void
swi_handler(trapframe_t *frame)
{
struct thread *td = curthread;
td->td_frame = frame;
td->td_pticks = 0;
/*
* Make sure the program counter is correctly aligned so we
* don't take an alignment fault trying to read the opcode.
*/
if (__predict_false(((frame->tf_pc - INSN_SIZE) & 3) != 0)) {
call_trapsignal(td, SIGILL, 0);
userret(td, frame);
return;
}
/*
* Enable interrupts if they were enabled before the exception.
* Since all syscalls *should* come from user mode it will always
* be safe to enable them, but check anyway.
*/
if (td->td_md.md_spinlock_count == 0) {
if (__predict_true(frame->tf_spsr & I32_bit) == 0)
enable_interrupts(I32_bit);
if (__predict_true(frame->tf_spsr & F32_bit) == 0)
enable_interrupts(F32_bit);
}
syscall(td, frame);
}
/*
* abort_debug() handles the following abort:
*
* FAULT_DEBUG - Debug Event
*
*/
static __inline void
abort_debug(struct trapframe *tf, u_int fsr, u_int prefetch, u_int usermode,
u_int far)
{
if (usermode) {
struct thread *td;
td = curthread;
call_trapsignal(td, SIGTRAP, TRAP_BRKPT, far);
userret(td, tf);
} else {
#ifdef KDB
kdb_trap(T_BREAKPOINT, 0, tf);
#else
printf("No debugger in kernel.\n");
#endif
}
}
static void
syscall(struct thread *td, trapframe_t *frame)
{
struct syscall_args sa;
int error;
#ifndef __ARM_EABI__
sa.insn = *(uint32_t *)(frame->tf_pc - INSN_SIZE);
switch (sa.insn & SWI_OS_MASK) {
case 0: /* XXX: we need our own one. */
break;
default:
call_trapsignal(td, SIGILL, 0);
userret(td, frame);
return;
}
#endif
sa.nap = 4;
error = syscallenter(td, &sa);
KASSERT(error != 0 || td->td_ar == NULL,
("returning from syscall with td_ar set!"));
syscallret(td, error, &sa);
}
/*
* void prefetch_abort_handler(struct trapframe *tf)
*
* Abort handler called when instruction execution occurs at
* a non existent or restricted (access permissions) memory page.
* If the address is invalid and we were in SVC mode then panic as
* the kernel should never prefetch abort.
* If the address is invalid and the page is mapped then the user process
* does no have read permission so send it a signal.
* Otherwise fault the page in and try again.
*/
static void
prefetch_abort_handler(struct trapframe *tf)
{
struct thread *td;
struct proc * p;
struct vm_map *map;
vm_offset_t fault_pc, va;
int error = 0;
struct ksig ksig;
#if 0
/* Update vmmeter statistics */
uvmexp.traps++;
#endif
#if 0
printf("prefetch abort handler: %p %p\n", (void*)tf->tf_pc,
(void*)tf->tf_usr_lr);
#endif
td = curthread;
p = td->td_proc;
PCPU_INC(cnt.v_trap);
if (TRAP_USERMODE(tf)) {
td->td_frame = tf;
if (td->td_cowgen != td->td_proc->p_cowgen)
thread_cow_update(td);
}
fault_pc = tf->tf_pc;
if (td->td_md.md_spinlock_count == 0) {
if (__predict_true(tf->tf_spsr & PSR_I) == 0)
enable_interrupts(PSR_I);
if (__predict_true(tf->tf_spsr & PSR_F) == 0)
enable_interrupts(PSR_F);
}
/* Prefetch aborts cannot happen in kernel mode */
if (__predict_false(!TRAP_USERMODE(tf)))
dab_fatal(tf, 0, tf->tf_pc, NULL, &ksig);
td->td_pticks = 0;
/* Ok validate the address, can only execute in USER space */
if (__predict_false(fault_pc >= VM_MAXUSER_ADDRESS ||
(fault_pc < VM_MIN_ADDRESS && vector_page == ARM_VECTORS_LOW))) {
ksig.signb = SIGSEGV;
ksig.code = 0;
goto do_trapsignal;
}
map = &td->td_proc->p_vmspace->vm_map;
va = trunc_page(fault_pc);
/*
* See if the pmap can handle this fault on its own...
