The next commit will introduce nfsd_open_local_fh() which returns an
nfsd_file structure. This commit exposes LOCALIO's required NFSD
symbols to the NFS client:
- Make nfsd_open_local_fh() symbol and other required NFSD symbols
available to NFS in a global 'nfs_to' nfsd_localio_operations
struct (global access suggested by Trond, nfsd_localio_operations
suggested by NeilBrown). The next commit will also introduce
nfsd_localio_ops_init() that init_nfsd() will call to initialize
'nfs_to'.
- Introduce nfsd_file_file() that provides access to nfsd_file's
backing file. Keeps nfsd_file structure opaque to NFS client (as
suggested by Jeff Layton).
- Introduce nfsd_file_put_local() that will put the reference to the
nfsd_file's associated nn->nfsd_serv and then put the reference to
the nfsd_file (as suggested by NeilBrown).
Suggested-by: Trond Myklebust <trond.myklebust@hammerspace.com> # nfs_to
Suggested-by: NeilBrown <neilb@suse.de> # nfsd_localio_operations
Suggested-by: Jeff Layton <jlayton@kernel.org> # nfsd_file_file
Signed-off-by: Mike Snitzer <snitzer@kernel.org>
Reviewed-by: NeilBrown <neilb@suse.de>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Anna Schumaker <anna.schumaker@oracle.com>
nfsd_file_acquire_local() can be used to look up a file by filehandle
without having a struct svc_rqst. This can be used by NFS LOCALIO to
allow the NFS client to bypass the NFS protocol to directly access a
file provided by the NFS server which is running in the same kernel.
In nfsd_file_do_acquire() care is taken to always use fh_verify() if
rqstp is not NULL (as is the case for non-LOCALIO callers). Otherwise
the non-LOCALIO callers will not supply the correct and required
arguments to __fh_verify (e.g. gssclient isn't passed).
Introduce fh_verify_local() wrapper around __fh_verify to make it
clear that LOCALIO is intended caller.
Also, use GC for nfsd_file returned by nfsd_file_acquire_local. GC
offers performance improvements if/when a file is reopened before
launderette cleans it from the filecache's LRU.
Suggested-by: Jeff Layton <jlayton@kernel.org> # use filecache's GC
Signed-off-by: NeilBrown <neilb@suse.de>
Co-developed-by: Mike Snitzer <snitzer@kernel.org>
Signed-off-by: Mike Snitzer <snitzer@kernel.org>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Anna Schumaker <anna.schumaker@oracle.com>
nfsd_file_put() in one thread can race with another thread doing
garbage collection (running nfsd_file_gc() -> list_lru_walk() ->
nfsd_file_lru_cb()):
* In nfsd_file_put(), nf->nf_ref is 1, so it tries to do nfsd_file_lru_add().
* nfsd_file_lru_add() returns true (with NFSD_FILE_REFERENCED bit set)
* garbage collector kicks in, nfsd_file_lru_cb() clears REFERENCED bit and
returns LRU_ROTATE.
* garbage collector kicks in again, nfsd_file_lru_cb() now decrements nf->nf_ref
to 0, runs nfsd_file_unhash(), removes it from the LRU and adds to the dispose
list [list_lru_isolate_move(lru, &nf->nf_lru, head)]
* nfsd_file_put() detects NFSD_FILE_HASHED bit is cleared, so it tries to remove
the 'nf' from the LRU [if (!nfsd_file_lru_remove(nf))]. The 'nf' has been added
to the 'dispose' list by nfsd_file_lru_cb(), so nfsd_file_lru_remove(nf) simply
treats it as part of the LRU and removes it, which leads to its removal from
the 'dispose' list.
* At this moment, 'nf' is unhashed with its nf_ref being 0, and not on the LRU.
nfsd_file_put() continues its execution [if (refcount_dec_and_test(&nf->nf_ref))],
as nf->nf_ref is already 0, nf->nf_ref is set to REFCOUNT_SATURATED, and the 'nf'
gets no chance of being freed.
nfsd_file_put() can also race with nfsd_file_cond_queue():
* In nfsd_file_put(), nf->nf_ref is 1, so it tries to do nfsd_file_lru_add().
