fusefs: cache negative lookups

The FUSE protocol includes a way for a server to tell the client that a
negative lookup response is cacheable for a certain amount of time.

PR:		236226
Sponsored by:	The FreeBSD Foundation
This commit is contained in:
Alan Somers 2019-04-09 21:22:02 +00:00
parent ccb75e4939
commit 44f10c6e40
Notes: svn2git 2020-12-20 02:59:44 +00:00
svn path=/projects/fuse2/; revision=346067
5 changed files with 66 additions and 84 deletions

View File

@ -189,22 +189,12 @@ fuse_internal_cache_attrs(struct vnode *vp, struct fuse_attr *attr,
struct mount *mp;
struct fuse_vnode_data *fvdat;
struct vattr *vp_cache_at;
struct timespec now, duration, timeout;
mp = vnode_mount(vp);
fvdat = VTOFUD(vp);
getnanouptime(&now);
/* "+ 2" is the bound of attr_valid_nsec + now.tv_nsec */
/* Why oh why isn't there a TIME_MAX defined? */
if (attr_valid >= INT_MAX || attr_valid + now.tv_sec + 2 >= INT_MAX) {
fvdat->attr_cache_timeout.sec = INT_MAX;
} else {
duration.tv_sec = attr_valid;
duration.tv_nsec = attr_valid_nsec;
timespecadd(&duration, &now, &timeout);
timespec2bintime(&timeout, &fvdat->attr_cache_timeout);
}
fuse_validity_2_bintime(attr_valid, attr_valid_nsec,
&fvdat->attr_cache_timeout);
vp_cache_at = VTOVA(vp);

View File

@ -150,6 +150,52 @@ fuse_iosize(struct vnode *vp)
return (vp->v_mount->mnt_stat.f_iosize);
}
/*
* Make a cacheable timeout in bintime format value based on a fuse_attr_out
* response
*/
static inline void
fuse_validity_2_bintime(uint64_t attr_valid, uint32_t attr_valid_nsec,
struct bintime *timeout)
{
struct timespec now, duration, timeout_ts;
getnanouptime(&now);
/* "+ 2" is the bound of attr_valid_nsec + now.tv_nsec */
/* Why oh why isn't there a TIME_MAX defined? */
if (attr_valid >= INT_MAX || attr_valid + now.tv_sec + 2 >= INT_MAX) {
timeout->sec = INT_MAX;
} else {
duration.tv_sec = attr_valid;
duration.tv_nsec = attr_valid_nsec;
timespecadd(&duration, &now, &timeout_ts);
timespec2bintime(&timeout_ts, timeout);
}
}
/*
* Make a cacheable timeout value in timespec format based on the fuse_entry_out
* response
*/
static inline void
fuse_validity_2_timespec(const struct fuse_entry_out *feo,
struct timespec *timeout)
{
struct timespec duration, now;
getnanouptime(&now);
/* "+ 2" is the bound of entry_valid_nsec + now.tv_nsec */
if (feo->entry_valid >= INT_MAX ||
feo->entry_valid + now.tv_sec + 2 >= INT_MAX) {
timeout->tv_sec = INT_MAX;
} else {
duration.tv_sec = feo->entry_valid;
duration.tv_nsec = feo->entry_valid_nsec;
timespecadd(&duration, &now, timeout);
}
}
/* access */
#define FVP_ACCESS_NOOP 0x01

View File

@ -291,20 +291,12 @@ fuse_vnode_get(struct mount *mp,
if (dvp != NULL && cnp != NULL && (cnp->cn_flags & MAKEENTRY) != 0 &&
feo != NULL &&
(feo->entry_valid != 0 || feo->entry_valid_nsec != 0)) {
struct timespec duration, now, timeout;
struct timespec timeout;
ASSERT_VOP_LOCKED(*vpp, "fuse_vnode_get");
ASSERT_VOP_LOCKED(dvp, "fuse_vnode_get");
getnanouptime(&now);
if (feo->entry_valid >= INT_MAX ||
feo->entry_valid + now.tv_sec + 2 >= INT_MAX) {
timeout.tv_sec = INT_MAX;
} else {
duration.tv_sec = feo->entry_valid;
duration.tv_nsec = feo->entry_valid_nsec;
timespecadd(&duration, &now, &timeout);
}
fuse_validity_2_timespec(feo, &timeout);
cache_enter_time(dvp, *vpp, cnp, &timeout, NULL);
}

View File

@ -706,6 +706,9 @@ fuse_vnop_lookup(struct vop_lookup_args *ap)
struct fuse_dispatcher fdi;
enum fuse_opcode op;
struct fuse_entry_out *feo = NULL;
struct fuse_attr_out *fao = NULL;
struct fuse_attr *fattr = NULL;
uint64_t nid;
struct fuse_access_param facp;
@ -802,15 +805,18 @@ fuse_vnop_lookup(struct vop_lookup_args *ap)
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)
* See PR 236226
*/
if (nid == 0) {
/* zero nodeid means ENOENT and cache it */
struct timespec timeout;
fdi.answ_stat = ENOENT;
lookup_err = ENOENT;
if (cnp->cn_flags & MAKEENTRY) {
feo = (struct fuse_entry_out *)fdi.answ;
fuse_validity_2_timespec(feo, &timeout);
cache_enter_time(dvp, *vpp, cnp, &timeout,
NULL);
}
} else if (nid == FUSE_ROOT_ID) {
lookup_err = EINVAL;
}
@ -849,24 +855,7 @@ fuse_vnop_lookup(struct vop_lookup_args *ap)
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) != 0) {
SDT_PROBE2(fuse, , vnops, trace, 1,
"inserting NULL into cache");
cache_enter(dvp, NULL, cnp);
}
#endif
err = ENOENT;
goto out;
@ -874,9 +863,6 @@ fuse_vnop_lookup(struct vop_lookup_args *ap)
/* !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 {
@ -1040,47 +1026,16 @@ fuse_vnop_lookup(struct vop_lookup_args *ap)
}
if (op == FUSE_GETATTR) {
struct fuse_attr_out *fao =
(struct fuse_attr_out*)fdi.answ;
fao = (struct fuse_attr_out*)fdi.answ;
fuse_internal_cache_attrs(*vpp,
&fao->attr, fao->attr_valid,
fao->attr_valid_nsec, NULL);
} else {
struct fuse_entry_out *feo =
(struct fuse_entry_out*)fdi.answ;
feo = (struct fuse_entry_out*)fdi.answ;
fuse_internal_cache_attrs(*vpp,
&feo->attr, feo->attr_valid,
feo->attr_valid_nsec, NULL);
}
/* 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) {

View File

@ -169,8 +169,7 @@ TEST_F(Lookup, entry_cache)
* If the daemon returns an error of 0 and an inode of 0, that's a flag for
* "ENOENT and cache it" with the given entry_timeout
*/
/* https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=236226 */
TEST_F(Lookup, DISABLED_entry_cache_negative)
TEST_F(Lookup, entry_cache_negative)
{
struct timespec entry_valid = {.tv_sec = TIME_T_MAX, .tv_nsec = 0};