*/
#ifdef DEBUG
last_fault_code = -1;
#endif
if (pmap_fault_fixup(map->pmap, va, VM_PROT_READ, 1))
goto out;
error = vm_fault(map, va, VM_PROT_READ | VM_PROT_EXECUTE,
VM_FAULT_NORMAL);
if (__predict_true(error == 0))
goto out;
if (error == ENOMEM) {
printf("VM: pid %d (%s), uid %d killed: "
"out of swap\n", td->td_proc->p_pid, td->td_name,
(td->td_proc->p_ucred) ?
td->td_proc->p_ucred->cr_uid : -1);
ksig.signb = SIGKILL;
} else {
ksig.signb = SIGSEGV;
}
ksig.code = 0;
do_trapsignal:
call_trapsignal(td, ksig.signb, ksig.code);
out:
userret(td, tf);
}
void
abort_handler(struct trapframe *tf, int type)
{
struct vm_map *map;
struct pcb *pcb;
struct thread *td;
u_int user, far, fsr;
vm_prot_t ftype;
void *onfault;
vm_offset_t va;
int error = 0;
struct ksig ksig;
struct proc *p;
if (type == 1)
return (prefetch_abort_handler(tf));
/* Grab FAR/FSR before enabling interrupts */
far = cpu_faultaddress();
fsr = cpu_faultstatus();
#if 0
printf("data abort: fault address=%p (from pc=%p lr=%p)\n",
(void*)far, (void*)tf->tf_pc, (void*)tf->tf_svc_lr);
#endif
/* Update vmmeter statistics */
#if 0
vmexp.traps++;
#endif
td = curthread;
p = td->td_proc;
PCPU_INC(cnt.v_trap);
/* Data abort came from user mode? */
user = TRAP_USERMODE(tf);
if (user) {
td->td_pticks = 0;
td->td_frame = tf;
if (td->td_cowgen != td->td_proc->p_cowgen)
thread_cow_update(td);
}
/* Grab the current pcb */
pcb = td->td_pcb;
/* Re-enable interrupts if they were enabled previously */
if (td->td_md.md_spinlock_count == 0) {
if (__predict_true(tf->tf_spsr & PSR_I) == 0)
enable_interrupts(PSR_I);
if (__predict_true(tf->tf_spsr & PSR_F) == 0)
enable_interrupts(PSR_F);
}
/* Invoke the appropriate handler, if necessary */
if (__predict_false(data_aborts[fsr & FAULT_TYPE_MASK].func != NULL)) {
if ((data_aborts[fsr & FAULT_TYPE_MASK].func)(tf, fsr, far,
td, &ksig)) {
goto do_trapsignal;
}
goto out;
}
/*
* At this point, we're dealing with one of the following data aborts:
*
* FAULT_TRANS_S - Translation -- Section
* FAULT_TRANS_P - Translation -- Page
* FAULT_DOMAIN_S - Domain -- Section
* FAULT_DOMAIN_P - Domain -- Page
* FAULT_PERM_S - Permission -- Section
* FAULT_PERM_P - Permission -- Page
*
* These are the main virtual memory-related faults signalled by
* the MMU.
*/
/*
* Make sure the Program Counter is sane. We could fall foul of
* someone executing Thumb code, in which case the PC might not
* be word-aligned. This would cause a kernel alignment fault
* further down if we have to decode the current instruction.
* XXX: It would be nice to be able to support Thumb at some point.
*/
if (__predict_false((tf->tf_pc & 3) != 0)) {
if (user) {
/*
* Give the user an illegal instruction signal.
*/
/* Deliver a SIGILL to the process */
ksig.signb = SIGILL;
ksig.code = 0;
goto do_trapsignal;
}
/*
* The kernel never executes Thumb code.
*/
printf("\ndata_abort_fault: Misaligned Kernel-mode "
"Program Counter\n");
dab_fatal(tf, fsr, far, td, &ksig);
}
va = trunc_page((vm_offset_t)far);
/*
* It is only a kernel address space fault iff:
* 1. user == 0 and
* 2. pcb_onfault not set or
* 3. pcb_onfault set and not LDRT/LDRBT/STRT/STRBT instruction.
*/
if (user == 0 && (va >= VM_MIN_KERNEL_ADDRESS ||
(va < VM_MIN_ADDRESS && vector_page == ARM_VECTORS_LOW)) &&
__predict_true((pcb->pcb_onfault == NULL ||
(ReadWord(tf->tf_pc) & 0x05200000) != 0x04200000))) {
map = kernel_map;
/* Was the fault due to the FPE/IPKDB ? */
if (__predict_false((tf->tf_spsr & PSR_MODE)==PSR_UND32_MODE)) {
/*
* Force exit via userret()
* This is necessary as the FPE is an extension to
* userland that actually runs in a priveledged mode
* but uses USR mode permissions for its accesses.
*/
user = 1;
ksig.signb = SIGSEGV;
ksig.code = 0;
goto do_trapsignal;
}
} else {
map = &td->td_proc->p_vmspace->vm_map;
}
/*
* We need to know whether the page should be mapped as R or R/W.
* On armv4, the fault status register does not indicate whether
* the access was a read or write. We know that a permission fault
* can only be the result of a write to a read-only location, so we
* can deal with those quickly. Otherwise we need to disassemble
* the faulting instruction to determine if it was a write.
