/**
* \internal Make sure the DBF table, which points to all the memory regions collected
* by CCCPSE, is as compact as possible. The goal is to cut the number of ELF
* Phdrs down as much as possible by merging adjoining memory regions. In order to do
* this, it needs to mark each DBFITEM, which is in protected memory, SPK=0. This
* function is therefore in SPK=0 for its duration, and restores SPK1 prior to return.
*
* \param[in] dbfHdr Address of the Java dump buffer as returned from CCCPSE.
*
* \return
* The number of program headers after merge
*/
static uintptr_t
adjustDBFTable(DBFHDR *dbfHdr)
{
DBFITEM *dbfTbl = (DBFITEM *) dbfHdr + 1; /* DBTI[0] starts right after dbfHdr */
DBFITEM *pred = dbfTbl; /* Predecessor DBTI */
/*
* For first pass only, set pred = curr. We never want to address zero.
* We'll use it only to test bitfields which aren't possible to be set
* for the very first item.
*/
DBFITEM *curr = dbfTbl; /* Current DBTI under examination */
DBFITEM *succ = pred + 1; /* Successor DBTI */
DBFITEM *limit = dbfTbl + (dbfHdr->ijavcnt); /* DBTI[n+1] address */
uintptr_t phnum = 0; /* Number of expected Elf64_Phdrs */
uintptr_t wGap; /* ACCUMULATOR: total ibc gap */
void *wVstart; /* WORK: new p_vaddr */
IDATA chunkSize = 0L; /* Semi-permanent per set ... size */
IDATA totalIBCSize = 0L; /* Total sizes of all writeable sets */
KEY0(); /* JDB's in SPK=0, get auth to write it.*/
while (curr < limit) {
memset(DBTITYPE, 0, sizeof(char) + (sizeof(DBTIGAP)));/* Clear DBTI tail */
calcGap(curr, succ, &wGap); /* Get GAP size. */
if (0 == wGap) {
/*
* This pair qualifies to be combined
*/
if (PSETSTRT || PSETMIDL) {
CSETMIDL = TRUE; /* It is either mid-set ... */
} else { /* ... or starts a new set. */
CSETSTRT = TRUE;
wVstart = DBTIVADDR; /* If it's a start, hold the start vaddr. */
chunkSize += DBTISIZ;
DBTIGAP = wGap;
}
} else {
/*
* This pair does not qualify to be combined. Is it last in set, or is its
* predecessor not part of this one?
*/
if (PSETMIDL || PSETSTRT) { /* It's the end of a set, so summarize */
CSETLAST = TRUE; /* the set's start adrs & size in this */
chunkSize += DBTISIZ; /* JDB index item, it's the one we'll */
DBTISIZ += chunkSize; /* be building the Elf64_Phdr from. */
chunkSize = 0L;
DBTIVADDR = wVstart; /* Reset starting address & size */
wVstart = 0L; /* accumulators. */
} else {
CSETSOLO = TRUE; /* This one is a standalone item. */
}
/*
* We will only write Elf64_Phdr items for those DBF items marked CSETLAST
* or CSETSOLO.
*/
if (CSETLAST || CSETSOLO) {
phnum += 1; /* We will write a Phdr because of this item*/
}
}
DBTIGAP = wGap; /* Save the gap for later analysis */
pred = curr; /* Advance our pair pointers */
curr = succ;
succ += 1;
}
UNKEY(); /* Restore SPK=1 */
return phnum; /* Return revised count to caller. */
}
/**
* \internal Build the z/TPF ELF-format core dump from the JDB built by CCCPSE.
*
* First figure out the absolute file path (starts with a '/' and
* ends with a file suffix) we are to write the dump to, then ensure we
* have enough room left over in the Java dump buffer left for us by
* CCCPSE. If we do not, return an error. Otherwise, section the buffer
* space off, building in order:
* - The ELF64 header
* - The ELF64 Program Header section
* - The ELF64 NOTE section
*
* Finally, write all the above to the indicated filename, and follow it
* with a copy of storage contents from the Java dump buffer, which is
* laid out in ascending address order, with the ELF64_Phdrs created to
* match the dump buffer. Expect a final file size measured in tens
* of megabytes.
*
* As a matter of protocol, set the first 8 bytes of the Java Dump
* Buffer (struct ijavdbf) to zeros so CCCPSE knows it's okay to
* write over its contents once we're done working with it.
*
* The arg block is used for the bi-directional passing of data. Those of its
* members which are pointers (most of them, in fact) are used mainly for
* output or at least have an output role. The fields of <arg> which are
* required are shown as input parameters below. The fields of <arg> which
* are used for output are detailed as return values.
