/**
* CHECK: On Altix we are forced to use SysV as shmalloc is collective. We
* may use a preallocated shmalloc memory, however, it may NOT still solve
* our problem...
* NOTE: "int memflg" option for future optimiztions.
*/
void PARMCI_Memget(size_t bytes, armci_meminfo_t *meminfo, int memflg) {
void *myptr=NULL;
void *armci_ptr=NULL; /* legal ARCMI ptr used in ARMCI data xfer ops*/
size_t size = bytes;
if(size<=0) armci_die("PARMCI_Memget: size must be > 0", (int)size);
if(meminfo==NULL) armci_die("PARMCI_Memget: Invalid arg #2 (NULL ptr)",0);
if(memflg!=0) armci_die("PARMCI_Memget: Invalid memflg", memflg);
if( !ARMCI_Uses_shm() )
{
/* fill the meminfo structure */
meminfo->armci_addr = armci_ptr;
meminfo->addr = myptr;
meminfo->size = size;
meminfo->cpid = armci_me;
/* meminfo->attr = NULL; */
}
else
{
armci_shmem_memget(meminfo, size);
}
if(DEBUG_){
printf("%d: PARMCI_Memget: addresses server=%p myptr=%p bytes=%ld\n",
armci_me, meminfo->armci_addr, meminfo->addr, bytes);
fflush(stdout);
}
}
void* PARMCI_Memat(armci_meminfo_t *meminfo, long memflg) {
void *ptr=NULL;
if(meminfo==NULL) armci_die("PARMCI_Memget: Invalid arg #2 (NULL ptr)",0);
if(memflg!=0) armci_die("PARMCI_Memget: Invalid memflg", memflg);
if(meminfo->cpid==armci_me) { ptr = meminfo->addr; return ptr; }
if( !ARMCI_Uses_shm())
{
ptr = meminfo->addr;
}
else
{
ptr = armci_shmem_memat(meminfo);
}
if(DEBUG_)
{
printf("%d:PARMCI_Memat: attached addr mptr=%p size=%ld\n",
armci_me, ptr, meminfo->size); fflush(stdout);
}
return ptr;
}
void ARMCI_Memctl(armci_meminfo_t *meminfo) {
if(meminfo==NULL) armci_die("PARMCI_Memget: Invalid arg #2 (NULL ptr)",0);
/* only the creator can delete the segment */
if(meminfo->cpid == armci_me)
{
if( !ARMCI_Uses_shm() )
{
void *ptr = meminfo->addr;
}
else
{
armci_shmem_memctl(meminfo);
}
}
meminfo->addr = NULL;
meminfo->armci_addr = NULL;
/* if(meminfo->attr!=NULL) free(meminfo->attr); */
}
/*\ acquire exclusive LOCK to MEMORY area <pstart,pend> owned by process "proc"
* . only one area can be locked at a time by the calling process
* . must unlock it with armci_unlockmem
\*/
void armci_lockmem(void *start, void *end, int proc)
{
#ifdef ARMCIX
ARMCIX_Lockmem (start, end, proc);
#else
register void* pstart, *pend;
register int slot, avail=0;
int turn=0, conflict=0;
memlock_t *memlock_table;
#if defined(CLUSTER) && !defined(SGIALTIX)
int lock = (proc-armci_clus_info[armci_clus_id(proc)].master)%NUM_LOCKS;
#else
int lock = 0;
#endif
#ifdef CORRECT_PTR
if(! *armci_use_memlock_table) {
/* if offset invalid, use dumb locking scheme ignoring addresses */
armci_lockmem_(start, end, proc);
return;
}
# ifndef SGIALTIX
/* when processes are attached to a shmem region at different addresses,
* addresses written to memlock table must be adjusted to the node master
*/
if(armci_mem_offset) {
start = armci_mem_offset + (char*)start;
end = armci_mem_offset + (char*)end;
}
# endif
#endif
if(DEBUG_) {
printf("%d: calling armci_lockmem for %d range %p -%p\n",
armci_me, proc, start,end);
fflush(stdout);
}
memlock_table = (memlock_t*)memlock_table_array[proc];
#ifdef ALIGN_ADDRESS
/* align address range on cache line boundary to avoid false sharing */
pstart = ALIGN_ADDRESS(start);
pend = CALGN -1 + ALIGN_ADDRESS(end);
#else
pstart=start;
pend =end;
#endif
#ifdef CRAY_SHMEM
{ /* adjust according the remote process raw address */
long bytes = (long) ((char*)pend-(char*)pstart);
extern void* armci_shmalloc_remote_addr(void *ptr, int proc);
pstart = armci_shmalloc_remote_addr(pstart, proc);
pend = (char*)pstart + bytes;
}
#endif
#ifdef SGIALTIX
if (proc == armci_me) {
pstart = shmem_ptr(pstart,armci_me);
pend = shmem_ptr(pend,armci_me);
}
/* In SGI Altix processes are attached to a shmem region at different
addresses. Addresses written to memlock table must be adjusted to
the node master
*/
if(ARMCI_Uses_shm()) {
int i, seg_id=-1;
size_t tile_size,offset;
void *start_addr, *end_addr;
for(i=0; i<seg_count; i++) {
tile_size = armci_memoffset_table[i].tile_size;
start_addr = (void*) ((char*)armci_memoffset_table[i].seg_addr +
proc*tile_size);
end_addr = (void*) ((char*)start_addr +
armci_memoffset_table[i].seg_size);
/* CHECK: because of too much "span" in armci_lockmem_patch in
* strided.c, it is not possible to have condition as (commented):*/
/*if(pstart>=start_addr && pend<=end_addr) {seg_id=i; break;}*/
if(pstart >= start_addr && pstart <= end_addr) {
seg_id=i;
break;
}
}
if(seg_id==-1) armci_die("armci_lockmem: Invalid segment", seg_id);
offset = armci_memoffset_table[seg_id].mem_offset;
pstart = ((char*)pstart + offset);
pend = ((char*)pend + offset);
}
#endif
while(1) {
NATIVE_LOCK(lock,proc);
armci_get(memlock_table, table, sizeof(table), proc);
/* armci_copy(memlock_table, table, sizeof(table));*/
/* inspect the table */
conflict = 0;
avail =-1;
for(slot = 0; slot < MAX_SLOTS; slot ++) {
/* nonzero starting address means the slot is occupied */
if(table[slot].start == NULL) {
/* remember a free slot to store address range */
avail = slot;
} else {
/*check for conflict: overlap between stored and current range*/
if( (pstart >= table[slot].start && pstart <= table[slot].end)
|| (pend >= table[slot].start && pend <= table[slot].end) ) {
conflict = 1;
break;
}
/*
printf("%d: locking %ld-%ld (%d) conflict\n",
armci_me, */
}
}
if(avail != -1 && !conflict) break;
NATIVE_UNLOCK(lock,proc);
armci_waitsome( ++turn );
}
/* we got the memory lock: enter address into the table */
table[avail].start = pstart;
table[avail].end = pend;
armci_put(table+avail,memlock_table+avail,sizeof(memlock_t),proc);
FENCE_NODE(proc);
NATIVE_UNLOCK(lock,proc);
locked_slot = avail;
#endif /* ! ARMCIX */
}
