// Code generation
address MethodHandles::generate_method_handle_interpreter_entry(MacroAssembler* _masm,
vmIntrinsics::ID iid) {
const bool not_for_compiler_entry = false; // this is the interpreter entry
assert(is_signature_polymorphic(iid), "expected invoke iid");
if (iid == vmIntrinsics::_invokeGeneric ||
iid == vmIntrinsics::_compiledLambdaForm) {
// Perhaps surprisingly, the symbolic references visible to Java are not directly used.
// They are linked to Java-generated adapters via MethodHandleNatives.linkMethod.
// They all allow an appendix argument.
__ hlt(); // empty stubs make SG sick
return NULL;
}
// rsi/r13: sender SP (must preserve; see prepare_to_jump_from_interpreted)
// rbx: Method*
// rdx: argument locator (parameter slot count, added to rsp)
// rcx: used as temp to hold mh or receiver
// rax, rdi: garbage temps, blown away
Register rdx_argp = rdx; // argument list ptr, live on error paths
Register rax_temp = rax;
Register rcx_mh = rcx; // MH receiver; dies quickly and is recycled
Register rbx_method = rbx; // eventual target of this invocation
// here's where control starts out:
__ align(CodeEntryAlignment);
address entry_point = __ pc();
if (VerifyMethodHandles) {
Label L;
BLOCK_COMMENT("verify_intrinsic_id {");
__ cmpb(Address(rbx_method, Method::intrinsic_id_offset_in_bytes()), (int) iid);
__ jcc(Assembler::equal, L);
if (iid == vmIntrinsics::_linkToVirtual ||
iid == vmIntrinsics::_linkToSpecial) {
// could do this for all kinds, but would explode assembly code size
trace_method_handle(_masm, "bad Method*::intrinsic_id");
}
__ STOP("bad Method*::intrinsic_id");
__ bind(L);
BLOCK_COMMENT("} verify_intrinsic_id");
}
// First task: Find out how big the argument list is.
Address rdx_first_arg_addr;
int ref_kind = signature_polymorphic_intrinsic_ref_kind(iid);
assert(ref_kind != 0 || iid == vmIntrinsics::_invokeBasic, "must be _invokeBasic or a linkTo intrinsic");
if (ref_kind == 0 || MethodHandles::ref_kind_has_receiver(ref_kind)) {
__ movptr(rdx_argp, Address(rbx_method, Method::const_offset()));
__ load_sized_value(rdx_argp,
Address(rdx_argp, ConstMethod::size_of_parameters_offset()),
sizeof(u2), /*is_signed*/ false);
// assert(sizeof(u2) == sizeof(Method::_size_of_parameters), "");
rdx_first_arg_addr = __ argument_address(rdx_argp, -1);
} else {
DEBUG_ONLY(rdx_argp = noreg);
}
if (!is_signature_polymorphic_static(iid)) {
__ movptr(rcx_mh, rdx_first_arg_addr);
DEBUG_ONLY(rdx_argp = noreg);
}
// rdx_first_arg_addr is live!
trace_method_handle_interpreter_entry(_masm, iid);
if (iid == vmIntrinsics::_invokeBasic) {
generate_method_handle_dispatch(_masm, iid, rcx_mh, noreg, not_for_compiler_entry);
} else {
// Adjust argument list by popping the trailing MemberName argument.
Register rcx_recv = noreg;
if (MethodHandles::ref_kind_has_receiver(ref_kind)) {
// Load the receiver (not the MH; the actual MemberName's receiver) up from the interpreter stack.
__ movptr(rcx_recv = rcx, rdx_first_arg_addr);
}
DEBUG_ONLY(rdx_argp = noreg);
Register rbx_member = rbx_method; // MemberName ptr; incoming method ptr is dead now
__ pop(rax_temp); // return address
__ pop(rbx_member); // extract last argument
__ push(rax_temp); // re-push return address
generate_method_handle_dispatch(_masm, iid, rcx_recv, rbx_member, not_for_compiler_entry);
}
return entry_point;
}
void RowScatter
( T alpha,
const ElementalMatrix<T>& A,
ElementalMatrix<T>& B )
{
DEBUG_ONLY(CSE cse("axpy_contract::RowScatter"))
AssertSameGrids( A, B );
if( A.Height() != B.Height() || A.Width() != B.Width() )
LogicError("Matrix sizes did not match");
if( !B.Participating() )
return;
const Int width = B.Width();
const Int colDiff = B.ColAlign()-A.ColAlign();
if( colDiff == 0 )
{
if( width == 1 )
{
const Int localHeight = B.LocalHeight();
const Int portionSize = mpi::Pad( localHeight );
//vector<T> buffer( portionSize );
vector<T> buffer;
buffer.reserve( portionSize );
// Reduce to rowAlign
const Int rowAlign = B.RowAlign();
mpi::Reduce
( A.LockedBuffer(), buffer.data(), portionSize,
rowAlign, B.RowComm() );
if( B.RowRank() == rowAlign )
{
axpy::util::InterleaveMatrixUpdate
( alpha, localHeight, 1,
buffer.data(), 1, localHeight,
B.Buffer(), 1, B.LDim() );
}
}
else
{
const Int rowStride = B.RowStride();
const Int rowAlign = B.RowAlign();
const Int localHeight = B.LocalHeight();
const Int localWidth = B.LocalWidth();
const Int maxLocalWidth = MaxLength(width,rowStride);
const Int portionSize = mpi::Pad( localHeight*maxLocalWidth );
const Int sendSize = rowStride*portionSize;
// Pack
//vector<T> buffer( sendSize );
vector<T> buffer;
buffer.reserve( sendSize );
copy::util::RowStridedPack
( localHeight, width,
rowAlign, rowStride,
A.LockedBuffer(), A.LDim(),
buffer.data(), portionSize );
// Communicate
mpi::ReduceScatter( buffer.data(), portionSize, B.RowComm() );
// Update with our received data
axpy::util::InterleaveMatrixUpdate
( alpha, localHeight, localWidth,
buffer.data(), 1, localHeight,
B.Buffer(), 1, B.LDim() );
}
}
else
{
#ifdef EL_UNALIGNED_WARNINGS
if( B.Grid().Rank() == 0 )
cerr << "Unaligned RowScatter" << endl;
#endif
const Int colRank = B.ColRank();
const Int colStride = B.ColStride();
const Int sendRow = Mod( colRank+colDiff, colStride );
const Int recvRow = Mod( colRank-colDiff, colStride );
const Int localHeight = B.LocalHeight();
const Int localHeightA = A.LocalHeight();
if( width == 1 )
{
//vector<T> buffer( localHeight+localHeightA );
vector<T> buffer;
buffer.reserve( localHeight+localHeightA );
T* sendBuf = &buffer[0];
T* recvBuf = &buffer[localHeightA];
// Reduce to rowAlign
const Int rowAlign = B.RowAlign();
mpi::Reduce
( A.LockedBuffer(), sendBuf, localHeightA, rowAlign, B.RowComm() );
if( B.RowRank() == rowAlign )
{
// Perform the realignment
mpi::SendRecv
( sendBuf, localHeightA, sendRow,
recvBuf, localHeight, recvRow, B.ColComm() );
axpy::util::InterleaveMatrixUpdate
( alpha, localHeight, 1,
recvBuf, 1, localHeight,
B.Buffer(), 1, B.LDim() );
}
}
else
{
const Int rowStride = B.RowStride();
const Int rowAlign = B.RowAlign();
const Int localWidth = B.LocalWidth();
const Int maxLocalWidth = MaxLength(width,rowStride);
const Int recvSize_RS = mpi::Pad( localHeightA*maxLocalWidth );
const Int sendSize_RS = rowStride * recvSize_RS;
const Int recvSize_SR = localHeight * localWidth;
//vector<T> buffer( recvSize_RS + Max(sendSize_RS,recvSize_SR) );
vector<T> buffer;
buffer.reserve( recvSize_RS + Max(sendSize_RS,recvSize_SR) );
T* firstBuf = &buffer[0];
T* secondBuf = &buffer[recvSize_RS];
// Pack
copy::util::RowStridedPack
( localHeightA, width,
rowAlign, rowStride,
A.LockedBuffer(), A.LDim(),
secondBuf, recvSize_RS );
// Reduce-scatter over each process row
mpi::ReduceScatter( secondBuf, firstBuf, recvSize_RS, B.RowComm() );
// Trade reduced data with the appropriate process row
mpi::SendRecv
( firstBuf, localHeightA*localWidth, sendRow,
secondBuf, localHeight*localWidth, recvRow, B.ColComm() );
// Update with our received data
axpy::util::InterleaveMatrixUpdate
( alpha, localHeight, localWidth,
secondBuf, 1, localHeight,
B.Buffer(), 1, B.LDim() );
}
}
}
BDM& BDM::operator=( const DistMatrix<T,STAR,VR,BLOCK>& A )
{
DEBUG_ONLY(CSE cse("[STAR,VC] = [STAR,VR]"))
LogicError("This routine is not yet written");
return *this;
}
inline void
UpdateImagPart( Real& alpha, const Real& beta )
{
DEBUG_ONLY(CallStackEntry cse("UpdateImagPart"))
LogicError("Nonsensical update");
}
DM& DM::operator=( const DistMatrix<T,STAR,MR>& A )
{
DEBUG_ONLY(CSE cse("[MC,MR] = [STAR,MR]"))
copy::ColFilter( A, *this );
return *this;
}
/**
* Add a client to the waiting queue for uploads.
*
* @param client address of the client that should be added to the waiting queue
*
* @param bIgnoreTimelimit don't check time limit to possibly ban the client.
*/
void CUploadQueue::AddClientToQueue(CUpDownClient* client, bool bIgnoreTimelimit)
{
//This is to keep users from abusing the limits we put on lowID callbacks.
//1)Check if we are connected to any network and that we are a lowID.
//(Although this check shouldn't matter as they wouldn't have found us..
// But, maybe I'm missing something, so it's best to check as a precaution.)
//2)Check if the user is connected to Kad. We do allow all Kad Callbacks.
//3)Check if the user is in our download list or a friend..
//We give these users a special pass as they are helping us..
//4)Are we connected to a server? If we are, is the user on the same server?
//TCP lowID callbacks are also allowed..
//5)If the queue is very short, allow anyone in as we want to make sure
//our upload is always used.
if (theApp.IsConnected()
&& theApp.IsFirewalled()
&& !client->GetKadPort()
&& client->GetDownloadState() == DS_NONE
&& !client->IsFriend()
&& theApp.serverconnect
&& !theApp.serverconnect->IsLocalServer(client->GetServerIP(),client->GetServerPort())
&& GetWaitingUserCount() > 50)
return;
client->AddAskedCount();
client->SetLastUpRequest();
if (!bIgnoreTimelimit)
client->AddRequestCount(client->GetUploadFileID());
if (client->IsBanned())
return;
uint16 cSameIP = 0;
// check for double
POSITION pos1, pos2;
for (pos1 = waitinglist.GetHeadPosition();( pos2 = pos1 ) != NULL;)
{
waitinglist.GetNext(pos1);
CUpDownClient* cur_client= waitinglist.GetAt(pos2);
if (cur_client == client)
{
if (client->m_bAddNextConnect && AcceptNewClient(client->m_bAddNextConnect))
{
//Special care is given to lowID clients that missed their upload slot
//due to the saving bandwidth on callbacks.