*
* Since the pthread_create() calls limits us to one pointer to type void as the
* function's input, and the function is required to return a void pointer, this
* function will take the address of the <arg> block as the input parameter, and
* will return a pointer to type char which represents the file name which now
* contains the full path name of the ELF-format core dump file. There are also
* fields in <arg> that will be updated as described below.
*
* \param[in][out] arg pointer to a user-created datatype 'args', declared
* in header file j9osdump_helpers.h
* \param[in] arg->OSFilename Pointer to scratch storage for a path +
* file name string. It will be presumed
* to be at least PATH_MAX+1 in size.
* \param[in] arg->wkspcSize 32-bit quantity representing the size in
* bytes of the wkSpace field, following.
* \param[in] arg->wkSpace 64-bit pointer to scratch workspace for
* use by this function. It is recommended
* to make it equal to PATH_MAX, since file &
* path names will be built in this space.
* \param[in][out] arg->flags Indicators as to what is present and what
* is not.
* \param[in] arg->sii Pointer to scratch storage to be used
* forever more as a siginfo_t block
* \param[in] arg->uct Pointer to scratch storage to be used
* forever more as a ucontext_t block
* \param[in] arg->sct Pointer to scratch storage to be used
* forever more as a sigcontext block
* \param[in] arg->portLibrary Pointer to an initialized OMRPortLibrary
* block. If there isn't one at call time,
* leave this value NULL and set flag
* J9ZTPF_NO_PORT_LIBRARY
* \param[in] arg->dibPtr Address of the DIB attached to the faulting
* UOW at post-interrupt time.
*
* \returns Pointer to a hz-terminated string representing the absolute path
* name at which the core dump file was written.
*
* \returns arg->sii Filled in.
* \returns arg->uct Filled in.
* \returns arg->sct Filled in.
* \returns arg->rc Final return code. Zero if successful
* (core file built), non-zero if not.
* \returns arg->OSFilename Same as the function return value.
*/
void *
ztpfBuildCoreFile(void *argv_block)
{
#define MOUNT_TFS 4 /* TPF Filesystem equate from imount.h */
args *arg = (args *) argv_block;
uint8_t *buffer, *endBuffer;
DBFHDR *dbfptr; /* Ptr to JDB's index header */
DIB *dibPtr = dibAddr(arg); /* Ptr to the Dump I'chg Block */
Elf64_Ehdr *ehdrp; /* Pointer to Elf64 file header */
Elf64_Phdr *phdrp; /* Pointer to Elf64_Phdr block */
Elf64_Nhdr *narea; /* Pointer to the ELF NOTE data */
char pathName[PATH_MAX]; /* Working buffer for core.* fname.*/
char *ofn = dumpFilename(arg); /* Output dump file path */
uint32_t ofd; /* File descriptor for output dump */
uintptr_t rc; /* Working return code d.o. */
uintptr_t wPtr; /* Working byte-sized pointer d.o. */
uint64_t phCount; /* Counter of Elf64_Phdrs required */
uint8_t *imageBuffer; /* Start of the JDB's ICB */
uint64_t imageBufferSize; /* Size of data in the ICB */
uint64_t octetsToWrite = 0UL; /* Count of bytes to write */
uint64_t octetsWritten = 0UL; /* Count of bytes written */
uint64_t spcAvailable;
/*
* If there is a Java dump buffer belonging to this process, then
* convert its contents into an Elf64 core dump file; otherwise
* return failure.
*/
if (!(dibPtr->dibjdb)) { /* No dump buffer? Not possible.*/
dumpFlags(arg) |= J9TPF_NO_JAVA_DUMPBUFFER;
returnCode (arg) = -1; /* Set error flags & bad RC ... */
return NULL; /* 'Bye. */
}
dbfptr = dibPtr->dibjdb; /* Pick up the dump buffer ptr */
if (0L == dbfptr->ijavcnt) { /* Did CCCPSE write us one? */
returnCode (arg) = -1; /* Nope. JDB is locked... */
dumpFlags(arg) |= J9TPF_JDUMPBUFFER_LOCKED;
return NULL; /* See ya next time. */
}
/*
* Calculate the start, end, and net length of the output buffer we'll
* use for the ELF-dictated start of the file content. The start address
* should, as a matter of good practice, start on a paragraph boundary.
