static void
pmap_inval_done(pmap_t pmap)
{
if (pmap != &kernel_pmap) {
atomic_add_acq_long(&pmap->pm_invgen, 1);
atomic_clear_int(&pmap->pm_active_lock, CPULOCK_EXCL);
}
crit_exit_id("inval");
}
static void
pmap_inval_init(pmap_t pmap)
{
cpulock_t olock;
cpulock_t nlock;
crit_enter_id("inval");
if (pmap != &kernel_pmap) {
for (;;) {
olock = pmap->pm_active_lock;
cpu_ccfence();
nlock = olock | CPULOCK_EXCL;
if (olock != nlock &&
atomic_cmpset_int(&pmap->pm_active_lock,
olock, nlock)) {
break;
}
lwkt_process_ipiq();
cpu_pause();
}
atomic_add_acq_long(&pmap->pm_invgen, 1);
}
}
/*
struct vnop_lookup_args {
struct vnodeop_desc *a_desc;
struct vnode *a_dvp;
struct vnode **a_vpp;
struct componentname *a_cnp;
};
*/
int
fuse_vnop_lookup(struct vop_lookup_args *ap)
{
struct vnode *dvp = ap->a_dvp;
struct vnode **vpp = ap->a_vpp;
struct componentname *cnp = ap->a_cnp;
struct thread *td = cnp->cn_thread;
struct ucred *cred = cnp->cn_cred;
int nameiop = cnp->cn_nameiop;
int flags = cnp->cn_flags;
int wantparent = flags & (LOCKPARENT | WANTPARENT);
int islastcn = flags & ISLASTCN;
struct mount *mp = vnode_mount(dvp);
int err = 0;
int lookup_err = 0;
struct vnode *vp = NULL;
struct fuse_dispatcher fdi;
enum fuse_opcode op;
uint64_t nid;
struct fuse_access_param facp;
FS_DEBUG2G("parent_inode=%ju - %*s\n",
(uintmax_t)VTOI(dvp), (int)cnp->cn_namelen, cnp->cn_nameptr);
if (fuse_isdeadfs(dvp)) {
*vpp = NULL;
return ENXIO;
}
if (!vnode_isdir(dvp)) {
return ENOTDIR;
}
if (islastcn && vfs_isrdonly(mp) && (nameiop != LOOKUP)) {
return EROFS;
}
/*
* We do access check prior to doing anything else only in the case
* when we are at fs root (we'd like to say, "we are at the first
* component", but that's not exactly the same... nevermind).
* See further comments at further access checks.
*/
bzero(&facp, sizeof(facp));
if (vnode_isvroot(dvp)) { /* early permission check hack */
if ((err = fuse_internal_access(dvp, VEXEC, &facp, td, cred))) {
return err;
}
}
if (flags & ISDOTDOT) {
nid = VTOFUD(dvp)->parent_nid;
if (nid == 0) {
return ENOENT;
}
fdisp_init(&fdi, 0);
op = FUSE_GETATTR;
goto calldaemon;
} else if (cnp->cn_namelen == 1 && *(cnp->cn_nameptr) == '.') {
nid = VTOI(dvp);
fdisp_init(&fdi, 0);
op = FUSE_GETATTR;
goto calldaemon;
} else if (fuse_lookup_cache_enable) {
err = cache_lookup(dvp, vpp, cnp, NULL, NULL);
switch (err) {
case -1: /* positive match */
atomic_add_acq_long(&fuse_lookup_cache_hits, 1);
return 0;
case 0: /* no match in cache */
atomic_add_acq_long(&fuse_lookup_cache_misses, 1);
break;
case ENOENT: /* negative match */
/* fall through */
default:
return err;
}
}
nid = VTOI(dvp);
fdisp_init(&fdi, cnp->cn_namelen + 1);
op = FUSE_LOOKUP;
calldaemon:
fdisp_make(&fdi, op, mp, nid, td, cred);
if (op == FUSE_LOOKUP) {
memcpy(fdi.indata, cnp->cn_nameptr, cnp->cn_namelen);
((char *)fdi.indata)[cnp->cn_namelen] = '\0';
}
lookup_err = fdisp_wait_answ(&fdi);
if ((op == FUSE_LOOKUP) && !lookup_err) { /* lookup call succeeded */
