int
vm_fault_hold(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
int fault_flags, vm_page_t *m_hold)
{
vm_prot_t prot;
long ahead, behind;
int alloc_req, era, faultcount, nera, reqpage, result;
boolean_t growstack, is_first_object_locked, wired;
int map_generation;
vm_object_t next_object;
vm_page_t marray[VM_FAULT_READ_MAX];
int hardfault;
struct faultstate fs;
struct vnode *vp;
int locked, error;
hardfault = 0;
growstack = TRUE;
PCPU_INC(cnt.v_vm_faults);
fs.vp = NULL;
faultcount = reqpage = 0;
RetryFault:;
/*
* Find the backing store object and offset into it to begin the
* search.
*/
fs.map = map;
result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
&fs.first_object, &fs.first_pindex, &prot, &wired);
if (result != KERN_SUCCESS) {
if (growstack && result == KERN_INVALID_ADDRESS &&
map != kernel_map) {
result = vm_map_growstack(curproc, vaddr);
if (result != KERN_SUCCESS)
return (KERN_FAILURE);
growstack = FALSE;
goto RetryFault;
}
return (result);
}
map_generation = fs.map->timestamp;
if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
panic("vm_fault: fault on nofault entry, addr: %lx",
(u_long)vaddr);
}
/*
* Make a reference to this object to prevent its disposal while we
* are messing with it. Once we have the reference, the map is free
* to be diddled. Since objects reference their shadows (and copies),
* they will stay around as well.
*
* Bump the paging-in-progress count to prevent size changes (e.g.
* truncation operations) during I/O. This must be done after
* obtaining the vnode lock in order to avoid possible deadlocks.
*/
VM_OBJECT_WLOCK(fs.first_object);
vm_object_reference_locked(fs.first_object);
vm_object_pip_add(fs.first_object, 1);
fs.lookup_still_valid = TRUE;
if (wired)
fault_type = prot | (fault_type & VM_PROT_COPY);
fs.first_m = NULL;
/*
* Search for the page at object/offset.
*/
fs.object = fs.first_object;
fs.pindex = fs.first_pindex;
while (TRUE) {
/*
* If the object is dead, we stop here
*/
if (fs.object->flags & OBJ_DEAD) {
unlock_and_deallocate(&fs);
return (KERN_PROTECTION_FAILURE);
}
/*
* See if page is resident
*/
fs.m = vm_page_lookup(fs.object, fs.pindex);
if (fs.m != NULL) {
/*
* check for page-based copy on write.
* We check fs.object == fs.first_object so
* as to ensure the legacy COW mechanism is
* used when the page in question is part of
* a shadow object. Otherwise, vm_page_cowfault()
* removes the page from the backing object,
* which is not what we want.
*/
vm_page_lock(fs.m);
if ((fs.m->cow) &&
(fault_type & VM_PROT_WRITE) &&
(fs.object == fs.first_object)) {
vm_page_cowfault(fs.m);
unlock_and_deallocate(&fs);
goto RetryFault;
}
/*
* Wait/Retry if the page is busy. We have to do this
* if the page is busy via either VPO_BUSY or
* vm_page_t->busy because the vm_pager may be using
* vm_page_t->busy for pageouts ( and even pageins if
* it is the vnode pager ), and we could end up trying
* to pagein and pageout the same page simultaneously.
*
* We can theoretically allow the busy case on a read
* fault if the page is marked valid, but since such
* pages are typically already pmap'd, putting that
* special case in might be more effort then it is
* worth. We cannot under any circumstances mess
* around with a vm_page_t->busy page except, perhaps,
* to pmap it.
*/
if ((fs.m->oflags & VPO_BUSY) || fs.m->busy) {
/*
* Reference the page before unlocking and
* sleeping so that the page daemon is less
* likely to reclaim it.