*/
buffer = (uint8_t *) (dbfptr->ijavbfp); /* Get buffer start as uint8_t ptr */
buffer = (uint8_t *) NEXT_PARA(buffer); /* Start it on next paragraph */
/* boundary. */
endBuffer = buffer + dbfptr->ijavbfl; /* Get bytes left in buffer */
spcAvailable = endBuffer - buffer; /* Get corrected count of bytes */
phCount = dbfptr->ijavcnt;
int numJavaPgms = sizeof(javaPgmsToDump) / 5;
octetsToWrite = sizeof(Elf64_Ehdr)
+ ((phCount + 1 + numJavaPgms) * sizeof(Elf64_Phdr)) +
NOTE_TABLE_SIZE;
/*
* Uh oh. Not enough free space remaining in the dump buffer. Now
* we've gotta go to the heap and see if there's enough there. Not
* much hope; but we have to try it.
*/
if (octetsToWrite > spcAvailable) { /* Check for available memory, */
buffer = malloc64(octetsToWrite); /* anywhere. We're desperate. */
if (!buffer) { /* If we can't buffer our I/Os, */
returnCode (arg) = -1; /* we are done. Indicate error */
errMsg(arg, "Cannot buffer dump file, out of memory");
dumpFlags(arg) |= J9TPF_OUT_OF_BUFFERSPACE;
KEY0();
dbfptr->ijavcnt = 0L; /* Unlock the JDB so CCCPSE can reuse it*/
UNKEY();
return NULL; /* See ya 'round. */
}
}
/*
* We're ready to write the file. First, we need a full path + filename
* to represent the "OS filename" we're going to write. Get that path
* from the final file name passed in the <tt>args</tt> block, then
* follow that with "core.%X.%d" with the TOD in the second node and
* the PID number in the third. In this way, we can identify the OS
* filename later.
*
* Next, we'll try to open it in CREATE+WRITE_ONLY mode. If it fails,
* halt; else write away!
*/
splitPathName(ofn, pathName);
/*
* The "OS filename" will look like ${path}/core.${TOD_in_hex}.${pid}
*/
sprintf(workSpace(arg), "core.%lX.%d", dibPtr->dstckf,
dumpSiginfo(arg)->si_pid);
{
int pathLen = strlen(pathName);
char ebcdicPath[pathLen + 1];
a2e_len(pathName, ebcdicPath, strlen(pathName));
int fsystype = pathconf(ebcdicPath, _TPF_PC_FS_TYPE);
if (fsystype == MOUNT_TFS) {
errMsg(arg, "Cannot use the tfs for java dumps: path used='%s'\n",
pathName);
returnCode (arg) = -1;
KEY0();
dbfptr->ijavcnt = 0L; /* Unlock the JDB so CCCPSE can reuse it*/
UNKEY();
return NULL;
//dir_path_name is within the TFS
}
}
strcat(pathName, workSpace(arg));
ofd = open(pathName, O_WRONLY | O_CREAT | O_EXCL);
if (-1 == ofd) {
errMsg(arg, "Cannot open() filename %s: %s\n", strerror(errno));
returnCode (arg) = -1;
KEY0();
dbfptr->ijavcnt = 0L; /* Unlock the JDB so CCCPSE can reuse it*/
UNKEY();
return NULL;
}
/*
* The file is named and opened. Now we have to go build its ELF-dictated
* parts. Set up pointers to the start of each sub-section, and then go
* build them. When that's complete, write the file in two parts: first,
* the ELF-dictated start, and then follow that with the data in the
* ICB.
*/
wPtr = (uintptr_t) buffer; /* Calc section addresses with single */
/* byte arithmetic */
ehdrp = (Elf64_Ehdr *) wPtr; /* Elf64_Ehdr pointer */
phdrp = (Elf64_Phdr *) (ehdrp + 1); /* Elf64_Phdr pointer */
buildELFHeader(ehdrp);
uint64_t numBytes = 0;
uint64_t pCount = buildELFPheaderBlk(phdrp, dbfptr, phCount, &numBytes);
KEY0(); /* Get into SPK 0, then write the */
ehdrp->e_phnum = pCount; /* absolutely final Phdr count and */
ehdrp->e_phoff = 0x40; /* offset of its table into place in */
UNKEY(); /* the Ehdr, then get back to SPK 1. */
wPtr = (uintptr_t) phdrp; /* Calculate the end address */
wPtr += ((phCount + 1 + numJavaPgms) * sizeof(Elf64_Phdr)); /* of the Elf64_Phdr section */
narea = (Elf64_Nhdr *) wPtr; /* Calculate the address of NOTE sec, */
buildELFNoteArea(narea, dibPtr); /* and go write it there. */
/*
* Write the ELF-dictated portion of the file first.
