A NFSv4.1/4.2 pNFS mount needs to do a
separate Open+LayoutGet RPC, so do not do
a Lookup+Open RPC for these mounts.
The Lookup+Open RPCs are still disabled,
until further testing is done, so this patch
has no effect at this time.
Commit 867c27c23a enabled the n_directio_opens code
in open/close, which sets/clears NNONCACHE, for
IO_APPEND. This code should not be enabled unless
newnfs_directio_enable is non-zero.
This patch reverts that part of commit 867c27c23a.
A future patch that fixes the case where the
file that is being written IO_APPEND is mmap()'d.
MFC after: 3 months
IO_APPEND writes have always been very slow over NFS, due to
the need to acquire an up to date file size after flushing
all writes to the NFS server.
This patch switches the IO_APPEND writes to use direct I/O,
bypassing the buffer cache. As such, flushing of writes
normally only occurs when the open(..O_APPEND..) is done.
It does imply that all writes must be done synchronously
and must be committed to stable storage on the file server
(NFSWRITE_FILESYNC).
For a simple test program that does 10,000 IO_APPEND writes
in a loop, performance improved significantly with this patch.
For a UFS exported file system, the test ran 12x faster.
This drops to 3x faster when the open(2)/close(2) are done
for each loop iteration.
For a ZFS exported file system, the test ran 40% faster.
The much smaller improvement may have been because the ZFS
file system I tested against does not have a ZIL log and
does have "sync" enabled.
Note that IO_APPEND write performance is still much slower
than when done on local file systems.
Although this is a simple patch, it does result in a
significant semantics change, so I have given it a
large MFC time.
Tested by: otis
MFC after: 3 months
As reported in PR#260343, nfs_allocate() did not check
the filesize rlimit. This patch adds that check.
PR: 260343
Reviewed by: asomers
MFC after: 2 weeks
Differential Revision: https://reviews.freebsd.org/D33422
The LookupOpen RPC reduces the number of Open RPCs
needed. Unfortunately, it breaks certain software
builds over NFS, so disable it until this is fixed.
The LookupOpen RPC is only used for NFSv4.1/4.2
mounts when the "oneopenown" mount option is
specified, so this should not affect many users.
VOP_FSYNC() asserts that the vnode is exclusively locked for NFS.
If we try to execute file with recently modified content, the assert is
triggered.
Reviewed by: rmacklem
Tested by: pho
Sponsored by: The FreeBSD Foundation
MFC after: 1 week
Differential revision: https://reviews.freebsd.org/D32999
When the NFSv4.2 server does a VOP_ALLOCATE(), it needs
the operation to be done for the RPC's credential and not
td_ucred. It also needs the writing to be done synchronously.
This patch adds "ioflag" and "cred" arguments to VOP_ALLOCATE()
and modifies vop_stdallocate() to use these arguments.
The VOP_ALLOCATE.9 man page will be patched separately.
Reviewed by: khng, kib
Differential Revision: https://reviews.freebsd.org/D32865
When a forced dismount is in progress, it is possible to
end up looping, retrying commits that fail.
This patch fixes the problem by pretending
that commits succeeded when a forced dismount is in prgress.
MFC after: 2 weeks
Similar to commit 2be417843a, I believe there could be a race between
the NFS client VOP_LOOKUP() and file Writing that could result in stale
file attributes being loaded into the NFS vnode by VOP_LOOKUP().
I have not been able to reproduce a failure due to this race, but
I believe that there are two possibilities:
The Lookup RPC happens while VOP_WRITE() is being executed and loads
stale file attributes after VOP_WRITE() returns when it has already
completed the Write/Commit RPC(s).
--> For this case, setting the local modify timestamp at the end of
VOP_WRITE() should ensure that stale file attributes are not loaded.
The Lookup RPC occurs after VOP_WRITE() has returned, while
asynchronous Write/Commit RPCs are in progress and then is
blocked by the vnode held by VOP_OPEN/VOP_CLOSE/VOP_FSYNC which
will flush writes via ncl_flush() or ncl_vinvalbuf(), clearing the
NMODIFIED flag (which indicates Writes-in-progress). The VOP_LOOKUP()
then acquires the NFS vnode lock and fills in stale file attributes.