*/
vm_page_aflag_set(fs.m, PGA_REFERENCED);
vm_page_unlock(fs.m);
if (fs.object != fs.first_object) {
if (!VM_OBJECT_TRYWLOCK(
fs.first_object)) {
VM_OBJECT_WUNLOCK(fs.object);
VM_OBJECT_WLOCK(fs.first_object);
VM_OBJECT_WLOCK(fs.object);
}
vm_page_lock(fs.first_m);
vm_page_free(fs.first_m);
vm_page_unlock(fs.first_m);
vm_object_pip_wakeup(fs.first_object);
VM_OBJECT_WUNLOCK(fs.first_object);
fs.first_m = NULL;
}
unlock_map(&fs);
if (fs.m == vm_page_lookup(fs.object,
fs.pindex)) {
vm_page_sleep_if_busy(fs.m, TRUE,
"vmpfw");
}
vm_object_pip_wakeup(fs.object);
VM_OBJECT_WUNLOCK(fs.object);
PCPU_INC(cnt.v_intrans);
vm_object_deallocate(fs.first_object);
goto RetryFault;
}
vm_page_remque(fs.m);
vm_page_unlock(fs.m);
/*
* Mark page busy for other processes, and the
* pagedaemon. If it still isn't completely valid
* (readable), jump to readrest, else break-out ( we
* found the page ).
*/
vm_page_busy(fs.m);
if (fs.m->valid != VM_PAGE_BITS_ALL)
goto readrest;
break;
}
/*
* Page is not resident, If this is the search termination
* or the pager might contain the page, allocate a new page.
*/
if (TRYPAGER || fs.object == fs.first_object) {
if (fs.pindex >= fs.object->size) {
unlock_and_deallocate(&fs);
return (KERN_PROTECTION_FAILURE);
}
/*
* Allocate a new page for this object/offset pair.
*
* Unlocked read of the p_flag is harmless. At
* worst, the P_KILLED might be not observed
* there, and allocation can fail, causing
* restart and new reading of the p_flag.
*/
fs.m = NULL;
if (!vm_page_count_severe() || P_KILLED(curproc)) {
#if VM_NRESERVLEVEL > 0
if ((fs.object->flags & OBJ_COLORED) == 0) {
fs.object->flags |= OBJ_COLORED;
fs.object->pg_color = atop(vaddr) -
fs.pindex;
}
#endif
alloc_req = P_KILLED(curproc) ?
VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
if (fs.object->type != OBJT_VNODE &&
fs.object->backing_object == NULL)
alloc_req |= VM_ALLOC_ZERO;
fs.m = vm_page_alloc(fs.object, fs.pindex,
alloc_req);
}
if (fs.m == NULL) {
unlock_and_deallocate(&fs);
VM_WAITPFAULT;
goto RetryFault;
} else if (fs.m->valid == VM_PAGE_BITS_ALL)
break;
}
readrest:
/*
* We have found a valid page or we have allocated a new page.
* The page thus may not be valid or may not be entirely
* valid.
*
* Attempt to fault-in the page if there is a chance that the
* pager has it, and potentially fault in additional pages
* at the same time.
*/
if (TRYPAGER) {
int rv;
u_char behavior = vm_map_entry_behavior(fs.entry);
if (behavior == MAP_ENTRY_BEHAV_RANDOM ||
P_KILLED(curproc)) {
behind = 0;
ahead = 0;
} else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
behind = 0;
ahead = atop(fs.entry->end - vaddr) - 1;
if (ahead > VM_FAULT_READ_AHEAD_MAX)
ahead = VM_FAULT_READ_AHEAD_MAX;
if (fs.pindex == fs.entry->next_read)
vm_fault_cache_behind(&fs,
VM_FAULT_READ_MAX);
} else {
/*
* If this is a sequential page fault, then
* arithmetically increase the number of pages
* in the read-ahead window. Otherwise, reset
* the read-ahead window to its smallest size.
