static void gpfs_free_all_locks(int fd)
{
int rc;
struct {
gpfsFcntlHeader_t header;
gpfsFreeRange_t release;
} release_all;
release_all.header.totalLength = sizeof(release_all);
release_all.header.fcntlVersion = GPFS_FCNTL_CURRENT_VERSION;
release_all.header.fcntlReserved = 0;
release_all.release.structLen = sizeof(release_all.release);
release_all.release.structType = GPFS_FREE_RANGE;
release_all.release.start = 0;
release_all.release.length = 0;
rc = gpfs_fcntl(fd, &release_all);
if (rc != 0) {
DBGV_FPRINTF(stderr,"GPFS fcntl release failed with rc=%d, errno=%d\n",
rc,errno);
}
}
static void scaleable_stat(ADIO_File fd)
{
struct stat64 bg_stat;
struct statfs bg_statfs;
int rank, rc;
char * dir;
long buf[2];
MPI_Comm_rank(fd->comm, &rank);
if (rank == fd->hints->ranklist[0]) {
/* Get the (real) underlying file system block size */
rc = stat64(fd->filename, &bg_stat);
if (rc >= 0)
{
buf[0] = bg_stat.st_blksize;
DBGV_FPRINTF(stderr,"Successful stat '%s'. Blocksize=%ld\n",
fd->filename,bg_stat.st_blksize);
}
else
{
DBGV_FPRINTF(stderr,"Stat '%s' failed with rc=%d, errno=%d\n",
fd->filename,rc,errno);
}
/* Get the (real) underlying file system type so we can
* plan our fsync scaling strategy */
rc = statfs(fd->filename,&bg_statfs);
if (rc >= 0)
{
DBGV_FPRINTF(stderr,"Successful statfs '%s'. Magic number=%#lX\n",
fd->filename,bg_statfs.f_type);
buf[1] = bg_statfs.f_type;
}
else
{
DBGV_FPRINTF(stderr,"Statfs '%s' failed with rc=%d, errno=%d\n",
fd->filename,rc,errno);
ADIO_FileSysType_parentdir(fd->filename, &dir);
rc = statfs(dir,&bg_statfs);
if (rc >= 0)
{
DBGV_FPRINTF(stderr,"Successful statfs '%s'. Magic number=%#lX\n",dir,bg_statfs.f_type);
buf[1] = bg_statfs.f_type;
}
else
{
/* Hmm. Guess we'll assume the worst-case, that it's not GPFS
* or BGLOCKLESSMPIO_F_TYPE (default PVFS2) below */
buf[1] = -1; /* bogus magic number */
DBGV_FPRINTF(stderr,"Statfs '%s' failed with rc=%d, errno=%d\n",dir,rc,errno);
}
free(dir);
}
}
/* now we can broadcast the stat/statfs data to everyone else */
if (fd->comm != MPI_COMM_SELF) { /* if indep open, there's no one to talk to*/
if (fd->agg_comm != MPI_COMM_NULL) /* deferred open: only a subset of
processes participate */
MPI_Bcast(buf, 2, MPI_LONG, fd->hints->ranklist[0], fd->agg_comm);
else
MPI_Bcast(buf, 2, MPI_LONG, fd->hints->ranklist[0], fd->comm);
}
bg_stat.st_blksize = buf[0];
bg_statfs.f_type = buf[1];
/* data from stat64 */
/* store the blksize in the file system specific storage */
((ADIOI_BG_fs*)fd->fs_ptr)->blksize = bg_stat.st_blksize;
/* data from statfs */
if ((bg_statfs.f_type == GPFS_SUPER_MAGIC) ||
(bg_statfs.f_type == bglocklessmpio_f_type))
{
((ADIOI_BG_fs*)fd->fs_ptr)->fsync_aggr =
ADIOI_BG_FSYNC_AGGREGATION_ENABLED;
/* Only one rank is an "fsync aggregator" because only one
* fsync is needed */
if (rank == fd->hints->ranklist[0])
{