*/
if (IS_PERMISSION_FAULT(fsr))
ftype = VM_PROT_WRITE;
else {
u_int insn = ReadWord(tf->tf_pc);
if (((insn & 0x0c100000) == 0x04000000) || /* STR/STRB */
((insn & 0x0e1000b0) == 0x000000b0) || /* STRH/STRD */
((insn & 0x0a100000) == 0x08000000)) { /* STM/CDT */
ftype = VM_PROT_WRITE;
} else {
if ((insn & 0x0fb00ff0) == 0x01000090) /* SWP */
ftype = VM_PROT_READ | VM_PROT_WRITE;
else
ftype = VM_PROT_READ;
}
}
/*
* See if the fault is as a result of ref/mod emulation,
* or domain mismatch.
*/
#ifdef DEBUG
last_fault_code = fsr;
#endif
if (td->td_critnest != 0 || WITNESS_CHECK(WARN_SLEEPOK | WARN_GIANTOK,
NULL, "Kernel page fault") != 0)
goto fatal_pagefault;
if (pmap_fault_fixup(vmspace_pmap(td->td_proc->p_vmspace), va, ftype,
user)) {
goto out;
}
onfault = pcb->pcb_onfault;
pcb->pcb_onfault = NULL;
error = vm_fault(map, va, ftype, VM_FAULT_NORMAL);
pcb->pcb_onfault = onfault;
if (__predict_true(error == 0))
goto out;
fatal_pagefault:
if (user == 0) {
if (pcb->pcb_onfault) {
tf->tf_r0 = error;
tf->tf_pc = (register_t)(intptr_t) pcb->pcb_onfault;
return;
}
printf("\nvm_fault(%p, %x, %x, 0) -> %x\n", map, va, ftype,
error);
dab_fatal(tf, fsr, far, td, &ksig);
}
if (error == ENOMEM) {
printf("VM: pid %d (%s), uid %d killed: "
"out of swap\n", td->td_proc->p_pid, td->td_name,
(td->td_proc->p_ucred) ?
td->td_proc->p_ucred->cr_uid : -1);
ksig.signb = SIGKILL;
} else {
ksig.signb = SIGSEGV;
}
ksig.code = 0;
do_trapsignal:
call_trapsignal(td, ksig.signb, ksig.code);
out:
/* If returning to user mode, make sure to invoke userret() */
if (user)
userret(td, tf);
}
/*
* void prefetch_abort_handler(trapframe_t *tf)
*
* Abort handler called when instruction execution occurs at
* a non existent or restricted (access permissions) memory page.
* If the address is invalid and we were in SVC mode then panic as
* the kernel should never prefetch abort.
* If the address is invalid and the page is mapped then the user process
* does no have read permission so send it a signal.
* Otherwise fault the page in and try again.
*/
void
prefetch_abort_handler(trapframe_t *tf)
{
struct thread *td;
struct proc * p;
struct vm_map *map;
vm_offset_t fault_pc, va;
int error = 0;
struct ksig ksig;
#if 0
/* Update vmmeter statistics */
uvmexp.traps++;
#endif
#if 0
printf("prefetch abort handler: %p %p\n", (void*)tf->tf_pc,
(void*)tf->tf_usr_lr);
#endif
td = curthread;
p = td->td_proc;
PCPU_INC(cnt.v_trap);
if (TRAP_USERMODE(tf)) {
td->td_frame = tf;
if (td->td_ucred != td->td_proc->p_ucred)
cred_update_thread(td);
}
fault_pc = tf->tf_pc;
if (td->td_md.md_spinlock_count == 0) {
if (__predict_true(tf->tf_spsr & I32_bit) == 0)
enable_interrupts(I32_bit);
if (__predict_true(tf->tf_spsr & F32_bit) == 0)
enable_interrupts(F32_bit);
}
/* See if the cpu state needs to be fixed up */
switch (prefetch_abort_fixup(tf, &ksig)) {
case ABORT_FIXUP_RETURN:
return;
case ABORT_FIXUP_FAILED:
/* Deliver a SIGILL to the process */
ksig.signb = SIGILL;
ksig.code = 0;
td->td_frame = tf;
goto do_trapsignal;
default:
break;
}
/* Prefetch aborts cannot happen in kernel mode */
if (__predict_false(!TRAP_USERMODE(tf)))
dab_fatal(tf, 0, tf->tf_pc, NULL, &ksig);
td->td_pticks = 0;
/* Ok validate the address, can only execute in USER space */
if (__predict_false(fault_pc >= VM_MAXUSER_ADDRESS ||
(fault_pc < VM_MIN_ADDRESS && vector_page == ARM_VECTORS_LOW))) {
ksig.signb = SIGSEGV;
ksig.code = 0;
goto do_trapsignal;
}
map = &td->td_proc->p_vmspace->vm_map;
va = trunc_page(fault_pc);
/*
* See if the pmap can handle this fault on its own...