*/
octetsWritten = writeDumpFile(ofd, buffer, octetsToWrite);
if (-1 == octetsWritten) {
errMsg(arg, "Error writing dump file %s: %s", ofn, strerror(errno));
dumpFlags(arg) |= J9TPF_FILE_SYSTEM_ERROR;
KEY0();
dbfptr->ijavcnt = 0L; /* Unlock the JDB so CCCPSE can reuse it*/
UNKEY();
return NULL;
}
/*
* Finish the output with the ICB portion of the Java Dump Buffer
*/
imageBuffer = (uint8_t *) cinfc_fast(CINFC_CMMJDB);
octetsWritten = writeDumpFile(ofd, imageBuffer, numBytes);
if (-1 == octetsWritten) {
errMsg(arg, "Error writing dump file %s: %s", ofn, strerror(errno));
dumpFlags(arg) |= J9TPF_FILE_SYSTEM_ERROR;
KEY0();
dbfptr->ijavcnt = 0L; /* Unlock the JDB so CCCPSE can reuse it*/
UNKEY();
return NULL;
}
int i = 0;
for (i = 0; i < numJavaPgms; i++) {
struct pat * pgmpat = progc(javaPgmsToDump[i], PROGC_PBI);
struct ifetch *pgmbase = pgmpat->patgca;
if (pgmbase != NULL) {
int offset = pgmbase->_iftch_txt_off + pgmbase->_iftch_txt_size
+ 0x1000;
char * text = ((char*) pgmbase) + offset;
uint64_t size = pgmpat->patpsize - offset;
octetsWritten = writeDumpFile(ofd, text, size);
if (-1 == octetsWritten) {
errMsg(arg, "Error writing dump file %s: %s", ofn,
strerror(errno));
dumpFlags(arg) |= J9TPF_FILE_SYSTEM_ERROR;
KEY0();
dbfptr->ijavcnt = 0L; /* Unlock the JDB so CCCPSE can reuse it*/
UNKEY();
return NULL;
}
}
}
/*
* That's all folks. Close the file and return the so-called "OS Filename"
* to the caller as if the OS had written it.
*/
rc = close(ofd); /* Only try to close() the file once */
if (-1 == rc) {
errMsg(arg, "I/O error attempting to close %s:%s\n", ofn,
strerror(errno));
KEY0();
dbfptr->ijavcnt = 0L; /* Unlock the JDB so CCCPSE can reuse it*/
UNKEY();
return NULL;
}
/*
* Make sure we have a recognizable IPC permission on the OS filename,
* then make sure the JDB buffer is unlocked in case CPSE needs it again.
*/
rc = chmod(workSpace(arg), S_IRUSR | S_IWUSR | S_IRGRP | S_IWGRP | S_IROTH);
KEY0();
dbfptr->ijavcnt = 0L; /* Unlock the JDB so CCCPSE can reuse it*/
UNKEY();
/*
* Return the filename to the caller. Remember that the buffer containing
* it is non-JVM-managed heap space and should be free()d back to it to
* avoid a mem leak.
*/
strcpy(dumpFilename(arg), ofn); /* Store it in the args block so the */
return (void *) ofn; /* caller always has it, and goback.*/
}
// . THIS Msg0 class must be alloc'd, i.e. not on the stack, etc.
// . if list is stored locally this tries to get it locally
// . otherwise tries to get the list from the network
// . returns false if blocked, true otherwise
// . sets g_errno on error
// . NOTE: i was having problems with queries being cached too long, you
// see the cache here is a NETWORK cache, so when the machines that owns
// the list updates it on disk it can't flush our cache... so use a small
// maxCacheAge of like , 30 seconds or so...
bool Msg0::getList ( long long hostId , // host to ask (-1 if none)
long ip , // info on hostId
short port ,
long maxCacheAge , // max cached age in seconds
bool addToCache , // add net recv'd list to cache?
char rdbId , // specifies the rdb
char *coll ,
RdbList *list ,
//key_t startKey ,
//key_t endKey ,
char *startKey ,
char *endKey ,
long minRecSizes , // use -1 for no max
void *state ,
void (* callback)(void *state ),//, RdbList *list ) ,
long niceness ,
bool doErrorCorrection ,
bool includeTree ,
bool doMerge ,
long firstHostId ,
long startFileNum ,
long numFiles ,
long timeout ,
long long syncPoint ,
long preferLocalReads ,
Msg5 *msg5 ,
Msg5 *msg5b ,
bool isRealMerge ,
//#ifdef SPLIT_INDEXDB
bool allowPageCache ,
bool forceLocalIndexdb ,
bool noSplit , // doIndexdbSplit ,
long forceParitySplit ) {
//#else
// bool allowPageCache ) {
//#endif
// this is obsolete! mostly, but we need it for PageIndexdb.cpp to
// show a "termlist" for a given query term in its entirety so you
// don't have to check each machine in the network. if this is true it
// means to query each split and merge the results together into a
// single unified termlist. only applies to indexdb/datedb.