if(thePrefs.GetLogUlDlEvents())
AddDebugLogLine(true, _T("Adding ****lowid when reconnecting. Client: %s"), client->DbgGetClientInfo());
client->m_bAddNextConnect = false;
RemoveFromWaitingQueue(client, true);
AddUpNextClient(_T("Adding ****lowid when reconnecting."), client);
return;
}
client->SendRankingInfo();
theApp.emuledlg->transferwnd->queuelistctrl.RefreshClient(client);
return;
}
else if ( client->Compare(cur_client) )
{
theApp.clientlist->AddTrackClient(client); // in any case keep track of this client
// another client with same ip:port or hash
// this happens only in rare cases, because same userhash / ip:ports are assigned to the right client on connecting in most cases
if (cur_client->credits != NULL && cur_client->credits->GetCurrentIdentState(cur_client->GetIP()) == IS_IDENTIFIED)
{
//cur_client has a valid secure hash, don't remove him
if (thePrefs.GetVerbose())
AddDebugLogLine(false,CString(GetResString(IDS_SAMEUSERHASH)),client->GetUserName(),cur_client->GetUserName(),client->GetUserName() );
return;
}
if (client->credits != NULL && client->credits->GetCurrentIdentState(client->GetIP()) == IS_IDENTIFIED)
{
//client has a valid secure hash, add him remove other one
if (thePrefs.GetVerbose())
AddDebugLogLine(false,CString(GetResString(IDS_SAMEUSERHASH)),client->GetUserName(),cur_client->GetUserName(),cur_client->GetUserName() );
RemoveFromWaitingQueue(pos2,true);
if (!cur_client->socket)
{
if(cur_client->Disconnected(_T("AddClientToQueue - same userhash 1")))
delete cur_client;
}
}
else
{
// remove both since we do not know who the bad one is
if (thePrefs.GetVerbose())
AddDebugLogLine(false,CString(GetResString(IDS_SAMEUSERHASH)),client->GetUserName(),cur_client->GetUserName(),"Both" );
RemoveFromWaitingQueue(pos2,true);
if (!cur_client->socket)
{
if(cur_client->Disconnected(_T("AddClientToQueue - same userhash 2")))
delete cur_client;
}
return;
}
}
else if (client->GetIP() == cur_client->GetIP())
{
// same IP, different port, different userhash
cSameIP++;
}
}
if (cSameIP >= 3)
{
// do not accept more than 3 clients from the same IP
if (thePrefs.GetVerbose())
DEBUG_ONLY( AddDebugLogLine(false,_T("%s's (%s) request to enter the queue was rejected, because of too many clients with the same IP"), client->GetUserName(), ipstr(client->GetConnectIP())) );
return;
}
else if (theApp.clientlist->GetClientsFromIP(client->GetIP()) >= 3)
{
if (thePrefs.GetVerbose())
DEBUG_ONLY( AddDebugLogLine(false,_T("%s's (%s) request to enter the queue was rejected, because of too many clients with the same IP (found in TrackedClientsList)"), client->GetUserName(), ipstr(client->GetConnectIP())) );
return;
}
// done
// statistic values
CKnownFile* reqfile = theApp.sharedfiles->GetFileByID((uchar*)client->GetUploadFileID());
if (reqfile)
reqfile->statistic.AddRequest();
// emule collection will bypass the queue
if (reqfile != NULL && CCollection::HasCollectionExtention(reqfile->GetFileName()) && reqfile->GetFileSize() < MAXPRIORITYCOLL_SIZE
&& !client->IsDownloading() && client->socket != NULL && client->socket->IsConnected())
{
client->SetCollectionUploadSlot(true);
RemoveFromWaitingQueue(client, true);
AddUpNextClient(_T("Collection Priority Slot"), client);
return;
}
else
client->SetCollectionUploadSlot(false);
// cap the list
// the queue limit in prefs is only a soft limit. Hard limit is 25% higher, to let in powershare clients and other
// high ranking clients after soft limit has been reached
uint32 softQueueLimit = thePrefs.GetQueueSize();
uint32 hardQueueLimit = thePrefs.GetQueueSize() + max(thePrefs.GetQueueSize()/4, 200);
// if soft queue limit has been reached, only let in high ranking clients
if ((uint32)waitinglist.GetCount() >= hardQueueLimit ||
(uint32)waitinglist.GetCount() >= softQueueLimit && // soft queue limit is reached
(client->IsFriend() && client->GetFriendSlot()) == false && // client is not a friend with friend slot
client->GetCombinedFilePrioAndCredit() < GetAverageCombinedFilePrioAndCredit()) { // and client has lower credits/wants lower prio file than average client in queue
// then block client from getting on queue
return;
}
if (client->IsDownloading())
{
// he's already downloading and wants probably only another file
if (thePrefs.GetDebugClientTCPLevel() > 0)
DebugSend("OP__AcceptUploadReq", client);
Packet* packet = new Packet(OP_ACCEPTUPLOADREQ,0);
theStats.AddUpDataOverheadFileRequest(packet->size);
client->socket->SendPacket(packet,true);
return;
}
if (waitinglist.IsEmpty() && AcceptNewClient())
{
AddUpNextClient(_T("Direct add with empty queue."), client);
}
else
{
waitinglist.AddTail(client);
client->SetUploadState(US_ONUPLOADQUEUE);
theApp.emuledlg->transferwnd->queuelistctrl.AddClient(client,true);
theApp.emuledlg->transferwnd->ShowQueueCount(waitinglist.GetCount());
client->SendRankingInfo();
}
}
OS_PAGE_SIZE_DECLARE
uint4 gdsfilext(uint4 blocks, uint4 filesize, boolean_t trans_in_prog)
{
sm_uc_ptr_t old_base[2], mmap_retaddr;
boolean_t was_crit, is_mm;
char buff[DISK_BLOCK_SIZE];
int result, save_errno, status;
uint4 new_bit_maps, bplmap, map, new_blocks, new_total, max_tot_blks;
uint4 jnl_status, to_wait, to_msg, wait_period;
gtm_uint64_t avail_blocks, mmap_sz;
off_t new_eof;
trans_num curr_tn;
unix_db_info *udi;
inctn_opcode_t save_inctn_opcode;
int4 prev_extend_blks_to_upgrd;
jnl_private_control *jpc;
jnl_buffer_ptr_t jbp;
DCL_THREADGBL_ACCESS;
assert(!IS_DSE_IMAGE);
assert((cs_addrs->nl == NULL) || (process_id != cs_addrs->nl->trunc_pid)); /* mu_truncate shouldn't extend file... */
assert(!process_exiting);
DEBUG_ONLY(old_base[0] = old_base[1] = NULL);
assert(!gv_cur_region->read_only);
udi = FILE_INFO(gv_cur_region);
is_mm = (dba_mm == cs_addrs->hdr->acc_meth);
# if !defined(MM_FILE_EXT_OK)
if (!udi->grabbed_access_sem && is_mm)
return (uint4)(NO_FREE_SPACE); /* should this be changed to show extension not allowed ? */
# endif
/* Both blocks and total blocks are unsigned ints so make sure we aren't asking for huge numbers that will
overflow and end up doing silly things.
*/
assert((blocks <= (MAXTOTALBLKS(cs_data) - cs_data->trans_hist.total_blks)) || WBTEST_ENABLED(WBTEST_FILE_EXTEND_ERROR));
if (!blocks)
return (uint4)(NO_FREE_SPACE); /* should this be changed to show extension not enabled ? */
bplmap = cs_data->bplmap;
/* New total of non-bitmap blocks will be number of current, non-bitmap blocks, plus new blocks desired
* There are (bplmap - 1) non-bitmap blocks per bitmap, so add (bplmap - 2) to number of non-bitmap blocks
* and divide by (bplmap - 1) to get total number of bitmaps for expanded database. (must round up in this
* manner as every non-bitmap block must have an associated bitmap)
* Current number of bitmaps is (total number of current blocks + bplmap - 1) / bplmap.
* Subtract current number of bitmaps from number needed for expanded database to get number of new bitmaps needed.
*/
new_bit_maps = DIVIDE_ROUND_UP(cs_data->trans_hist.total_blks
- DIVIDE_ROUND_UP(cs_data->trans_hist.total_blks, bplmap) + blocks, bplmap - 1)
- DIVIDE_ROUND_UP(cs_data->trans_hist.total_blks, bplmap);
new_blocks = blocks + new_bit_maps;
assert(0 < (int)new_blocks);
if (new_blocks + cs_data->trans_hist.total_blks > MAXTOTALBLKS(cs_data))
{
assert(FALSE);
send_msg_csa(CSA_ARG(cs_addrs) VARLSTCNT(1) ERR_TOTALBLKMAX);
return (uint4)(NO_FREE_SPACE);
}
if (0 != (save_errno = disk_block_available(udi->fd, &avail_blocks, FALSE)))
{
send_msg_csa(CSA_ARG(cs_addrs) VARLSTCNT(5) ERR_DBFILERR, 2, DB_LEN_STR(gv_cur_region), save_errno);
rts_error_csa(CSA_ARG(cs_addrs) VARLSTCNT(5) ERR_DBFILERR, 2, DB_LEN_STR(gv_cur_region), save_errno);
} else
{
if (!(gtmDebugLevel & GDL_IgnoreAvailSpace))
{ /* Bypass this space check if debug flag above is on. Allows us to create a large sparce DB
* in space it could never fit it if wasn't sparse. Needed for some tests.
*/
avail_blocks = avail_blocks / (cs_data->blk_size / DISK_BLOCK_SIZE);
if ((blocks * EXTEND_WARNING_FACTOR) > avail_blocks)
{
if (blocks > (uint4)avail_blocks)
{
SETUP_THREADGBL_ACCESS;
if (!ANTICIPATORY_FREEZE_ENABLED(cs_addrs))
return (uint4)(NO_FREE_SPACE);
else
send_msg_csa(CSA_ARG(cs_addrs) VARLSTCNT(6) MAKE_MSG_WARNING(ERR_NOSPACEEXT), 4,
DB_LEN_STR(gv_cur_region), new_blocks, (uint4)avail_blocks);
} else
send_msg_csa(CSA_ARG(cs_addrs) VARLSTCNT(5) ERR_DSKSPACEFLOW, 3, DB_LEN_STR(gv_cur_region),
(uint4)(avail_blocks - ((new_blocks <= avail_blocks) ? new_blocks : 0)));
}
}
}
/* From here on, we need to use GDSFILEXT_CLNUP before returning to the caller */
was_crit = cs_addrs->now_crit;
assert(!cs_addrs->hold_onto_crit || was_crit);
/* If we are coming from mupip_extend (which gets crit itself) we better have waited for any unfreezes to occur.
* If we are coming from online rollback (when that feature is available), we will come in holding crit and in
* the final retry. In that case too, we expect to have waited for unfreezes to occur in the caller itself.
* Therefore if we are coming in holding crit from MUPIP, we expect the db to be unfrozen so no need to wait for
* freeze.
* If we are coming from GT.M and final retry (in which case we come in holding crit) we expect to have waited
* for any unfreezes (by invoking tp_crit_all_regions) to occur (TP or non-TP) before coming into this
* function. However, there is one exception. In the final retry, if tp_crit_all_regions notices that
* at least one of the participating regions did ONLY READs, it will not wait for any freeze on THAT region
* to complete before grabbing crit. Later, in the final retry, if THAT region did an update which caused
* op_tcommit to invoke bm_getfree->gdsfilext, then we would have come here with a frozen region on which
* we hold crit.
*/
assert(!was_crit || !cs_data->freeze || (dollar_tlevel && (CDB_STAGNATE <= t_tries)));
/*
* If we are in the final retry and already hold crit, it is possible that csa->nl->wc_blocked is also set to TRUE
* (by a concurrent process in phase2 which encountered an error in the midst of commit and secshr_db_clnup
* finished the job for it). In this case we do NOT want to invoke wcs_recover as that will update the "bt"
* transaction numbers without correspondingly updating the history transaction numbers (effectively causing
* a cdb_sc_blkmod type of restart). Therefore do NOT call grab_crit (which unconditionally invokes wcs_recover)
* if we already hold crit.
*/
if (!was_crit)
{
for ( ; ; )
{
grab_crit(gv_cur_region);
if (!cs_data->freeze && !IS_REPL_INST_FROZEN)
break;
rel_crit(gv_cur_region);
while (cs_data->freeze || IS_REPL_INST_FROZEN)
hiber_start(1000);
}
} else if ((cs_data->freeze || IS_REPL_INST_FROZEN) && dollar_tlevel)
{ /* We don't want to continue with file extension as explained above. Hence return with an error code which
* op_tcommit will recognize (as a cdb_sc_needcrit type of restart) and restart accordingly.
*/
assert(CDB_STAGNATE <= t_tries);
GDSFILEXT_CLNUP;
return (uint4)(FINAL_RETRY_FREEZE_PROG);
}
assert(cs_addrs->ti->total_blks == cs_data->trans_hist.total_blks);
if (cs_data->trans_hist.total_blks != filesize)
{ /* Somebody else has already extended it, since we are in crit, this is trust-worthy. However, in case of MM,
* we still need to remap the database
*/
assert((cs_data->trans_hist.total_blks > filesize) GTM_TRUNCATE_ONLY( || !is_mm));
/* For BG, someone else could have truncated or extended - we have no idea */
GDSFILEXT_CLNUP;
return (SS_NORMAL);
}
inline void ComputeFactorCommMeta
( DistSymmInfo& info, bool computeFactRecvInds )
{
DEBUG_ONLY(CallStackEntry cse("ComputeFactorCommMeta"))
info.distNodes[0].factorMeta.Empty();
const Int numDist = info.distNodes.size();
for( Int s=1; s<numDist; ++s )
{
DistSymmNodeInfo& node = info.distNodes[s];
const int teamSize = mpi::CommSize( node.comm );
const DistSymmNodeInfo& childNode = info.distNodes[s-1];
// Fill factorMeta.numChildSendInds
FactorCommMeta& commMeta = node.factorMeta;
commMeta.Empty();
const int gridHeight = node.grid->Height();
const int gridWidth = node.grid->Width();
const int childGridHeight = childNode.grid->Height();
const int childGridWidth = childNode.grid->Width();
const int childGridRow = childNode.grid->Row();
const int childGridCol = childNode.grid->Col();
const Int mySize = childNode.size;
const Int updateSize = childNode.lowerStruct.size();
commMeta.numChildSendInds.resize( teamSize );
elem::MemZero( &commMeta.numChildSendInds[0], teamSize );
const std::vector<Int>& myRelInds =
( childNode.onLeft ? node.leftRelInds : node.rightRelInds );
{
const Int colAlign = mySize % childGridHeight;
const Int rowAlign = mySize % childGridWidth;
const Int colShift =
Shift( childGridRow, colAlign, childGridHeight );
const Int rowShift =
Shift( childGridCol, rowAlign, childGridWidth );
const Int localHeight =
Length( updateSize, colShift, childGridHeight );
const Int localWidth =
Length( updateSize, rowShift, childGridWidth );
for( Int jChildLoc=0; jChildLoc<localWidth; ++jChildLoc )
{
const Int jChild = rowShift + jChildLoc*childGridWidth;
const int destGridCol = myRelInds[jChild] % gridWidth;
Int localColShift;
if( colShift > jChild )
localColShift = 0;
else if( (jChild-colShift) % childGridHeight == 0 )
localColShift = (jChild-colShift)/childGridHeight;
else
localColShift = (jChild-colShift)/childGridHeight + 1;
for( Int iChildLoc=localColShift;
iChildLoc<localHeight; ++iChildLoc )
{
const Int iChild = colShift + iChildLoc*childGridHeight;
if( iChild >= jChild )
{
const int destGridRow = myRelInds[iChild] % gridHeight;
const int destRank = destGridRow+destGridCol*gridHeight;
++commMeta.numChildSendInds[destRank];
}
}
}
}
// Optionally compute the recv indices for the factorization.