*/
#ifdef DEBUG
last_fault_code = -1;
#endif
if (pmap_fault_fixup(map->pmap, va, VM_PROT_READ, 1))
goto out;
if (map != kernel_map) {
PROC_LOCK(p);
p->p_lock++;
PROC_UNLOCK(p);
}
error = vm_fault(map, va, VM_PROT_READ | VM_PROT_EXECUTE,
VM_FAULT_NORMAL);
if (map != kernel_map) {
PROC_LOCK(p);
p->p_lock--;
PROC_UNLOCK(p);
}
if (__predict_true(error == 0))
goto out;
if (error == ENOMEM) {
printf("VM: pid %d (%s), uid %d killed: "
"out of swap\n", td->td_proc->p_pid, td->td_name,
(td->td_proc->p_ucred) ?
td->td_proc->p_ucred->cr_uid : -1);
ksig.signb = SIGKILL;
} else {
ksig.signb = SIGSEGV;
}
ksig.code = 0;
do_trapsignal:
call_trapsignal(td, ksig.signb, ksig.code);
out:
userret(td, tf);
}
/*
* Abort handler.
*
* FAR, FSR, and everything what can be lost after enabling
* interrupts must be grabbed before the interrupts will be
* enabled. Note that when interrupts will be enabled, we
* could even migrate to another CPU ...
*
* TODO: move quick cases to ASM
*/
void
abort_handler(struct trapframe *tf, int prefetch)
{
struct thread *td;
vm_offset_t far, va;
int idx, rv;
uint32_t fsr;
struct ksig ksig;
struct proc *p;
struct pcb *pcb;
struct vm_map *map;
struct vmspace *vm;
vm_prot_t ftype;
bool usermode;
#ifdef INVARIANTS
void *onfault;
#endif
td = curthread;
fsr = (prefetch) ? cp15_ifsr_get(): cp15_dfsr_get();
#if __ARM_ARCH >= 7
far = (prefetch) ? cp15_ifar_get() : cp15_dfar_get();
#else
far = (prefetch) ? TRAPF_PC(tf) : cp15_dfar_get();
#endif
idx = FSR_TO_FAULT(fsr);
usermode = TRAPF_USERMODE(tf); /* Abort came from user mode? */
if (usermode)
td->td_frame = tf;
CTR6(KTR_TRAP, "%s: fsr %#x (idx %u) far %#x prefetch %u usermode %d",
__func__, fsr, idx, far, prefetch, usermode);
/*
* Firstly, handle aborts that are not directly related to mapping.
*/
if (__predict_false(idx == FAULT_EA_IMPREC)) {
abort_imprecise(tf, fsr, prefetch, usermode);
return;
}
if (__predict_false(idx == FAULT_DEBUG)) {
abort_debug(tf, fsr, prefetch, usermode, far);
return;
}
/*
* ARM has a set of unprivileged load and store instructions
* (LDRT/LDRBT/STRT/STRBT ...) which are supposed to be used in other
* than user mode and OS should recognize their aborts and behave
* appropriately. However, there is no way how to do that reasonably
* in general unless we restrict the handling somehow.
*
* For now, these instructions are used only in copyin()/copyout()
* like functions where usermode buffers are checked in advance that
* they are not from KVA space. Thus, no action is needed here.
*/
#ifdef ARM_NEW_PMAP
rv = pmap_fault(PCPU_GET(curpmap), far, fsr, idx, usermode);
if (rv == 0) {
return;
} else if (rv == EFAULT) {
call_trapsignal(td, SIGSEGV, SEGV_MAPERR, far);
userret(td, tf);
return;
}
#endif
/*
* Now, when we handled imprecise and debug aborts, the rest of
* aborts should be really related to mapping.
*/
PCPU_INC(cnt.v_trap);
#ifdef KDB
if (kdb_active) {
kdb_reenter();
goto out;
}
#endif
if (__predict_false((td->td_pflags & TDP_NOFAULTING) != 0)) {
/*
* Due to both processor errata and lazy TLB invalidation when
* access restrictions are removed from virtual pages, memory
* accesses that are allowed by the physical mapping layer may
* nonetheless cause one spurious page fault per virtual page.
* When the thread is executing a "no faulting" section that
* is bracketed by vm_fault_{disable,enable}_pagefaults(),
* every page fault is treated as a spurious page fault,
* unless it accesses the same virtual address as the most
* recent page fault within the same "no faulting" section.
*/
if (td->td_md.md_spurflt_addr != far ||
(td->td_pflags & TDP_RESETSPUR) != 0) {
td->td_md.md_spurflt_addr = far;
td->td_pflags &= ~TDP_RESETSPUR;
tlb_flush_local(far & ~PAGE_MASK);
return;
}
} else {
/*
* If we get a page fault while in a critical section, then
* it is most likely a fatal kernel page fault. The kernel
* is already going to panic trying to get a sleep lock to
* do the VM lookup, so just consider it a fatal trap so the
* kernel can print out a useful trap message and even get
* to the debugger.
*
* If we get a page fault while holding a non-sleepable
* lock, then it is most likely a fatal kernel page fault.
* If WITNESS is enabled, then it's going to whine about
* bogus LORs with various VM locks, so just skip to the
* fatal trap handling directly.