--> Setting the local modify timestamp in ncl_flsuh() and ncl_vinvalbuf()
when they clear NMODIFIED should ensure that stale file attributes
are not loaded.
This patch does the above.
PR: 259071
Reviewed by: asomers
MFC after: 2 weeks
Differential Revision: https://reviews.freebsd.org/D32677
Commit 2be417843a added n_localmodtime, which is used by Lookup
and ReaddirPlus to check to see if the file attributes in an RPC
reply might be stale. This patch sets n_localmodtime in Deallocate.
Done as a separate commit, since Deallocate is not in stable/13.
PR: 259071
Reviewed by: asomers
Differential Revision: https://reviews.freebsd.org/D32635
Testing with it, there appears to be a race between Lookup
and VOPs like Setattr-of-size, where Lookup ends up loading
stale attributes (including what might be the wrong file size)
into the NFS vnode's attribute cache.
The race occurs when the modifying VOP (which holds a lock
on the vnode), blocks the acquisition of the vnode in Lookup,
after the RPC (with now potentially stale attributes).
Here's what seems to happen:
Child Parent
does stat(), which does
VOP_LOOKUP(), doing the Lookup
RPC with the directory vnode
locked, acquiring file attributes
valid at this point in time
blocks waiting for locked file does ftruncate(), which
vnode does VOP_SETATTR() of Size,
changing the file's size
while holding an exclusive
lock on the file's vnode
releases the vnode lock
acquires file vnode and fills in
now stale attributes including
the old wrong Size
does a read() which returns
wrong data size
This patch fixes the problem by saving a timestamp in the NFS vnode
in the VOPs that modify the file (Setattr-of-size, Allocate).
Then lookup/readdirplus compares that timestamp with the time just
before starting the RPC after it has acquired the file's vnode.
If the modifying RPC occurred during the Lookup, the attributes
in the RPC reply are discarded, since they might be stale.
With this patch the test program works as expected.
Note that the test program does not fail on a July stable/12,
although this race is in the NFS client code. I suspect a
fairly recent change to the name caching code exposed this
bug.
PR: 259071
Reviewed by: asomers
MFC after: 2 weeks
Differential Revision: https://reviews.freebsd.org/D32635
Without this patch, nfs_allocate() fell back on using vop_stdallocate()
for NFS mounts without Allocate operation support. This was incorrect,
since some file systems, such as ZFS, cannot do allocate via
vop_stdallocate(), which uses writes to try and allocate blocks.
Also, fix nfs_allocate() to return EINVAL when mounts cannot do Allocate,
since that is the correct error for posix_fallocate(2).
Note that Allocate is only supported by some NFSv4.2 servers.
MFC after: 2 weeks
Unlike Copy, the NFSv4.2 Allocate and Deallocate operations do not
allow a reply with partial completion. As such, the only way to
limit the time the operation takes to provide a reasonable RPC RTT
is to limit the size of the allocation/deallocation in the NFSv4.2
client.
This patch uses the sysctl vfs.nfs.maxalloclen to set
the limit on the size of the Deallocate operation.
There is no way to know how long a server will take to do an
deallocate operation, but 64Mbytes results in a reasonable
RPC RTT for the slow hardware I test on.
For an 8Gbyte deallocation, the elapsed time for doing it in 64Mbyte
chunks was the same (within margin of variability) as the
elapsed time taken for a single large deallocation
operation for a FreeBSD server with a UFS file system.
Unlike Copy, the NFSv4.2 Allocate and Deallocate operations do not
allow a reply with partial completion. As such, the only way to
limit the time the operation takes to provide a reasonable RPC RTT
is to limit the size of the allocation/deallocation in the NFSv4.2
client.
This patch adds a sysctl called vfs.nfs.maxalloclen to set
the limit on the size of the Allocate operation.
There is no way to know how long a server will take to do an
allocate operation, but 64Mbytes results in a reasonable
RPC RTT for the slow hardware I test on, so that is what
the default value for vfs.nfs.maxalloclen is set to.
For an 8Gbyte allocation, the elapsed time for doing it in 64Mbyte
chunks was the same as the elapsed time taken for a single large
allocation operation for a FreeBSD server with a UFS file system.