// This is optional since it requires a nontrivial amount of storage.
if( computeFactRecvInds )
ComputeFactRecvInds( node, childNode );
}
}
/* Note we don't increment fast_lock_count as part of getting the latch and decrement it when releasing it because ROLLBACK
* can hold onto this latch for a long while and can do updates in this duration and we should NOT have a non-zero fast_lock_count
* as many places like t_begin/dsk_read have asserts to this effect. It is okay to NOT increment fast_lock_count as ROLLBACK
* anyways have logic to disable interrupts the moment it starts doing database updates.
*/
boolean_t grab_gtmsource_srv_latch(sm_global_latch_ptr_t latch, uint4 max_timeout_in_secs, uint4 onln_rlbk_action)
{
int spins, maxspins, retries, max_retries;
unix_db_info *udi;
sgmnt_addrs *repl_csa;
boolean_t cycle_mismatch;
assert(!have_crit(CRIT_HAVE_ANY_REG));
udi = FILE_INFO(jnlpool.jnlpool_dummy_reg);
repl_csa = &udi->s_addrs;
maxspins = num_additional_processors ? MAX_LOCK_SPINS(LOCK_SPINS, num_additional_processors) : 1;
max_retries = max_timeout_in_secs * 4 * 1000; /* outer-loop : X minutes, 1 loop in 4 is sleep of 1 ms */
for (retries = max_retries - 1; 0 < retries; retries--)
{
for (spins = maxspins; 0 < spins; spins--)
{
assert(latch->u.parts.latch_pid != process_id); /* We better not hold it if trying to get it */
if (GET_SWAPLOCK(latch))
{
DEBUG_ONLY(locknl = repl_csa->nl); /* Use the journal pool to maintain lock history */
LOCK_HIST("OBTN", latch, process_id, retries);
DEBUG_ONLY(locknl = NULL);
if (jnlpool.repl_inst_filehdr->file_corrupt && !jgbl.onlnrlbk)
{
/* Journal pool indicates an abnormally terminated online rollback. Cannot continue until
* the rollback command is re-run to bring the journal pool/file and instance file to a
* consistent state.
*/
/* No need to release the latch before rts_error (mupip_exit_handler will do it for us) */
rts_error(VARLSTCNT(8) ERR_REPLREQROLLBACK, 2, LEN_AND_STR(udi->fn),
ERR_TEXT, 2, LEN_AND_LIT("file_corrupt field in instance file header is set to"
" TRUE"));
}
cycle_mismatch = (repl_csa->onln_rlbk_cycle != jnlpool.jnlpool_ctl->onln_rlbk_cycle);
assert((ASSERT_NO_ONLINE_ROLLBACK != onln_rlbk_action) || !cycle_mismatch);
if ((HANDLE_CONCUR_ONLINE_ROLLBACK == onln_rlbk_action) && cycle_mismatch)
{
assert(is_src_server);
SYNC_ONLN_RLBK_CYCLES;
gtmsource_onln_rlbk_clnup(); /* side-effect : sets gtmsource_state */
rel_gtmsource_srv_latch(latch);
}
return TRUE;
}
}
if (retries & 0x3)
{ /* On all but every 4th pass, do a simple rel_quant */
rel_quant();
} else
{
/* On every 4th pass, we bide for awhile */
wcs_sleep(LOCK_SLEEP);
if (RETRY_CASLATCH_CUTOFF == (retries % LOCK_TRIES))
performCASLatchCheck(latch, TRUE);
}
}
DUMP_LOCKHIST();
assert(FALSE);
assert(jnlpool.gtmsource_local && jnlpool.gtmsource_local->gtmsource_pid);
rts_error(VARLSTCNT(5) ERR_SRVLCKWT2LNG, 2, max_timeout_in_secs, jnlpool.gtmsource_local->gtmsource_pid);
return FALSE; /* to keep the compiler happy */
}
int send_mesg2gtmsecshr(unsigned int code, unsigned int id, char *path, int path_len)
{
int client_sockfd, create_server_status, fcntl_res;
int req_code, wait_count = 0;
int recv_len, send_len;
ssize_t num_chars_recvd, num_chars_sent;
int save_errno, ret_code = 0, init_ret_code = 0;
int loop_count = 0;
int recv_complete, send_complete;
boolean_t retry = FALSE;
size_t server_proc_len;
int semop_res;
int selstat, status;
char *recv_ptr, *send_ptr;
struct sockaddr_un server_proc;
struct sembuf sop[4];
fd_set wait_on_fd;
gtmsecshr_mesg mesg;
TID timer_id;
int4 msec_timeout;
char *gtm_tmp_ptr;
struct stat stat_buf;
struct shmid_ds shm_info;
int len;
DCL_THREADGBL_ACCESS;
SETUP_THREADGBL_ACCESS;
DBGGSSHR((LOGFLAGS, "secshr_client: New send request\n"));
if (!gtm_dist_ok_to_use)
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(6) ERR_GTMDISTUNVERIF, 4, STRLEN(gtm_dist), gtm_dist,
gtmImageNames[image_type].imageNameLen, gtmImageNames[image_type].imageName);
/* Create communication key (hash of release name) if it has not already been done */
if (0 == TREF(gtmsecshr_comkey))
{
STR_HASH((char *)gtm_release_name, gtm_release_name_len, TREF(gtmsecshr_comkey), 0);
}
timer_id = (TID)send_mesg2gtmsecshr;
if (!gtmsecshr_file_check_done)
{
len = STRLEN(gtm_dist);
memcpy(gtmsecshr_path, gtm_dist, len);
gtmsecshr_path[len] = '/';
memcpy(gtmsecshr_path + len + 1, GTMSECSHR_EXECUTABLE, STRLEN(GTMSECSHR_EXECUTABLE));
gtmsecshr_pathname.addr = gtmsecshr_path;
gtmsecshr_pathname.len = len + 1 + STRLEN(GTMSECSHR_EXECUTABLE);
assertpro(GTM_PATH_MAX > gtmsecshr_pathname.len);
gtmsecshr_pathname.addr[gtmsecshr_pathname.len] = '\0';
if (-1 == Stat(gtmsecshr_pathname.addr, &stat_buf))
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(8) ERR_SYSCALL, 5,
LEN_AND_LIT("stat"), CALLFROM, errno);
if ((ROOTUID != stat_buf.st_uid)
|| !(stat_buf.st_mode & S_ISUID)
|| (0 != ACCESS(gtmsecshr_pathname.addr, (X_OK))))
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(1) ERR_GTMSECSHRPERM);
gtmsecshr_file_check_done = TRUE;
}
if (!gtmsecshr_sock_init_done && (0 < (init_ret_code = gtmsecshr_sock_init(CLIENT)))) /* Note assignment */
return init_ret_code;
DEBUG_ONLY(mesg.usesecshr = TREF(gtm_usesecshr)); /* Flag ignored in PRO build */
while (MAX_COMM_ATTEMPTS >= loop_count)
{ /* first, try the sendto */
req_code = mesg.code = code;
send_len = (int4)(GTM_MESG_HDR_SIZE);
if (REMOVE_FILE == code)
{
assert(GTM_PATH_MAX > path_len); /* Name is not user supplied so simple check */
memcpy(mesg.mesg.path, path, path_len);
send_len += path_len;
} else if (FLUSH_DB_IPCS_INFO == code)
{
assert(GTM_PATH_MAX > db_ipcs.fn_len);
memcpy(&mesg.mesg.db_ipcs, &db_ipcs, (offsetof(ipcs_mesg, fn[0]) + db_ipcs.fn_len + 1));
/* Most of the time file length is much smaller than GTM_PATH_MAX */
send_len += offsetof(ipcs_mesg, fn[0]);
send_len += mesg.mesg.db_ipcs.fn_len + 1;
} else
{
mesg.mesg.id = id;
send_len += SIZEOF(mesg.mesg.id);
}
DBGGSSHR((LOGFLAGS, "secshr_client: loop %d frm-pid: %d to-pid: %d send_len: %d code: %d\n", loop_count,
process_id, id, send_len, code));
mesg.comkey = TREF(gtmsecshr_comkey); /* Version communication key */
mesg.pid = process_id; /* Process id of client */
mesg.seqno = ++cur_seqno;
send_ptr = (char *)&mesg;
send_complete = FALSE;
SENDTO_SOCK(gtmsecshr_sockfd, send_ptr, send_len, 0, (struct sockaddr *)>msecshr_sock_name,
(GTM_SOCKLEN_TYPE)gtmsecshr_sockpath_len, num_chars_sent); /* This form handles EINTR internally */
save_errno = errno;
DBGGSSHR((LOGFLAGS, "secshr_client: sendto rc: %d errno: %d (only important if rc=-1)\n", (int)num_chars_sent,
save_errno));
if (0 >= num_chars_sent)
{ /* SENDTO_SOCK failed - start server and attempt to resend */
if ((EISCONN == save_errno) || (EBADF == save_errno))
{
gtmsecshr_sock_cleanup(CLIENT);
gtmsecshr_sock_init(CLIENT);
wcs_backoff(loop_count + 1);
DBGGSSHR((LOGFLAGS, "secshr_client: Connection error, reset socket\n"));
} else
{
if (0 < loop_count)
/* No message unless attempted server start at least once */
send_msg_csa(CSA_ARG(NULL) VARLSTCNT(11) ERR_GTMSECSHRSRVF, 4,
RTS_ERROR_TEXT("Client"), process_id,
loop_count - 1, ERR_TEXT, 2,
RTS_ERROR_TEXT("sendto to gtmsecshr failed"), save_errno);
START_SERVER;
DBGGSSHR((LOGFLAGS, "secshr_client: sendto() failed - restarting server\n"));
}
loop_count++;
continue;
}
SETUP_FOR_RECV; /* Sets timer, recvcomplete = FALSE */
do
{ /* Note RECVFROM does not loop on EINTR return codes so must be handled. Note also we only expect
* to receive the message header back as an acknowledgement.
*/
num_chars_recvd = RECVFROM(gtmsecshr_sockfd, recv_ptr, GTM_MESG_HDR_SIZE, 0, (struct sockaddr *)0,
(GTM_SOCKLEN_TYPE *)0);
save_errno = errno;
DBGGSSHR((LOGFLAGS, "secshr_client: recvfrom rc: %d errno: %d (only important if rc=-1)\n",
(int)num_chars_recvd, save_errno));
if (0 <= num_chars_recvd)
{ /* Message received - make sure it is large enough to have set seqno before we do anything
* to rely on it.