*/
if (td->td_critnest != 0 ||
WITNESS_CHECK(WARN_SLEEPOK | WARN_GIANTOK, NULL,
"Kernel page fault") != 0) {
abort_fatal(tf, idx, fsr, far, prefetch, td, &ksig);
return;
}
}
/* Re-enable interrupts if they were enabled previously. */
if (td->td_md.md_spinlock_count == 0) {
if (__predict_true(tf->tf_spsr & PSR_I) == 0)
enable_interrupts(PSR_I);
if (__predict_true(tf->tf_spsr & PSR_F) == 0)
enable_interrupts(PSR_F);
}
p = td->td_proc;
if (usermode) {
td->td_pticks = 0;
if (td->td_cowgen != p->p_cowgen)
thread_cow_update(td);
}
/* Invoke the appropriate handler, if necessary. */
if (__predict_false(aborts[idx].func != NULL)) {
if ((aborts[idx].func)(tf, idx, fsr, far, prefetch, td, &ksig))
goto do_trapsignal;
goto out;
}
/*
* Don't pass faulting cache operation to vm_fault(). We don't want
* to handle all vm stuff at this moment.
*/
pcb = td->td_pcb;
if (__predict_false(pcb->pcb_onfault == cachebailout)) {
tf->tf_r0 = far; /* return failing address */
tf->tf_pc = (register_t)pcb->pcb_onfault;
return;
}
/* Handle remaining I-cache aborts. */
if (idx == FAULT_ICACHE) {
if (abort_icache(tf, idx, fsr, far, prefetch, td, &ksig))
goto do_trapsignal;
goto out;
}
/*
* At this point, we're dealing with one of the following aborts:
*
* FAULT_TRAN_xx - Translation
* FAULT_PERM_xx - Permission
*
* These are the main virtual memory-related faults signalled by
* the MMU.
*/
/* fusubailout is used by [fs]uswintr to avoid page faulting. */
pcb = td->td_pcb;
if (__predict_false(pcb->pcb_onfault == fusubailout)) {
tf->tf_r0 = EFAULT;
tf->tf_pc = (register_t)pcb->pcb_onfault;
return;
}
va = trunc_page(far);
if (va >= KERNBASE) {
/*
* Don't allow user-mode faults in kernel address space.
*/
if (usermode)
goto nogo;
map = kernel_map;
} else {
/*
* This is a fault on non-kernel virtual memory. If curproc
* is NULL or curproc->p_vmspace is NULL the fault is fatal.
*/
vm = (p != NULL) ? p->p_vmspace : NULL;
if (vm == NULL)
goto nogo;
map = &vm->vm_map;
if (!usermode && (td->td_intr_nesting_level != 0 ||
pcb->pcb_onfault == NULL)) {
abort_fatal(tf, idx, fsr, far, prefetch, td, &ksig);
return;
}
}
ftype = (fsr & FSR_WNR) ? VM_PROT_WRITE : VM_PROT_READ;
if (prefetch)
ftype |= VM_PROT_EXECUTE;
#ifdef DEBUG
last_fault_code = fsr;
#endif
#ifndef ARM_NEW_PMAP
if (pmap_fault_fixup(vmspace_pmap(td->td_proc->p_vmspace), va, ftype,
usermode)) {
goto out;
}
#endif
#ifdef INVARIANTS
onfault = pcb->pcb_onfault;
pcb->pcb_onfault = NULL;
#endif
/* Fault in the page. */
rv = vm_fault(map, va, ftype, VM_FAULT_NORMAL);
#ifdef INVARIANTS
pcb->pcb_onfault = onfault;
#endif
if (__predict_true(rv == KERN_SUCCESS))
goto out;
nogo:
if (!usermode) {
if (td->td_intr_nesting_level == 0 &&
pcb->pcb_onfault != NULL) {
tf->tf_r0 = rv;
tf->tf_pc = (int)pcb->pcb_onfault;
return;
}
CTR2(KTR_TRAP, "%s: vm_fault() failed with %d", __func__, rv);
abort_fatal(tf, idx, fsr, far, prefetch, td, &ksig);
return;
}
ksig.sig = SIGSEGV;
ksig.code = (rv == KERN_PROTECTION_FAILURE) ? SEGV_ACCERR : SEGV_MAPERR;
ksig.addr = far;
do_trapsignal:
call_trapsignal(td, ksig.sig, ksig.code, ksig.addr);
out:
if (usermode)
userret(td, tf);
}
/*
* void prefetch_abort_handler(trapframe_t *tf)
*
* Abort handler called when instruction execution occurs at
* a non existent or restricted (access permissions) memory page.
* If the address is invalid and we were in SVC mode then panic as
* the kernel should never prefetch abort.
* If the address is invalid and the page is mapped then the user process
* does no have read permission so send it a signal.
* Otherwise fault the page in and try again.