//if ( doIndexdbSplit ) { char *xx = NULL; *xx = 0; }
// note this because if caller is wrong it hurts performance major!!
//if ( doIndexdbSplit )
// logf(LOG_DEBUG,"net: doing msg0 with indexdb split true");
// warning
if ( ! coll ) log(LOG_LOGIC,"net: NULL collection. msg0.");
//if ( doIndexdbSplit ) { char *xx=NULL;*xx=0; }
// reset the list they passed us
list->reset();
// get keySize of rdb
m_ks = getKeySizeFromRdbId ( rdbId );
// if startKey > endKey, don't read anything
//if ( startKey > endKey ) return true;
if ( KEYCMP(startKey,endKey,m_ks)>0 ) { char *xx=NULL;*xx=0; }//rettrue
// . reset hostid if it is dead
// . this is causing UOR queries to take forever when we have a dead
if ( hostId >= 0 && g_hostdb.isDead ( hostId ) ) hostId = -1;
// no longer accept negative minrecsize
if ( minRecSizes < 0 ) {
g_errno = EBADENGINEER;
log(LOG_LOGIC,
"net: msg0: Negative minRecSizes no longer supported.");
char *xx=NULL;*xx=0;
return true;
}
// debug msg
//if ( niceness != 0 ) log("HEY start");
// ensure startKey last bit clear, endKey last bit set
//if ( (startKey.n0 & 0x01) == 0x01 )
// log("Msg0::getList: warning startKey lastbit set");
//if ( (endKey.n0 & 0x01) == 0x00 )
// log("Msg0::getList: warning endKey lastbit clear");
// remember these
m_state = state;
m_callback = callback;
m_list = list;
m_hostId = hostId;
m_niceness = niceness;
//m_ip = ip;
//m_port = port;
m_addToCache = addToCache;
// . these define our request 100%
//m_startKey = startKey;
//m_endKey = endKey;
KEYSET(m_startKey,startKey,m_ks);
KEYSET(m_endKey,endKey,m_ks);
m_minRecSizes = minRecSizes;
m_rdbId = rdbId;
m_coll = coll;
m_isRealMerge = isRealMerge;
m_allowPageCache = allowPageCache;
// . group to ask is based on the first key
// . we only do 1 group per call right now
// . groupMask must turn on higher bits first (count downwards kinda)
// . titledb and spiderdb use special masks to get groupId
// if diffbot.cpp is reading spiderdb from each shard we have to
// get groupid from hostid here lest we core in getGroupId() below.
// it does that for dumping spiderdb to the client browser. they
// can download the whole enchilada.
if ( hostId >= 0 && m_rdbId == RDB_SPIDERDB )
m_shardNum = 0;
// did they force it? core until i figure out what this is
else if ( forceParitySplit >= 0 )
//m_groupId = g_hostdb.getGroupId ( forceParitySplit );
m_shardNum = forceParitySplit;
else
//m_groupId = getGroupId ( m_rdbId , startKey , ! noSplit );
m_shardNum = getShardNum ( m_rdbId , startKey );
// if we are looking up a termlist in posdb that is split by termid and
// not the usual docid then we have to set this posdb key bit that tells
// us that ...
if ( noSplit && m_rdbId == RDB_POSDB )
m_shardNum = g_hostdb.getShardNumByTermId ( startKey );
// how is this used?
//if ( forceLocalIndexdb ) m_groupId = g_hostdb.m_groupId;
if ( forceLocalIndexdb ) m_shardNum = getMyShardNum();
// . store these parameters
// . get a handle to the rdb in case we can satisfy locally
// . returns NULL and sets g_errno on error
QUICKPOLL((m_niceness));
Rdb *rdb = getRdbFromId ( m_rdbId );
if ( ! rdb ) return true;
// we need the fixedDataSize
m_fixedDataSize = rdb->getFixedDataSize();
m_useHalfKeys = rdb->useHalfKeys();
// . debug msg
// . Msg2 does this when checking for a cached compound list.
