static void
ztpfDeriveUcontext( ucontext_t *uscPtr ) {
DIB * pDib = ecbp2()->ce2dib;
DBFHDR * pJdb = (DBFHDR *)(pDib->dibjdb);
DBFITEM * pDbi = NULL;
void * lowcore = NULL;
if( pJdb ) { /* If JDB contents are for this thread */
if( pJdb->ijavcnt ) { /* and the JDB is unlocked */
pDbi = (DBFITEM *)(pJdb+1); /* Point at the JDB base */
lowcore = (pDbi->ijavsbuf); /* and its low core copy. */
}
else { /* Otherwise use what's in the DIB */
lowcore = pDib->dilow; /* Use the low core pointer from DIB */
}
}
/*
* Fill in the uc_mcontext, uc_flags and uc_sigmask fields.
* Leave uc_link alone, it has no meaning in or for z/TPF.
*/
uscPtr->uc_mcontext.psw.mask = pDib->dispw[0]; /* PSW lo half */
uscPtr->uc_mcontext.psw.addr = pDib->dispw[1]; /* PSW hi half */
memcpy( uscPtr->uc_mcontext.gregs,
pDib->direg, sizeof(pDib->direg) ); /*GPRs*/
memcpy( uscPtr->uc_mcontext.fpregs.fprs,
pDib->dfreg, sizeof(pDib->dfreg) ); /*FPRs*/
uscPtr->uc_flags = ecbp2()->ce2thnum; /* Put the TID in ucontext->uc_flags */
sigprocmask( SIG_SETMASK, NULL, &(uscPtr->uc_sigmask) ); /* Our mask is the old mask */
return;
}
void
ztpf_nonpreemptible_helper(void)
{
safeStorage *s = allocateDurableStorage();
uint8_t holdkey;
sigset_t newmask;
sigset_t oldmask;
/*
* Go translate the TPF-ish data available at interrupt time into UNIX-ish
* formatted blocks in preparation to call the master synchronous signal
* handler for the JVM's structured signal handling architecture. This
* storage comes from a static storage segment in DJPR.
*/
s->pDIB = (DIB *) (ecbp2()->ce2dib);
s->argv.dibPtr = s->pDIB;
translateInterruptContexts(&(s->argv));
/*
* Only proceed if (1) the signal is currently NOT blocked, and (2)
* if no other interrupt has queued itself. In the interim, if
* applicable, sleep for 50 ms. IOW, serialize all synchronous
* signals -- we have to, we only have one JDB.
*/
sigemptyset(&newmask); /* Ensure our 'constant' mask word is zeros */
sigaddset(&newmask, s->siginfo.si_signo); /* Build a constant signal mask word */
do {
PROC_LOCK(&(s->pPROC->iproc_JVMLock), holdkey);
if (s->pPROC->iproc_tdibptr) { /* Are we blocked because we're already working */
PROC_UNLOCK(&(s->pPROC->iproc_JVMLock), holdkey); /* another signal? If so,unlock */
usleep(50000); /* and sleep for 50 ms. */
}
pthread_sigmask(SIG_BLOCK, NULL, &oldmask); /* Test the signal mask again. */
} while ((newmask & oldmask) || (s->pPROC->iproc_tdibptr));
/*
* We are released to go process this signal. Do it. The lock we also hold
* on our little subsection of the PROC block is also still in effect.
* But first, stash the DIB pointer in the IPROC block, then unlock it.
*/
s->pPROC->iproc_tdibptr = (uint64_t) s->pDIB; /* Stash the faulting DIB pointer */
PROC_UNLOCK(&(s->pPROC->iproc_JVMLock), holdkey);
/*
* Call the master structured signal handler for synchronous signals.
* It's not possible that an async signal can get here. The signal handler will
* either hand off execution to the JIT, or will run the dump agents. Do not
* expect a return; else the normal SIG_DFL (bye, bye) rule applies.
*/
masterSynchSignalHandler(s->siginfo.si_signo, &(s->siginfo), &(s->ucontext),
s->pDIB->dbrevn);
/*
* If execution returns to this point, the signal handler returned
* here. This means that the thread accepts the consequences of a
* return from the SIG_DFL treatment of a fatal signal, which means
* an end to the process.