*/
void
prefetch_abort_handler(trapframe_t *tf)
{
struct lwp *l;
struct pcb *pcb __diagused;
struct vm_map *map;
vaddr_t fault_pc, va;
ksiginfo_t ksi;
int error, user;
UVMHIST_FUNC(__func__);
UVMHIST_CALLED(maphist);
/* Update vmmeter statistics */
curcpu()->ci_data.cpu_ntrap++;
l = curlwp;
pcb = lwp_getpcb(l);
if ((user = TRAP_USERMODE(tf)) != 0)
LWP_CACHE_CREDS(l, l->l_proc);
/*
* Enable IRQ's (disabled by the abort) This always comes
* from user mode so we know interrupts were not disabled.
* But we check anyway.
*/
KASSERT(!TRAP_USERMODE(tf) || (tf->tf_spsr & IF32_bits) == 0);
if (__predict_true((tf->tf_spsr & I32_bit) != IF32_bits))
restore_interrupts(tf->tf_spsr & IF32_bits);
/* See if the CPU state needs to be fixed up */
switch (prefetch_abort_fixup(tf)) {
case ABORT_FIXUP_RETURN:
KASSERT(!TRAP_USERMODE(tf) || (tf->tf_spsr & IF32_bits) == 0);
return;
case ABORT_FIXUP_FAILED:
/* Deliver a SIGILL to the process */
KSI_INIT_TRAP(&ksi);
ksi.ksi_signo = SIGILL;
ksi.ksi_code = ILL_ILLOPC;
ksi.ksi_addr = (uint32_t *)(intptr_t) tf->tf_pc;
lwp_settrapframe(l, tf);
goto do_trapsignal;
default:
break;
}
/* Prefetch aborts cannot happen in kernel mode */
if (__predict_false(!user))
dab_fatal(tf, 0, tf->tf_pc, NULL, NULL);
/* Get fault address */
fault_pc = tf->tf_pc;
lwp_settrapframe(l, tf);
UVMHIST_LOG(maphist, " (pc=0x%x, l=0x%x, tf=0x%x)",
fault_pc, l, tf, 0);
/* Ok validate the address, can only execute in USER space */
if (__predict_false(fault_pc >= VM_MAXUSER_ADDRESS ||
(fault_pc < VM_MIN_ADDRESS && vector_page == ARM_VECTORS_LOW))) {
KSI_INIT_TRAP(&ksi);
ksi.ksi_signo = SIGSEGV;
ksi.ksi_code = SEGV_ACCERR;
ksi.ksi_addr = (uint32_t *)(intptr_t) fault_pc;
ksi.ksi_trap = fault_pc;
goto do_trapsignal;
}
map = &l->l_proc->p_vmspace->vm_map;
va = trunc_page(fault_pc);
/*
* See if the pmap can handle this fault on its own...
*/
#ifdef DEBUG
last_fault_code = -1;
#endif
if (pmap_fault_fixup(map->pmap, va, VM_PROT_READ|VM_PROT_EXECUTE, 1)) {
UVMHIST_LOG (maphist, " <- emulated", 0, 0, 0, 0);
goto out;
}
#ifdef DIAGNOSTIC
if (__predict_false(curcpu()->ci_intr_depth > 0)) {
printf("\nNon-emulated prefetch abort with intr_depth > 0\n");
dab_fatal(tf, 0, tf->tf_pc, NULL, NULL);
}
#endif
KASSERT(pcb->pcb_onfault == NULL);
error = uvm_fault(map, va, VM_PROT_READ|VM_PROT_EXECUTE);
if (__predict_true(error == 0)) {
UVMHIST_LOG (maphist, " <- uvm", 0, 0, 0, 0);
goto out;
}
KSI_INIT_TRAP(&ksi);
UVMHIST_LOG (maphist, " <- fatal (%d)", error, 0, 0, 0);
if (error == ENOMEM) {
printf("UVM: pid %d (%s), uid %d killed: "
"out of swap\n", l->l_proc->p_pid, l->l_proc->p_comm,
l->l_cred ? kauth_cred_geteuid(l->l_cred) : -1);
ksi.ksi_signo = SIGKILL;
} else
ksi.ksi_signo = SIGSEGV;
ksi.ksi_code = SEGV_MAPERR;
ksi.ksi_addr = (uint32_t *)(intptr_t) fault_pc;
ksi.ksi_trap = fault_pc;
do_trapsignal:
call_trapsignal(l, tf, &ksi);
out:
KASSERT(!TRAP_USERMODE(tf) || (tf->tf_spsr & IF32_bits) == 0);
userret(l);
}
void
data_abort_handler(trapframe_t *tf)
{
struct vm_map *map;
struct lwp * const l = curlwp;
struct cpu_info * const ci = curcpu();
u_int far, fsr;
vm_prot_t ftype;
void *onfault;
vaddr_t va;
int error;
ksiginfo_t ksi;
UVMHIST_FUNC(__func__);
UVMHIST_CALLED(maphist);
/* Grab FAR/FSR before enabling interrupts */
far = cpu_faultaddress();
fsr = cpu_faultstatus();
/* Update vmmeter statistics */
ci->ci_data.cpu_ntrap++;
/* Re-enable interrupts if they were enabled previously */
KASSERT(!TRAP_USERMODE(tf) || (tf->tf_spsr & IF32_bits) == 0);
if (__predict_true((tf->tf_spsr & IF32_bits) != IF32_bits))