*/
behind = atop(vaddr - fs.entry->start);
if (behind > VM_FAULT_READ_BEHIND)
behind = VM_FAULT_READ_BEHIND;
ahead = atop(fs.entry->end - vaddr) - 1;
era = fs.entry->read_ahead;
if (fs.pindex == fs.entry->next_read) {
nera = era + behind;
if (nera > VM_FAULT_READ_AHEAD_MAX)
nera = VM_FAULT_READ_AHEAD_MAX;
behind = 0;
if (ahead > nera)
ahead = nera;
if (era == VM_FAULT_READ_AHEAD_MAX)
vm_fault_cache_behind(&fs,
VM_FAULT_CACHE_BEHIND);
} else if (ahead > VM_FAULT_READ_AHEAD_MIN)
ahead = VM_FAULT_READ_AHEAD_MIN;
if (era != ahead)
fs.entry->read_ahead = ahead;
}
/*
* Call the pager to retrieve the data, if any, after
* releasing the lock on the map. We hold a ref on
* fs.object and the pages are VPO_BUSY'd.
*/
unlock_map(&fs);
if (fs.object->type == OBJT_VNODE) {
vp = fs.object->handle;
if (vp == fs.vp)
goto vnode_locked;
else if (fs.vp != NULL) {
vput(fs.vp);
fs.vp = NULL;
}
locked = VOP_ISLOCKED(vp);
if (locked != LK_EXCLUSIVE)
locked = LK_SHARED;
/* Do not sleep for vnode lock while fs.m is busy */
error = vget(vp, locked | LK_CANRECURSE |
LK_NOWAIT, curthread);
if (error != 0) {
vhold(vp);
release_page(&fs);
unlock_and_deallocate(&fs);
error = vget(vp, locked | LK_RETRY |
LK_CANRECURSE, curthread);
vdrop(vp);
fs.vp = vp;
KASSERT(error == 0,
("vm_fault: vget failed"));
goto RetryFault;
}
fs.vp = vp;
}
vnode_locked:
KASSERT(fs.vp == NULL || !fs.map->system_map,
("vm_fault: vnode-backed object mapped by system map"));
/*
* now we find out if any other pages should be paged
* in at this time this routine checks to see if the
* pages surrounding this fault reside in the same
* object as the page for this fault. If they do,
* then they are faulted in also into the object. The
* array "marray" returned contains an array of
* vm_page_t structs where one of them is the
* vm_page_t passed to the routine. The reqpage
* return value is the index into the marray for the
* vm_page_t passed to the routine.
*
* fs.m plus the additional pages are VPO_BUSY'd.
*/
faultcount = vm_fault_additional_pages(
fs.m, behind, ahead, marray, &reqpage);
rv = faultcount ?
vm_pager_get_pages(fs.object, marray, faultcount,
reqpage) : VM_PAGER_FAIL;
if (rv == VM_PAGER_OK) {
/*
* Found the page. Leave it busy while we play
* with it.
*/
/*
* Relookup in case pager changed page. Pager
* is responsible for disposition of old page
* if moved.
*/
fs.m = vm_page_lookup(fs.object, fs.pindex);
if (!fs.m) {
unlock_and_deallocate(&fs);
goto RetryFault;
}
hardfault++;
break; /* break to PAGE HAS BEEN FOUND */
}
/*
* Remove the bogus page (which does not exist at this
* object/offset); before doing so, we must get back
* our object lock to preserve our invariant.
*
* Also wake up any other process that may want to bring
* in this page.
*
* If this is the top-level object, we must leave the
* busy page to prevent another process from rushing
* past us, and inserting the page in that object at
* the same time that we are.
*/
if (rv == VM_PAGER_ERROR)
printf("vm_fault: pager read error, pid %d (%s)\n",
curproc->p_pid, curproc->p_comm);
/*
* Data outside the range of the pager or an I/O error
*/
/*
* XXX - the check for kernel_map is a kludge to work
* around having the machine panic on a kernel space
* fault w/ I/O error.