// compound lists do not actually exist, they are merges of smaller
// UOR'd lists.
if ( maxCacheAge != 0 && ! addToCache && (numFiles > 0 || includeTree))
log(LOG_LOGIC,"net: msg0: "
"Weird. check but don't add... rdbid=%li.",(long)m_rdbId);
// set this here since we may not call msg5 if list not local
//m_list->setFixedDataSize ( m_fixedDataSize );
// . now that we do load balancing we don't want to do a disk lookup
// even if local if we are merging or dumping
// . UNLESS g_conf.m_preferLocalReads is true
if ( preferLocalReads == -1 )
preferLocalReads = g_conf.m_preferLocalReads;
// . always prefer local for full split clusterdb
// . and keep the tfndb/titledb lookups in the same stripe
// . so basically we can't do biased caches if fully split
//if ( g_conf.m_fullSplit ) preferLocalReads = true;
preferLocalReads = true;
// it it stored locally?
bool isLocal = ( m_hostId == -1 && //g_hostdb.m_groupId == m_groupId );
m_shardNum == getMyShardNum() );
// only do local lookups if this is true
if ( ! preferLocalReads ) isLocal = false;
/*
m_numSplit = 1;
if ( g_hostdb.m_indexSplits > 1 &&
( rdbId == RDB_POSDB || rdbId==RDB_DATEDB)&&
! forceLocalIndexdb && doIndexdbSplit ) {
isLocal = false;
//m_numSplit = INDEXDB_SPLIT;
m_numSplit = g_hostdb.m_indexSplits;
char *xx=NULL;*xx=0;
}
*/
/*
long long singleDocIdQuery = 0LL;
if ( rdbId == RDB_POSDB ) {
long long d1 = g_posdb.getDocId(m_startKey);
long long d2 = g_posdb.getDocId(m_endKey);
if ( d1+1 == d2 ) singleDocIdQuery = d1;
}
// . try the LOCAL termlist cache
// . so when msg2 is evaluating a gbdocid:| query and it has to
// use msg0 to go across the network to get the same damn termlist
// over and over again for the same docid, this will help alot.
// . ideally it'd be nice if the seo pipe in xmldoc.cpp can try to
// send the same gbdocid:xxxx docids to the same hosts. maybe hash
// based on docid into the list of hosts and if that host is busy
// just chain until we find someone not busy.
if ( singleDocIdQuery &&
getListFromTermListCache ( coll,
m_startKey,
m_endKey,
maxCacheAge,
list ) )
// found!
return true;
*/
// but always local if only one host
if ( g_hostdb.getNumHosts() == 1 ) isLocal = true;
// force a msg0 if doing a docid restrictive query like
// gbdocid:xxxx|<query> so we call cacheTermLists()
//if ( singleDocIdQuery ) isLocal = false;
// . if the group is local then do it locally
// . Msg5::getList() returns false if blocked, true otherwise
// . Msg5::getList() sets g_errno on error
// . don't do this if m_hostId was specified
if ( isLocal ) { // && !g_conf.m_interfaceMachine ) {
if ( msg5 ) {
m_msg5 = msg5;
m_deleteMsg5 = false;
}
else {
try { m_msg5 = new ( Msg5 ); }
catch ( ... ) {
g_errno = ENOMEM;
log("net: Local alloc for disk read failed "
"while tring to read data for %s. "
"Trying remote request.",
getDbnameFromId(m_rdbId));
goto skip;
}
mnew ( m_msg5 , sizeof(Msg5) , "Msg0" );
m_deleteMsg5 = true;
}
QUICKPOLL(m_niceness);
// same for msg5b
if ( msg5b ) {
m_msg5b = msg5b;
m_deleteMsg5b = false;
}
/*
else if ( m_rdbId == RDB_TITLEDB ) {
try { m_msg5b = new ( Msg5 ); }
catch ( ... ) {
g_errno = ENOMEM;
log("net: Local alloc for disk read failed "
"while tring to read data for %s. "
"Trying remote request. 2.",
getDbnameFromId(m_rdbId));
goto skip;
}
mnew ( m_msg5b , sizeof(Msg5) , "Msg0b" );
m_deleteMsg5b = true;
}
*/
QUICKPOLL(m_niceness);
if ( ! m_msg5->getList ( rdbId,
coll ,
m_list ,
m_startKey ,
m_endKey ,
m_minRecSizes ,
includeTree , // include Tree?
addToCache , // addToCache?
