vm_page_t
shmem_read_mapping_page(vm_object_t object, vm_pindex_t pindex)
{
vm_page_t m;
int rv;
VM_OBJECT_LOCK_ASSERT_OWNED(object);
m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
if (m->valid != VM_PAGE_BITS_ALL) {
if (vm_pager_has_page(object, pindex)) {
rv = vm_pager_get_page(object, &m, 1);
m = vm_page_lookup(object, pindex);
if (m == NULL)
return ERR_PTR(-ENOMEM);
if (rv != VM_PAGER_OK) {
vm_page_free(m);
return ERR_PTR(-ENOMEM);
}
} else {
pmap_zero_page(VM_PAGE_TO_PHYS(m));
m->valid = VM_PAGE_BITS_ALL;
m->dirty = 0;
}
}
vm_page_wire(m);
vm_page_wakeup(m);
return (m);
}
/*
* Obtain a page pointer array and lock all pages into system memory. A segmentation violation will
* occur here if the calling user does not have access to the submitted address.
*/
static int
via_lock_all_dma_pages(drm_via_sg_info_t *vsg, drm_via_dmablit_t *xfer)
{
unsigned long first_pfn = VIA_PFN(xfer->mem_addr);
vm_page_t m;
int i;
vsg->num_pages = VIA_PFN(xfer->mem_addr +
(xfer->num_lines * xfer->mem_stride -1)) - first_pfn + 1;
if (NULL == (vsg->pages = malloc(sizeof(vm_page_t) * vsg->num_pages,
DRM_MEM_DRIVER, M_NOWAIT)))
return -ENOMEM;
vsg->state = dr_via_pages_alloc;
if (vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
(vm_offset_t)xfer->mem_addr, vsg->num_pages * PAGE_SIZE,
VM_PROT_READ | VM_PROT_WRITE, vsg->pages, vsg->num_pages) < 0)
return -EACCES;
for (i = 0; i < vsg->num_pages; i++) {
m = vsg->pages[i];
vm_page_lock(m);
vm_page_wire(m);
vm_page_unhold(m);
vm_page_unlock(m);
}
vsg->state = dr_via_pages_locked;
DRM_DEBUG("DMA pages locked\n");
return 0;
}
/*
* Allocate new wired pages in an object.
* The object is assumed to be mapped into the kernel map or
* a submap.
*/
void
kmem_alloc_pages(
vm_object_t object,
vm_offset_t offset,
vm_offset_t start,
vm_offset_t end,
vm_prot_t protection)
{
/*
* Mark the pmap region as not pageable.
*/
pmap_pageable(kernel_pmap, start, end, FALSE);
while (start < end) {
vm_page_t mem;
vm_object_lock(object);
/*
* Allocate a page
*/
while ((mem = vm_page_alloc(object, offset))
== VM_PAGE_NULL) {
vm_object_unlock(object);
VM_PAGE_WAIT((void (*)()) 0);
vm_object_lock(object);
}
/*
* Wire it down
*/
vm_page_lock_queues();
vm_page_wire(mem);
vm_page_unlock_queues();
vm_object_unlock(object);
/*
* Enter it in the kernel pmap
*/
PMAP_ENTER(kernel_pmap, start, mem,
protection, TRUE);
vm_object_lock(object);
PAGE_WAKEUP_DONE(mem);
vm_object_unlock(object);
start += PAGE_SIZE;
offset += PAGE_SIZE;
}
}
static void
wire_ddp_buffer(struct ddp_buffer *db)
{
int i;
vm_page_t p;
for (i = 0; i < db->npages; i++) {
p = db->pages[i];
vm_page_lock(p);
vm_page_wire(p);
vm_page_unhold(p);
vm_page_unlock(p);
}
}
/*
* Remap wired pages in an object into a new region.
* The object is assumed to be mapped into the kernel map or
* a submap.