*/
if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
(rv == VM_PAGER_BAD)) {
vm_page_lock(fs.m);
vm_page_free(fs.m);
vm_page_unlock(fs.m);
fs.m = NULL;
unlock_and_deallocate(&fs);
return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
}
if (fs.object != fs.first_object) {
vm_page_lock(fs.m);
vm_page_free(fs.m);
vm_page_unlock(fs.m);
fs.m = NULL;
/*
* XXX - we cannot just fall out at this
* point, m has been freed and is invalid!
*/
}
}
/*
* We get here if the object has default pager (or unwiring)
* or the pager doesn't have the page.
*/
if (fs.object == fs.first_object)
fs.first_m = fs.m;
/*
* Move on to the next object. Lock the next object before
* unlocking the current one.
*/
fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
next_object = fs.object->backing_object;
if (next_object == NULL) {
/*
* If there's no object left, fill the page in the top
* object with zeros.
*/
if (fs.object != fs.first_object) {
vm_object_pip_wakeup(fs.object);
VM_OBJECT_WUNLOCK(fs.object);
fs.object = fs.first_object;
fs.pindex = fs.first_pindex;
fs.m = fs.first_m;
VM_OBJECT_WLOCK(fs.object);
}
fs.first_m = NULL;
/*
* Zero the page if necessary and mark it valid.
*/
if ((fs.m->flags & PG_ZERO) == 0) {
pmap_zero_page(fs.m);
} else {
PCPU_INC(cnt.v_ozfod);
}
PCPU_INC(cnt.v_zfod);
fs.m->valid = VM_PAGE_BITS_ALL;
break; /* break to PAGE HAS BEEN FOUND */
} else {
KASSERT(fs.object != next_object,
("object loop %p", next_object));
VM_OBJECT_WLOCK(next_object);
vm_object_pip_add(next_object, 1);
if (fs.object != fs.first_object)
vm_object_pip_wakeup(fs.object);
VM_OBJECT_WUNLOCK(fs.object);
fs.object = next_object;
}
}
KASSERT((fs.m->oflags & VPO_BUSY) != 0,
("vm_fault: not busy after main loop"));
/*
* PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
* is held.]
*/
/*
* If the page is being written, but isn't already owned by the
* top-level object, we have to copy it into a new page owned by the
* top-level object.
*/
if (fs.object != fs.first_object) {
/*
* We only really need to copy if we want to write it.
*/
if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
/*
* This allows pages to be virtually copied from a
* backing_object into the first_object, where the
* backing object has no other refs to it, and cannot
* gain any more refs. Instead of a bcopy, we just
* move the page from the backing object to the
* first object. Note that we must mark the page
* dirty in the first object so that it will go out
* to swap when needed.
*/
is_first_object_locked = FALSE;
if (
/*
* Only one shadow object
*/
(fs.object->shadow_count == 1) &&
/*
* No COW refs, except us
*/
(fs.object->ref_count == 1) &&
/*
* No one else can look this object up
*/
(fs.object->handle == NULL) &&
/*
* No other ways to look the object up
*/
((fs.object->type == OBJT_DEFAULT) ||
(fs.object->type == OBJT_SWAP)) &&
(is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) &&
/*
* We don't chase down the shadow chain
*/
fs.object == fs.first_object->backing_object) {
/*
* get rid of the unnecessary page
*/
vm_page_lock(fs.first_m);
vm_page_free(fs.first_m);
vm_page_unlock(fs.first_m);
/*
* grab the page and put it into the
* process'es object. The page is
* automatically made dirty.
*/
vm_page_lock(fs.m);
vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
vm_page_unlock(fs.m);
vm_page_busy(fs.m);
fs.first_m = fs.m;
fs.m = NULL;
PCPU_INC(cnt.v_cow_optim);
} else {
/*
* Oh, well, lets copy it.
*/
pmap_copy_page(fs.m, fs.first_m);
fs.first_m->valid = VM_PAGE_BITS_ALL;
if (wired && (fault_flags &
VM_FAULT_CHANGE_WIRING) == 0) {
vm_page_lock(fs.first_m);
vm_page_wire(fs.first_m);
vm_page_unlock(fs.first_m);
vm_page_lock(fs.m);
vm_page_unwire(fs.m, FALSE);
vm_page_unlock(fs.m);
}
/*
* We no longer need the old page or object.