*/
if ((GTM_MESG_HDR_SIZE <= num_chars_recvd) && (mesg.seqno == cur_seqno)
&& (TREF(gtmsecshr_comkey) == mesg.comkey))
recv_complete = TRUE;
else
{ /* Message too short or not correct sequence */
cancel_timer(timer_id);
/* Print True/False for the possibilities we failed */
DBGGSSHR((LOGFLAGS, "secshr_client: Message incorrect - chars: %d, seq: %d\n",
(GTM_MESG_HDR_SIZE <= num_chars_recvd), (mesg.seqno == cur_seqno)));
SETUP_FOR_RECV;
continue;
}
} else
{ /* Something untoward happened */
if (client_timer_popped)
break;
if (EINTR == save_errno) /* Had an irrelevant interrupt - ignore */
continue;
if (EBADF == save_errno)
break;
send_msg_csa(CSA_ARG(NULL) VARLSTCNT(11) ERR_GTMSECSHRSRVF, 4,
RTS_ERROR_TEXT("Client"), process_id, loop_count - 1, ERR_TEXT, 2,
RTS_ERROR_TEXT("recvfrom from gtmsecshr failed"), save_errno);
if ((ECONNRESET == save_errno) || (ENOTCONN == save_errno))
{
num_chars_recvd = 0;
break;
}
gtmsecshr_sock_cleanup(CLIENT);
return save_errno;
}
} while (!recv_complete);
cancel_timer(timer_id);
if (client_timer_popped || (EBADF == save_errno) || (0 == num_chars_recvd))
{ /* Timeout, connection issues, bad descriptor block - retry */
gtmsecshr_sock_cleanup(CLIENT);
gtmsecshr_sock_init(CLIENT);
retry = TRUE;
if (client_timer_popped)
{
START_SERVER;
DBGGSSHR((LOGFLAGS, "secshr_client: Read timer popped - restarting server\n"));
} else
DBGGSSHR((LOGFLAGS, "secshr_client: Read error - socket reset, retrying\n"));
loop_count++;
continue;
}
/* Response to *our* latest message available */
assert(recv_complete);
if (ret_code = mesg.code) /* Warning - assignment */
{
DBGGSSHR((LOGFLAGS, "secshr_client: non-zero response from gtmsecshr - request: %d retcode: %d\n",
req_code, ret_code));
if (INVALID_COMKEY == ret_code)
{ /* Comkey mismatch means for a different version of GT.M - we will not handle it */
send_msg_csa(CSA_ARG(NULL) VARLSTCNT(13) ERR_GTMSECSHRSRVFIL, 7, RTS_ERROR_TEXT("Client"),
process_id, mesg.pid, req_code, RTS_ERROR_TEXT(mesg.mesg.path),
ERR_TEXT, 2, RTS_ERROR_STRING("Communicating with wrong GT.M version"));
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(13) ERR_GTMSECSHRSRVFIL, 7, RTS_ERROR_TEXT("Client"),
process_id, mesg.pid, req_code, RTS_ERROR_TEXT(mesg.mesg.path),
ERR_TEXT, 2, RTS_ERROR_STRING("Communicating with wrong GT.M version"));
break; /* rts_error should not return */
}
switch(req_code)
{
case REMOVE_FILE:
/* Called from mutex_sock_init(). Path (and length) contain null terminator byte.
* See if file still exists (may have been deleted by earlier attempt). Caller
* handles actual error.
*/
if ((-1 != Stat(path, &stat_buf)) || (ENOENT != ret_code))
send_msg_csa(CSA_ARG(NULL) VARLSTCNT(14) ERR_GTMSECSHRSRVFIL, 7,
RTS_ERROR_TEXT("Client"),
process_id, mesg.pid, req_code, RTS_ERROR_TEXT(mesg.mesg.path),
ERR_TEXT, 2, RTS_ERROR_STRING(secshr_fail_mesg_code[req_code]),
mesg.code);
else
ret_code = 0; /* File is gone so this or a previous try actually worked */
break;
case REMOVE_SEM:
/* See if semaphore still eixsts (may have been removed by earlier attempt that
* got a reply confused or lost). If not there, no error. Else error to op-log.
*/
if ((-1 != semctl(id, 0, GETVAL)) && !SEM_REMOVED(errno))
send_msg_csa(CSA_ARG(NULL) VARLSTCNT(13) ERR_GTMSECSHRSRVFID, 6,
RTS_ERROR_TEXT("Client"),
process_id, mesg.pid, req_code, mesg.mesg.id, ERR_TEXT, 2,
RTS_ERROR_STRING(secshr_fail_mesg_code[req_code]),
mesg.code);
else
ret_code = 0; /* File is gone so this or a previous try actually worked */
case REMOVE_SHM:
/* See if shmem still eixsts (may have been removed by earlier attempt that
* got a reply confused or lost). If not there, no error. Else error to op-log.
* Note -
*/
if ((-1 != shmctl(id, IPC_STAT, &shm_info)) && !SEM_REMOVED(errno))
send_msg_csa(CSA_ARG(NULL) VARLSTCNT(13) ERR_GTMSECSHRSRVFID, 6,
RTS_ERROR_TEXT("Client"),
process_id, mesg.pid, req_code, mesg.mesg.id, ERR_TEXT, 2,
RTS_ERROR_STRING(secshr_fail_mesg_code[req_code]),
mesg.code);
else
ret_code = 0; /* File is gone so this or a previous try actually worked */
break;
case FLUSH_DB_IPCS_INFO: /* Errors handled by caller */
break;
default:
if (EPERM != mesg.code && EACCES != mesg.code)
send_msg_csa(CSA_ARG(NULL) VARLSTCNT(13) ERR_GTMSECSHRSRVFID, 6,
RTS_ERROR_TEXT("Client"),
process_id, mesg.pid, req_code, mesg.mesg.id, ERR_TEXT, 2,
RTS_ERROR_STRING(secshr_fail_mesg_code[req_code]),
mesg.code);
break;
}
}
break;
}
if (MAX_COMM_ATTEMPTS < loop_count)
{
ret_code = -1;
gtm_putmsg_csa(CSA_ARG(NULL) VARLSTCNT(10) ERR_GTMSECSHRSRVF, 4,
RTS_ERROR_TEXT("Client"), process_id, loop_count - 1,
ERR_TEXT, 2, RTS_ERROR_TEXT("Unable to communicate with gtmsecshr"));
/* If gtm_tmp is not defined, show default path */
if (gtm_tmp_ptr = GETENV("gtm_tmp"))
send_msg_csa(CSA_ARG(NULL) VARLSTCNT(8) ERR_GTMSECSHRTMPPATH, 2,
RTS_ERROR_TEXT(gtm_tmp_ptr),
ERR_TEXT, 2, RTS_ERROR_TEXT("(from $gtm_tmp)"));
else
send_msg_csa(CSA_ARG(NULL) VARLSTCNT(4)
ERR_GTMSECSHRTMPPATH, 2, RTS_ERROR_TEXT("/tmp"));
}
if (ONETIMESOCKET == init_ret_code)
gtmsecshr_sock_cleanup(CLIENT);
return ret_code;
}
inline void ComputeMultiVecCommMeta( DistSymmInfo& info )
{
DEBUG_ONLY(CallStackEntry cse("ComputeMultiVecCommMeta"))
// Handle the interface node
info.distNodes[0].multiVecMeta.Empty();
info.distNodes[0].multiVecMeta.localSize = info.localNodes.back().size;
// Handle the truly distributed nodes
const int numDist = info.distNodes.size();
for( int s=1; s<numDist; ++s )
{
DistSymmNodeInfo& node = info.distNodes[s];
const int teamSize = mpi::CommSize( node.comm );
const int teamRank = mpi::CommRank( node.comm );
const DistSymmNodeInfo& childNode = info.distNodes[s-1];
const int childTeamSize = mpi::CommSize( childNode.comm );
const int childTeamRank = mpi::CommRank( childNode.comm );
const bool inFirstTeam = ( childTeamRank == teamRank );
const bool leftIsFirst = ( childNode.onLeft==inFirstTeam );
const int leftTeamSize =
( childNode.onLeft ? childTeamSize : teamSize-childTeamSize );
const int rightTeamSize = teamSize - leftTeamSize;
const int leftTeamOff = ( leftIsFirst ? 0 : rightTeamSize );
const int rightTeamOff = ( leftIsFirst ? leftTeamSize : 0 );
const std::vector<Int>& myRelInds =
( childNode.onLeft ? node.leftRelInds : node.rightRelInds );
// Fill numChildSendInds
MultiVecCommMeta& commMeta = node.multiVecMeta;
commMeta.Empty();
commMeta.numChildSendInds.resize( teamSize );
elem::MemZero( &commMeta.numChildSendInds[0], teamSize );
const Int updateSize = childNode.lowerStruct.size();
{
const Int align = childNode.size % childTeamSize;
const Int shift = Shift( childTeamRank, align, childTeamSize );
const Int localHeight = Length( updateSize, shift, childTeamSize );
for( Int iChildLoc=0; iChildLoc<localHeight; ++iChildLoc )
{
const Int iChild = shift + iChildLoc*childTeamSize;
const int destRank = myRelInds[iChild] % teamSize;
++commMeta.numChildSendInds[destRank];
}
}
const Int numLeftInds = node.leftRelInds.size();
const Int numRightInds = node.rightRelInds.size();
std::vector<Int> leftInds, rightInds;
for( Int i=0; i<numLeftInds; ++i )
if( node.leftRelInds[i] % teamSize == teamRank )
leftInds.push_back( i );
for( Int i=0; i<numRightInds; ++i )
if( node.rightRelInds[i] % teamSize == teamRank )
rightInds.push_back( i );
//
// Compute the solve recv indices
//
commMeta.childRecvInds.resize( teamSize );
// Compute the recv indices for the left child
const Int numLeftSolveInds = leftInds.size();
for( Int iPre=0; iPre<numLeftSolveInds; ++iPre )
{
const Int iChild = leftInds[iPre];
const Int iFront = node.leftRelInds[iChild];
const Int iFrontLoc = (iFront-teamRank) / teamSize;
const int childRank = (node.leftSize+iChild) % leftTeamSize;
const int frontRank = leftTeamOff + childRank;
commMeta.childRecvInds[frontRank].push_back(iFrontLoc);
}
// Compute the recv indices for the right child
const Int numRightSolveInds = rightInds.size();
for( Int iPre=0; iPre<numRightSolveInds; ++iPre )
{
const Int iChild = rightInds[iPre];
const Int iFront = node.rightRelInds[iChild];
const Int iFrontLoc = (iFront-teamRank) / teamSize;
const int childRank = (node.rightSize+iChild) % rightTeamSize;
const int frontRank = rightTeamOff + childRank;
commMeta.childRecvInds[frontRank].push_back(iFrontLoc);
}
commMeta.localSize = Length(node.size,teamRank,teamSize);
}
}
void Her( UpperOrLower uplo, Base<T> alpha, const Matrix<T>& x, Matrix<T>& A )
{
DEBUG_ONLY(CSE cse("Her"))
Syr( uplo, T(alpha), x, A, true );
}
BDM& BDM::operator=( const BlockMatrix<T>& A )
{
DEBUG_ONLY(CSE cse("[CIRC,CIRC] = ABDM"))
copy::Gather( A, *this );
return *this;
}
void MethodHandles::generate_method_handle_dispatch(MacroAssembler* _masm,
vmIntrinsics::ID iid,
Register receiver_reg,
Register member_reg,
bool for_compiler_entry) {
assert(is_signature_polymorphic(iid), "expected invoke iid");
Register rbx_method = rbx; // eventual target of this invocation
// temps used in this code are not used in *either* compiled or interpreted calling sequences
#ifdef _LP64
Register temp1 = rscratch1;
Register temp2 = rscratch2;
Register temp3 = rax;
if (for_compiler_entry) {
assert(receiver_reg == (iid == vmIntrinsics::_linkToStatic ? noreg : j_rarg0), "only valid assignment");
assert_different_registers(temp1, j_rarg0, j_rarg1, j_rarg2, j_rarg3, j_rarg4, j_rarg5);
assert_different_registers(temp2, j_rarg0, j_rarg1, j_rarg2, j_rarg3, j_rarg4, j_rarg5);
assert_different_registers(temp3, j_rarg0, j_rarg1, j_rarg2, j_rarg3, j_rarg4, j_rarg5);
}
#else
Register temp1 = (for_compiler_entry ? rsi : rdx);
Register temp2 = rdi;
Register temp3 = rax;
if (for_compiler_entry) {
assert(receiver_reg == (iid == vmIntrinsics::_linkToStatic ? noreg : rcx), "only valid assignment");
assert_different_registers(temp1, rcx, rdx);
assert_different_registers(temp2, rcx, rdx);
assert_different_registers(temp3, rcx, rdx);
}
#endif
else {
assert_different_registers(temp1, temp2, temp3, saved_last_sp_register()); // don't trash lastSP
}
assert_different_registers(temp1, temp2, temp3, receiver_reg);
assert_different_registers(temp1, temp2, temp3, member_reg);
if (iid == vmIntrinsics::_invokeBasic) {
// indirect through MH.form.vmentry.vmtarget
jump_to_lambda_form(_masm, receiver_reg, rbx_method, temp1, for_compiler_entry);
} else {
// The method is a member invoker used by direct method handles.
if (VerifyMethodHandles) {
// make sure the trailing argument really is a MemberName (caller responsibility)
verify_klass(_masm, member_reg, SystemDictionary::WK_KLASS_ENUM_NAME(java_lang_invoke_MemberName),
"MemberName required for invokeVirtual etc.");
}
Address member_clazz( member_reg, NONZERO(java_lang_invoke_MemberName::clazz_offset_in_bytes()));
Address member_vmindex( member_reg, NONZERO(java_lang_invoke_MemberName::vmindex_offset_in_bytes()));
Address member_vmtarget( member_reg, NONZERO(java_lang_invoke_MemberName::vmtarget_offset_in_bytes()));
Register temp1_recv_klass = temp1;
if (iid != vmIntrinsics::_linkToStatic) {
__ verify_oop(receiver_reg);
if (iid == vmIntrinsics::_linkToSpecial) {
// Don't actually load the klass; just null-check the receiver.
__ null_check(receiver_reg);
} else {
// load receiver klass itself
__ null_check(receiver_reg, oopDesc::klass_offset_in_bytes());
__ load_klass(temp1_recv_klass, receiver_reg);
__ verify_klass_ptr(temp1_recv_klass);
}
BLOCK_COMMENT("check_receiver {");
// The receiver for the MemberName must be in receiver_reg.
// Check the receiver against the MemberName.clazz
if (VerifyMethodHandles && iid == vmIntrinsics::_linkToSpecial) {
// Did not load it above...
