When dealing with high I/O throughput, we could run out of buffer
space if the filesystem was not consuming requests fast enough.
Here we slightly raise the buffer size, and we make it tunable
through the PERFUSE_BUFSIZE environment variable so that we can
cope with higher requirement later.
While there, document PERFUSE_OPTIONS environment variable.
Some filesystems do not provide inode numbers through readdir (FUSE mounts
without -o use_ino). We therefore have to lookup each directory entry to
get the missing numbers.
dot and double-dot are exceptions, as we already know the values. Moreover,
the lookup code does not expect to get requests for dot and will abort
perfused(8) when it gets some. In order to fix that, we just check for
dot and double-dot special case and use the known values instead of sending
a lookup.
FUSE filesystems do not expect to get metadata updates for [amc]time
and size, they updates the value on their own after operations.
The PUFFS PUFFS_KFLAG_NOFLUSH_META option prevents regular metadata cache
flushes to the filesystem , and libperfuse uses it to match Linux FUSE
behavior.
While there, fix a bug in SETATTR: do not update kernel metadata cache
from SETATTR reply when the request is asynchronous, as we do not have
the reply yet.
FUSE filesystems assume that SETATTR with atime is the result of utiimes()
being called. As a result, atime and mtime will be updated. This happens
with MooseFS and glusterFS. atime is supposed to be updated by the
filesystem itself when it gets read operations.
We fix the problem in SETATTR operations by
1) do not create a mtime update when we have an atime update (and vice
versa), just fill the fields to avoid the filesystem restting the
missing field to Epoch, but do not pretend we want to update it.
2) If the change is limited to atime, iscard it, as updates should be
done by READ operations
3) Kernel part of PUFFS has been fixed to make sure reads on empty file
are sent to the filesystem:
http://mail-index.netbsd.org/source-changes/2015/01/13/msg062364.html
Thanks to Tom Ivar Helbekkmo for reporting this issue.
- access denied is EPERM and not EACCES
- access to file owned by someone else in a sticy-bit directory should
be allowed for the sticy-bit directory owner
- setting sticky-bit on a non directory should produce EFTYPE
- implement PATHCONF method as much as we can.
to implement FUSE's OPEN_IO_DIRECT, by which the filesystem tells the kernel
that read/write to the file should bypass the page cache.
Remove a warning about read beyond EOF which will now normally appear when
page cache is bypassed.
able to fetch inode number. We perfom an addtional lookup
on each file to get it.
In that case, do not lookup .. from root, as it breaks
out of the filesystem and hits NULL pointers.
1) with recent FUSE, when lookup returns a null ino, it means ENOENT
2) odd corner case that caused a bug on dd if=test of=test conv=notrunc
This caused the file to be open first ro, then rw. A logic bug in
perfuse_node_open caused it to skip the second operation, whereas
it should open for writing, and store the write FH without touching
the read FH.
parent, keeping them active, and allowing to lookup .. without sending
a request to the filesystem.
Enable the featuure for perfused, as this is how FUSE works.
The scheme used to retreive known nodes on lookup was flawed, as it only
used parent and name. This produced a different cookie for the same file
if it was renamed, when looking up ../ or when dealing with multiple files
associated with the same name through link(2).
We therefore abandon the use of node name and introduce hashed lists of
inodes. This causes a huge rewrite of reclaim code, which do not attempt
to keep parents allocated until all their children are reclaimed
- Fix race conditions in reclaim
There are a few situations where we issue multiple FUSE operations for
a PUFFS operation. On reclaim, we therefore have to wait for all FUSE
operation to complete, not just the current exchanges. We do this by
introducing node reference count with node_ref() and node_rele().
- Detect data loss caused by FAF
VOP_PUTPAGES causes FAF writes where the kernel does not check the
operation result. At least issue a warning on error.
- Enjoy FAF shortcut on setattr
No need to wait for the result if the kernel does not want it. There is
however an exception for setattr that touch the size, we need to wait
for completion because we have other operations queued for after the
resize.
- Fix fchmod() on write-open file
fchmod() on a node open with write privilege will send setattr with both mode and size set. This confuses some FUSE filesystem. Therefore we send two FUSE operations, one for mode, and one for size.
- Remove node TTL handling for netbsd-5 for simplicity sake. The code
still builds on netbsd-5 but does not have the node TTL feature anymore.
It works fine with kernel support on netbsd-6.
attribute and TTL fora newly created node. Instead extend puffs_newinfo
and add puffs_newinfo_setva() and puffs_newinfo_setttl()
- Remove node_mk_common_final in libperfuse. It used to set uid/gid for
a newly created vnode but has been made redundant along time ago since
uid and gid are properly set in FUSE header.
- In libperfuse, check for corner case where opc = 0 on INACTIVE and RECLAIM (how is it possible? Check for it to avoid a crash anyway)
- In libperfuse, make sure we unlimit RLIMIT_AS and RLIMIT_DATA so that
we do notrun out of memory because the kernel is lazy at reclaiming vnodes.
- In libperfuse, cleanup style of perfuse_destroy_pn()
FUSE TTL on name and attributes are used. This save many PUFFS
operations and improves performances.
PUFFS_KFLAG_CACHE_FS_TTL is #ifdef'ed in many places for now so that
libperfuse can still be used on netbsd-5.
many times it looked up. All reclaims but the last one must be ignored,
otherwise we discard a node which will still get operations. We therefore
have to keep track of lookup/reclaim count and hnour reclaims only when
the count reaches zero.
christos