*/
release_page(&fs);
}
/*
* fs.object != fs.first_object due to above
* conditional
*/
vm_object_pip_wakeup(fs.object);
VM_OBJECT_WUNLOCK(fs.object);
/*
* Only use the new page below...
*/
fs.object = fs.first_object;
fs.pindex = fs.first_pindex;
fs.m = fs.first_m;
if (!is_first_object_locked)
VM_OBJECT_WLOCK(fs.object);
PCPU_INC(cnt.v_cow_faults);
curthread->td_cow++;
} else {
prot &= ~VM_PROT_WRITE;
}
}
/*
* We must verify that the maps have not changed since our last
* lookup.
*/
if (!fs.lookup_still_valid) {
vm_object_t retry_object;
vm_pindex_t retry_pindex;
vm_prot_t retry_prot;
if (!vm_map_trylock_read(fs.map)) {
release_page(&fs);
unlock_and_deallocate(&fs);
goto RetryFault;
}
fs.lookup_still_valid = TRUE;
if (fs.map->timestamp != map_generation) {
result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
&fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
/*
* If we don't need the page any longer, put it on the inactive
* list (the easiest thing to do here). If no one needs it,
* pageout will grab it eventually.
*/
if (result != KERN_SUCCESS) {
release_page(&fs);
unlock_and_deallocate(&fs);
/*
* If retry of map lookup would have blocked then
* retry fault from start.
*/
if (result == KERN_FAILURE)
goto RetryFault;
return (result);
}
if ((retry_object != fs.first_object) ||
(retry_pindex != fs.first_pindex)) {
release_page(&fs);
unlock_and_deallocate(&fs);
goto RetryFault;
}
/*
* Check whether the protection has changed or the object has
* been copied while we left the map unlocked. Changing from
* read to write permission is OK - we leave the page
* write-protected, and catch the write fault. Changing from
* write to read permission means that we can't mark the page
* write-enabled after all.
*/
prot &= retry_prot;
}
}
/*
* If the page was filled by a pager, update the map entry's
* last read offset. Since the pager does not return the
* actual set of pages that it read, this update is based on
* the requested set. Typically, the requested and actual
* sets are the same.
*
* XXX The following assignment modifies the map
* without holding a write lock on it.
*/
if (hardfault)
fs.entry->next_read = fs.pindex + faultcount - reqpage;
if ((prot & VM_PROT_WRITE) != 0 ||
(fault_flags & VM_FAULT_DIRTY) != 0) {
vm_object_set_writeable_dirty(fs.object);
/*
* If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
* if the page is already dirty to prevent data written with
* the expectation of being synced from not being synced.
* Likewise if this entry does not request NOSYNC then make
* sure the page isn't marked NOSYNC. Applications sharing
* data should use the same flags to avoid ping ponging.
*/
if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
if (fs.m->dirty == 0)
fs.m->oflags |= VPO_NOSYNC;
} else {
fs.m->oflags &= ~VPO_NOSYNC;
}
/*
* If the fault is a write, we know that this page is being
* written NOW so dirty it explicitly to save on
* pmap_is_modified() calls later.
*
* Also tell the backing pager, if any, that it should remove
* any swap backing since the page is now dirty.
*/
if (((fault_type & VM_PROT_WRITE) != 0 &&
(fault_flags & VM_FAULT_CHANGE_WIRING) == 0) ||
(fault_flags & VM_FAULT_DIRTY) != 0) {
vm_page_dirty(fs.m);
vm_pager_page_unswapped(fs.m);
}
}
/*
* Page had better still be busy
*/
KASSERT(fs.m->oflags & VPO_BUSY,
("vm_fault: page %p not busy!", fs.m));
/*
* Page must be completely valid or it is not fit to
* map into user space. vm_pager_get_pages() ensures this.