__ load_klass(temp1_recv_klass, receiver_reg);
__ verify_klass_ptr(temp1_recv_klass);
}
if (VerifyMethodHandles && iid != vmIntrinsics::_linkToInterface) {
Label L_ok;
Register temp2_defc = temp2;
__ load_heap_oop(temp2_defc, member_clazz);
load_klass_from_Class(_masm, temp2_defc);
__ verify_klass_ptr(temp2_defc);
__ check_klass_subtype(temp1_recv_klass, temp2_defc, temp3, L_ok);
// If we get here, the type check failed!
__ STOP("receiver class disagrees with MemberName.clazz");
__ bind(L_ok);
}
BLOCK_COMMENT("} check_receiver");
}
if (iid == vmIntrinsics::_linkToSpecial ||
iid == vmIntrinsics::_linkToStatic) {
DEBUG_ONLY(temp1_recv_klass = noreg); // these guys didn't load the recv_klass
}
// Live registers at this point:
// member_reg - MemberName that was the trailing argument
// temp1_recv_klass - klass of stacked receiver, if needed
// rsi/r13 - interpreter linkage (if interpreted)
// rcx, rdx, rsi, rdi, r8, r8 - compiler arguments (if compiled)
Label L_incompatible_class_change_error;
switch (iid) {
case vmIntrinsics::_linkToSpecial:
if (VerifyMethodHandles) {
verify_ref_kind(_masm, JVM_REF_invokeSpecial, member_reg, temp3);
}
__ movptr(rbx_method, member_vmtarget);
break;
case vmIntrinsics::_linkToStatic:
if (VerifyMethodHandles) {
verify_ref_kind(_masm, JVM_REF_invokeStatic, member_reg, temp3);
}
__ movptr(rbx_method, member_vmtarget);
break;
case vmIntrinsics::_linkToVirtual:
{
// same as TemplateTable::invokevirtual,
// minus the CP setup and profiling:
if (VerifyMethodHandles) {
verify_ref_kind(_masm, JVM_REF_invokeVirtual, member_reg, temp3);
}
// pick out the vtable index from the MemberName, and then we can discard it:
Register temp2_index = temp2;
__ movptr(temp2_index, member_vmindex);
if (VerifyMethodHandles) {
Label L_index_ok;
__ cmpl(temp2_index, 0);
__ jcc(Assembler::greaterEqual, L_index_ok);
__ STOP("no virtual index");
__ BIND(L_index_ok);
}
// Note: The verifier invariants allow us to ignore MemberName.clazz and vmtarget
// at this point. And VerifyMethodHandles has already checked clazz, if needed.
// get target Method* & entry point
__ lookup_virtual_method(temp1_recv_klass, temp2_index, rbx_method);
break;
}
case vmIntrinsics::_linkToInterface:
{
// same as TemplateTable::invokeinterface
// (minus the CP setup and profiling, with different argument motion)
if (VerifyMethodHandles) {
verify_ref_kind(_masm, JVM_REF_invokeInterface, member_reg, temp3);
}
Register temp3_intf = temp3;
__ load_heap_oop(temp3_intf, member_clazz);
load_klass_from_Class(_masm, temp3_intf);
__ verify_klass_ptr(temp3_intf);
Register rbx_index = rbx_method;
__ movptr(rbx_index, member_vmindex);
if (VerifyMethodHandles) {
Label L;
__ cmpl(rbx_index, 0);
__ jcc(Assembler::greaterEqual, L);
__ STOP("invalid vtable index for MH.invokeInterface");
__ bind(L);
}
// given intf, index, and recv klass, dispatch to the implementation method
__ lookup_interface_method(temp1_recv_klass, temp3_intf,
// note: next two args must be the same:
rbx_index, rbx_method,
temp2,
L_incompatible_class_change_error);
break;
}
default:
fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
break;
}
// Live at this point:
// rbx_method
// rsi/r13 (if interpreted)
// After figuring out which concrete method to call, jump into it.
// Note that this works in the interpreter with no data motion.
// But the compiled version will require that rcx_recv be shifted out.
__ verify_method_ptr(rbx_method);
jump_from_method_handle(_masm, rbx_method, temp1, for_compiler_entry);
if (iid == vmIntrinsics::_linkToInterface) {
__ bind(L_incompatible_class_change_error);
__ jump(RuntimeAddress(StubRoutines::throw_IncompatibleClassChangeError_entry()));
}
}
}
DM& DM::operator=( const DistMatrix<T,STAR,VC>& A )
{
DEBUG_ONLY(CSE cse("[STAR,STAR] = [STAR,VC]"))
copy::RowAllGather( A, *this );
return *this;
}
SafeProduct<F> SafeDeterminant( const ElementalMatrix<F>& A )
{
DEBUG_ONLY(CSE cse("SafeDeterminant"))
DistMatrix<F> B( A );
return det::LUPartialPiv( B );
}
void MakeHermitian( UpperOrLower uplo, DistSparseMatrix<T>& A )
{
DEBUG_ONLY(CallStackEntry cse("MakeHermitian"))
MakeSymmetric( uplo, A, true );
}
// Code generation
address MethodHandles::generate_method_handle_interpreter_entry(MacroAssembler* _masm,
vmIntrinsics::ID iid) {
const bool not_for_compiler_entry = false; // this is the interpreter entry
assert(is_signature_polymorphic(iid), "expected invoke iid");
if (iid == vmIntrinsics::_invokeGeneric ||
iid == vmIntrinsics::_compiledLambdaForm) {
// Perhaps surprisingly, the symbolic references visible to Java are not directly used.
// They are linked to Java-generated adapters via MethodHandleNatives.linkMethod.
// They all allow an appendix argument.
__ stop("Should not reach here"); // empty stubs make SG sick
return NULL;
}
Register argbase = CC_INTERP_ONLY(R17_tos) NOT_CC_INTERP(R15_esp); // parameter (preserved)
Register argslot = R3;
Register temp1 = R6;
Register param_size = R7;
// here's where control starts out:
__ align(CodeEntryAlignment);
address entry_point = __ pc();
if (VerifyMethodHandles) {
assert(Method::intrinsic_id_size_in_bytes() == 2, "assuming Method::_intrinsic_id is u2");
Label L;
BLOCK_COMMENT("verify_intrinsic_id {");
__ load_sized_value(temp1, Method::intrinsic_id_offset_in_bytes(), R19_method,
sizeof(u2), /*is_signed*/ false);
__ cmpwi(CCR1, temp1, (int) iid);
__ beq(CCR1, L);
if (iid == vmIntrinsics::_linkToVirtual ||
iid == vmIntrinsics::_linkToSpecial) {
// could do this for all kinds, but would explode assembly code size
trace_method_handle(_masm, "bad Method*:intrinsic_id");
}
__ stop("bad Method*::intrinsic_id");
__ BIND(L);
BLOCK_COMMENT("} verify_intrinsic_id");
}
// First task: Find out how big the argument list is.
int ref_kind = signature_polymorphic_intrinsic_ref_kind(iid);
assert(ref_kind != 0 || iid == vmIntrinsics::_invokeBasic, "must be _invokeBasic or a linkTo intrinsic");
if (ref_kind == 0 || MethodHandles::ref_kind_has_receiver(ref_kind)) {
__ ld(param_size, in_bytes(Method::const_offset()), R19_method);
__ load_sized_value(param_size, in_bytes(ConstMethod::size_of_parameters_offset()), param_size,
sizeof(u2), /*is_signed*/ false);
// assert(sizeof(u2) == sizeof(ConstMethod::_size_of_parameters), "");
} else {
DEBUG_ONLY(param_size = noreg);
}
Register tmp_mh = noreg;
if (!is_signature_polymorphic_static(iid)) {
__ ld(tmp_mh = temp1, __ argument_offset(param_size, param_size, 0), argbase);
DEBUG_ONLY(param_size = noreg);
}
if (TraceMethodHandles) {
if (tmp_mh != noreg) {
__ mr(R23_method_handle, tmp_mh); // make stub happy
}
trace_method_handle_interpreter_entry(_masm, iid);
}
if (iid == vmIntrinsics::_invokeBasic) {
generate_method_handle_dispatch(_masm, iid, tmp_mh, noreg, not_for_compiler_entry);
} else {
// Adjust argument list by popping the trailing MemberName argument.
Register tmp_recv = noreg;
if (MethodHandles::ref_kind_has_receiver(ref_kind)) {
// Load the receiver (not the MH; the actual MemberName's receiver) up from the interpreter stack.
__ ld(tmp_recv = temp1, __ argument_offset(param_size, param_size, 0), argbase);
DEBUG_ONLY(param_size = noreg);
}
Register R19_member = R19_method; // MemberName ptr; incoming method ptr is dead now
__ ld(R19_member, RegisterOrConstant((intptr_t)8), argbase);
__ add(argbase, Interpreter::stackElementSize, argbase);
generate_method_handle_dispatch(_masm, iid, tmp_recv, R19_member, not_for_compiler_entry);
}
return entry_point;
}
inline void
Snapshot
( const Matrix<Int>& preimage,
const Matrix<Real>& estimates,
const Matrix<Int>& itCounts,
Int numIts,
bool deflate,
SnapshotCtrl& snapCtrl )
{
DEBUG_ONLY(CSE cse("pspec::Snapshot"));
auto logMap = []( Real alpha ) { return Log(alpha); };
if( snapCtrl.realSize != 0 && snapCtrl.imagSize != 0 )
{
const bool numSave =
( snapCtrl.numSaveFreq > 0 &&
snapCtrl.numSaveCount >= snapCtrl.numSaveFreq );
const bool imgSave =
( snapCtrl.imgSaveFreq > 0 &&
snapCtrl.imgSaveCount >= snapCtrl.imgSaveFreq );
const bool imgDisp =
( snapCtrl.imgDispFreq > 0 &&
snapCtrl.imgDispCount >= snapCtrl.imgDispFreq );
Matrix<Real> invNorms, estMap;
Matrix<Int> itCountsReord, itCountMap;
if( numSave || imgSave || imgDisp )
{
invNorms = estimates;
if( deflate )
RestoreOrdering( preimage, invNorms );
ReshapeIntoGrid
( snapCtrl.realSize, snapCtrl.imagSize, invNorms, estMap );
if( snapCtrl.itCounts )
{
itCountsReord = itCounts;
if( deflate )
RestoreOrdering( preimage, itCountsReord );
ReshapeIntoGrid
( snapCtrl.realSize, snapCtrl.imagSize, itCountsReord,
itCountMap );
}
}
if( numSave )
{
ostringstream os;
os << snapCtrl.numBase << "_" << numIts;
Write( estMap, os.str(), snapCtrl.numFormat );
if( snapCtrl.itCounts )
Write( itCountMap, os.str()+"_counts", snapCtrl.numFormat );
snapCtrl.numSaveCount = 0;
}
if( imgSave || imgDisp )
EntrywiseMap( estMap, function<Real(Real)>(logMap) );
if( imgSave )
{
ostringstream os;
os << snapCtrl.imgBase << "_" << numIts;
Write( estMap, os.str(), snapCtrl.imgFormat );
if( snapCtrl.itCounts )
Write( itCountMap, os.str()+"_counts", snapCtrl.imgFormat );
auto colorMap = GetColorMap();
SetColorMap( GRAYSCALE_DISCRETE );
Write( estMap, os.str()+"_discrete", snapCtrl.imgFormat );
SetColorMap( colorMap );
snapCtrl.imgSaveCount = 0;
}
if( imgDisp )
{
ostringstream os;
os << snapCtrl.imgBase << "_" << numIts;
Display( estMap, os.str() );
if( snapCtrl.itCounts )
Display( itCountMap, os.str()+"_counts" );
auto colorMap = GetColorMap();
SetColorMap( GRAYSCALE_DISCRETE );
Display( estMap, os.str()+"_discrete" );
SetColorMap( colorMap );
snapCtrl.imgDispCount = 0;
}
}
}
void MethodHandles::generate_method_handle_dispatch(MacroAssembler* _masm,
vmIntrinsics::ID iid,
Register receiver_reg,
Register member_reg,
bool for_compiler_entry) {
assert(is_signature_polymorphic(iid), "expected invoke iid");
Register temp1 = (for_compiler_entry ? R25_tmp5 : R7);
Register temp2 = (for_compiler_entry ? R22_tmp2 : R8);
Register temp3 = (for_compiler_entry ? R23_tmp3 : R9);
Register temp4 = (for_compiler_entry ? R24_tmp4 : R10);
if (receiver_reg != noreg) assert_different_registers(temp1, temp2, temp3, temp4, receiver_reg);
if (member_reg != noreg) assert_different_registers(temp1, temp2, temp3, temp4, member_reg);
if (iid == vmIntrinsics::_invokeBasic) {
// indirect through MH.form.vmentry.vmtarget
jump_to_lambda_form(_masm, receiver_reg, R19_method, temp1, temp2, for_compiler_entry);
} else {
// The method is a member invoker used by direct method handles.
if (VerifyMethodHandles) {
// make sure the trailing argument really is a MemberName (caller responsibility)
verify_klass(_masm, member_reg, SystemDictionary::WK_KLASS_ENUM_NAME(MemberName_klass),
temp1, temp2,
"MemberName required for invokeVirtual etc.");
}
Register temp1_recv_klass = temp1;
if (iid != vmIntrinsics::_linkToStatic) {
__ verify_oop(receiver_reg);
if (iid == vmIntrinsics::_linkToSpecial) {
// Don't actually load the klass; just null-check the receiver.
