back to the lever before r343030.  For 64-bit machines reduce it slightly,
too.  Together with r343030 I bumped the limit up to the value we use at
Netflix to serve 100 Gbit/s of sendfile traffic, and it probably isn't a
good default.

Provide a loader tunable to change vnode pager pbufs count. Document it.

ViewVC

The cached fvdat->filesize is indepedent of the (mostly unused)
cached_attrs, and we failed to update it when a cached (but perhaps
inactive) vnode was found during VOP_LOOKUP to have a different size than
cached.

As noted in the code comment, this can occur in distributed filesystems or
with other kinds of irregular file behavior (anything is possible in FUSE).

We do something similar in fuse_vnop_getattr already.

PR:		230258 (as reported in description; other issues explored in
			comments are not all resolved)
Reported by:	MooseFS FreeBSD Team <freebsd AT moosefs.com>
Submitted by:	Jakub Kruszona-Zawadzki <acid AT moosefs.com> (earlier version)

ViewVC

At least prior to 7.23 (which adds FUSE_WRITEBACK_CACHE), the FUSE protocol
specifies only clean data to be cached.

Prior to this change, we implement and default to writeback caching.  This
is ok enough for local only filesystems without hardlinks, but violates the
general design contract with FUSE and breaks distributed filesystems or
concurrent access to hardlinks of the same inode.

In this change, add cache mode as an extension of cache enable/disable.  The
new modes are UC (was: cache disabled), WT (default), and WB (was: cache
enabled).

For now, WT caching is implemented as write-around, which meets the goal of
only caching clean data.  WT can be better than WA for workloads that
frequently read data that was recently written, but WA is trivial to
implement.  Note that this has no effect on O_WRONLY-opened files, which
were already coerced to write-around.

Refs:
  * https://sourceforge.net/p/fuse/mailman/message/8902254/
  * https://github.com/vgough/encfs/issues/315

PR:		230258 (inspired by)

ViewVC

Most users of fuse_vnode_setsize() set the cached fvdat->filesize and update
the buf cache bounds as a result of either a read from the underlying FUSE
filesystem, or as part of a write-through type operation (like truncate =>
VOP_SETATTR).  In these cases, do not set the FN_SIZECHANGE flag, which
indicates that an inode's data is dirty (in particular, that the local buf
cache and fvdat->filesize have dirty extended data).

PR:		230258 (related)

ViewVC

The FUSE protocol demands that kernel implementations cache user filesystem
path components (lookup/cnp data) for a maximum period of time in the range
of [0, ULONG_MAX] seconds.  In practice, typical requests are for 0, 1, or
10 seconds; or "a long time" to represent indefinite caching.

Historically, FreeBSD FUSE has ignored this client directive entirely.  This
works fine for local-only filesystems, but causes consistency issues with
multi-writer network filesystems.

For now, respect 0 second cache TTLs and do not cache such metadata.
Non-zero metadata caching TTLs in the range [0.000000001, ULONG_MAX] seconds
are still cached indefinitely, because it is unclear how a userspace
filesystem could do anything sensible with those semantics even if
implemented.

Pass fuse_entry_out to fuse_vnode_get when available and only cache lookup
if the user filesystem did not set a zero second TTL.

PR:		230258 (inspired by; does not fix)

ViewVC

The FUSE protocol demands that kernel implementations cache user filesystem
file attributes (vattr data) for a maximum period of time in the range of
[0, ULONG_MAX] seconds.  In practice, typical requests are for 0, 1, or 10
seconds; or "a long time" to represent indefinite caching.

Historically, FreeBSD FUSE has ignored this client directive entirely.  This
works fine for local-only filesystems, but causes consistency issues with
multi-writer network filesystems.

For now, respect 0 second cache TTLs and do not cache such metadata.
Non-zero metadata caching TTLs in the range [0.000000001, ULONG_MAX] seconds
are still cached indefinitely, because it is unclear how a userspace
filesystem could do anything sensible with those semantics even if
implemented.

In the future, as an optimization, we should implement notify_inval_entry,
etc, which provide userspace filesystems a way of evicting the kernel cache.

One potentially bogus access to invalid cached attribute data was left in
fuse_io_strategy.  It is restricted behind the undocumented and non-default
"vfs.fuse.fix_broken_io" sysctl or "brokenio" mount option; maybe these are
deadcode and can be eliminated?

Some minor APIs changed to facilitate this:

1. Attribute cache validity is tracked in FUSE inodes ("fuse_vnode_data").

2. cache_attrs() respects the provided TTL and only caches in the FUSE
inode if TTL > 0.  It also grows an "out" argument, which, if non-NULL,
stores the translated fuse_attr (even if not suitable for caching).

3. FUSE VTOVA(vp) returns NULL if the vnode's cache is invalid, to help
avoid programming mistakes.

4. A VOP_LINK check for potential nlink overflow prior to invoking the FUSE
link op was weakened (only performed when we have a valid attr cache).  The
check is racy in a multi-writer network filesystem anyway -- classic TOCTOU.
We have to trust any userspace filesystem that rejects local caching to
account for it correctly.

PR:		230258 (inspired by; does not fix)

ViewVC

Suggested by:	kevans
Reviewed by:	kevans

ViewVC

ViewVC

ViewVC