*/
exit(0);
/* UNREACHABLE */
}
int os_check_server(char *server) {
#ifndef USE_TPF_DAEMON
int rv;
int *current_acn;
if((rv = inetd_getServerStatus(server)) == INETD_SERVER_STATUS_INACTIVE)
return 1;
else {
current_acn = (int *)cinfc_fast(CINFC_CMMACNUM);
if(ecbp2()->ce2acn != *current_acn)
return 1;
}
#endif
return 0;
}
void
ztpf_preemptible_helper(void)
{
safeStorage *s = allocateDurableStorage();
translateInterruptContexts(&(s->argv));
s->argv.flags = J9TPF_NO_PORT_LIBRARY;
s->pDIB = ecbp2()->ce2dib;
s->argv.dibPtr = s->pDIB;
masterSynchSignalHandler(s->siginfo.si_signo, &(s->siginfo), &(s->ucontext),
s->pDIB->dbrevn);
/*
* Like its "harder" cousin, below, if the signal handler returns here, it is
* acknowledging that it wants to exit. Otherwise, don't expect a return. In
* the normal case, the faulting thread gets returned to the J9 thread pool
* so there's no chance of an accidental return here.
*/
exit(0);
/* UNREACHABLE */
}
/**
* \internal
*
* \details Build the data (in the user-passed <TT>buffer</TT> address)
* block to which an Elf64_Phdr will point. Its role is to extract
* information passed to it in CE2DIB and IJAVDBF and put them in the
* proper places. Sometimes these values require translation to put
* them into proper contextual meaning, for example, a hardware PIC code
* can infer a signal value. As long as we get close enough, it's good.
* The contents of the NOTE section are very poorly documented.
*
* \note The final sub-section, tagged <i><tt>NT_ZTPF_SERRC</tt></i>, is invented
* for this routine's purposes: its role is to express the SE number associated
* with the already-written traditional SERRC dump as requested during design. Its
* existence may require addition of a method to the Java class which processes
* ELF NOTEs in <b><tt>jdmpview</tt></b> - there is no ELF NOTE tag otherwise
* suited for this purpose. Otherwise, this information stays buried.
*
* This function returns the total size of the buffer as its Elf64_Phdr
* must report it.
*
* \param[in] buffer Writeable memory large enough to house the NOTE
* section (see \ref NOTE_TABLE_SIZE, above).
* \param[in] dibPtr Address of the CPSE<->JVM Dump Interchange Block
* belonging to the faulting or requesting Unit Of Work.
* \returns
* Size, in bytes, of the NOTE table section as written.
*/
static uintptr_t
buildELFNoteArea(Elf64_Nhdr *buffer, DIB *dibPtr)
{
uintptr_t tableSize = NOTE_TABLE_SIZE; /* Return value */
/* Pointers to the different required NOTE subsections, in sequential
* order of appearance: PRSTATUS, PRPSINFO, AUXV, FPREGSET, and
* S390_LAST_BREAK. The NOTE subsections are a little bizarre: each has a
* header of 0x0c bytes in length, followed by a displayable label of
* what it is. In general, these headers end up being 0x14 bytes in length
* because their data needs to be fullword-aligned. So, in general, the ptr
* prefixed with "nsh_" is the lower of the two addresses, being the
* Elf64_Nhdr pointer, and the un-prefixed pointer is where the data goes.
* Kinda screwy, but so is the NOTE section itself when there are a number
* of subsections in sequence.
*/
Elf64_Nhdr *nsh_prstat; /* NT_PRSTATUS NOTE header */
prstatus_t *prstat; /* PRSTATUS NOTE data ptr */
Elf64_Nhdr *nsh_prpsin; /* NT_PRPSINFO NOTE header */
prpsinfo_t *prinfo; /* PRPSINFO NOTE data ptr */
Elf64_Nhdr *nsh_auxvxx; /* NT_AUXV NOTE header */
AUXV *auxv; /* NT_AUXV NOTE data ptr */
Elf64_Nhdr *nsh_fpregs; /* NT_FPREGSET NOTE header */
fpregset_t *fpregs; /* NT_FPREGSET data ptr */
Elf64_Nhdr *nsh_lastbr; /* NT_S390_LAST_BREAK NOTE hdr */
uint16_t signalNumber;
siginfo_t wSiginfo;
DIB *holdDibPtr;
DBFHDR *dbfPtr = dibPtr->dibjdb;
/*
* The ERI sits as the first JDB index item. We have to bump its pointer
* past the DBFHDR, which just so happens to be the same length as a DBFITEM
* itself.
*/
DBFITEM *eriRef = (DBFITEM *) dbfPtr + 1;
struct cderi *eriPtr = (struct cderi *) (eriRef->ijavsbuf);
/*
* z/TPF keeps its process information in a unique IPROC block, address it.