* nfsd_file_lru_add() sets REFERENCED bit and returns true.
* Some userland application runs 'exportfs -f' or something like that, which triggers
__nfsd_file_cache_purge() -> nfsd_file_cond_queue().
* In nfsd_file_cond_queue(), it runs [if (!nfsd_file_unhash(nf))], unhash is done
successfully.
* nfsd_file_cond_queue() runs [if (!nfsd_file_get(nf))], now nf->nf_ref goes to 2.
* nfsd_file_cond_queue() runs [if (nfsd_file_lru_remove(nf))], it succeeds.
* nfsd_file_cond_queue() runs [if (refcount_sub_and_test(decrement, &nf->nf_ref))]
(with "decrement" being 2), so the nf->nf_ref goes to 0, the 'nf' is added to the
dispose list [list_add(&nf->nf_lru, dispose)]
* nfsd_file_put() detects NFSD_FILE_HASHED bit is cleared, so it tries to remove
the 'nf' from the LRU [if (!nfsd_file_lru_remove(nf))], although the 'nf' is not
in the LRU, but it is linked in the 'dispose' list, nfsd_file_lru_remove() simply
treats it as part of the LRU and removes it. This leads to its removal from
the 'dispose' list!
* Now nf->ref is 0, unhashed. nfsd_file_put() continues its execution and set
nf->nf_ref to REFCOUNT_SATURATED.
As shown in the above analysis, using nf_lru for both the LRU list and dispose list
can cause the leaks. This patch adds a new list_head nf_gc in struct nfsd_file, and uses
it for the dispose list. This does not fix the nfsd_file leaking issue completely.
Signed-off-by: Youzhong Yang <youzhong@gmail.com>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
The work of closing a file can have non-trivial cost. Doing it in a
separate work queue thread means that cost isn't imposed on the nfsd
threads and an imbalance can be created. This can result in files being
queued for the work queue more quickly that the work queue can process
them, resulting in unbounded growth of the queue and memory exhaustion.
To avoid this work imbalance that exhausts memory, this patch moves all
closing of files into the nfsd threads. This means that when the work
imposes a cost, that cost appears where it would be expected - in the
work of the nfsd thread. A subsequent patch will ensure the final
__fput() is called in the same (nfsd) thread which calls filp_close().
Files opened for NFSv3 are never explicitly closed by the client and are
kept open by the server in the "filecache", which responds to memory
pressure, is garbage collected even when there is no pressure, and
sometimes closes files when there is particular need such as for rename.
These files currently have filp_close() called in a dedicated work
queue, so their __fput() can have no effect on nfsd threads.
This patch discards the work queue and instead has each nfsd thread call
flip_close() on as many as 8 files from the filecache each time it acts
on a client request (or finds there are no pending client requests). If
there are more to be closed, more threads are woken. This spreads the
work of __fput() over multiple threads and imposes any cost on those
threads.
The number 8 is somewhat arbitrary. It needs to be greater than 1 to
ensure that files are closed more quickly than they can be added to the
cache. It needs to be small enough to limit the per-request delays that
will be imposed on clients when all threads are busy closing files.
Signed-off-by: NeilBrown <neilb@suse.de>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
While we were converting the nfs4_file hashtable to use the kernel's
resizable hashtable data structure, Neil Brown observed that the
list variant (rhltable) would be better for managing nfsd_file items
as well. The nfsd_file hash table will contain multiple entries for
the same inode -- these should be kept together on a list. And, it
could be possible for exotic or malicious client behavior to cause
the hash table to resize itself on every insertion.
A nice simplification is that rhltable_lookup() can return a list
that contains only nfsd_file items that match a given inode, which
enables us to eliminate specialized hash table helper functions and
use the default functions provided by the rhashtable implementation).
Since we are now storing nfsd_file items for the same inode on a
single list, that effectively reduces the number of hash entries
that have to be tracked in the hash table. The mininum bucket count
is therefore lowered.
Light testing with fstests generic/531 show no regressions.