nid = ((struct fuse_entry_out *)fdi.answ)->nodeid;
if (!nid) {
/*
* zero nodeid is the same as "not found",
* but it's also cacheable (which we keep
* keep on doing not as of writing this)
*/
lookup_err = ENOENT;
} else if (nid == FUSE_ROOT_ID) {
lookup_err = EINVAL;
}
}
if (lookup_err &&
(!fdi.answ_stat || lookup_err != ENOENT || op != FUSE_LOOKUP)) {
fdisp_destroy(&fdi);
return lookup_err;
}
/* lookup_err, if non-zero, must be ENOENT at this point */
if (lookup_err) {
if ((nameiop == CREATE || nameiop == RENAME) && islastcn
/* && directory dvp has not been removed */ ) {
if (vfs_isrdonly(mp)) {
err = EROFS;
goto out;
}
#if 0 /* THINK_ABOUT_THIS */
if ((err = fuse_internal_access(dvp, VWRITE, cred, td, &facp))) {
goto out;
}
#endif
/*
* Possibly record the position of a slot in the
* directory large enough for the new component name.
* This can be recorded in the vnode private data for
* dvp. Set the SAVENAME flag to hold onto the
* pathname for use later in VOP_CREATE or VOP_RENAME.
*/
cnp->cn_flags |= SAVENAME;
err = EJUSTRETURN;
goto out;
}
/* Consider inserting name into cache. */
/*
* No we can't use negative caching, as the fs
* changes are out of our control.
* False positives' falseness turns out just as things
* go by, but false negatives' falseness doesn't.
* (and aiding the caching mechanism with extra control
* mechanisms comes quite close to beating the whole purpose
* caching...)
*/
#if 0
if ((cnp->cn_flags & MAKEENTRY) && nameiop != CREATE) {
FS_DEBUG("inserting NULL into cache\n");
cache_enter(dvp, NULL, cnp);
}
#endif
err = ENOENT;
goto out;
} else {
/* !lookup_err */
struct fuse_entry_out *feo = NULL;
struct fuse_attr *fattr = NULL;
if (op == FUSE_GETATTR) {
fattr = &((struct fuse_attr_out *)fdi.answ)->attr;
} else {
feo = (struct fuse_entry_out *)fdi.answ;
fattr = &(feo->attr);
}
/*
* If deleting, and at end of pathname, return parameters
* which can be used to remove file. If the wantparent flag
* isn't set, we return only the directory, otherwise we go on
* and lock the inode, being careful with ".".
*/
if (nameiop == DELETE && islastcn) {
/*
* Check for write access on directory.
*/
facp.xuid = fattr->uid;
facp.facc_flags |= FACCESS_STICKY;
err = fuse_internal_access(dvp, VWRITE, &facp, td, cred);
facp.facc_flags &= ~FACCESS_XQUERIES;
if (err) {
goto out;
}
if (nid == VTOI(dvp)) {
vref(dvp);
*vpp = dvp;
} else {
err = fuse_vnode_get(dvp->v_mount, nid, dvp,
&vp, cnp, IFTOVT(fattr->mode));
if (err)
goto out;
*vpp = vp;
}
/*
* Save the name for use in VOP_RMDIR and VOP_REMOVE
* later.
*/
cnp->cn_flags |= SAVENAME;
goto out;
}
/*
* If rewriting (RENAME), return the inode and the
* information required to rewrite the present directory
* Must get inode of directory entry to verify it's a
* regular file, or empty directory.
*/
if (nameiop == RENAME && wantparent && islastcn) {
#if 0 /* THINK_ABOUT_THIS */
if ((err = fuse_internal_access(dvp, VWRITE, cred, td, &facp))) {
goto out;
}
#endif
/*
* Check for "."