*/
void
kmem_remap_pages(
vm_object_t object,
vm_offset_t offset,
vm_offset_t start,
vm_offset_t end,
vm_prot_t protection)
{
/*
* Mark the pmap region as not pageable.
*/
pmap_pageable(kernel_pmap, start, end, FALSE);
while (start < end) {
vm_page_t mem;
vm_object_lock(object);
/*
* Find a page
*/
if ((mem = vm_page_lookup(object, offset)) == VM_PAGE_NULL)
panic("kmem_remap_pages");
/*
* Wire it down (again)
*/
vm_page_lock_queues();
vm_page_wire(mem);
vm_page_unlock_queues();
vm_object_unlock(object);
/*
* Enter it in the kernel pmap. The page isn't busy,
* but this shouldn't be a problem because it is wired.
*/
PMAP_ENTER(kernel_pmap, start, mem,
protection, TRUE);
start += PAGE_SIZE;
offset += PAGE_SIZE;
}
}
/*
* Given a user pointer to a page of user memory, return an sf_buf for the
* page. Because we may be requesting quite a few sf_bufs, prefer failure to
* deadlock and use SFB_NOWAIT.
*/
static struct sf_buf *
zbuf_sfbuf_get(struct vm_map *map, vm_offset_t uaddr)
{
struct sf_buf *sf;
vm_page_t pp;
if (vm_fault_quick_hold_pages(map, uaddr, PAGE_SIZE, VM_PROT_READ |
VM_PROT_WRITE, &pp, 1) < 0)
return (NULL);
vm_page_lock(pp);
vm_page_wire(pp);
vm_page_unhold(pp);
vm_page_unlock(pp);
sf = sf_buf_alloc(pp, SFB_NOWAIT);
if (sf == NULL) {
zbuf_page_free(pp);
return (NULL);
}
return (sf);
}
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);
}
/*
* Routine:
* vm_fault_copy_entry
* Function:
* Create new shadow object backing dst_entry with private copy of
* all underlying pages. When src_entry is equal to dst_entry,
* function implements COW for wired-down map entry. Otherwise,
* it forks wired entry into dst_map.
*
* In/out conditions:
* The source and destination maps must be locked for write.
* The source map entry must be wired down (or be a sharing map
* entry corresponding to a main map entry that is wired down).
*/
void
vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
vm_ooffset_t *fork_charge)
{
vm_object_t backing_object, dst_object, object, src_object;
vm_pindex_t dst_pindex, pindex, src_pindex;
vm_prot_t access, prot;
vm_offset_t vaddr;
vm_page_t dst_m;
vm_page_t src_m;
boolean_t src_readonly, upgrade;
#ifdef lint
src_map++;
#endif /* lint */
upgrade = src_entry == dst_entry;
src_object = src_entry->object.vm_object;
src_pindex = OFF_TO_IDX(src_entry->offset);
src_readonly = (src_entry->protection & VM_PROT_WRITE) == 0;
/*
* Create the top-level object for the destination entry. (Doesn't
* actually shadow anything - we copy the pages directly.)
*/
dst_object = vm_object_allocate(OBJT_DEFAULT,
OFF_TO_IDX(dst_entry->end - dst_entry->start));
#if VM_NRESERVLEVEL > 0
dst_object->flags |= OBJ_COLORED;
dst_object->pg_color = atop(dst_entry->start);
#endif
VM_OBJECT_WLOCK(dst_object);
KASSERT(upgrade || dst_entry->object.vm_object == NULL,
("vm_fault_copy_entry: vm_object not NULL"));
dst_entry->object.vm_object = dst_object;
dst_entry->offset = 0;
dst_object->charge = dst_entry->end - dst_entry->start;
if (fork_charge != NULL) {
KASSERT(dst_entry->cred == NULL,
("vm_fault_copy_entry: leaked swp charge"));
dst_object->cred = curthread->td_ucred;
crhold(dst_object->cred);
*fork_charge += dst_object->charge;
} else {
dst_object->cred = dst_entry->cred;
dst_entry->cred = NULL;
}
access = prot = dst_entry->protection;
/*
* If not an upgrade, then enter the mappings in the pmap as
* read and/or execute accesses. Otherwise, enter them as
* write accesses.