MFC after: 2 weeks
This patch adds a VOP_DEALLOCATE() to the NFS client.
For NFSv4.2 servers that support the Deallocate operation,
it is used. Otherwise, it falls back on calling
vop_stddeallocate().
Reviewed by: kib
Differential Revision: https://reviews.freebsd.org/D31640
This patch adds a Lookup+Open compound RPC to the NFSv4.1/4.2
NFS client, which can be used by nfs_lookup() so that a
subsequent Open RPC is not required.
It uses the cn_flags OPENREAD, OPENWRITE added by commit c18c74a87c.
This reduced the number of RPCs by about 15% for a kernel
build over NFS.
For now, use of Lookup+Open is only done when the "oneopenown"
mount option is used. It may be possible for Lookup+Open to
be used for non-oneopenown NFSv4.1/4.2 mounts, but that will
require extensive further testing to determine if it works.
While here, I've added the changes to the nfscommon module
that are needed to implement the Deallocate NFSv4.2 operation.
This avoids needing another cycle of changes to the internal
KAPI between the NFS modules.
This commit has changed the internal KAPI between the NFS
modules and, as such, all need to be rebuilt from sources.
I have not bumped __FreeBSD_version, since it was bumped a
few days ago.
For NFSv4.1/4.2, if the "oneopenown" mount option is used,
there is, at most, only one open stateid for each NFS vnode.
When an open stateid for a file is acquired, set a pointer to
the open structure in the NFS vnode. This pointer can be used to
acquire the open stateid without searching the open linked list
when the following is true:
- No delegations have been issued for the file. Since delegations
can outlive an NFS vnode for a file, use the global
NFSMNTP_DELEGISSUED flag on the mount to determine this.
- No lock stateid has been issued for the file. To determine
this, a new NFS vnode flag called NMIGHTBELOCKED is set when a lock
stateid is issued, which can then be tested.
When this open structure pointer can be used, it avoids the need to
acquire the NFSCLSTATELOCK() and searching the open structure list for
an open. The NFSCLSTATELOCK() can be highly contended when there are
a lot of opens issued for the NFSv4.1/4.2 mount.
This patch only affects NFSv4.1/4.2 mounts when the "oneopenown"
mount option is used.
MFC after: 2 weeks
There is a NFSv4 file attribute called TimeCreate
that can be used for va_birthtime.
r362175 added some support for use of TimeCreate.
This patch completes support of va_birthtime by adding
support for setting this attribute to the server.
It also eanbles the client to
acquire and set the attribute for a NFSv4
server that supports the attribute.
Reviewed by: markj
MFC after: 2 weeks
Differential Revision: https://reviews.freebsd.org/D30156
When loading attributes from the cache, the NFS client is careful to
copy only the fields that it initialized. After fetching attributes
from the server, however, it would copy the entire vattr structure
initialized from the RPC response, so uninitialized stack bytes would
end up being copied to userspace. In particular, va_birthtime (v2 and
v3) and va_gen (v3) had this problem.
Use a common subroutine to copy fields provided by the NFS client, and
ensure that we provide a dummy va_gen for the v3 case.
Reviewed by: rmacklem
Reported by: KMSAN
MFC after: 1 week
Sponsored by: The FreeBSD Foundation
Differential Revision: https://reviews.freebsd.org/D30090
During code inspection I noticed that the n_direofoffset field
of the NFS node was being manipulated without any lock being
held to make it SMP safe.
This patch adds locking of the NFS node's mutex around
handling of n_direofoffset to make it SMP safe.
I have not seen any failure that could be attributed to n_direofoffset
being manipulated concurrently by multiple processors, but I think this
is possible, since directories are read with shared vnode
locking, plus locks only on individual buffer cache blocks.
However, there have been as yet unexplained issues w.r.t reading
large directories over NFS that could have conceivably been caused
by concurrent manipulation of n_direofoffset.
MFC after: 2 weeks
Commit 3fe2c68ba2 dealt with a panic in cache_enter_time() where
the vnode referred to the directory argument.
It would also be possible to get these panics if a broken
NFS server were to return the directory as an new object being
created within the directory or in a Lookup reply.