Suggested-by: Neil Brown <neilb@suse.de>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Commit fb70bf124b ("NFSD: Instantiate a struct file when creating a
regular NFSv4 file") added the ability to cache an open fd over a
compound. There are a couple of problems with the way this currently
works:
It's racy, as a newly-created nfsd_file can end up with its PENDING bit
cleared while the nf is hashed, and the nf_file pointer is still zeroed
out. Other tasks can find it in this state and they expect to see a
valid nf_file, and can oops if nf_file is NULL.
Also, there is no guarantee that we'll end up creating a new nfsd_file
if one is already in the hash. If an extant entry is in the hash with a
valid nf_file, nfs4_get_vfs_file will clobber its nf_file pointer with
the value of op_file and the old nf_file will leak.
Fix both issues by making a new nfsd_file_acquirei_opened variant that
takes an optional file pointer. If one is present when this is called,
we'll take a new reference to it instead of trying to open the file. If
the nfsd_file already has a valid nf_file, we'll just ignore the
optional file and pass the nfsd_file back as-is.
Also rework the tracepoints a bit to allow for an "opened" variant and
don't try to avoid counting acquisitions in the case where we already
have a cached open file.
Fixes: fb70bf124b ("NFSD: Instantiate a struct file when creating a regular NFSv4 file")
Cc: Trond Myklebust <trondmy@hammerspace.com>
Reported-by: Stanislav Saner <ssaner@redhat.com>
Reported-and-Tested-by: Ruben Vestergaard <rubenv@drcmr.dk>
Reported-and-Tested-by: Torkil Svensgaard <torkil@drcmr.dk>
Signed-off-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
NFSv4 operations manage the lifetime of nfsd_file items they use by
means of NFSv4 OPEN and CLOSE. Hence there's no need for them to be
garbage collected.
Introduce a mechanism to enable garbage collection for nfsd_file
items used only by NFSv2/3 callers.
Note that the change in nfsd_file_put() ensures that both CLOSE and
DELEGRETURN will actually close out and free an nfsd_file on last
reference of a non-garbage-collected file.
Link: https://bugzilla.linux-nfs.org/show_bug.cgi?id=394
Suggested-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Tested-by: Jeff Layton <jlayton@kernel.org>
Reviewed-by: NeilBrown <neilb@suse.de>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
This reverts commit 5e138c4a75.
That commit attempted to make files available to other users as soon
as all NFSv4 clients were done with them, rather than waiting until
the filecache LRU had garbage collected them.
It gets the reference counting wrong, for one thing.
But it also misses that DELEGRETURN should release a file in the
same fashion. In fact, any nfsd_file_put() on an file held open
by an NFSv4 client needs potentially to release the file
immediately...
Clear the way for implementing that idea.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Reviewed-by: NeilBrown <neilb@suse.de>
Use DEFINE_SHOW_ATTRIBUTE helper macro to simplify the code.
Signed-off-by: ChenXiaoSong <chenxiaosong2@huawei.com>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
The documenting comment for struct nf_file states:
/*
* A representation of a file that has been opened by knfsd. These are hashed
* in the hashtable by inode pointer value. Note that this object doesn't
* hold a reference to the inode by itself, so the nf_inode pointer should
* never be dereferenced, only used for comparison.
*/
Replace the two existing dereferences to make the comment always
true.
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
The last close of a file should enable other accessors to open and
use that file immediately. Leaving the file open in the filecache
prevents other users from accessing that file until the filecache
garbage-collects the file -- sometimes that takes several seconds.
Reported-by: Wang Yugui <wangyugui@e16-tech.com>
Link: https://bugzilla.linux-nfs.org/show_bug.cgi?387
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Add code to initialize and tear down an rhashtable. The rhashtable
is not used yet.
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
The value in this field can always be computed from nf_inode, thus
it is no longer used.
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
This is a measure of how long items stay in the filecache, to help
assess how efficient the cache is.
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
We had a report from the spring Bake-a-thon of data corruption in some
nfstest_interop tests. Looking at the traces showed the NFS server
allowing a v3 WRITE to proceed while a read delegation was still
outstanding.
Currently, we only set NFSD_FILE_BREAK_* flags if
NFSD_MAY_NOT_BREAK_LEASE was set when we call nfsd_file_alloc.
NFSD_MAY_NOT_BREAK_LEASE was intended to be set when finding files for
COMMIT ops, where we need a writeable filehandle but don't need to
break read leases.