__ null_check_throw(receiver_reg, -1, temp1, EXCEPTION_ENTRY);
} else {
// load receiver klass itself
__ null_check_throw(receiver_reg, oopDesc::klass_offset_in_bytes(), temp1, EXCEPTION_ENTRY);
__ load_klass(temp1_recv_klass, receiver_reg);
__ verify_klass_ptr(temp1_recv_klass);
}
BLOCK_COMMENT("check_receiver {");
// The receiver for the MemberName must be in receiver_reg.
// Check the receiver against the MemberName.clazz
if (VerifyMethodHandles && iid == vmIntrinsics::_linkToSpecial) {
// Did not load it above...
__ load_klass(temp1_recv_klass, receiver_reg);
__ verify_klass_ptr(temp1_recv_klass);
}
if (VerifyMethodHandles && iid != vmIntrinsics::_linkToInterface) {
Label L_ok;
Register temp2_defc = temp2;
__ load_heap_oop_not_null(temp2_defc, NONZERO(java_lang_invoke_MemberName::clazz_offset_in_bytes()), member_reg, temp3);
load_klass_from_Class(_masm, temp2_defc, temp3, temp4);
__ verify_klass_ptr(temp2_defc);
__ check_klass_subtype(temp1_recv_klass, temp2_defc, temp3, temp4, L_ok);
// If we get here, the type check failed!
__ stop("receiver class disagrees with MemberName.clazz");
__ BIND(L_ok);
}
BLOCK_COMMENT("} check_receiver");
}
if (iid == vmIntrinsics::_linkToSpecial ||
iid == vmIntrinsics::_linkToStatic) {
DEBUG_ONLY(temp1_recv_klass = noreg); // these guys didn't load the recv_klass
}
// Live registers at this point:
// member_reg - MemberName that was the trailing argument
// temp1_recv_klass - klass of stacked receiver, if needed
// O5_savedSP - interpreter linkage (if interpreted)
// O0..O5 - compiler arguments (if compiled)
Label L_incompatible_class_change_error;
switch (iid) {
case vmIntrinsics::_linkToSpecial:
if (VerifyMethodHandles) {
verify_ref_kind(_masm, JVM_REF_invokeSpecial, member_reg, temp2);
}
__ ld(R19_method, NONZERO(java_lang_invoke_MemberName::vmtarget_offset_in_bytes()), member_reg);
break;
case vmIntrinsics::_linkToStatic:
if (VerifyMethodHandles) {
verify_ref_kind(_masm, JVM_REF_invokeStatic, member_reg, temp2);
}
__ ld(R19_method, NONZERO(java_lang_invoke_MemberName::vmtarget_offset_in_bytes()), member_reg);
break;
case vmIntrinsics::_linkToVirtual:
{
// same as TemplateTable::invokevirtual,
// minus the CP setup and profiling:
if (VerifyMethodHandles) {
verify_ref_kind(_masm, JVM_REF_invokeVirtual, member_reg, temp2);
}
// pick out the vtable index from the MemberName, and then we can discard it:
Register temp2_index = temp2;
__ ld(temp2_index, NONZERO(java_lang_invoke_MemberName::vmindex_offset_in_bytes()), member_reg);
if (VerifyMethodHandles) {
Label L_index_ok;
__ cmpdi(CCR1, temp2_index, 0);
__ bge(CCR1, L_index_ok);
__ stop("no virtual index");
__ BIND(L_index_ok);
}
// Note: The verifier invariants allow us to ignore MemberName.clazz and vmtarget
// at this point. And VerifyMethodHandles has already checked clazz, if needed.
// get target Method* & entry point
__ lookup_virtual_method(temp1_recv_klass, temp2_index, R19_method);
break;
}
case vmIntrinsics::_linkToInterface:
{
// same as TemplateTable::invokeinterface
// (minus the CP setup and profiling, with different argument motion)
if (VerifyMethodHandles) {
verify_ref_kind(_masm, JVM_REF_invokeInterface, member_reg, temp2);
}
Register temp2_intf = temp2;
__ load_heap_oop_not_null(temp2_intf, NONZERO(java_lang_invoke_MemberName::clazz_offset_in_bytes()), member_reg, temp3);
load_klass_from_Class(_masm, temp2_intf, temp3, temp4);
__ verify_klass_ptr(temp2_intf);
Register vtable_index = R19_method;
__ ld(vtable_index, NONZERO(java_lang_invoke_MemberName::vmindex_offset_in_bytes()), member_reg);
if (VerifyMethodHandles) {
Label L_index_ok;
__ cmpdi(CCR1, vtable_index, 0);
__ bge(CCR1, L_index_ok);
__ stop("invalid vtable index for MH.invokeInterface");
__ BIND(L_index_ok);
}
// given intf, index, and recv klass, dispatch to the implementation method
__ lookup_interface_method(temp1_recv_klass, temp2_intf,
// note: next two args must be the same:
vtable_index, R19_method,
temp3, temp4,
L_incompatible_class_change_error);
break;
}
default:
fatal("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid));
break;
}
// Live at this point:
// R19_method
// O5_savedSP (if interpreted)
// After figuring out which concrete method to call, jump into it.
// Note that this works in the interpreter with no data motion.
// But the compiled version will require that rcx_recv be shifted out.
__ verify_method_ptr(R19_method);
jump_from_method_handle(_masm, R19_method, temp1, temp2, for_compiler_entry);
if (iid == vmIntrinsics::_linkToInterface) {
__ BIND(L_incompatible_class_change_error);
__ load_const_optimized(temp1, StubRoutines::throw_IncompatibleClassChangeError_entry());
__ mtctr(temp1);
__ bctr();
}
}
}
inline void
SetImagPart( Real& alpha, const Real& beta )
{
DEBUG_ONLY(CallStackEntry cse("SetImagPart"))
LogicError("Nonsensical assignment");
}
* You should have received a copy of the GNU General Public License version
* 2 along with this work; if not, write to the Free Software Foundation,
* Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
*
* Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
* or visit www.oracle.com if you need additional information or have any
* questions.
*
*/
// Platform-specific definitions for method handles.
// These definitions are inlined into class MethodHandles.
// Adapters
enum /* platform_dependent_constants */ {
adapter_code_size = NOT_LP64(23000 DEBUG_ONLY(+ 40000)) LP64_ONLY(35000 DEBUG_ONLY(+ 50000))
};
// Additional helper methods for MethodHandles code generation:
public:
static void load_klass_from_Class(MacroAssembler* _masm, Register klass_reg, Register temp_reg, Register temp2_reg);
static void verify_klass(MacroAssembler* _masm,
Register obj_reg, SystemDictionary::WKID klass_id,
Register temp_reg, Register temp2_reg,
const char* error_message = "wrong klass") NOT_DEBUG_RETURN;
static void verify_method_handle(MacroAssembler* _masm, Register mh_reg,
Register temp_reg, Register temp2_reg) {
verify_klass(_masm, mh_reg, SystemDictionary::WK_KLASS_ENUM_NAME(java_lang_invoke_MethodHandle),
temp_reg, temp2_reg,
/* Operation - The destination variable becomes a new alias of the data pointed to by the source container variable:
*
* 1) Index into the variable name table to get the variable name for the new alias (destination).
* 2) Create the hash for the variable and look up in hash table (adding if not yet created).
* 3) See if hash entry has a value pointer. If yes and same as source var, we are done (NOOP).
* 4) Set hash entry to point to new lv_val.
* 5) Whichever symval of source container or the alias it pointed to has the largest address, that symval is the
* earliest symval affected by this alias action. Mark all interveening symvals (end points included) as having
* had alias activity via MARK_ALIAS_ACTIVE macro. This is so we can properly handle issues in an exlusive NEW
* situation involving multiple symvals and potential aliasing between them.
*
* Note that this opcode's function is very similar to "op_setalsctin2als" but is necessarily different because the
* source container is in a temporary mval passed back through a function rather than the lv_val that "op_setalsctin2als"
* deals with. Consequently, the amount of verification we can do is reduced. But this is acceptable due to the checks
* done by "unw_retarg" and "op_exfunretals" which pre-processed this value for us. There is also different reference
* count maintenance to do than the "op_setalsctin2als" opcode. With substantially more work to reorganize how SET
* operates, it would likely be possible to combine these functions but the way things are structured now, all the
* set functions plus "op_sto" share the same API so adding a parm to one means adding a useless parm to all 6 of
* them which is not acceptable so we end up duplicating portions of code here.
*/
void op_setfnretin2als(mval *srcmv, int destindx)
{
ht_ent_mname *tabent;
mname_entry *varname;
lv_val *srclvc, *dstlv;
int4 srcsymvlvl;
boolean_t added;
error_def(ERR_ALIASEXPECTED);
assert(alias_retarg == srcmv);
assert(srcmv);
assert(srcmv->mvtype & MV_ALIASCONT);
/* Verify is a temp mval */
assert((char *)srcmv >= (char *)frame_pointer->temps_ptr
&& (char *)srcmv < ((char *)frame_pointer->temps_ptr + (SIZEOF(char *) * frame_pointer->temp_mvals)));
srclvc = (lv_val *)srcmv->str.addr;
assert(srclvc);
assert(LV_IS_BASE_VAR(srclvc)); /* Verify base var */
assert(srclvc->stats.trefcnt >= srclvc->stats.crefcnt);
assert(1 <= srclvc->stats.crefcnt); /* Verify we have an existing container reference */
srcsymvlvl = LV_SYMVAL(srclvc)->symvlvl; /* lv_val may go away below so record symlvl */
varname = &(((mname_entry *)frame_pointer->vartab_ptr)[destindx]);
DEBUG_ONLY(added = FALSE);
/* Find hash table entry */
if (NULL == (tabent = (ht_ent_mname *)frame_pointer->l_symtab[destindx])) /* note tabent assignment */
{ /* No fast path to hash table entry -- look it up the hard(er) way */
varname = &(((mname_entry *)frame_pointer->vartab_ptr)[destindx]);
added = add_hashtab_mname_symval(&curr_symval->h_symtab, varname, NULL, &tabent);
}
assert(tabent);
assert(tabent || added);
dstlv = (lv_val *)tabent->value;
if (NULL == dstlv && curr_symval->tp_save_all)
{ /* dstlv does not exist yet we need to be able to save a previous "novalue" lvval in case a TPRESTART
needs to restore the value. Create a var so its undefined status can be saved.
*/
lv_newname(tabent, curr_symval);
dstlv = (lv_val *)tabent->value;
assert(dstlv);
}
/* No need to increment before dstlv processing to prevent removal of last reference to srclvc in this case because
* the increment has already been done in "unw_retarg".
*/
if (dstlv)
{
assert(LV_IS_BASE_VAR(dstlv)); /* Verify base var */
if (dollar_tlevel && NULL != dstlv->tp_var && !dstlv->tp_var->var_cloned)
TP_VAR_CLONE(dstlv);
assert(0 < dstlv->stats.trefcnt);
assert(0 < (dstlv->stats.trefcnt - dstlv->stats.crefcnt)); /* Make sure there is one non-container reference */
DECR_TREFCNT(dstlv);
assert(dstlv->stats.trefcnt >= dstlv->stats.crefcnt);
if (0 == dstlv->stats.trefcnt)
{ /* Non alias -- make room for an alias to live here instead */
lv_kill(dstlv, DOTPSAVE_TRUE, DO_SUBTREE_TRUE);
LV_FREESLOT(dstlv);
} /* Else alias pointer in the hash table is just replaced below */
}
DECR_CREFCNT(srclvc); /* In "unw_retarg" we incremented for a container but is now "just" an alias so get rid of
* the container count from the temp return parm */
frame_pointer->l_symtab[destindx] = tabent;
DBGRFCT((stderr, "op_setfnret2als: hte 0x"lvaddr" being reset from 0x"lvaddr" to 0x"lvaddr"\n",
tabent, tabent->value, srclvc));
tabent->value = (void *)srclvc;
/* These symvals have had alias activity - Note the possibility of re-marking srcsymvlvl is not necessarily re-doing
* the mark done by "unw_retarg" since the source lv_val may have been re-created if it was originally in an xnew'd
* symtab which popped during the return.
