The actual bug is not yet addressed as it will get much easier after other
problems are addressed (most notably rename contract).
The only affected in-tree consumer is realpath. Everyone else happens to be
performing lookups within a mount point, having a side effect of ni_dvp being
set to mount point's root vnode in the worst case.
Reported by: pho
This reduces struct namecache by sizeof(void *).
Negative side is that we have to find the previous element (if any) when
removing an entry, but since we normally don't expect collisions it should be
fine.
Note this adds cache_get_hash calls which can be eliminated.
It used to be sizeof of the given struct to accomodate for 32 bit mips
doing 64 bit loads, but the same can be achieved with requireing just
64 bit alignment.
While here reorder struct namecache so that most commonly used fields
are closer.
This removes flag setting/unsetting carried over from regular lookup.
Flags still get for compatibility when falling back.
Note .. and . handling can get partially folded together.
This allows making half-constructed entries visible to the lockless lookup,
which now can check for either "not yet fully constructed" and "no longer valid"
state.
This will be used for .. lookup.
cache_purge locklessly checks whether the vnode at hand has any namecache
entries. This can race with a concurrent purge which managed to remove
the last entry, but may not be done touching the vnode.
Make sure we observe the relevant vnode lock as not taken before proceeding
with vgone.
Paired with the fact that doomed vnodes cannnot receive entries this restores
the invariant that there are no namecache-related writing users past cache_purge
in vgone.
Reported by: pho
This significantly speeds up path lookup, Cascade Lake doing access(2) on ufs
on /usr/obj/usr/src/amd64.amd64/sys/GENERIC/vnode_if.c, ops/s:
before: 2535298
after: 2797621
Over +10%.
The reversed order of computation here does not seem to matter for hash
distribution.
Reviewed by: kib
Differential Revision: https://reviews.freebsd.org/D25921
This makes the realpath syscall operational with the new lookup. Note that the
walk to obtain the full path name still takes locks.
Tested by: pho
Differential Revision: https://reviews.freebsd.org/D23917
Provides full scalability as long as all visited filesystems support the
lookup and terminal vnodes are different.
Inner workings are explained in the comment above cache_fplookup.
Capabilities and fd-relative lookups are not supported and will result in
immediate fallback to regular code.
Symlinks, ".." in the path, mount points without support for lockless lookup
and mismatched counters will result in an attempt to get a reference to the
directory vnode and continue in regular lookup. If this fails, the entire
operation is aborted and regular lookup starts from scratch. However, care is
taken that data is not copied again from userspace.
Sample benchmark:
incremental -j 104 bzImage on tmpfs:
before: 142.96s user 1025.63s system 4924% cpu 23.731 total
after: 147.36s user 313.40s system 3216% cpu 14.326 total
Sample microbenchmark: access calls to separate files in /tmpfs, 104 workers, ops/s:
before: 2165816
after: 151216530
Reviewed by: kib
Tested by: pho (in a patchset)
Differential Revision: https://reviews.freebsd.org/D25578
Previously it would check 4, 3, 2, 1 lists. In practice by the time
it is getting called all lists have some elements and consequently
this does not result in new evictions.
Nonetheless, the code is clearer.
Tested by: pho