It doesn't make any sense to consult that flag when allocating a file
since the file may be used on subsequent calls where we do want to break
the lease (and the usage of it here seems to be reverse from what it
should be anyway).
Also, after calling nfsd_open_break_lease, we don't want to clear the
BREAK_* bits. A lease could end up being set on it later (more than
once) and we need to be able to break those leases as well.
This means that the NFSD_FILE_BREAK_* flags now just mirror
NFSD_MAY_{READ,WRITE} flags, so there's no need for them at all. Just
drop those flags and unconditionally call nfsd_open_break_lease every
time.
Reported-by: Olga Kornieskaia <kolga@netapp.com>
Link: https://bugzilla.redhat.com/show_bug.cgi?id=2107360
Fixes: 65294c1f2c (nfsd: add a new struct file caching facility to nfsd)
Cc: <stable@vger.kernel.org> # 5.4.x : bb283ca18d NFSD: Clean up the show_nf_flags() macro
Cc: <stable@vger.kernel.org> # 5.4.x
Signed-off-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
There have been reports of races that cause NFSv4 OPEN(CREATE) to
return an error even though the requested file was created. NFSv4
does not provide a status code for this case.
To mitigate some of these problems, reorganize the NFSv4
OPEN(CREATE) logic to allocate resources before the file is actually
created, and open the new file while the parent directory is still
locked.
Two new APIs are added:
+ Add an API that works like nfsd_file_acquire() but does not open
the underlying file. The OPEN(CREATE) path can use this API when it
already has an open file.
+ Add an API that is kin to dentry_open(). NFSD needs to create a
file and grab an open "struct file *" atomically. The
alloc_empty_file() has to be done before the inode create. If it
fails (for example, because the NFS server has exceeded its
max_files limit), we avoid creating the file and can still return
an error to the NFS client.
BugLink: https://bugzilla.linux-nfs.org/show_bug.cgi?id=382
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Tested-by: JianHong Yin <jiyin@redhat.com>
The nfsd_file nf_rwsem is currently being used to separate file write
and commit instances to ensure that we catch errors and apply them to
the correct write/commit.
We can improve scalability at the expense of a little accuracy (some
extra false positives) by replacing the nf_rwsem with more careful
use of the errseq_t mechanism to track errors across the different
operations.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
[ cel: rebased on zero-verifier fix ]
Use the 'refcount_t' type instead of 'atomic_t' for improved
refcounting safety.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
Strictly speaking, a stable write error needs to reflect the
write + the commit of that write (and only that write). To
ensure that we don't pick up the write errors from other
writebacks, add a rw_semaphore to provide exclusion.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
Ensure that if the filecache laundrette gets stuck, it only affects
the knfsd instances of one container.
The notifier callbacks can be called from various contexts so avoid
using synchonous filesystem operations that might deadlock.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
Ensure that we can safely clear out the file cache entries when the
nfs server is shut down on a container. Otherwise, the file cache
may end up pinning the mounts.
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
Currently, NFSv2/3 reads and writes have to open a file, do the read or
write and then close it again for each RPC. This is highly inefficient,
especially when the underlying filesystem has a relatively slow open
routine.
This patch adds a new open file cache to knfsd. Rather than doing an
open for each RPC, the read/write handlers can call into this cache to
see if there is one already there for the correct filehandle and
NFS_MAY_READ/WRITE flags.
If there isn't an entry, then we create a new one and attempt to
perform the open. If there is, then we wait until the entry is fully
instantiated and return it if it is at the end of the wait. If it's
not, then we attempt to take over construction.
Since the main goal is to speed up NFSv2/3 I/O, we don't want to
close these files on last put of these objects. We need to keep them
around for a little while since we never know when the next READ/WRITE
will come in.
Cache entries have a hardcoded 1s timeout, and we have a recurring
workqueue job that walks the cache and purges any entries that have
expired.
Signed-off-by: Jeff Layton <jeff.layton@primarydata.com>
Signed-off-by: Weston Andros Adamson <dros@primarydata.com>
Signed-off-by: Richard Sharpe <richard.sharpe@primarydata.com>
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