*/
MARK_ALIAS_ACTIVE(MIN(srcsymvlvl, LV_SYMVAL(srclvc)->symvlvl));
active_lv = (lv_val *)NULL; /* if we get here, subscript set was successful. clear active_lv to avoid later
cleanup problems */
alias_retarg = NULL;
}
int gtmsource()
{
int status, log_init_status, waitpid_res, save_errno;
char print_msg[1024], tmpmsg[1024];
gd_region *reg, *region_top;
sgmnt_addrs *csa, *repl_csa;
boolean_t all_files_open, isalive;
pid_t pid, ppid, procgp;
seq_num read_jnl_seqno, jnl_seqno;
unix_db_info *udi;
gtmsource_local_ptr_t gtmsource_local;
boolean_t this_side_std_null_coll;
int null_fd, rc;
memset((uchar_ptr_t)&jnlpool, 0, SIZEOF(jnlpool_addrs));
call_on_signal = gtmsource_sigstop;
ESTABLISH_RET(gtmsource_ch, SS_NORMAL);
if (-1 == gtmsource_get_opt())
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(1) ERR_MUPCLIERR);
if (gtmsource_options.shut_down)
{ /* Wait till shutdown time nears even before going to "jnlpool_init". This is because the latter will return
* with the ftok semaphore and access semaphore held and we do not want to be holding those locks (while
* waiting for the user specified timeout to expire) as that will affect new GTM processes and/or other
* MUPIP REPLIC commands that need these locks for their function.
*/
if (0 < gtmsource_options.shutdown_time)
{
repl_log(stdout, TRUE, TRUE, "Waiting for %d seconds before signalling shutdown\n",
gtmsource_options.shutdown_time);
LONG_SLEEP(gtmsource_options.shutdown_time);
} else
repl_log(stdout, TRUE, TRUE, "Signalling shutdown immediate\n");
} else if (gtmsource_options.start)
{
repl_log(stdout, TRUE, TRUE, "Initiating START of source server for secondary instance [%s]\n",
gtmsource_options.secondary_instname);
}
if (gtmsource_options.activate && (ROOTPRIMARY_SPECIFIED == gtmsource_options.rootprimary))
{ /* MUPIP REPLIC -SOURCE -ACTIVATE -UPDOK has been specified. We need to open the gld and db regions now
* in case this is a secondary -> primary transition. This is so we can later switch journal files in all
* journaled regions when the transition actually happens inside "gtmsource_rootprimary_init". But since
* we have not yet done a "jnlpool_init", we dont know if updates are disabled in it or not. Although we
* need to do the gld/db open only if updates are currently disabled in the jnlpool, we do this always
* because once we do a jnlpool_init, we will come back with the ftok on the jnlpool held and that has
* issues with later db open since we will try to hold the db ftok as part of db open and the ftok logic
* currently has assumptions that a process holds only one ftok at any point in time.
*/
assert(NULL == gd_header);
gvinit();
all_files_open = region_init(FALSE);
if (!all_files_open)
{
gtm_putmsg_csa(CSA_ARG(NULL) VARLSTCNT(1) ERR_NOTALLDBOPN);
gtmsource_exit(ABNORMAL_SHUTDOWN);
}
}
jnlpool_init(GTMSOURCE, gtmsource_options.start, &is_jnlpool_creator);
/* is_jnlpool_creator == TRUE ==> this process created the journal pool
* is_jnlpool_creator == FALSE ==> journal pool already existed and this process simply attached to it.
*/
if (gtmsource_options.shut_down)
gtmsource_exit(gtmsource_shutdown(FALSE, NORMAL_SHUTDOWN) - NORMAL_SHUTDOWN);
else if (gtmsource_options.activate)
gtmsource_exit(gtmsource_mode_change(GTMSOURCE_MODE_ACTIVE_REQUESTED) - NORMAL_SHUTDOWN);
else if (gtmsource_options.deactivate)
gtmsource_exit(gtmsource_mode_change(GTMSOURCE_MODE_PASSIVE_REQUESTED) - NORMAL_SHUTDOWN);
else if (gtmsource_options.checkhealth)
gtmsource_exit(gtmsource_checkhealth() - NORMAL_SHUTDOWN);
else if (gtmsource_options.changelog)
gtmsource_exit(gtmsource_changelog() - NORMAL_SHUTDOWN);
else if (gtmsource_options.showbacklog)
gtmsource_exit(gtmsource_showbacklog() - NORMAL_SHUTDOWN);
else if (gtmsource_options.stopsourcefilter)
gtmsource_exit(gtmsource_stopfilter() - NORMAL_SHUTDOWN);
else if (gtmsource_options.jnlpool)
gtmsource_exit(gtmsource_jnlpool() - NORMAL_SHUTDOWN);
else if (gtmsource_options.losttncomplete)
gtmsource_exit(gtmsource_losttncomplete() - NORMAL_SHUTDOWN);
else if (gtmsource_options.needrestart)
gtmsource_exit(gtmsource_needrestart() - NORMAL_SHUTDOWN);
else if (gtmsource_options.showfreeze)
gtmsource_exit(gtmsource_showfreeze() - NORMAL_SHUTDOWN);
else if (gtmsource_options.setfreeze)
gtmsource_exit(gtmsource_setfreeze() - NORMAL_SHUTDOWN);
else if (!gtmsource_options.start)
{
assert(CLI_PRESENT == cli_present("STATSLOG"));
gtmsource_exit(gtmsource_statslog() - NORMAL_SHUTDOWN);
}
assert(gtmsource_options.start);
# ifndef REPL_DEBUG_NOBACKGROUND
/* Set "child_server_running" to FALSE before forking off child. Wait for it to be set to TRUE by the child. */
gtmsource_local = jnlpool.gtmsource_local;
gtmsource_local->child_server_running = FALSE;
FORK(pid);
if (0 > pid)
{
save_errno = errno;
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(7) ERR_JNLPOOLSETUP, 0,
ERR_TEXT, 2, RTS_ERROR_LITERAL("Could not fork source server"), save_errno);
} else if (0 < pid)
{ /* Parent. Wait until child sets "child_server_running" to FALSE. That is an indication that the child
* source server has completed its initialization phase and is all set so the parent command can return.
*/
while (isalive = is_proc_alive(pid, 0)) /* note : intended assignment */
{
if (gtmsource_local->child_server_running)
break;
/* To take care of reassignment of PIDs, the while condition should be && with the condition
* (PPID of pid == process_id)
*/
SHORT_SLEEP(GTMSOURCE_WAIT_FOR_SRV_START);
WAITPID(pid, &status, WNOHANG, waitpid_res); /* Release defunct child if dead */
}
if (isalive)
{ /* Child process is alive and started with no issues */
if (0 != (save_errno = rel_sem(SOURCE, JNL_POOL_ACCESS_SEM)))
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(7) ERR_JNLPOOLSETUP, 0,
ERR_TEXT, 2, RTS_ERROR_LITERAL("Error in rel_sem"), save_errno);
ftok_sem_release(jnlpool.jnlpool_dummy_reg, TRUE, TRUE);
} else
{ /* Child source server process errored out at startup and is no longer alive.
* If we were the one who created the journal pool, let us clean it up.
*/
repl_log(stdout, TRUE, TRUE, "Source server startup failed. See source server log file\n");
if (is_jnlpool_creator)
status = gtmsource_shutdown(TRUE, NORMAL_SHUTDOWN);
}
/* If the parent is killed (or crashes) between the fork and exit, checkhealth may not detect that startup
* is in progress - parent forks and dies, the system will release sem 0 and 1, checkhealth might test the
* value of sem 1 before the child grabs sem 1.
*/
gtmsource_exit(isalive ? SRV_ALIVE : SRV_ERR);
}
/* Point stdin to /dev/null */
OPENFILE("/dev/null", O_RDONLY, null_fd);
if (0 > null_fd)
rts_error_csa(CSA_ARG(NULL) ERR_REPLERR, RTS_ERROR_LITERAL("Failed to open /dev/null for read"), errno, 0);
FCNTL3(null_fd, F_DUPFD, 0, rc);
if (0 > rc)
rts_error_csa(CSA_ARG(NULL) ERR_REPLERR, RTS_ERROR_LITERAL("Failed to set stdin to /dev/null"), errno, 0);
CLOSEFILE(null_fd, rc);
if (0 > rc)
rts_error_csa(CSA_ARG(NULL) ERR_REPLERR, RTS_ERROR_LITERAL("Failed to close /dev/null"), errno, 0);
/* The parent process (source server startup command) will be holding the ftok semaphore and jnlpool access semaphore
* at this point. The variables that indicate this would have been copied over to the child during the fork. This will
* make the child think it is actually holding them as well when actually it is not. Reset those variables in the child
* to ensure they do not misrepresent the holder of those semaphores.
*/
ftok_sem_reg = NULL;
udi = FILE_INFO(jnlpool.jnlpool_dummy_reg);
assert(udi->grabbed_ftok_sem);
udi->grabbed_ftok_sem = FALSE;
assert(holds_sem[SOURCE][JNL_POOL_ACCESS_SEM]);
holds_sem[SOURCE][JNL_POOL_ACCESS_SEM] = FALSE;
assert(!holds_sem[SOURCE][SRC_SERV_COUNT_SEM]);
/* Start child source server initialization */
is_src_server = TRUE;
OPERATOR_LOG_MSG;
process_id = getpid();
/* Reinvoke secshr related initialization with the child's pid */
INVOKE_INIT_SECSHR_ADDRS;
/* Initialize mutex socket, memory semaphore etc. before any "grab_lock" is done by this process on the journal pool.
* Note that the initialization would already have been done by the parent receiver startup command but we need to
* redo the initialization with the child process id.
*/
assert(mutex_per_process_init_pid && (mutex_per_process_init_pid != process_id));
mutex_per_process_init();
START_HEARTBEAT_IF_NEEDED;
ppid = getppid();
log_init_status = repl_log_init(REPL_GENERAL_LOG, >msource_log_fd, gtmsource_options.log_file);
assert(SS_NORMAL == log_init_status);
repl_log_fd2fp(>msource_log_fp, gtmsource_log_fd);
if (-1 == (procgp = setsid()))
send_msg_csa(CSA_ARG(NULL) VARLSTCNT(7) ERR_JNLPOOLSETUP, 0, ERR_TEXT, 2,
RTS_ERROR_LITERAL("Source server error in setsid"), errno);
# endif /* REPL_DEBUG_NOBACKGROUND */
if (ZLIB_CMPLVL_NONE != gtm_zlib_cmp_level)
gtm_zlib_init(); /* Open zlib shared library for compression/decompression */
REPL_DPRINT1("Setting up regions\n");
gvinit();
/* We use the same code dse uses to open all regions but we must make sure they are all open before proceeding. */
all_files_open = region_init(FALSE);
if (!all_files_open)
{
gtm_putmsg_csa(CSA_ARG(NULL) VARLSTCNT(1) ERR_NOTALLDBOPN);
gtmsource_exit(ABNORMAL_SHUTDOWN);
}
/* Determine primary side null subscripts collation order */
/* Also check whether all regions have same null collation order */
this_side_std_null_coll = -1;
for (reg = gd_header->regions, region_top = gd_header->regions + gd_header->n_regions; reg < region_top; reg++)
{
csa = &FILE_INFO(reg)->s_addrs;
if (this_side_std_null_coll != csa->hdr->std_null_coll)
{
if (-1 == this_side_std_null_coll)
this_side_std_null_coll = csa->hdr->std_null_coll;
else
{
gtm_putmsg_csa(CSA_ARG(NULL) VARLSTCNT(1) ERR_NULLCOLLDIFF);
gtmsource_exit(ABNORMAL_SHUTDOWN);
}
}
if (!REPL_ALLOWED(csa) && JNL_ALLOWED(csa))
{
gtm_putmsg_csa(CSA_ARG(NULL) VARLSTCNT(4) ERR_REPLOFFJNLON, 2, DB_LEN_STR(reg));
gtmsource_exit(ABNORMAL_SHUTDOWN);
}
if (reg->read_only && REPL_ALLOWED(csa))
{
gtm_putmsg_csa(CSA_ARG(NULL) VARLSTCNT(6) ERR_JNLPOOLSETUP, 0, ERR_TEXT, 2,
RTS_ERROR_LITERAL("Source Server does not have write permissions to one or "
"more database files that are replicated"));
gtmsource_exit(ABNORMAL_SHUTDOWN);
}
}
/* Initialize source server alive/dead state related fields in "gtmsource_local" before the ftok semaphore is released */
gtmsource_local->gtmsource_pid = process_id;
gtmsource_local->gtmsource_state = GTMSOURCE_START;
if (is_jnlpool_creator)
{
DEBUG_ONLY(jnlpool.jnlpool_ctl->jnlpool_creator_pid = process_id);
gtmsource_seqno_init(this_side_std_null_coll);
if (ROOTPRIMARY_SPECIFIED == gtmsource_options.rootprimary)
{ /* Created the journal pool as a root primary. Append a history record to the replication instance file.
* Invoke the function "gtmsource_rootprimary_init" to do that.
*/
gtmsource_rootprimary_init(jnlpool.jnlpool_ctl->jnl_seqno);
}
}
/* after this point we can no longer have the case where all the regions are unreplicated/non-journaled. */
# ifndef REPL_DEBUG_NOBACKGROUND
/* It is necessary for every process that is using the ftok semaphore to increment the counter by 1. This is used
* by the last process that shuts down to delete the ftok semaphore when it notices the counter to be 0.
* Note that the parent source server startup command would have done an increment of the ftok counter semaphore
* for the replication instance file. But the source server process (the child) that comes here would not have done
* that. Do that while the parent is still holding on to the ftok semaphore waiting for our okay.