((ADIOI_BG_fs*)fd->fs_ptr)->fsync_aggr |=
ADIOI_BG_FSYNC_AGGREGATOR;
DBG_FPRINTF(stderr,"fsync aggregator %d\n",rank);
}
else
; /* aggregation enabled but this rank is not an aggregator*/
}
else
; /* Other filesystems default to no fsync aggregation */
}
void ADIOI_GPFS_Open(ADIO_File fd, int *error_code)
{
int perm, old_mask, amode, rank, rc;
static char myname[] = "ADIOI_GPFS_OPEN";
/* set internal variables for tuning environment variables */
ad_gpfs_get_env_vars();
if (fd->perm == ADIO_PERM_NULL) {
old_mask = umask(022);
umask(old_mask);
perm = old_mask ^ 0666;
}
else perm = fd->perm;
amode = 0;
if (fd->access_mode & ADIO_CREATE)
amode = amode | O_CREAT;
if (fd->access_mode & ADIO_RDONLY)
amode = amode | O_RDONLY;
if (fd->access_mode & ADIO_WRONLY)
amode = amode | O_WRONLY;
if (fd->access_mode & ADIO_RDWR)
amode = amode | O_RDWR;
if (fd->access_mode & ADIO_EXCL)
amode = amode | O_EXCL;
#ifdef ADIOI_MPE_LOGGING
MPE_Log_event(ADIOI_MPE_open_a, 0, NULL);
#endif
fd->fd_sys = open(fd->filename, amode, perm);
#ifdef ADIOI_MPE_LOGGING
MPE_Log_event(ADIOI_MPE_open_b, 0, NULL);
#endif
DBG_FPRINTF(stderr,"open('%s',%#X,%#X) rc=%d, errno=%d\n",fd->filename,amode,perm,fd->fd_sys,errno);
fd->fd_direct = -1;
if (gpfsmpio_devnullio == 1) {
fd->null_fd = open("/dev/null", O_RDWR);
} else {
fd->null_fd = -1;
}
if ((fd->fd_sys != -1) && (fd->access_mode & ADIO_APPEND))
fd->fp_ind = fd->fp_sys_posn = lseek(fd->fd_sys, 0, SEEK_END);
if(fd->fd_sys != -1)
{
fd->blksize = 1048576; /* default to 1M */
#ifdef ADIOI_MPE_LOGGING
MPE_Log_event(ADIOI_MPE_stat_a, 0, NULL);
#endif
/* in this fs-specific routine, we might not be called over entire
* communicator (deferred open). Collect statistics on one process.
* ADIOI_GEN_Opencoll (common-code caller) will take care of the
* broadcast */
MPI_Comm_rank(fd->comm, &rank);
if ((rank == fd->hints->ranklist[0]) || (fd->comm == MPI_COMM_SELF)) {
struct stat64 gpfs_statbuf;
/* Get the (real) underlying file system block size */
rc = stat64(fd->filename, &gpfs_statbuf);
if (rc >= 0)
{
fd->blksize = gpfs_statbuf.st_blksize;
DBGV_FPRINTF(stderr,"Successful stat '%s'. Blocksize=%ld\n",
fd->filename,gpfs_statbuf.st_blksize);
}
else
{
DBGV_FPRINTF(stderr,"Stat '%s' failed with rc=%d, errno=%d\n",
fd->filename,rc,errno);
}
}
/* all other ranks have incorrect fd->blocksize, but ADIOI_GEN_Opencoll
* will take care of that in both standard and deferred-open case */
#ifdef ADIOI_MPE_LOGGING
MPE_Log_event(ADIOI_MPE_stat_b, 0, NULL);
#endif
#ifdef HAVE_GPFS_FCNTL_H
/* in parallel workload, might be helpful to immediately release block
* tokens. Or, system call overhead will outweigh any benefits... */
if (getenv("ROMIO_GPFS_FREE_LOCKS")!=NULL)
gpfs_free_all_locks(fd->fd_sys);
#endif
}
if (fd->fd_sys == -1) {
*error_code = ADIOI_Err_create_code(myname, fd->filename, errno);
}
else *error_code = MPI_SUCCESS;
}