This patch adds checks to avoid the panics and logs
messages to indicate that the server is broken for the
file object creation cases.
Reviewd by: kib
MFC after: 2 weeks
Differential Revision: https://reviews.freebsd.org/D28987
successful RPC.
Without this patch, the NFSv4.2 VOP_COPY_FILE_RANGE() client call would
loop until the copy "len" was completed. The problem with doing this is
that it might take a considerable time to complete for a large "len".
By returning after a single successful Copy RPC that copied some of the
data, the application that did the copy_file_range(2) syscall will be
more responsive to signal delivery for large "len" copies.
These macro definitions are no longer needed as the NFS OSX port is long
dead. The vfs_statfs macro conflicts with the vfsops field of the same
name.
Submitted by: shivank@
Reviewed by: rmacklem
MFC after: 2 weeks
Sponsored by: Google, Inc. (GSoC 2020)
Differential Revision: https://reviews.freebsd.org/D25263
I did not realize that zero length attributes are allowed, but they are.
This patch fixes the NFSv4.2 client and server to handle zero length
extended attributes correctly.
Submitted by: Frank van der Linden <fllinden@amazon.com> (earlier version)
Reported by: Frank van der Linden <fllinder@amazon.com>
O_SEARCH is defined by POSIX [0] to open a directory for searching, skipping
permissions checks on the directory itself after the initial open(). This is
close to the semantics we've historically applied for O_EXEC on a directory,
which is UB according to POSIX. Conveniently, O_SEARCH on a file is also
explicitly undefined behavior according to POSIX, so O_EXEC would be a fine
choice. The spec goes on to state that O_SEARCH and O_EXEC need not be
distinct values, but they're not defined to be the same value.
This was pointed out as an incompatibility with other systems that had made
its way into libarchive, which had assumed that O_EXEC was an alias for
O_SEARCH.
This defines compatibility O_SEARCH/FSEARCH (equivalent to O_EXEC and FEXEC
respectively) and expands our UB for O_EXEC on a directory. O_EXEC on a
directory is checked in vn_open_vnode already, so for completeness we add a
NOEXECCHECK when O_SEARCH has been specified on the top-level fd and do not
re-check that when descending in namei.
[0] https://pubs.opengroup.org/onlinepubs/9699919799/
Reviewed by: kib
Differential Revision: https://reviews.freebsd.org/D23247
Filesystems which want to use it in limited capacity can employ the
VOP_UNLOCK_FLAGS macro.
Reviewed by: kib (previous version)
Differential Revision: https://reviews.freebsd.org/D21427
This patch adds support for NFSv4.2 (RFC-7862) and Extended Attributes
(RFC-8276) to the NFS client and server.
NFSv4.2 is comprised of several optional features that can be supported
in addition to NFSv4.1. This patch adds the following optional features:
- posix_fadvise(POSIX_FADV_WILLNEED/POSIX_FADV_DONTNEED)
- posix_fallocate()
- intra server file range copying via the copy_file_range(2) syscall
--> Avoiding data tranfer over the wire to/from the NFS client.
- lseek(SEEK_DATA/SEEK_HOLE)
- Extended attribute syscalls for "user" namespace attributes as defined
by RFC-8276.
Although this patch is fairly large, it should not affect support for
the other versions of NFS. However it does add two new sysctls that allow
a sysadmin to limit which minor versions of NFSv4 a server supports, allowing
a sysadmin to disable NFSv4.2.
Unfortunately, when the NFS stats structure was last revised, it was assumed
that there would be no additional operations added beyond what was
specified in RFC-7862. However RFC-8276 did add additional operations,
forcing the NFS stats structure to revised again. It now has extra unused
entries in all arrays, so that future extensions to NFSv4.2 can be
accomodated without revising this structure again.
A future commit will update nfsstat(1) to report counts for the new NFSv4.2
specific operations/procedures.
This patch affects the internal interface between the nfscommon, nfscl and
nfsd modules and, as such, they all must be upgraded simultaneously.
I will do a version bump (although arguably not needed), due to this.
This code has survived a "make universe" but has not been built with a
recent GCC. If you encounter build problems, please email me.