maxCacheAge ,
startFileNum ,
numFiles ,
this ,
gotListWrapper2 ,
niceness ,
doErrorCorrection ,
NULL , // cacheKeyPtr
0 , // retryNum
-1 , // maxRetries
true , // compensateForMerge
syncPoint ,
NULL,//m_msg5b ,
m_isRealMerge ,
m_allowPageCache ) ) return false;
// nuke it
reset();
return true;
}
skip:
// debug msg
if ( g_conf.m_logDebugQuery )
log(LOG_DEBUG,"net: msg0: Sending request for data to "
"shard=%lu listPtr=%li minRecSizes=%li termId=%llu "
//"startKey.n1=%lx,n0=%llx (niceness=%li)",
"startKey.n1=%llx,n0=%llx (niceness=%li)",
//g_hostdb.makeHostId ( m_groupId ) ,
m_shardNum,
(long)m_list,
m_minRecSizes, g_posdb.getTermId(m_startKey) ,
//m_startKey.n1,m_startKey.n0 , (long)m_niceness);
KEY1(m_startKey,m_ks),KEY0(m_startKey),
(long)m_niceness);
char *replyBuf = NULL;
long replyBufMaxSize = 0;
bool freeReply = true;
// adjust niceness for net transmission
bool realtime = false;
//if ( minRecSizes + 32 < TMPBUFSIZE ) realtime = true;
// if we're niceness 0 we need to pre-allocate for reply since it
// might be received within the asynchronous signal handler which
// cannot call mmalloc()
if ( realtime ) { // niceness <= 0 || netnice == 0 ) {
// . we should not get back more than minRecSizes bytes since
// we are now performing merges
// . it should not slow things down too much since the hashing
// is 10 times slower than merging anyhow...
// . CAUTION: if rdb is not fixed-datasize then this will
// not work for us! it can exceed m_minRecSizes.
replyBufMaxSize = m_minRecSizes ;
// . get a little extra to fix the error where we ask for 64
// but get 72
// . where is that coming from?
// . when getting titleRecs we often exceed the minRecSizes
// . ?Msg8? was having trouble. was short 32 bytes sometimes.
replyBufMaxSize += 36;
// why add ten percent?
//replyBufMaxSize *= 110 ;
//replyBufMaxSize /= 100 ;
// make a buffer to hold the reply
//#ifdef SPLIT_INDEXDB
/*
if ( m_numSplit > 1 ) {
m_replyBufSize = replyBufMaxSize * m_numSplit;
replyBuf = (char *) mmalloc(m_replyBufSize, "Msg0");
m_replyBuf = replyBuf;
freeReply = false;
}
else
*/
//#endif
replyBuf = (char *) mmalloc(replyBufMaxSize , "Msg0");
// g_errno is set and we return true if it failed
if ( ! replyBuf ) {
log("net: Failed to pre-allocate %li bytes to hold "
"data read remotely from %s: %s.",
replyBufMaxSize,getDbnameFromId(m_rdbId),
mstrerror(g_errno));
return true;
}
}
// . make a request with the info above (note: not in network order)
// . IMPORTANT!!!!! if you change this change
// Multicast.cpp::sleepWrapper1 too!!!!!!!!!!!!
// no, not anymore, we commented out that request peeking code
char *p = m_request;
*(long long *) p = syncPoint ; p += 8;
//*(key_t *) p = m_startKey ; p += sizeof(key_t);
//*(key_t *) p = m_endKey ; p += sizeof(key_t);
*(long *) p = m_minRecSizes ; p += 4;
*(long *) p = startFileNum ; p += 4;
*(long *) p = numFiles ; p += 4;
*(long *) p = maxCacheAge ; p += 4;
*p = m_rdbId ; p++;
*p = addToCache ; p++;
*p = doErrorCorrection; p++;
*p = includeTree ; p++;
*p = (char)niceness ; p++;
*p = (char)m_allowPageCache; p++;
KEYSET(p,m_startKey,m_ks); ; p+=m_ks;
KEYSET(p,m_endKey,m_ks); ; p+=m_ks;
// NULL terminated collection name
strcpy ( p , coll ); p += gbstrlen ( coll ); *p++ = '\0';
m_requestSize = p - m_request;
// ask an individual host for this list if hostId is NOT -1
if ( m_hostId != -1 ) {
// get Host
Host *h = g_hostdb.getHost ( m_hostId );
if ( ! h ) {
g_errno = EBADHOSTID;
log(LOG_LOGIC,"net: msg0: Bad hostId of %lli.",
m_hostId);
return true;
}
// if niceness is 0, use the higher priority udpServer
UdpServer *us ;
unsigned short port;
QUICKPOLL(m_niceness);
//if ( niceness <= 0 || netnice == 0 ) {
//if ( realtime ) {
// us = &g_udpServer2; port = h->m_port2; }
//else {
