- 31 Jul, 2014 36 commits
-
-
Jeff Layton authored
...to better match other functions that deal with open/lock stateids. Signed-off-by:
Jeff Layton <jlayton@primarydata.com> Signed-off-by:
J. Bruce Fields <bfields@redhat.com>
-
Jeff Layton authored
When we remove the client_mutex, we'll have a potential race between FREE_STATEID and CLOSE. The root of the problem is that we are walking the st_locks list, dropping the spinlock and then trying to release the persistent reference to the lockstateid. In between, a FREE_STATEID call can come along and take the lock, find the stateid and then try to put the reference. That leads to a double put. Fix this by not releasing the cl_lock in order to release each lock stateid. Use put_generic_stateid_locked to unhash them and gather them onto a list, and free_ol_stateid_reaplist to free any that end up on the list. Signed-off-by:
-
Jeff Layton authored
Releasing an openowner is a bit inefficient as it can potentially thrash the cl_lock if you have a lot of stateids attached to it. Once we remove the client_mutex, it'll also potentially be dangerous to do this. Add some functions to make it easier to defer the part of putting a generic stateid reference that needs to be done outside the cl_lock while doing the parts that must be done while holding it under a single lock. First we unhash each open stateid. Then we call put_generic_stateid_locked which will put the reference to an nfs4_ol_stateid. If it turns out to be the last reference, it'll go ahead and remove the stid from the IDR tree and put it onto the reaplist using the st_locks list_head. Then, after dropping the lock we'll call free_ol_stateid_reaplist to walk the list of stateids that are fully unhashed and ready to be freed, and free each of them. This function can sleep, so it must be done outside any spinlocks. Signed-off-by:
-
Jeff Layton authored
Once we remove the client_mutex, it'll be possible for the sc_type of a lock stateid to change after it's found and checked, but before we can go to destroy it. If that happens, we can end up putting the persistent reference to the stateid more than once, and unhash it more than once. Fix this by unhashing the lock stateid prior to dropping the cl_lock but after finding it. Signed-off-by:
-
Jeff Layton authored
Reduce the cl_lock trashing in destroy_lockowner. Unhash all of the lockstateids on the lockowner's list. Put the reference under the lock and see if it was the last one. If so, then add it to a private list to be destroyed after we drop the lock. Signed-off-by:
-
Jeff Layton authored
Once we remove the client_mutex, we'll need to properly protect the stateowner reference counts using the cl_lock. Signed-off-by:
-
Jeff Layton authored
Do more within the main loop, and simplify the function a bit. Also, there's no need to take a stateowner reference unless we're going to call release_lockowner. Signed-off-by:
-
Trond Myklebust authored
Preparation for removing the client_mutex. Convert the open owner hash table into a per-client table and protect it using the nfs4_client->cl_lock spin lock. Signed-off-by:
Trond Myklebust <trond.myklebust@primarydata.com> Signed-off-by:
-
Trond Myklebust authored
Once we remove client mutex protection, we'll need to ensure that stateowner lookup and creation are atomic between concurrent compounds. Ensure that alloc_init_lock_stateowner checks the hashtable under the client_lock before adding a new element. Signed-off-by:
-
Trond Myklebust authored
Once we remove client mutex protection, we'll need to ensure that stateowner lookup and creation are atomic between concurrent compounds. Ensure that alloc_init_open_stateowner checks the hashtable under the client_lock before adding a new element. Signed-off-by:
-
Jeff Layton authored
Once we remove client_mutex protection, it'll be possible to have an in-flight operation using an openstateid when a CLOSE call comes in. If that happens, we can't just put the sc_file reference and clear its pointer without risking an oops. Fix this by ensuring that v4.0 CLOSE operations wait for the refcount to drop before proceeding to do so. Signed-off-by:
-
Jeff Layton authored