Relnotes: yes
The current vnode layout is not smp-friendly by having frequently read data
avoidably sharing cachelines with very frequently modified fields. In
particular v_iflag inspected for VI_DOOMED can be found in the same line with
v_usecount. Instead make it available in the same cacheline as the v_op, v_data
and v_type which all get read all the time.
v_type is avoidably 4 bytes while the necessary data will easily fit in 1.
Shrinking it frees up 3 bytes, 2 of which get used here to introduce a new
flag field with a new value: VIRF_DOOMED.
Reviewed by: kib, jeff
Differential Revision: https://reviews.freebsd.org/D22715
We might race with reclaim, and then this is no longer a nfs vnode, in
which case we do not need to handle deferred vnode_pager_setsize()
either.
Reported by: rk@ronald.org
PR: 242184
Sponsored by: The FreeBSD Foundation
MFC after: 3 days
flag and use the same system.
This enables further fault locking improvements by allowing more faults to
proceed with a shared lock.
Reviewed by: kib
Tested by: pho
Differential Revision: https://reviews.freebsd.org/D22116
Make the nfsclient always call vnode_pager_setsize() with the vnode
exclusively locked. This ensures that page fault always can find the
backing page if the object size check succeeded. Set VV_VMSIZEVNLOCK
flag on NFS nodes.
The main offender breaking the interface in nfsclient is
nfs_loadattrcache(), which is used whenever server responded with
updated attributes, which can happen on non-changing operations as
well. Also, iod threads only have buffers locked (and even that is
LK_KERNPROC), but they still may call nfs_loadattrcache() on RPC
response.
Instead of immediately calling vnode_pager_setsize() if server
response indicated changed file size, but the vnode is not exclusively
locked, set a new node flag NVNSETSZSKIP. When the vnode exclusively
locked, or when we can temporary upgrade the lock to exclusive, call
vnode_pager_setsize(), by providing the nfsclient VOP_LOCK() implementation.
Tested by: pho
Discussed with: rmacklem
Sponsored by: The FreeBSD Foundation
MFC after: 2 weeks
Differential revision: https://reviews.freebsd.org/D21883
For a long time, some places in the NFS code have locked/unlocked the
NFS node lock with the macros NFSLOCKNODE()/NFSUNLOCKNODE() whereas
others have simply used mtx_lock()/mtx_unlock().
Since the NFS node mutex needs to change to an sx lock so it can be held when
vnode_pager_setsize() is called, replace all occurrences of mtx_lock/mtx_unlock
with the macros to simply making the change to an sx lock in future commit.
There is no semantic change as a result of this commit.
I am not sure if the change to an sx lock will be MFC'd soon, so I put
an MFC of 1 week on this commit so that it could be MFC'd with that commit.
Suggested by: kib
MFC after: 1 week
These were fully neutered in r177676 (2008), but not removed at the time for
unclear reasons. They're totally dead code, so go ahead and yank them now.
No functional change.
This silly code segment has existed in the sources since it was brought
into FreeBSD 10 years ago. I honestly have no idea why this was done.
It was possible that I thought that it might have been better to not
set B_ASYNC for the "else" case, but I can't remember.
Anyhow, this patch gets rid of the if/else that does the same thing
either way, since it looks silly and upsets a static analyser.
This will have no semantic effect on the NFS client.
PR: 238167
vtruncbuf takes a "struct ucred*" argument. AFAICT, it's been unused ever
since that function was first added in r34611. Remove it. Also, remove some
"struct ucred" arguments from fuse and nfs functions that were only used by
vtruncbuf.
Reviewed by: cem
MFC after: 2 weeks
Sponsored by: The FreeBSD Foundation
Differential Revision: https://reviews.freebsd.org/D20377
The more appropriate place to do the flushing is VOP_OPEN(). This was
uncovered because VOP_SET_TEXT() is now called with the vnode'
vm_object rlocked, which is incompatible with the flush operations.
After the move, there is no need for NFS-specific VOP_SET_TEXT
overload.
Sponsored by: The FreeBSD Foundation
MFC after: 30 days
kern_execve() locks text vnode exclusive to be able to set and clear
VV_TEXT flag. VV_TEXT is mutually exclusive with the v_writecount > 0
condition.