us = &g_udpServer ; port = h->m_port ;
// . returns false on error and sets g_errno, true otherwise
// . calls callback when reply is received (or error)
// . we return true if it returns false
if ( ! us->sendRequest ( m_request ,
m_requestSize ,
0x00 , // msgType
h->m_ip ,
port ,
m_hostId ,
NULL , // the slotPtr
this ,
gotSingleReplyWrapper ,
timeout ,
-1 , // backoff
-1 , // maxwait
replyBuf ,
replyBufMaxSize ,
m_niceness ) ) // cback niceness
return true;
// return false cuz it blocked
return false;
}
// timing debug
if ( g_conf.m_logTimingNet )
m_startTime = gettimeofdayInMilliseconds();
else
m_startTime = 0;
//if ( m_rdbId == RDB_INDEXDB ) log("Msg0:: getting remote indexlist. "
// "termId=%llu, "
// "groupNum=%lu",
// g_indexdb.getTermId(m_startKey) ,
// g_hostdb.makeHostId ( m_groupId ) );
/*
// make the cache key so we can see what remote host cached it, if any
char cacheKey[MAX_KEY_BYTES];
//key_t cacheKey = makeCacheKey ( startKey ,
makeCacheKey ( startKey ,
endKey ,
includeTree ,
minRecSizes ,
startFileNum ,
numFiles ,
cacheKey ,
m_ks );
*/
// . get the top long of the key
// . i guess this will work for 128 bit keys... hmmmmm
long keyTop = hash32 ( (char *)startKey , m_ks );
/*
// allocate space
if ( m_numSplit > 1 ) {
long need = m_numSplit * sizeof(Multicast) ;
char *buf = (char *)mmalloc ( need,"msg0mcast" );
if ( ! buf ) return true;
m_mcasts = (Multicast *)buf;
for ( long i = 0; i < m_numSplit ; i++ )
m_mcasts[i].constructor();
}
*/
// . otherwise, multicast to a host in group "groupId"
// . returns false and sets g_errno on error
// . calls callback on completion
// . select first host to send to in group based on upper 32 bits
// of termId (m_startKey.n1)
//#ifdef SPLIT_INDEXDB
// . need to send out to all the indexdb split hosts
m_numRequests = 0;
m_numReplies = 0;
//for ( long i = 0; i < m_numSplit; i++ ) {
QUICKPOLL(m_niceness);
//long gr;
char *buf;
/*
if ( m_numSplit > 1 ) {
gr = g_indexdb.getSplitGroupId ( baseGroupId, i );
buf = &replyBuf[i*replyBufMaxSize];
}
else {
*/
//gr = m_groupId;
buf = replyBuf;
//}
// get the multicast
Multicast *m = &m_mcast;
//if ( m_numSplit > 1 ) m = &m_mcasts[i];
if ( ! m->send ( m_request ,
//#else
// if ( ! m_mcast.send ( m_request ,
//#endif
m_requestSize,
0x00 , // msgType 0x00
false , // does multicast own request?
m_shardNum ,
//#ifdef SPLIT_INDEXDB
// gr , // group + offset
//#else
// m_groupId , // group to send to (groupKey)
//#endif
false , // send to whole group?
//m_startKey.n1, // key is passed on startKey
keyTop , // key is passed on startKey
this , // state data
NULL , // state data
gotMulticastReplyWrapper0 ,
timeout , // timeout in seconds (was 30)
niceness ,
realtime ,
firstHostId ,
//#ifdef SPLIT_INDEXDB
// &replyBuf[i*replyBufMaxSize] ,
//#else
// replyBuf ,
//#endif
buf ,
replyBufMaxSize ,
freeReply , // free reply buf?
true , // do disk load balancing?
maxCacheAge ,
//(key_t *)cacheKey ,
// multicast uses it for determining the best
// host to send the request to when doing
// disk load balancing. if the host has our
// data cached, then it will probably get to
// handle the request. for now let's just assume
// this is a 96-bit key. TODO: fix...
0 , // *(key_t *)cacheKey ,
rdbId ,
minRecSizes ) ) {
log("net: Failed to send request for data from %s in shard "
"#%lu over network: %s.",
getDbnameFromId(m_rdbId),m_shardNum, mstrerror(g_errno));
// no, multicast will free this when it is destroyed
//if (replyBuf) mfree ( replyBuf , replyBufMaxSize , "Msg22" );
// but speed it up
//#ifdef SPLIT_INDEXDB
m_errno = g_errno;
m->reset();
if ( m_numRequests > 0 )
return false;
//#else
// m_mcast.reset();
//#endif
return true;
}
//#ifdef SPLIT_INDEXDB
m_numRequests++;
//#endif
// we blocked
return false;
}