*
* A writeable large page mapping is only created if all of
* the constituent small page mappings are modified. Marking
* PTEs as modified on inception allows promotion to happen
* without taking potentially large number of soft faults.
*/
if (!upgrade)
access &= ~VM_PROT_WRITE;
/*
* Loop through all of the virtual pages within the entry's
* range, copying each page from the source object to the
* destination object. Since the source is wired, those pages
* must exist. In contrast, the destination is pageable.
* Since the destination object does share any backing storage
* with the source object, all of its pages must be dirtied,
* regardless of whether they can be written.
*/
for (vaddr = dst_entry->start, dst_pindex = 0;
vaddr < dst_entry->end;
vaddr += PAGE_SIZE, dst_pindex++) {
/*
* Allocate a page in the destination object.
*/
do {
dst_m = vm_page_alloc(dst_object, dst_pindex,
VM_ALLOC_NORMAL);
if (dst_m == NULL) {
VM_OBJECT_WUNLOCK(dst_object);
VM_WAIT;
VM_OBJECT_WLOCK(dst_object);
}
} while (dst_m == NULL);
/*
* Find the page in the source object, and copy it in.
* (Because the source is wired down, the page will be in
* memory.)
*/
VM_OBJECT_WLOCK(src_object);
object = src_object;
pindex = src_pindex + dst_pindex;
while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
src_readonly &&
(backing_object = object->backing_object) != NULL) {
/*
* Allow fallback to backing objects if we are reading.
*/
VM_OBJECT_WLOCK(backing_object);
pindex += OFF_TO_IDX(object->backing_object_offset);
VM_OBJECT_WUNLOCK(object);
object = backing_object;
}
if (src_m == NULL)
panic("vm_fault_copy_wired: page missing");
pmap_copy_page(src_m, dst_m);
VM_OBJECT_WUNLOCK(object);
dst_m->valid = VM_PAGE_BITS_ALL;
dst_m->dirty = VM_PAGE_BITS_ALL;
VM_OBJECT_WUNLOCK(dst_object);
/*
* Enter it in the pmap. If a wired, copy-on-write
* mapping is being replaced by a write-enabled
* mapping, then wire that new mapping.
*/
pmap_enter(dst_map->pmap, vaddr, access, dst_m, prot, upgrade);
/*
* Mark it no longer busy, and put it on the active list.
*/
VM_OBJECT_WLOCK(dst_object);
if (upgrade) {
vm_page_lock(src_m);
vm_page_unwire(src_m, 0);
vm_page_unlock(src_m);
vm_page_lock(dst_m);
vm_page_wire(dst_m);
vm_page_unlock(dst_m);
} else {
vm_page_lock(dst_m);
vm_page_activate(dst_m);
vm_page_unlock(dst_m);
}
vm_page_wakeup(dst_m);
}
VM_OBJECT_WUNLOCK(dst_object);
if (upgrade) {
dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
vm_object_deallocate(src_object);
}
}
int
socow_setup(struct mbuf *m0, struct uio *uio)
{
struct sf_buf *sf;
vm_page_t pp;
struct iovec *iov;
struct vmspace *vmspace;
struct vm_map *map;
vm_offset_t offset, uva;
socow_stats.attempted++;
vmspace = curproc->p_vmspace;
map = &vmspace->vm_map;
uva = (vm_offset_t) uio->uio_iov->iov_base;
offset = uva & PAGE_MASK;
/*
* Verify that access to the given address is allowed from user-space.