*/
if (!ftok_sem_incrcnt(jnlpool.jnlpool_dummy_reg))
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(1) ERR_JNLPOOLSETUP);
/* Increment the source server count semaphore */
status = incr_sem(SOURCE, SRC_SERV_COUNT_SEM);
if (0 != status)
{
save_errno = errno;
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(7) ERR_JNLPOOLSETUP, 0, ERR_TEXT, 2,
RTS_ERROR_LITERAL("Counter semaphore increment failure in child source server"), save_errno);
}
# else
if (0 != (save_errno = rel_sem_immediate(SOURCE, JNL_POOL_ACCESS_SEM)))
{
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(7) ERR_JNLPOOLSETUP, 0, ERR_TEXT, 2,
RTS_ERROR_LITERAL("Error in rel_sem_immediate"), save_errno);
}
# endif /* REPL_DEBUG_NOBACKGROUND */
gtmsource_srv_count++;
gtmsource_local->child_server_running = TRUE; /* At this point, the parent startup command will stop waiting for child */
gtm_event_log_init();
/* Log source server startup command line first */
SPRINTF(tmpmsg, "%s %s\n", cli_lex_in_ptr->argv[0], cli_lex_in_ptr->in_str);
repl_log(gtmsource_log_fp, TRUE, TRUE, tmpmsg);
SPRINTF(tmpmsg, "GTM Replication Source Server with Pid [%d] started for Secondary Instance [%s]",
process_id, gtmsource_local->secondary_instname);
sgtm_putmsg(print_msg, VARLSTCNT(4) ERR_REPLINFO, 2, LEN_AND_STR(tmpmsg));
repl_log(gtmsource_log_fp, TRUE, TRUE, print_msg);
if (is_jnlpool_creator)
{
repl_log(gtmsource_log_fp, TRUE, TRUE, "Created jnlpool with shmid = [%d] and semid = [%d]\n",
jnlpool.repl_inst_filehdr->jnlpool_shmid, jnlpool.repl_inst_filehdr->jnlpool_semid);
} else
repl_log(gtmsource_log_fp, TRUE, TRUE, "Attached to existing jnlpool with shmid = [%d] and semid = [%d]\n",
jnlpool.repl_inst_filehdr->jnlpool_shmid, jnlpool.repl_inst_filehdr->jnlpool_semid);
gtm_event_log(GTM_EVENT_LOG_ARGC, "MUPIP", "REPLINFO", print_msg);
# ifdef GTM_TLS
if (REPL_TLS_REQUESTED)
{
repl_do_tls_init(gtmsource_log_fp);
assert(REPL_TLS_REQUESTED || PLAINTEXT_FALLBACK);
}
# endif
if (jnlpool.jnlpool_ctl->freeze)
{
last_seen_freeze_flag = jnlpool.jnlpool_ctl->freeze;
sgtm_putmsg(print_msg, VARLSTCNT(3) ERR_REPLINSTFROZEN, 1, jnlpool.repl_inst_filehdr->inst_info.this_instname);
repl_log(gtmsource_log_fp, TRUE, FALSE, print_msg);
sgtm_putmsg(print_msg, VARLSTCNT(3) ERR_REPLINSTFREEZECOMMENT, 1, jnlpool.jnlpool_ctl->freeze_comment);
repl_log(gtmsource_log_fp, TRUE, TRUE, print_msg);
}
gtmsource_local->jnlfileonly = gtmsource_options.jnlfileonly;
do
{ /* If mode is passive, go to sleep. Wakeup every now and then and check to see if I have to become active. */
gtmsource_state = gtmsource_local->gtmsource_state = GTMSOURCE_START;
if ((gtmsource_local->mode == GTMSOURCE_MODE_PASSIVE) && (gtmsource_local->shutdown == NO_SHUTDOWN))
{
gtmsource_poll_actions(FALSE);
SHORT_SLEEP(GTMSOURCE_WAIT_FOR_MODE_CHANGE);
continue;
}
if (GTMSOURCE_MODE_PASSIVE == gtmsource_local->mode)
{ /* Shutdown initiated */
assert(gtmsource_local->shutdown == SHUTDOWN);
sgtm_putmsg(print_msg, VARLSTCNT(4) ERR_REPLINFO, 2,
RTS_ERROR_LITERAL("GTM Replication Source Server Shutdown signalled"));
repl_log(gtmsource_log_fp, TRUE, TRUE, print_msg);
gtm_event_log(GTM_EVENT_LOG_ARGC, "MUPIP", "REPLINFO", print_msg);
break;
}
gtmsource_poll_actions(FALSE);
if (GTMSOURCE_CHANGING_MODE == gtmsource_state)
continue;
if (GTMSOURCE_MODE_ACTIVE_REQUESTED == gtmsource_local->mode)
gtmsource_local->mode = GTMSOURCE_MODE_ACTIVE;
SPRINTF(tmpmsg, "GTM Replication Source Server now in ACTIVE mode using port %d", gtmsource_local->secondary_port);
sgtm_putmsg(print_msg, VARLSTCNT(4) ERR_REPLINFO, 2, LEN_AND_STR(tmpmsg));
repl_log(gtmsource_log_fp, TRUE, TRUE, print_msg);
gtm_event_log(GTM_EVENT_LOG_ARGC, "MUPIP", "REPLINFO", print_msg);
DEBUG_ONLY(repl_csa = &FILE_INFO(jnlpool.jnlpool_dummy_reg)->s_addrs;)
assert(!repl_csa->hold_onto_crit); /* so it is ok to invoke "grab_lock" and "rel_lock" unconditionally */
grab_lock(jnlpool.jnlpool_dummy_reg, TRUE, HANDLE_CONCUR_ONLINE_ROLLBACK);
if (GTMSOURCE_HANDLE_ONLN_RLBK == gtmsource_state)
{
repl_log(gtmsource_log_fp, TRUE, TRUE, "Starting afresh due to ONLINE ROLLBACK\n");
repl_log(gtmsource_log_fp, TRUE, TRUE, "REPL INFO - Current Jnlpool Seqno : %llu\n",
jnlpool.jnlpool_ctl->jnl_seqno);
continue;
}
QWASSIGN(gtmsource_local->read_addr, jnlpool.jnlpool_ctl->write_addr);
gtmsource_local->read = jnlpool.jnlpool_ctl->write;
gtmsource_local->read_state = gtmsource_local->jnlfileonly ? READ_FILE : READ_POOL;
read_jnl_seqno = gtmsource_local->read_jnl_seqno;
assert(read_jnl_seqno <= jnlpool.jnlpool_ctl->jnl_seqno);
if (read_jnl_seqno < jnlpool.jnlpool_ctl->jnl_seqno)
{
gtmsource_local->read_state = READ_FILE;
QWASSIGN(gtmsource_save_read_jnl_seqno, jnlpool.jnlpool_ctl->jnl_seqno);
gtmsource_pool2file_transition = TRUE; /* so that we read the latest gener jnl files */
}
rel_lock(jnlpool.jnlpool_dummy_reg);
if (SS_NORMAL != (status = gtmsource_alloc_tcombuff()))
rts_error_csa(CSA_ARG(NULL) VARLSTCNT(7) ERR_REPLCOMM, 0, ERR_TEXT, 2,
RTS_ERROR_LITERAL("Error allocating initial tcom buffer space. Malloc error"), status);
gtmsource_filter = NO_FILTER;
if ('\0' != gtmsource_local->filter_cmd[0])
{
if (SS_NORMAL == (status = repl_filter_init(gtmsource_local->filter_cmd)))
gtmsource_filter |= EXTERNAL_FILTER;
else
gtmsource_exit(ABNORMAL_SHUTDOWN);
}
gtmsource_process();
/* gtmsource_process returns only when mode needs to be changed to PASSIVE */
assert(gtmsource_state == GTMSOURCE_CHANGING_MODE);
gtmsource_ctl_close();
gtmsource_free_msgbuff();
gtmsource_free_tcombuff();
gtmsource_free_filter_buff();
gtmsource_stop_heartbeat();
if (FD_INVALID != gtmsource_sock_fd)
repl_close(>msource_sock_fd);
if (gtmsource_filter & EXTERNAL_FILTER)
repl_stop_filter();
} while (TRUE);
void ColScatter
( T alpha,
const ElementalMatrix<T>& A,
ElementalMatrix<T>& B )
{
DEBUG_ONLY(CSE cse("axpy_contract::ColScatter"))
AssertSameGrids( A, B );
if( A.Height() != B.Height() || A.Width() != B.Width() )
LogicError("A and B must be the same size");
#ifdef EL_VECTOR_WARNINGS
if( A.Width() == 1 && B.Grid().Rank() == 0 )
{
cerr <<
"The vector version of ColScatter does not"
" yet have a vector version implemented, but it would only "
"require a modification of the vector version of RowScatter"
<< endl;
}
#endif
#ifdef EL_CACHE_WARNINGS
if( A.Width() != 1 && B.Grid().Rank() == 0 )
{
cerr <<
"axpy_contract::ColScatter potentially causes a large "
"amount of cache-thrashing. If possible, avoid it by forming the "
"(conjugate-)transpose of the [* ,V] matrix instead." << endl;
}
#endif
if( !B.Participating() )
return;
const Int height = B.Height();
const Int localHeight = B.LocalHeight();
const Int localWidth = B.LocalWidth();
const Int colAlign = B.ColAlign();
const Int colStride = B.ColStride();
const Int rowDiff = B.RowAlign()-A.RowAlign();
// TODO: Allow for modular equivalence if possible
if( rowDiff == 0 )
{
const Int maxLocalHeight = MaxLength(height,colStride);
const Int recvSize = mpi::Pad( maxLocalHeight*localWidth );
const Int sendSize = colStride*recvSize;
//vector<T> buffer( sendSize );
vector<T> buffer;
buffer.reserve( sendSize );
// Pack
copy::util::ColStridedPack
( height, localWidth,
colAlign, colStride,
A.LockedBuffer(), A.LDim(),
buffer.data(), recvSize );
// Communicate
mpi::ReduceScatter( buffer.data(), recvSize, B.ColComm() );
// Update with our received data
axpy::util::InterleaveMatrixUpdate
( alpha, localHeight, localWidth,
buffer.data(), 1, localHeight,
B.Buffer(), 1, B.LDim() );
}
else
{
#ifdef EL_UNALIGNED_WARNINGS
if( B.Grid().Rank() == 0 )
cerr << "Unaligned ColScatter" << endl;
#endif
const Int localWidthA = A.LocalWidth();
const Int maxLocalHeight = MaxLength(height,colStride);
const Int recvSize_RS = mpi::Pad( maxLocalHeight*localWidthA );
const Int sendSize_RS = colStride*recvSize_RS;
const Int recvSize_SR = localHeight*localWidth;
//vector<T> buffer( recvSize_RS + Max(sendSize_RS,recvSize_SR) );
vector<T> buffer;
buffer.reserve( recvSize_RS + Max(sendSize_RS,recvSize_SR) );
T* firstBuf = &buffer[0];
T* secondBuf = &buffer[recvSize_RS];
// Pack
copy::util::ColStridedPack
( height, localWidth,
colAlign, colStride,
A.LockedBuffer(), A.LDim(),
secondBuf, recvSize_RS );
// Reduce-scatter over each col
mpi::ReduceScatter( secondBuf, firstBuf, recvSize_RS, B.ColComm() );
// Trade reduced data with the appropriate col
const Int sendCol = Mod( B.RowRank()+rowDiff, B.RowStride() );
const Int recvCol = Mod( B.RowRank()-rowDiff, B.RowStride() );
mpi::SendRecv
( firstBuf, localHeight*localWidthA, sendCol,
secondBuf, localHeight*localWidth, recvCol, B.RowComm() );
// Update with our received data
axpy::util::InterleaveMatrixUpdate
( alpha, localHeight, localWidth,
secondBuf, 1, localHeight,
B.Buffer(), 1, B.LDim() );
}
}
DM& DM::operator=( const DistMatrix<T,CIRC,CIRC>& A )
{
DEBUG_ONLY(CSE cse("[STAR,STAR] = [CIRC,CIRC]"))
copy::Scatter( A, *this );
return *this;
}
BDM& BDM::operator=( const BDM& A )
{
DEBUG_ONLY(CSE cse("[STAR,VC] = [STAR,VC]"))
copy::Translate( A, *this );
return *this;
}
DM& DM::operator=( const DistMatrix<T,MR,STAR>& A )
{
DEBUG_ONLY(CSE cse("[STAR,STAR] = [MR,STAR]"))
copy::ColAllGather( A, *this );
return *this;
}
BDM& BDM::operator=( const DistMatrix<T,STAR,STAR,BLOCK>& A )
{
DEBUG_ONLY(CSE cse("[STAR,VC] = [STAR,STAR]"))
copy::RowFilter( A, *this );
return *this;
}
Base<F> LogBarrier( UpperOrLower uplo, const ElementalMatrix<F>& A )
{
DEBUG_ONLY(CSE cse("LogBarrier"))
SafeProduct<Base<F>> safeDet = SafeHPDDeterminant( uplo, A );
return -safeDet.kappa*safeDet.n;
}