The change removes VV_TEXT, replacing it with the condition
v_writecount <= -1, and puts v_writecount under the vnode interlock.
Each text reference decrements v_writecount. To clear the text
reference when the segment is unmapped, it is recorded in the
vm_map_entry backed by the text file as MAP_ENTRY_VN_TEXT flag, and
v_writecount is incremented on the map entry removal
The operations like VOP_ADD_WRITECOUNT() and VOP_SET_TEXT() check that
v_writecount does not contradict the desired change. vn_writecheck()
is now racy and its use was eliminated everywhere except access.
Atomic check for writeability and increment of v_writecount is
performed by the VOP. vn_truncate() now increments v_writecount
around VOP_SETATTR() call, lack of which is arguably a bug on its own.
nullfs bypasses v_writecount to the lower vnode always, so nullfs
vnode has its own v_writecount correct, and lower vnode gets all
references, since object->handle is always lower vnode.
On the text vnode' vm object dealloc, the v_writecount value is reset
to zero, and deadfs vop_unset_text short-circuit the operation.
Reclamation of lowervp always reclaims all nullfs vnodes referencing
lowervp first, so no stray references are left.
Reviewed by: markj, trasz
Tested by: mjg, pho
Sponsored by: The FreeBSD Foundation
MFC after: 1 month
Differential revision: https://reviews.freebsd.org/D19923
Prior to this patch, nfs_advlock() did NFSVOPUNLOCK(); return (error);
in many places. This patch replaces these code sequenences with a "goto out;"
and does the NFSVOPUNLOCK(); return (error); at the end of the function
in order to make the vnode locking simpler.
This patch does not change the semantics of nfs_advlock().
Suggested by: kib
Reviewed by: kib
Differential Revision: https://reviews.freebsd.org/D17853
A crash was reported where the crash occurred in nfs_advlock() when the
NFS_ISV4(vp) macro was being executed. This was caused by the vnode
being VI_DOOMED due to a forced dismount in progress.
This patch fixes the problem by locking the vnode before executing the
NFS_ISV4() macro.
Tested by: rlibby
PR: 232673
Reviewed by: kib
MFC after: 2 weeks
Differential Revision: https://reviews.freebsd.org/D17757
Use bypass to catch any NFS VOP dispatch and route it through the
wrapper which does sigdeferstop() and then dispatches original
VOP. NFS does not need a bypass below it, which is not supported.
The vop offset in the vop_vector is added since otherwise it is
impossible to get vop_op_t from the internal table, and I did not
wanted to create the layered fs only to wrap NFS VOPs.
VFS_OP()s wrap is straightforward.
Requested and reviewed by: mjg (previous version)
Tested by: pho
Sponsored by: The FreeBSD Foundation
Differential revision: https://reviews.freebsd.org/D17658
Mechanically replace uses of MALLOC/FREE with appropriate invocations of
malloc(9) / free(9) (a series of sed expressions). Something like:
* MALLOC(a, b, ... -> a = malloc(...
* FREE( -> free(
* free((caddr_t) -> free(
No functional change.
For now, punt on modifying contrib ipfilter code, leaving a definition of
the macro in its KMALLOC().
Reported by: jhb
Reviewed by: cy, imp, markj, rmacklem
Sponsored by: Dell EMC Isilon
Differential Revision: https://reviews.freebsd.org/D14035
When allocating memory through malloc(9), we always expect the amount of
memory requested to be unsigned as a negative value would either stand for
an error or an overflow.
Unsign some values, found when considering the use of mallocarray(9), to
avoid unnecessary casting. Also consider that indexes should be of
at least the same size/type as the upper limit they pretend to index.
MFC after: 3 weeks
Uses of mallocarray(9).
The use of mallocarray(9) has rocketed the required swap to build FreeBSD.
This is likely caused by the allocation size attributes which put extra pressure
on the compiler.
Given that most of these checks are superfluous we have to choose better
where to use mallocarray(9). We still have more uses of mallocarray(9) but
hopefully this is enough to bring swap usage to a reasonable level.
Reported by: wosch
PR: 225197