*/
if (vm_fault_quick((caddr_t)uva, VM_PROT_READ) < 0)
return (0);
/*
* verify page is mapped & not already wired for i/o
*/
pp = pmap_extract_and_hold(map->pmap, uva, VM_PROT_READ);
if (pp == NULL) {
socow_stats.fail_not_mapped++;
return(0);
}
/*
* set up COW
*/
vm_page_lock(pp);
if (vm_page_cowsetup(pp) != 0) {
vm_page_unhold(pp);
vm_page_unlock(pp);
return (0);
}
/*
* wire the page for I/O
*/
vm_page_wire(pp);
vm_page_unhold(pp);
vm_page_unlock(pp);
/*
* Allocate an sf buf
*/
sf = sf_buf_alloc(pp, SFB_CATCH);
if (sf == NULL) {
vm_page_lock(pp);
vm_page_cowclear(pp);
vm_page_unwire(pp, 0);
/*
* Check for the object going away on us. This can
* happen since we don't hold a reference to it.
* If so, we're responsible for freeing the page.
*/
if (pp->wire_count == 0 && pp->object == NULL)
vm_page_free(pp);
vm_page_unlock(pp);
socow_stats.fail_sf_buf++;
return(0);
}
/*
* attach to mbuf
*/
MEXTADD(m0, sf_buf_kva(sf), PAGE_SIZE, socow_iodone,
(void*)sf_buf_kva(sf), sf, M_RDONLY, EXT_SFBUF);
m0->m_len = PAGE_SIZE - offset;
m0->m_data = (caddr_t)sf_buf_kva(sf) + offset;
socow_stats.success++;
iov = uio->uio_iov;
iov->iov_base = (char *)iov->iov_base + m0->m_len;
iov->iov_len -= m0->m_len;
uio->uio_resid -= m0->m_len;
uio->uio_offset += m0->m_len;
if (iov->iov_len == 0) {
uio->uio_iov++;
uio->uio_iovcnt--;
}
return(m0->m_len);
}
int
proc_rwmem(struct proc *p, struct uio *uio)
{
struct vmspace *vm;
vm_map_t map;
vm_object_t object = NULL;
vm_offset_t pageno = 0; /* page number */
vm_prot_t reqprot;
vm_offset_t kva;
int error, writing;
GIANT_REQUIRED;
/*
* if the vmspace is in the midst of being deallocated or the
* process is exiting, don't try to grab anything. The page table
* usage in that process can be messed up.
*/
vm = p->p_vmspace;
if ((p->p_flag & P_WEXIT))
return (EFAULT);
if (vm->vm_refcnt < 1)
return (EFAULT);
++vm->vm_refcnt;
/*
* The map we want...
*/
map = &vm->vm_map;
writing = uio->uio_rw == UIO_WRITE;
reqprot = writing ? (VM_PROT_WRITE | VM_PROT_OVERRIDE_WRITE) :
VM_PROT_READ;
kva = kmem_alloc_pageable(kernel_map, PAGE_SIZE);
/*
* Only map in one page at a time. We don't have to, but it
* makes things easier. This way is trivial - right?
*/
do {
vm_map_t tmap;
vm_offset_t uva;
int page_offset; /* offset into page */
vm_map_entry_t out_entry;
vm_prot_t out_prot;
boolean_t wired;
vm_pindex_t pindex;
u_int len;
vm_page_t m;
object = NULL;
uva = (vm_offset_t)uio->uio_offset;
/*
* Get the page number of this segment.
*/
pageno = trunc_page(uva);
page_offset = uva - pageno;
/*
* How many bytes to copy
*/
len = min(PAGE_SIZE - page_offset, uio->uio_resid);
/*
* Fault the page on behalf of the process
*/
error = vm_fault(map, pageno, reqprot, VM_FAULT_NORMAL);
if (error) {
error = EFAULT;
break;
}
/*
* Now we need to get the page. out_entry, out_prot, wired,
* and single_use aren't used. One would think the vm code
* would be a *bit* nicer... We use tmap because
* vm_map_lookup() can change the map argument.