*/
struct iproc *pIPROC = iproc_process; /* Ptr to this PID's IPROC block */
struct iproc *pPIPROC; /* Ptr to mom's PID IPROC block */
/*
* The datatypes inside each NOTE section and subsection are so diverse
* as to make the commonly-used typecasting approach cumbersome and difficult
* to understand. Often working pointer fields which have a length of 1 for
* address arithmetic uses are the easiest or best way to do this yet still
* have legible code. We use all three of the following data objects in that
* vein.
*/
uint8_t *wSrc; /* General uint8_t address pointer */
uint8_t *wDest; /* General uint8_t address pointer */
uint8_t *wPtr; /* General uint8_t address pointer */
/*
* Build the NT_PRSTATUS subsection first
*/
KEY0();
memset(buffer, 0, tableSize); /* Clear the entire buffer */
nsh_prstat = buffer; /* Start the data sec hdr ptr. */
prstat = (prstatus_t *) NOTE_DATA_OFFSET(nsh_prstat); /* Point at its data*/
nsh_prstat->n_namesz = 5; /* Size of "CORE" = 5 */
nsh_prstat->n_descsz = sizeof(prstatus_t);
nsh_prstat->n_type = NT_PRSTATUS; /* Type = 1 */
memcpy(nsh_prstat + 1, "CORE", 5); /* Label this one "CORE". */
/*
* The "current signal" value is inferred from the h/w PIC value lookup.
* Unfortunately, ztpfDeriveSiginfo() is coded to use the DIB hung off
* page 2 of the ECB, not the passed DIB. So hold on to the real
* DIB address for this ECB in a safe place, and stuff the passed dibPtr
* into page 2 of the ECB before the call, then replace it after the
* call.
*/
holdDibPtr = ecbp2()->ce2dib; /* Substitute the passed DIB */
ecbp2()->ce2dib = dibPtr; /* for the real one, then ... */
UNKEY(); /* reset the SPK to prob state */
ztpfDeriveSiginfo(&wSiginfo); /* While we go build a long */
KEY0(); /* siginfo_t from which to */
prstat->info.si_signo = wSiginfo.si_signo; /* feed the short one. After */
prstat->info.si_code = wSiginfo.si_code; /* we've returned, get back */
prstat->info.si_errno = wSiginfo.si_errno; /* into supv state prot key. */
prstat->cursig = wSiginfo.si_errno;
ecbp2()->ce2dib = holdDibPtr; /* Replace the real DIB pointer */
/*
* The relative PID values come from the process block .... Just read
* them dirty; don't bother with a lock at first, see if it works before
* we take a chance on losing ctl with a lock
*/
prstat->pid = pIPROC->iproc_pid; /* My PID */
prstat->pgid = pIPROC->iproc_pgid; /* My PGID */
prstat->ppid = pIPROC->iproc_pptr->iproc_pid; /* Mom's PID */
/* No concept of a session ID */
/*
* The System z TOD clock ticks once every 244 picoseconds. To get to accurate
* nanoseconds, we'd need to divide its value by 4.09836. Since we can't
* divide by a fractional quantity without going to floating point (and
* thereby introducing approximation error & burning more cycles), we
* usually trade precision for speed by shifting right 2 bits (dividing by 4)
* to yield nanoseconds from a little less than a quarter of a ns. That's
* considered close enough. But now that we need _micro_second (ns * 1000)
* precision, we'd be amplifying the error. And since we have to divide by 1000
* anyway, we might as well divide by 4098 (would shifting right 12 bits, for
* 4096, be close enough? No. That would cause us to lose a full millisecond in
* precision error!) to be accurate. So that's what we're doing below, since
* we have to eat the cycles in the division operation anyway -- might as well
* use the right divisor. Never mind rounding up with the remainder; we're
* already generating a larger result than reality because we can't go
* fractional. Does it really matter? Nah. But if you're gonna do it, you might
* as well do it right ....
*/
prstat->stime.tv_usec = ecbptr()->ce1istim / 4098; /* Used CPU time */
wDest = (char *) &(prstat->reg); /* Point at offset +0 of reg fld */
/*
* The PGM/O PSW is followed by a 16-wide array of ULONGs which
* represent the values of all 16 general registers, 0 through 15,
* as they were at the time the PGM/N PSW fired. Prepare to copy
* them into the NT_PRSTATUS section, lth=128+16 (144).