restore_interrupts(tf->tf_spsr & IF32_bits);
/* Get the current lwp structure */
UVMHIST_LOG(maphist, " (l=%#x, far=%#x, fsr=%#x",
l, far, fsr, 0);
UVMHIST_LOG(maphist, " tf=%#x, pc=%#x)",
tf, tf->tf_pc, 0, 0);
/* Data abort came from user mode? */
bool user = (TRAP_USERMODE(tf) != 0);
if (user)
LWP_CACHE_CREDS(l, l->l_proc);
/* Grab the current pcb */
struct pcb * const pcb = lwp_getpcb(l);
curcpu()->ci_abt_evs[fsr & FAULT_TYPE_MASK].ev_count++;
/* Invoke the appropriate handler, if necessary */
if (__predict_false(data_aborts[fsr & FAULT_TYPE_MASK].func != NULL)) {
#ifdef DIAGNOSTIC
printf("%s: data_aborts fsr=0x%x far=0x%x\n",
__func__, fsr, far);
#endif
if ((data_aborts[fsr & FAULT_TYPE_MASK].func)(tf, fsr, far,
l, &ksi))
goto do_trapsignal;
goto out;
}
/*
* At this point, we're dealing with one of the following data aborts:
*
* FAULT_TRANS_S - Translation -- Section
* FAULT_TRANS_P - Translation -- Page
* FAULT_DOMAIN_S - Domain -- Section
* FAULT_DOMAIN_P - Domain -- Page
* FAULT_PERM_S - Permission -- Section
* FAULT_PERM_P - Permission -- Page
*
* These are the main virtual memory-related faults signalled by
* the MMU.
*/
/* fusubailout is used by [fs]uswintr to avoid page faulting */
if (__predict_false(pcb->pcb_onfault == fusubailout)) {
tf->tf_r0 = EFAULT;
tf->tf_pc = (intptr_t) pcb->pcb_onfault;
return;
}
if (user) {
lwp_settrapframe(l, tf);
}
/*
* Make sure the Program Counter is sane. We could fall foul of
* someone executing Thumb code, in which case the PC might not
* be word-aligned. This would cause a kernel alignment fault
* further down if we have to decode the current instruction.
*/
#ifdef THUMB_CODE
/*
* XXX: It would be nice to be able to support Thumb in the kernel
* at some point.
*/
if (__predict_false(!user && (tf->tf_pc & 3) != 0)) {
printf("\n%s: Misaligned Kernel-mode Program Counter\n",
__func__);
dab_fatal(tf, fsr, far, l, NULL);
}
#else
if (__predict_false((tf->tf_pc & 3) != 0)) {
if (user) {
/*
* Give the user an illegal instruction signal.
*/
/* Deliver a SIGILL to the process */
KSI_INIT_TRAP(&ksi);
ksi.ksi_signo = SIGILL;
ksi.ksi_code = ILL_ILLOPC;
ksi.ksi_addr = (uint32_t *)(intptr_t) far;
ksi.ksi_trap = fsr;
goto do_trapsignal;
}
/*
* The kernel never executes Thumb code.
*/
printf("\n%s: Misaligned Kernel-mode Program Counter\n",
__func__);
dab_fatal(tf, fsr, far, l, NULL);
}
#endif
/* See if the CPU state needs to be fixed up */
switch (data_abort_fixup(tf, fsr, far, l)) {
case ABORT_FIXUP_RETURN:
return;
case ABORT_FIXUP_FAILED:
/* Deliver a SIGILL to the process */
KSI_INIT_TRAP(&ksi);
ksi.ksi_signo = SIGILL;
ksi.ksi_code = ILL_ILLOPC;
ksi.ksi_addr = (uint32_t *)(intptr_t) far;
ksi.ksi_trap = fsr;
goto do_trapsignal;
default:
break;
}
va = trunc_page((vaddr_t)far);
/*
* It is only a kernel address space fault iff:
* 1. user == 0 and
* 2. pcb_onfault not set or
* 3. pcb_onfault set and not LDRT/LDRBT/STRT/STRBT instruction.
*/
if (!user && (va >= VM_MIN_KERNEL_ADDRESS ||
(va < VM_MIN_ADDRESS && vector_page == ARM_VECTORS_LOW)) &&
__predict_true((pcb->pcb_onfault == NULL ||
(read_insn(tf->tf_pc, false) & 0x05200000) != 0x04200000))) {
map = kernel_map;
/* Was the fault due to the FPE/IPKDB ? */
if (__predict_false((tf->tf_spsr & PSR_MODE)==PSR_UND32_MODE)) {
KSI_INIT_TRAP(&ksi);
ksi.ksi_signo = SIGSEGV;
ksi.ksi_code = SEGV_ACCERR;
ksi.ksi_addr = (uint32_t *)(intptr_t) far;
ksi.ksi_trap = fsr;
/*
* Force exit via userret()
* This is necessary as the FPE is an extension to
* userland that actually runs in a priveledged mode
* but uses USR mode permissions for its accesses.