*/
KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
("vm_fault: page %p partially invalid", fs.m));
VM_OBJECT_WUNLOCK(fs.object);
/*
* Put this page into the physical map. We had to do the unlock above
* because pmap_enter() may sleep. We don't put the page
* back on the active queue until later so that the pageout daemon
* won't find it (yet).
*/
pmap_enter(fs.map->pmap, vaddr, fault_type, fs.m, prot, wired);
if ((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 && wired == 0)
vm_fault_prefault(fs.map->pmap, vaddr, fs.entry);
VM_OBJECT_WLOCK(fs.object);
vm_page_lock(fs.m);
/*
* If the page is not wired down, then put it where the pageout daemon
* can find it.
*/
if (fault_flags & VM_FAULT_CHANGE_WIRING) {
if (wired)
vm_page_wire(fs.m);
else
vm_page_unwire(fs.m, 1);
} else
vm_page_activate(fs.m);
if (m_hold != NULL) {
*m_hold = fs.m;
vm_page_hold(fs.m);
}
vm_page_unlock(fs.m);
vm_page_wakeup(fs.m);
/*
* Unlock everything, and return
*/
unlock_and_deallocate(&fs);
if (hardfault) {
PCPU_INC(cnt.v_io_faults);
curthread->td_ru.ru_majflt++;
} else
curthread->td_ru.ru_minflt++;
return (KERN_SUCCESS);
}
/*
* Original vm_pageout_oom, will be called if LRU pageout_oom will fail
*/
static void
original_vm_pageout_oom(int shortage)
{
struct proc *p, *bigproc;
vm_offset_t size, bigsize;
struct thread *td;
struct vmspace *vm;
/*
* We keep the process bigproc locked once we find it to keep anyone
* from messing with it; however, there is a possibility of
* deadlock if process B is bigproc and one of it's child processes
* attempts to propagate a signal to B while we are waiting for A's
* lock while walking this list. To avoid this, we don't block on
* the process lock but just skip a process if it is already locked.
*/
bigproc = NULL;
bigsize = 0;
sx_slock(&allproc_lock);
FOREACH_PROC_IN_SYSTEM(p) {
int breakout;
if (PROC_TRYLOCK(p) == 0)
continue;
/*
* If this is a system, protected or killed process, skip it.
*/
if (p->p_state != PRS_NORMAL ||
(p->p_flag & (P_INEXEC | P_PROTECTED | P_SYSTEM)) ||
(p->p_pid == 1) || P_KILLED(p) ||
((p->p_pid < 48) && (swap_pager_avail != 0))) {
PROC_UNLOCK(p);
continue;
}
/*
* If the process is in a non-running type state,
* don't touch it. Check all the threads individually.
*/
breakout = 0;
FOREACH_THREAD_IN_PROC(p, td) {
thread_lock(td);
if (!TD_ON_RUNQ(td) &&
!TD_IS_RUNNING(td) &&
!TD_IS_SLEEPING(td)) {
thread_unlock(td);
breakout = 1;
break;
}
thread_unlock(td);
}
if (breakout) {
PROC_UNLOCK(p);
continue;
}
/*
* get the process size
*/
vm = vmspace_acquire_ref(p);
if (vm == NULL) {
PROC_UNLOCK(p);
continue;
}
if (!vm_map_trylock_read(&vm->vm_map)) {
vmspace_free(vm);
PROC_UNLOCK(p);
continue;
}
size = vmspace_swap_count(vm);
vm_map_unlock_read(&vm->vm_map);
if (shortage == VM_OOM_MEM)
size += vmspace_resident_count(vm);
vmspace_free(vm);
/*
* if the this process is bigger than the biggest one
* remember it.
*/
if (size > bigsize) {
if (bigproc != NULL)
PROC_UNLOCK(bigproc);
bigproc = p;
bigsize = size;
} else
PROC_UNLOCK(p);
}