*/
tmap = map;
error = vm_map_lookup(&tmap, pageno, reqprot, &out_entry,
&object, &pindex, &out_prot, &wired);
if (error) {
error = EFAULT;
/*
* Make sure that there is no residue in 'object' from
* an error return on vm_map_lookup.
*/
object = NULL;
break;
}
m = vm_page_lookup(object, pindex);
/* Allow fallback to backing objects if we are reading */
while (m == NULL && !writing && object->backing_object) {
pindex += OFF_TO_IDX(object->backing_object_offset);
object = object->backing_object;
m = vm_page_lookup(object, pindex);
}
if (m == NULL) {
error = EFAULT;
/*
* Make sure that there is no residue in 'object' from
* an error return on vm_map_lookup.
*/
object = NULL;
vm_map_lookup_done(tmap, out_entry);
break;
}
/*
* Wire the page into memory
*/
vm_page_lock_queues();
vm_page_wire(m);
vm_page_unlock_queues();
/*
* We're done with tmap now.
* But reference the object first, so that we won't loose
* it.
*/
vm_object_reference(object);
vm_map_lookup_done(tmap, out_entry);
pmap_qenter(kva, &m, 1);
/*
* Now do the i/o move.
*/
error = uiomove((caddr_t)(kva + page_offset), len, uio);
pmap_qremove(kva, 1);
/*
* release the page and the object
*/
vm_page_lock_queues();
vm_page_unwire(m, 1);
vm_page_unlock_queues();
vm_object_deallocate(object);
object = NULL;
} while (error == 0 && uio->uio_resid > 0);
if (object)
vm_object_deallocate(object);
kmem_free(kernel_map, kva, PAGE_SIZE);
vmspace_free(vm);
return (error);
}
/*
* vm_contig_pg_alloc:
*
* Allocate contiguous pages from the VM. This function does not
* map the allocated pages into the kernel map, otherwise it is
* impossible to make large allocations (i.e. >2G).
*
* Malloc()'s data structures have been used for collection of
* statistics and for allocations of less than a page.
*/
static int
vm_contig_pg_alloc(unsigned long size, vm_paddr_t low, vm_paddr_t high,
unsigned long alignment, unsigned long boundary, int mflags)
{
int i, q, start, pass;
vm_offset_t phys;
vm_page_t pga = vm_page_array;
vm_page_t m;
int pqtype;
size = round_page(size);
if (size == 0)
panic("vm_contig_pg_alloc: size must not be 0");
if ((alignment & (alignment - 1)) != 0)
panic("vm_contig_pg_alloc: alignment must be a power of 2");
if ((boundary & (boundary - 1)) != 0)
panic("vm_contig_pg_alloc: boundary must be a power of 2");
/*
* See if we can get the pages from the contiguous page reserve
* alist. The returned pages will be allocated and wired but not
* busied.
*/
m = vm_page_alloc_contig(low, high, alignment, boundary, size);
if (m)
return (m - &pga[0]);
/*
* Three passes (0, 1, 2). Each pass scans the VM page list for
* free or cached pages. After each pass if the entire scan failed
* we attempt to flush inactive pages and reset the start index back
* to 0. For passes 1 and 2 we also attempt to flush active pages.
*/
start = 0;
for (pass = 0; pass < 3; pass++) {
/*
* Find first page in array that is free, within range,
* aligned, and such that the boundary won't be crossed.
*/
again:
for (i = start; i < vmstats.v_page_count; i++) {
m = &pga[i];
phys = VM_PAGE_TO_PHYS(m);
pqtype = m->queue - m->pc;
if (((pqtype == PQ_FREE) || (pqtype == PQ_CACHE)) &&
(phys >= low) && (phys < high) &&
((phys & (alignment - 1)) == 0) &&
(((phys ^ (phys + size - 1)) & ~(boundary - 1)) == 0) &&
m->busy == 0 && m->wire_count == 0 &&
m->hold_count == 0 &&
(m->flags & (PG_BUSY | PG_NEED_COMMIT)) == 0)
{
break;
}
}
/*
* If we cannot find the page in the given range, or we have
* crossed the boundary, call the vm_contig_pg_clean() function
* for flushing out the queues, and returning it back to
* normal state.