*/
if (nid == VTOI(dvp)) {
err = EISDIR;
goto out;
}
err = fuse_vnode_get(vnode_mount(dvp),
nid,
dvp,
&vp,
cnp,
IFTOVT(fattr->mode));
if (err) {
goto out;
}
*vpp = vp;
/*
* Save the name for use in VOP_RENAME later.
*/
cnp->cn_flags |= SAVENAME;
goto out;
}
if (flags & ISDOTDOT) {
struct mount *mp;
int ltype;
/*
* Expanded copy of vn_vget_ino() so that
* fuse_vnode_get() can be used.
*/
mp = dvp->v_mount;
ltype = VOP_ISLOCKED(dvp);
err = vfs_busy(mp, MBF_NOWAIT);
if (err != 0) {
vfs_ref(mp);
VOP_UNLOCK(dvp, 0);
err = vfs_busy(mp, 0);
vn_lock(dvp, ltype | LK_RETRY);
vfs_rel(mp);
if (err)
goto out;
if ((dvp->v_iflag & VI_DOOMED) != 0) {
err = ENOENT;
vfs_unbusy(mp);
goto out;
}
}
VOP_UNLOCK(dvp, 0);
err = fuse_vnode_get(vnode_mount(dvp),
nid,
NULL,
&vp,
cnp,
IFTOVT(fattr->mode));
vfs_unbusy(mp);
vn_lock(dvp, ltype | LK_RETRY);
if ((dvp->v_iflag & VI_DOOMED) != 0) {
if (err == 0)
vput(vp);
err = ENOENT;
}
if (err)
goto out;
*vpp = vp;
} else if (nid == VTOI(dvp)) {
vref(dvp);
*vpp = dvp;
} else {
err = fuse_vnode_get(vnode_mount(dvp),
nid,
dvp,
&vp,
cnp,
IFTOVT(fattr->mode));
if (err) {
goto out;
}
fuse_vnode_setparent(vp, dvp);
*vpp = vp;
}
if (op == FUSE_GETATTR) {
cache_attrs(*vpp, (struct fuse_attr_out *)fdi.answ);
} else {
cache_attrs(*vpp, (struct fuse_entry_out *)fdi.answ);
}
/* Insert name into cache if appropriate. */
/*
* Nooo, caching is evil. With caching, we can't avoid stale
* information taking over the playground (cached info is not
* just positive/negative, it does have qualitative aspects,
* too). And a (VOP/FUSE)_GETATTR is always thrown anyway, when
* walking down along cached path components, and that's not
* any cheaper than FUSE_LOOKUP. This might change with
* implementing kernel side attr caching, but... In Linux,
* lookup results are not cached, and the daemon is bombarded
* with FUSE_LOOKUPS on and on. This shows that by design, the
* daemon is expected to handle frequent lookup queries
* efficiently, do its caching in userspace, and so on.
*
* So just leave the name cache alone.
*/
/*
* Well, now I know, Linux caches lookups, but with a
* timeout... So it's the same thing as attribute caching:
* we can deal with it when implement timeouts.
*/
#if 0
if (cnp->cn_flags & MAKEENTRY) {
cache_enter(dvp, *vpp, cnp);
}
#endif
}
out:
if (!lookup_err) {
/* No lookup error; need to clean up. */
if (err) { /* Found inode; exit with no vnode. */
if (op == FUSE_LOOKUP) {
fuse_internal_forget_send(vnode_mount(dvp), td, cred,
nid, 1);
}
fdisp_destroy(&fdi);
return err;
} else {
#ifndef NO_EARLY_PERM_CHECK_HACK
if (!islastcn) {
/*
* We have the attributes of the next item
* *now*, and it's a fact, and we do not
* have to do extra work for it (ie, beg the
* daemon), and it neither depends on such
* accidental things like attr caching. So
* the big idea: check credentials *now*,
* not at the beginning of the next call to
* lookup.
