static inline void
unlock_map(struct faultstate *fs)
{
if (fs->lookup_still_valid) {
vm_map_lookup_done(fs->map, fs->entry);
fs->lookup_still_valid = FALSE;
}
}
/*
* Identify the physical page mapped at the given kernel virtual
* address. Insert this physical page into the given address space at
* the given virtual address, replacing the physical page, if any,
* that already exists there.
*/
static int
vm_pgmoveco(vm_map_t mapa, vm_offset_t kaddr, vm_offset_t uaddr)
{
vm_map_t map = mapa;
vm_page_t kern_pg, user_pg;
vm_object_t uobject;
vm_map_entry_t entry;
vm_pindex_t upindex;
vm_prot_t prot;
boolean_t wired;
KASSERT((uaddr & PAGE_MASK) == 0,
("vm_pgmoveco: uaddr is not page aligned"));
/*
* Herein the physical page is validated and dirtied. It is
* unwired in sf_buf_mext().
*/
kern_pg = PHYS_TO_VM_PAGE(vtophys(kaddr));
kern_pg->valid = VM_PAGE_BITS_ALL;
KASSERT(kern_pg->queue == PQ_NONE && kern_pg->wire_count == 1,
("vm_pgmoveco: kern_pg is not correctly wired"));
if ((vm_map_lookup(&map, uaddr,
VM_PROT_WRITE, &entry, &uobject,
&upindex, &prot, &wired)) != KERN_SUCCESS) {
return(EFAULT);
}
VM_OBJECT_LOCK(uobject);
retry:
if ((user_pg = vm_page_lookup(uobject, upindex)) != NULL) {
if (vm_page_sleep_if_busy(user_pg, TRUE, "vm_pgmoveco"))
goto retry;
vm_page_lock_queues();
pmap_remove_all(user_pg);
vm_page_free(user_pg);
} else {
/*
* Even if a physical page does not exist in the
* object chain's first object, a physical page from a
* backing object may be mapped read only.
*/
if (uobject->backing_object != NULL)
pmap_remove(map->pmap, uaddr, uaddr + PAGE_SIZE);
vm_page_lock_queues();
}
vm_page_insert(kern_pg, uobject, upindex);
vm_page_dirty(kern_pg);
vm_page_unlock_queues();
VM_OBJECT_UNLOCK(uobject);
vm_map_lookup_done(map, entry);
return(KERN_SUCCESS);
}
/*
* We require the caller to unmap the entire entry. This allows us to
* safely decrement shm_kmappings when a mapping is removed.
*/
int
shm_unmap(struct file *fp, void *mem, size_t size)
{
struct shmfd *shmfd;
vm_map_entry_t entry;
vm_offset_t kva, ofs;
vm_object_t obj;
vm_pindex_t pindex;
vm_prot_t prot;
boolean_t wired;
vm_map_t map;
int rv;
if (fp->f_type != DTYPE_SHM)
return (EINVAL);
shmfd = fp->f_data;
kva = (vm_offset_t)mem;
ofs = kva & PAGE_MASK;
kva = trunc_page(kva);
size = round_page(size + ofs);
map = kernel_map;
rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry,
&obj, &pindex, &prot, &wired);
if (rv != KERN_SUCCESS)
return (EINVAL);
if (entry->start != kva || entry->end != kva + size) {
vm_map_lookup_done(map, entry);
return (EINVAL);
}
vm_map_lookup_done(map, entry);
if (obj != shmfd->shm_object)
return (EINVAL);
vm_map_remove(map, kva, kva + size);
VM_OBJECT_LOCK(obj);
KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped"));
shmfd->shm_kmappings--;
VM_OBJECT_UNLOCK(obj);
return (0);
}
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);
}
static int
privcmd_ioctl(struct cdev *dev, unsigned long cmd, caddr_t arg,
int mode, struct thread *td)
{
int error, i;
switch (cmd) {
case IOCTL_PRIVCMD_HYPERCALL: {
struct ioctl_privcmd_hypercall *hcall;
hcall = (struct ioctl_privcmd_hypercall *)arg;
#ifdef __amd64__
/*
* The hypervisor page table walker will refuse to access
* user-space pages if SMAP is enabled, so temporary disable it
* while performing the hypercall.