Change it so that only openstateids hold persistent references to openowners. References can still be held by compounds in progress. With this, we can get rid of NFS4_OO_NEW. It's possible that we will create a new openowner in the process of doing the open, but something later fails. In the meantime, another task could find that openowner and start using it on a successful open. If that occurs we don't necessarily want to tear it down, just put the reference that the failing compound holds. Signed-off-by:
-
Jeff Layton authored
Ensure that lockowner references are only held by lockstateids and operations that are in-progress. With this, we can get rid of release_lockowner_if_empty, which will be racy once we remove client_mutex protection. Signed-off-by:
-
Trond Myklebust authored
A necessary step toward client_mutex removal. Signed-off-by:
-
Jeff Layton authored
Allow stateowners to be unhashed and destroyed when the last reference is put. The unhashing must be idempotent. In a future patch, we'll add some locking around it, but for now it's only protected by the client_mutex. Signed-off-by:
-
Jeff Layton authored
Ensure that when finding or creating a lockowner, that we get a reference to it. For now, we also take an extra reference when a lockowner is created that can be put when release_lockowner is called, but we'll remove that in a later patch once we change how references are held. Since we no longer destroy lockowners in the event of an error in nfsd4_lock, we must change how the seqid gets bumped in the lk_is_new case. Instead of doing so on creation, do it manually in nfsd4_lock. Signed-off-by:
-
Jeff Layton authored
We don't want to rely on the client_mutex for protection in the case of NFSv4 open owners. Instead, we add a mutex that will only be taken for NFSv4.0 state mutating operations, and that will be released once the entire compound is done. Also, ensure that nfsd4_cstate_assign_replay/nfsd4_cstate_clear_replay take a reference to the stateowner when they are using it for NFSv4.0 open and lock replay caching. Signed-off-by:
-
Jeff Layton authored
The way stateowners are managed today is somewhat awkward. They need to be explicitly destroyed, even though the stateids reference them. This will be particularly problematic when we remove the client_mutex. We may create a new stateowner and attempt to open a file or set a lock, and have that fail. In the meantime, another RPC may come in that uses that same stateowner and succeed. We can't have the first task tearing down the stateowner in that situation. To fix this, we need to change how stateowners are tracked altogether. Refcount them and only destroy them once all stateids that reference them have been destroyed. This patch starts by adding the refcounting necessary to do that. Signed-off-by:
-
Trond Myklebust authored
Allow nfs4_find_stateid_by_type to take the stateid reference, while still holding the &cl->cl_lock. Necessary step toward client_mutex removal. Signed-off-by:
-
Trond Myklebust authored
Allow nfs4_lookup_stateid to take the stateid reference, instead of having all the callers do so. Signed-off-by:
-
Trond Myklebust authored
Allow nfs4_preprocess_seqid_op to take the stateid reference, instead of having all the callers do so. Signed-off-by:
-
Trond Myklebust authored
Ensure that all the callers put the open stateid after use. Necessary step toward client_mutex removal. Signed-off-by:
-
Trond Myklebust authored
Ensure that nfsd4_open_confirm() keeps a reference to the open stateid until it is done working with it. Necessary step toward client_mutex removal. Signed-off-by:
-
Trond Myklebust authored
Prepare nfsd4_close for a future where nfs4_preprocess_seqid_op() hands it a fully referenced open stateid. Necessary step toward client_mutex removal. Signed-off-by:
-
Trond Myklebust authored
Ensure that nfsd4_process_open2() keeps a reference to the open stateid until it is done working with it. Necessary step toward client_mutex removal. Signed-off-by:
-
Trond Myklebust authored
Ensure that nfsd4_process_open2() keeps a reference to the delegation stateid until it is done working with it. Necessary step toward client_mutex removal. Signed-off-by:
-
Trond Myklebust authored