*
* The first item of the lookup chain (fs root)
* won't be checked then here, of course, as
* its never "the next". But go and see that
* the root is taken care about at the very
* beginning of this function.
*
* Now, given we want to do the access check
* this way, one might ask: so then why not
* do the access check just after fetching
* the inode and its attributes from the
* daemon? Why bother with producing the
* corresponding vnode at all if something
* is not OK? We know what's the deal as
* soon as we get those attrs... There is
* one bit of info though not given us by
* the daemon: whether his response is
* authorative or not... His response should
* be ignored if something is mounted over
* the dir in question. But that can be
* known only by having the vnode...
*/
int tmpvtype = vnode_vtype(*vpp);
bzero(&facp, sizeof(facp));
/*the early perm check hack */
facp.facc_flags |= FACCESS_VA_VALID;
if ((tmpvtype != VDIR) && (tmpvtype != VLNK)) {
err = ENOTDIR;
}
if (!err && !vnode_mountedhere(*vpp)) {
err = fuse_internal_access(*vpp, VEXEC, &facp, td, cred);
}
if (err) {
if (tmpvtype == VLNK)
FS_DEBUG("weird, permission error with a symlink?\n");
vput(*vpp);
*vpp = NULL;
}
}
#endif
}
}
fdisp_destroy(&fdi);
return err;
}
int
sys_vmm_guest_sync_addr(struct vmm_guest_sync_addr_args *uap)
{
int error = 0;
cpulock_t olock;
cpulock_t nlock;
cpumask_t mask;
long val;
struct proc *p = curproc;
if (p->p_vmm == NULL)
return ENOSYS;
crit_enter_id("vmm_inval");
/*
* Acquire CPULOCK_EXCL, spin while we wait. This will prevent
* any other cpu trying to use related VMMs to wait for us.
*/
KKASSERT(CPUMASK_TESTMASK(p->p_vmm_cpumask, mycpu->gd_cpumask) == 0);
for (;;) {
olock = p->p_vmm_cpulock & ~CPULOCK_EXCL;
cpu_ccfence();
nlock = olock | CPULOCK_EXCL;
if (atomic_cmpset_int(&p->p_vmm_cpulock, olock, nlock))
break;
lwkt_process_ipiq();
cpu_pause();
}
/*
* Wait for other cpu's to exit VMM mode (for this vkernel). No
* new cpus will enter VMM mode while we hold the lock. New waiters
* may turn-up though so the wakeup() later on has to be
* unconditional.
*
* We must test on p_vmm_cpulock's counter, not the mask, because
* VMM entries will set the mask bit unconditionally first
* (interlocking our IPI below) and then conditionally bump the
* counter.
*/
if (olock & CPULOCK_CNTMASK) {
mask = p->p_vmm_cpumask;
CPUMASK_ANDMASK(mask, mycpu->gd_other_cpus);
lwkt_send_ipiq_mask(mask, vmm_exit_vmm, NULL);
while (p->p_vmm_cpulock & CPULOCK_CNTMASK) {
lwkt_process_ipiq();
cpu_pause();
}
}
#ifndef _KERNEL_VIRTUAL
/*
* Ensure that any new entries into VMM mode using
* vmm's managed under this process will issue a
* INVEPT before resuming.
*/
atomic_add_acq_long(&p->p_vmspace->vm_pmap.pm_invgen, 1);
#endif
/*
* Make the requested modification, wakeup any waiters.
*/
if (uap->srcaddr) {
copyin(uap->srcaddr, &val, sizeof(long));
copyout(&val, uap->dstaddr, sizeof(long));
}
/*
* VMMs on remote cpus will not be re-entered until we
* clear the lock.
*/
atomic_clear_int(&p->p_vmm_cpulock, CPULOCK_EXCL);
#if 0
wakeup(&p->p_vmm_cpulock);
#endif
crit_exit_id("vmm_inval");
return error;
}