*/
if (cpu_stdext_feature & CPUID_STDEXT_SMAP)
stac();
#endif
error = privcmd_hypercall(hcall->op, hcall->arg[0],
hcall->arg[1], hcall->arg[2], hcall->arg[3], hcall->arg[4]);
#ifdef __amd64__
if (cpu_stdext_feature & CPUID_STDEXT_SMAP)
clac();
#endif
if (error >= 0) {
hcall->retval = error;
error = 0;
} else {
error = xen_translate_error(error);
hcall->retval = 0;
}
break;
}
case IOCTL_PRIVCMD_MMAPBATCH: {
struct ioctl_privcmd_mmapbatch *mmap;
vm_map_t map;
vm_map_entry_t entry;
vm_object_t mem;
vm_pindex_t pindex;
vm_prot_t prot;
boolean_t wired;
struct xen_add_to_physmap_range add;
xen_ulong_t *idxs;
xen_pfn_t *gpfns;
int *errs, index;
struct privcmd_map *umap;
uint16_t num;
mmap = (struct ioctl_privcmd_mmapbatch *)arg;
if ((mmap->num == 0) ||
((mmap->addr & PAGE_MASK) != 0)) {
error = EINVAL;
break;
}
map = &td->td_proc->p_vmspace->vm_map;
error = vm_map_lookup(&map, mmap->addr, VM_PROT_NONE, &entry,
&mem, &pindex, &prot, &wired);
if (error != KERN_SUCCESS) {
error = EINVAL;
break;
}
if ((entry->start != mmap->addr) ||
(entry->end != mmap->addr + (mmap->num * PAGE_SIZE))) {
vm_map_lookup_done(map, entry);
error = EINVAL;
break;
}
vm_map_lookup_done(map, entry);
if ((mem->type != OBJT_MGTDEVICE) ||
(mem->un_pager.devp.ops != &privcmd_pg_ops)) {
error = EINVAL;
break;
}
umap = mem->handle;
add.domid = DOMID_SELF;
add.space = XENMAPSPACE_gmfn_foreign;
add.foreign_domid = mmap->dom;
/*
* The 'size' field in the xen_add_to_physmap_range only
* allows for UINT16_MAX mappings in a single hypercall.
*/
num = MIN(mmap->num, UINT16_MAX);
idxs = malloc(sizeof(*idxs) * num, M_PRIVCMD, M_WAITOK);
gpfns = malloc(sizeof(*gpfns) * num, M_PRIVCMD, M_WAITOK);
errs = malloc(sizeof(*errs) * num, M_PRIVCMD, M_WAITOK);
set_xen_guest_handle(add.idxs, idxs);
set_xen_guest_handle(add.gpfns, gpfns);
set_xen_guest_handle(add.errs, errs);
/* Allocate a bitset to store broken page mappings. */
umap->err = BITSET_ALLOC(mmap->num, M_PRIVCMD,
M_WAITOK | M_ZERO);
for (index = 0; index < mmap->num; index += num) {
num = MIN(mmap->num - index, UINT16_MAX);
add.size = num;
error = copyin(&mmap->arr[index], idxs,
sizeof(idxs[0]) * num);
if (error != 0)
goto mmap_out;
for (i = 0; i < num; i++)
gpfns[i] = atop(umap->phys_base_addr +
(i + index) * PAGE_SIZE);
bzero(errs, sizeof(*errs) * num);
error = HYPERVISOR_memory_op(
XENMEM_add_to_physmap_range, &add);
if (error != 0) {
error = xen_translate_error(error);
goto mmap_out;
}
for (i = 0; i < num; i++) {
if (errs[i] != 0) {
errs[i] = xen_translate_error(errs[i]);
/* Mark the page as invalid. */
BIT_SET(mmap->num, index + i,
umap->err);
}
}
error = copyout(errs, &mmap->err[index],
sizeof(errs[0]) * num);
if (error != 0)
goto mmap_out;
}
umap->mapped = true;
mmap_out:
free(idxs, M_PRIVCMD);
free(gpfns, M_PRIVCMD);
free(errs, M_PRIVCMD);
if (!umap->mapped)
free(umap->err, M_PRIVCMD);
break;
}
default:
error = ENOSYS;
break;
}
return (error);
}