*/
user = true;
goto do_trapsignal;
}
} else {
map = &l->l_proc->p_vmspace->vm_map;
}
/*
* We need to know whether the page should be mapped as R or R/W.
* Before ARMv6, the MMU did not give us the info as to whether the
* fault was caused by a read or a write.
*
* However, we know that a permission fault can only be the result of
* a write to a read-only location, so we can deal with those quickly.
*
* Otherwise we need to disassemble the instruction responsible to
* determine if it was a write.
*/
if (CPU_IS_ARMV6_P() || CPU_IS_ARMV7_P()) {
ftype = (fsr & FAULT_WRITE) ? VM_PROT_WRITE : VM_PROT_READ;
} else if (IS_PERMISSION_FAULT(fsr)) {
ftype = VM_PROT_WRITE;
} else {
#ifdef THUMB_CODE
/* Fast track the ARM case. */
if (__predict_false(tf->tf_spsr & PSR_T_bit)) {
u_int insn = read_thumb_insn(tf->tf_pc, user);
u_int insn_f8 = insn & 0xf800;
u_int insn_fe = insn & 0xfe00;
if (insn_f8 == 0x6000 || /* STR(1) */
insn_f8 == 0x7000 || /* STRB(1) */
insn_f8 == 0x8000 || /* STRH(1) */
insn_f8 == 0x9000 || /* STR(3) */
insn_f8 == 0xc000 || /* STM */
insn_fe == 0x5000 || /* STR(2) */
insn_fe == 0x5200 || /* STRH(2) */
insn_fe == 0x5400) /* STRB(2) */
ftype = VM_PROT_WRITE;
else
ftype = VM_PROT_READ;
}
else
#endif
{
u_int insn = read_insn(tf->tf_pc, user);
if (((insn & 0x0c100000) == 0x04000000) || /* STR[B] */
((insn & 0x0e1000b0) == 0x000000b0) || /* STR[HD]*/
((insn & 0x0a100000) == 0x08000000) || /* STM/CDT*/
((insn & 0x0f9000f0) == 0x01800090)) /* STREX[BDH] */
ftype = VM_PROT_WRITE;
else if ((insn & 0x0fb00ff0) == 0x01000090)/* SWP */
ftype = VM_PROT_READ | VM_PROT_WRITE;
else
ftype = VM_PROT_READ;
}
}
/*
* See if the fault is as a result of ref/mod emulation,
* or domain mismatch.
*/
#ifdef DEBUG
last_fault_code = fsr;
#endif
if (pmap_fault_fixup(map->pmap, va, ftype, user)) {
UVMHIST_LOG(maphist, " <- ref/mod emul", 0, 0, 0, 0);
goto out;
}
if (__predict_false(curcpu()->ci_intr_depth > 0)) {
if (pcb->pcb_onfault) {
tf->tf_r0 = EINVAL;
tf->tf_pc = (register_t)(intptr_t) pcb->pcb_onfault;
return;
}
printf("\nNon-emulated page fault with intr_depth > 0\n");
dab_fatal(tf, fsr, far, l, NULL);
}
onfault = pcb->pcb_onfault;
pcb->pcb_onfault = NULL;
error = uvm_fault(map, va, ftype);
pcb->pcb_onfault = onfault;
if (__predict_true(error == 0)) {
if (user)
uvm_grow(l->l_proc, va); /* Record any stack growth */
else
ucas_ras_check(tf);
UVMHIST_LOG(maphist, " <- uvm", 0, 0, 0, 0);
goto out;
}
if (user == 0) {
if (pcb->pcb_onfault) {
tf->tf_r0 = error;
tf->tf_pc = (register_t)(intptr_t) pcb->pcb_onfault;
return;
}
printf("\nuvm_fault(%p, %lx, %x) -> %x\n", map, va, ftype,
error);
dab_fatal(tf, fsr, far, l, NULL);
}
KSI_INIT_TRAP(&ksi);
if (error == ENOMEM) {
printf("UVM: pid %d (%s), uid %d killed: "
"out of swap\n", l->l_proc->p_pid, l->l_proc->p_comm,
l->l_cred ? kauth_cred_geteuid(l->l_cred) : -1);
ksi.ksi_signo = SIGKILL;
} else
ksi.ksi_signo = SIGSEGV;
ksi.ksi_code = (error == EACCES) ? SEGV_ACCERR : SEGV_MAPERR;
ksi.ksi_addr = (uint32_t *)(intptr_t) far;
ksi.ksi_trap = fsr;
UVMHIST_LOG(maphist, " <- error (%d)", error, 0, 0, 0);
do_trapsignal:
call_trapsignal(l, tf, &ksi);
out:
/* If returning to user mode, make sure to invoke userret() */
if (user)
userret(l);
}