*/
wSrc = (char *) (dibPtr->dispw); /* Point @ PSW & general regs */
memcpy(wDest, wSrc, 144); /* Preserve all that data. */
prstat->fpvalid = TRUE; /* Yes, we have floating point */
/*
* NT_PRPSINFO is next
*/
nsh_prpsin = (Elf64_Nhdr *) (prstat + 1); /* Point to next NOTE hdr */
prinfo = (prpsinfo_t *) NOTE_DATA_OFFSET(nsh_prpsin);
nsh_prpsin->n_namesz = 5; /* Size of "CORE" = 5 */
nsh_prpsin->n_descsz = sizeof(prpsinfo_t);
nsh_prpsin->n_type = NT_PRPSINFO; /* Type = 3 */
memcpy(nsh_prpsin + 1, "CORE", 5); /* Set "CORE" name ... */
prinfo->char_state = 'R'; /* We were 'R'unning. */
prinfo->flag = 0x00400440L; /* Flags are constant */
prinfo->uid = pIPROC->iproc_uid; /* User id & */
prinfo->gid = pIPROC->iproc_gid; /* group id. */
memcpy(&prinfo->pid, &prstat->pid, 16); /* Process ID set again */
strcpy(prinfo->fname, "main"); /* Just like Linux does it */
wSrc = getenv("IBM_JAVA_COMMAND_LINE");
if (wSrc) {
strncpy(prinfo->psargs, wSrc, PRARGSZ);
prinfo->psargs[PRARGSZ - 1] = '\0';
}
/*
* NT_AUXV after that ... we fake it (we have no ELF protocols involved
* in task switching and thus have no AUXV structure or meaning).
*/
nsh_auxvxx = (Elf64_Nhdr *) (prinfo + 1); /* Next NOTE header */
auxv = (AUXV *) NOTE_DATA_OFFSET(nsh_auxvxx); /* and data region ptr. */
nsh_auxvxx->n_namesz = 5; /* Size of "CORE" = 5 */
nsh_auxvxx->n_descsz = sizeof(AUXV);
nsh_auxvxx->n_type = NT_AUXV; /* Type = 6 */
memcpy(nsh_auxvxx + 1, "CORE", 5); /* Set "CORE" name ... */
memset(auxv, 0, sizeof(AUXV)); /* Fill it out. */
/*
* ... then NT_FPREGSET
*/
nsh_fpregs = (Elf64_Nhdr *) (auxv + 1); /* Next NOTE header */
fpregs = (fpregset_t *) NOTE_DATA_OFFSET(nsh_fpregs); /* and data ptr */
nsh_fpregs->n_namesz = 5; /* Size of "CORE" = 5 */
nsh_fpregs->n_descsz = sizeof(fpregset_t);
nsh_fpregs->n_type = NT_FPREGSET; /* Type = 2 */
memcpy(nsh_fpregs + 1, "CORE", 5); /* Set "CORE" name ... */
/*
* We don't care what the real FPCR contained. We're setting the mask
* only to state that 16 fprs exist. The real FPCR is too hard to get,
* anyway.
*/
fpregs->fpcr = 0x00800000;
wSrc = (uint8_t *) &(dibPtr->dfreg); /* Get the adrs of the fp */
wDest = (uint8_t *) &(fpregs->fpreg); /* regs, then copy them */
memcpy(wDest, wSrc, sizeof(fpregs->fpreg)); /* into the note area. */
/*
* ... then NT_S390_LAST_BREAK
*/
nsh_lastbr = (Elf64_Nhdr *) (fpregs + 1);
nsh_lastbr->n_namesz = 6; /* Size of "LINUX" = 6 */
nsh_lastbr->n_descsz = 8; /* Always 0x008 bytes. */
nsh_lastbr->n_type = 0x306; /* Type = NT_S390_LAST_BR */
wDest = (uint8_t *) NOTE_DATA_OFFSET(nsh_lastbr);
memcpy(nsh_lastbr + 1, "LINUX", 6); /* Set "LINUX" name */
wSrc = (char *) (dibPtr->dbrevn); /* Get NOTE data source ptr */
memcpy(wDest, wSrc, 8); /* and move it in. */
UNKEY();
return tableSize; /* We're done. */
/*
* ... and finally, NT_ZTPF_SERRC (made up; we'll need to add the Java
* class code to display this)
*/
wPtr += 8;
nsh_lastbr = (Elf64_Nhdr *) (wDest + 8);
nsh_lastbr->n_namesz = 5; /* Size of "ZTPF" = 5 */
nsh_lastbr->n_descsz = 68; /* Always 0x044 bytes. */
nsh_lastbr->n_type = 0x500; /* Type = 500 */
memcpy(nsh_lastbr + 1, "ZTPF", 5); /* Set "ZTPF" name ... */
wDest += 0x14; /* and point at data */