*/
if ((i == vmstats.v_page_count) ||
((VM_PAGE_TO_PHYS(&pga[i]) + size) > high)) {
/*
* Best effort flush of all inactive pages.
* This is quite quick, for now stall all
* callers, even if they've specified M_NOWAIT.
*/
for (q = 0; q < PQ_L2_SIZE; ++q) {
vm_contig_pg_clean(PQ_INACTIVE + q,
vmstats.v_inactive_count);
lwkt_yield();
}
/*
* Best effort flush of active pages.
*
* This is very, very slow.
* Only do this if the caller has agreed to M_WAITOK.
*
* If enough pages are flushed, we may succeed on
* next (final) pass, if not the caller, contigmalloc(),
* will fail in the index < 0 case.
*/
if (pass > 0 && (mflags & M_WAITOK)) {
for (q = 0; q < PQ_L2_SIZE; ++q) {
vm_contig_pg_clean(PQ_ACTIVE + q,
vmstats.v_active_count);
}
lwkt_yield();
}
/*
* We're already too high in the address space
* to succeed, reset to 0 for the next iteration.
*/
start = 0;
continue; /* next pass */
}
start = i;
/*
* Check successive pages for contiguous and free.
*
* (still in critical section)
*/
for (i = start + 1; i < (start + size / PAGE_SIZE); i++) {
m = &pga[i];
pqtype = m->queue - m->pc;
if ((VM_PAGE_TO_PHYS(&m[0]) !=
(VM_PAGE_TO_PHYS(&m[-1]) + PAGE_SIZE)) ||
((pqtype != PQ_FREE) && (pqtype != PQ_CACHE)) ||
m->busy || m->wire_count ||
m->hold_count ||
(m->flags & (PG_BUSY | PG_NEED_COMMIT)))
{
start++;
goto again;
}
}
/*
* Try to allocate the pages, wiring them as we go.
*
* (still in critical section)
*/
for (i = start; i < (start + size / PAGE_SIZE); i++) {
m = &pga[i];
if (vm_page_busy_try(m, TRUE)) {
vm_contig_pg_free(start,
(i - start) * PAGE_SIZE);
start++;
goto again;
}
pqtype = m->queue - m->pc;
if (pqtype == PQ_CACHE &&
m->hold_count == 0 &&
m->wire_count == 0 &&
(m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) == 0) {
vm_page_protect(m, VM_PROT_NONE);
KKASSERT((m->flags & PG_MAPPED) == 0);
KKASSERT(m->dirty == 0);
vm_page_free(m);
--i;
continue; /* retry the page */
}
if (pqtype != PQ_FREE || m->hold_count) {
vm_page_wakeup(m);
vm_contig_pg_free(start,
(i - start) * PAGE_SIZE);
start++;
goto again;
}
KKASSERT((m->valid & m->dirty) == 0);
KKASSERT(m->wire_count == 0);
KKASSERT(m->object == NULL);
vm_page_unqueue_nowakeup(m);
m->valid = VM_PAGE_BITS_ALL;
if (m->flags & PG_ZERO)
vm_page_zero_count--;
KASSERT(m->dirty == 0,
("vm_contig_pg_alloc: page %p was dirty", m));
KKASSERT(m->wire_count == 0);
KKASSERT(m->busy == 0);
/*
* Clear all flags except PG_BUSY, PG_ZERO, and
* PG_WANTED, then unbusy the now allocated page.
*/
vm_page_flag_clear(m, ~(PG_BUSY | PG_SBUSY |
PG_ZERO | PG_WANTED));
vm_page_wire(m);
vm_page_wakeup(m);
}
/*
* Our job is done, return the index page of vm_page_array.
*/
return (start); /* aka &pga[start] */
}
/*
* Failed.
*/
return (-1);
}
