen, status); + if (HYD_pmcd_pmip.user_global.auto_cleanup) { HYDU_ASSERT(!closed, status); } Modified: mpich2/trunk/src/pm/hydra/utils/sock/sock.c =================================================================== --- mpich2/trunk/src/pm/hydra/utils/sock/sock.c 2012-11-30 07:48:54 UTC (rev 10697) +++ mpich2/trunk/src/pm/hydra/utils/sock/sock.c 2012-11-30 07:48:57 UTC (rev 10698) @@ -306,7 +306,7 @@ goto fn_exit; } -static HYD_status set_nonblock(int fd) +HYD_status HYDU_sock_set_nonblock(int fd) { int flags; HYD_status status = HYD_SUCCESS; @@ -343,10 +343,10 @@ (*fwd_hash)->next = NULL; - status = set_nonblock(in); + status = HYDU_sock_set_nonblock(in); HYDU_ERR_POP(status, "unable to set out-socket to non-blocking\n"); - status = set_nonblock(out); + status = HYDU_sock_set_nonblock(out); HYDU_ERR_POP(status, "unable to set out-socket to non-blocking\n"); fn_exit:

1 0