sprintf(wDest, "SE-%4.4d OPR-%c%6.6X PGM-%4.4s %5.5c %8.8s", eriPtr->eriseq,
eriPtr->eridpfx, eriPtr->eridnum, eriPtr->eridate, eriPtr->eritime);
UNKEY();
return tableSize; /* We're done. */
}
void
ztpfDeriveSiginfo( siginfo_t *build ) {
uint16_t pic;
uint16_t ilc;
uint8_t dxc;
sigvpair_t *translated = NULL;
void *lowcore = NULL;
DIB *pDib = ecbp2()->ce2dib;
DBFHDR *pDbf = (DBFHDR *)(pDib->dibjdb);
DBFITEM *pDbi;
struct cderi *pEri = NULL;
struct iproc *procPtr = iproc_process;
/*
* If we have a java dump buffer (ijavdbf) which belongs to this
* thread AND that buffer is not locked, then use it. We prefer
* the low core image in the java dump buffer to the copy in the
* DIB; but use the DIB's copy if there's no buffer.
*/
if( pDbf && pDbf->ijavcnt ) { /* We have the full Error Recording Info, use it. */
pDbi = (DBFITEM *)pDbf+1; /* We have a bunch of void ptrs */
pEri = (struct cderi *)(pDbi->ijavsbuf); /* to cast through ... ugh... */
lowcore = pEri; /* And struct cderi is incomplete ... */
lowcore += (unsigned long int)0x310; /* ai yi yi Lucy. */
}
else { /* All we have is the DIB, so use it instead. */
lowcore = pDib->dilow; /* Fixup low core pointer */
pEri = NULL; /* And indicate there's no full ERI. */
}
pic = s390FetchPIC( lowcore ); /* Translate the PIC into si_signo & si_code */
translated = (sigvpair_t *)&pic2pair[pic];
build->si_signo = translated->si_signo;
build->si_code = translated->si_code;
/*
* This could just as well have been an "if-then-else" construct, but
* it was felt wise to leave it in case testing revealed something else
* that requires analysis.
*/
switch( build->si_signo ) {
case SIGFPE:
if( 0 == build->si_code ) { /* If we need to interpret DXC bits, */
/*
* Returning a good si_code value means we have to interrogate the
* DXC value, so go get a normalized version of it.
*/
dxc = s390FetchDXC( lowcore );
/*
* "Sluice" through the acceptable values of the DXC for known si_code
* values. If we don't find one, call it SI_KERNEL for lack of any
* better idea.
*/
if( 0x09 == dxc )
build->si_code = FPE_INTDIV;
else if( 0x08 == dxc )
build->si_code = FPE_INTOVF;
else if( 0x40 == dxc )
build->si_code = FPE_FLTDIV;
else if( 0x20 == dxc )
build->si_code = FPE_FLTOVF;
else if( 0x10 == dxc )
build->si_code = FPE_FLTUND;
else if( 0x08 == dxc )
build->si_code = FPE_FLTRES;
else if( 0x80 == dxc )
build->si_code = FPE_FLTINV;
else
build->si_code = SI_KERNEL; /* Unknown synchronous signal */
}
break;
/*
* SIGSEGV signals need si_addr set to the addr that failed translation
* It's in the ERI, but not in the DIB. So if we have an ERI, use it.
*/
case SIGSEGV:
if( build->si_code == SEGV_MAPERR && pEri ) {
uint64_t *address;
lowcore = pEri; /* The failing vaddr is at ISVTRXAD */
address = lowcore+(unsigned long int)0x1b8; /* at offset 0x1B8 into the ERI*/
build->si_addr = (void *)*address; /* Move the failed address. */
} /* Otherwise leave it zero. */
break;
case SIGILL:
build->si_addr = (char *)pDib->dispw[1]; /* Failing PSW $IP */
break;
default:
break;
}
build->si_pid = procPtr->iproc_pid; /* All of them need PID and */
build->si_uid = procPtr->iproc_uid; /* UID values, with TID & errno */
build->si_ztpf_tid = ecbp2()->ce2thnum; /* values, too. */
build->si_errno = errno;
return;
}