Ensure that nfs4_open_delegation() keeps a reference to the delegation stateid until it is done working with it. Necessary step toward client_mutex removal. Signed-off-by:
-
Trond Myklebust authored
Ensure that nfsd4_locku() keeps a reference to the lock stateid until it is done working with it. Necessary step toward client_mutex removal. Signed-off-by:
-
Trond Myklebust authored
Ensure that nfsd4_lock() references the lock stateid while it is manipulating it. Not currently necessary, but will be once the client_mutex is removed. Signed-off-by:
-
Jeff Layton authored
Hold the cl_lock over the bulk of these functions. In addition to ensuring that they aren't freed prematurely, this will also help prevent a potential race that could be introduced later. Once we remove the client_mutex, it'll be possible for FREE_STATEID and CLOSE to race and for both to try to put the "persistent" reference to the stateid. Signed-off-by:
-
Jeff Layton authored
Preparation for removal of the client_mutex. Currently, no lock aside from the client_mutex is held when calling find_lock_state. Ensure that the cl_lock is held by adding a lockdep assertion. Once we remove the client_mutex, it'll be possible for another thread to race in and insert a lock state for the same file after we search but before we insert a new one. Ensure that doesn't happen by redoing the search after allocating a new stid that we plan to insert. If one is found just put the one that was allocated. Signed-off-by:
-
Jeff Layton authored
Change to using the clp->cl_lock for this. For now, there's a lot of cl_lock thrashing, but in later patches we'll eliminate that and close the potential races that can occur when releasing the cl_lock while walking the lists. For now, the client_mutex prevents those races. Signed-off-by:
-
Jeff Layton authored
Releasing locks when we unhash the stateid instead of doing so only when the stateid is actually released will be problematic in later patches when we need to protect the unhashing with spinlocks. Move it into the sc_free operation instead. Signed-off-by:
Christoph Hellwig <hch@lst.de> Signed-off-by:
-
Jeff Layton authored
Currently, this is serialized by the client_mutex, which is slated for removal. Add finer-grained locking here. Also, do some cleanup around find_stateid to prepare for taking references. Signed-off-by:
Benny Halevy <bhalevy@primarydata.com> Signed-off-by:
-
Trond Myklebust authored
All stateids are associated with a nfs4_file. Let's consolidate. Replace delegation->dl_file with the dl_stid.sc_file, and nfs4_ol_stateid->st_file with st_stid.sc_file. Signed-off-by:
-
Trond Myklebust authored
When we remove the client_mutex, we'll need to be able to ensure that these objects aren't destroyed while we're not holding locks. Add a ->free() callback to the struct nfs4_stid, so that we can release a reference to the stid without caring about the contents. Signed-off-by:
-
- 30 Jul, 2014 1 commit
-
-
Jeff Layton authored
It's possible for nfsd to fail opening a file that it has just created. When that happens, we throw a WARN but it doesn't include any info about the error code. Print the status code to give us a bit more info. Our QA group hit some of these warnings under some very heavy stress testing. My suspicion is that they hit the file-max limit, but it's hard to know for sure. Go ahead and add a -ENFILE mapping to nfserr_serverfault to make the error more distinct (and correct). Signed-off-by:
-
- 29 Jul, 2014 3 commits
-
-
Trond Myklebust authored
Signed-off-by:
-
Trond Myklebust authored
If requests are queued in the socket inbuffer waiting for an svc_tcp_has_wspace() requirement to be satisfied, then we do not want to clear the SOCK_NOSPACE flag until we've satisfied that requirement. Signed-off-by:
-
Trond Myklebust authored
Ensure that all calls to svc_xprt_enqueue() except svc_xprt_received() check the value of XPT_BUSY, before attempting to grab spinlocks etc. This is to avoid situations such as the following "perf" trace, which shows heavy contention on the pool spinlock: 54.15% nfsd [kernel.kallsyms] [k] _raw_spin_lock_bh | --- _raw_spin_lock_bh | |--71.43%-- svc_xprt_enqueue | | | |--50.31%-- svc_reserve | | | |--31.35%-- svc_xprt_received | | | |--18.34%-- svc_tcp_data_ready ... Signed-off-by:
-
