nfsd: add a new struct file caching facility to nfsd
authorJeff Layton <jeff.layton@primarydata.com>
Sun, 18 Aug 2019 18:18:48 +0000 (14:18 -0400)
committerJ. Bruce Fields <bfields@redhat.com>
Mon, 19 Aug 2019 15:00:39 +0000 (11:00 -0400)
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>
fs/nfsd/Kconfig
fs/nfsd/Makefile
fs/nfsd/export.c
fs/nfsd/filecache.c [new file with mode: 0644]
fs/nfsd/filecache.h [new file with mode: 0644]
fs/nfsd/nfssvc.c
fs/nfsd/trace.h
fs/nfsd/vfs.c
fs/nfsd/vfs.h

index d25f6bbe7006594065807f0462571999458ee618..bff8456220e00da4cddcbb39eca9ff334a0ec461 100644 (file)
@@ -3,6 +3,7 @@ config NFSD
        tristate "NFS server support"
        depends on INET
        depends on FILE_LOCKING
+       depends on FSNOTIFY
        select LOCKD
        select SUNRPC
        select EXPORTFS
index 2bfb58eefad1c2e015178140fc5c9566cf9348fb..6a40b1afe7031a9fe6f58c485fc122366e337d81 100644 (file)
@@ -11,7 +11,8 @@ obj-$(CONFIG_NFSD)    += nfsd.o
 nfsd-y                 += trace.o
 
 nfsd-y                         += nfssvc.o nfsctl.o nfsproc.o nfsfh.o vfs.o \
-                          export.o auth.o lockd.o nfscache.o nfsxdr.o stats.o
+                          export.o auth.o lockd.o nfscache.o nfsxdr.o \
+                          stats.o filecache.o
 nfsd-$(CONFIG_NFSD_FAULT_INJECTION) += fault_inject.o
 nfsd-$(CONFIG_NFSD_V2_ACL) += nfs2acl.o
 nfsd-$(CONFIG_NFSD_V3) += nfs3proc.o nfs3xdr.o
index baa01956a5b3dda259202a0486e99f89a3543376..052fac64b57819c38d50d6b7001ed206af094856 100644 (file)
@@ -22,6 +22,7 @@
 #include "nfsfh.h"
 #include "netns.h"
 #include "pnfs.h"
+#include "filecache.h"
 
 #define NFSDDBG_FACILITY       NFSDDBG_EXPORT
 
@@ -232,6 +233,17 @@ static struct cache_head *expkey_alloc(void)
                return NULL;
 }
 
+static void expkey_flush(void)
+{
+       /*
+        * Take the nfsd_mutex here to ensure that the file cache is not
+        * destroyed while we're in the middle of flushing.
+        */
+       mutex_lock(&nfsd_mutex);
+       nfsd_file_cache_purge();
+       mutex_unlock(&nfsd_mutex);
+}
+
 static const struct cache_detail svc_expkey_cache_template = {
        .owner          = THIS_MODULE,
        .hash_size      = EXPKEY_HASHMAX,
@@ -244,6 +256,7 @@ static const struct cache_detail svc_expkey_cache_template = {
        .init           = expkey_init,
        .update         = expkey_update,
        .alloc          = expkey_alloc,
+       .flush          = expkey_flush,
 };
 
 static int
diff --git a/fs/nfsd/filecache.c b/fs/nfsd/filecache.c
new file mode 100644 (file)
index 0000000..a2fcb25
--- /dev/null
@@ -0,0 +1,885 @@
+/*
+ * Open file cache.
+ *
+ * (c) 2015 - Jeff Layton <jeff.layton@primarydata.com>
+ */
+
+#include <linux/hash.h>
+#include <linux/slab.h>
+#include <linux/hash.h>
+#include <linux/file.h>
+#include <linux/sched.h>
+#include <linux/list_lru.h>
+#include <linux/fsnotify_backend.h>
+#include <linux/fsnotify.h>
+#include <linux/seq_file.h>
+
+#include "vfs.h"
+#include "nfsd.h"
+#include "nfsfh.h"
+#include "filecache.h"
+#include "trace.h"
+
+#define NFSDDBG_FACILITY       NFSDDBG_FH
+
+/* FIXME: dynamically size this for the machine somehow? */
+#define NFSD_FILE_HASH_BITS                   12
+#define NFSD_FILE_HASH_SIZE                  (1 << NFSD_FILE_HASH_BITS)
+#define NFSD_LAUNDRETTE_DELAY               (2 * HZ)
+
+#define NFSD_FILE_LRU_RESCAN                (0)
+#define NFSD_FILE_SHUTDOWN                  (1)
+#define NFSD_FILE_LRU_THRESHOLD                     (4096UL)
+#define NFSD_FILE_LRU_LIMIT                 (NFSD_FILE_LRU_THRESHOLD << 2)
+
+/* We only care about NFSD_MAY_READ/WRITE for this cache */
+#define NFSD_FILE_MAY_MASK     (NFSD_MAY_READ|NFSD_MAY_WRITE)
+
+struct nfsd_fcache_bucket {
+       struct hlist_head       nfb_head;
+       spinlock_t              nfb_lock;
+       unsigned int            nfb_count;
+       unsigned int            nfb_maxcount;
+};
+
+static DEFINE_PER_CPU(unsigned long, nfsd_file_cache_hits);
+
+static struct kmem_cache               *nfsd_file_slab;
+static struct kmem_cache               *nfsd_file_mark_slab;
+static struct nfsd_fcache_bucket       *nfsd_file_hashtbl;
+static struct list_lru                 nfsd_file_lru;
+static long                            nfsd_file_lru_flags;
+static struct fsnotify_group           *nfsd_file_fsnotify_group;
+static atomic_long_t                   nfsd_filecache_count;
+static struct delayed_work             nfsd_filecache_laundrette;
+
+enum nfsd_file_laundrette_ctl {
+       NFSD_FILE_LAUNDRETTE_NOFLUSH = 0,
+       NFSD_FILE_LAUNDRETTE_MAY_FLUSH
+};
+
+static void
+nfsd_file_schedule_laundrette(enum nfsd_file_laundrette_ctl ctl)
+{
+       long count = atomic_long_read(&nfsd_filecache_count);
+
+       if (count == 0 || test_bit(NFSD_FILE_SHUTDOWN, &nfsd_file_lru_flags))
+               return;
+
+       /* Be more aggressive about scanning if over the threshold */
+       if (count > NFSD_FILE_LRU_THRESHOLD)
+               mod_delayed_work(system_wq, &nfsd_filecache_laundrette, 0);
+       else
+               schedule_delayed_work(&nfsd_filecache_laundrette, NFSD_LAUNDRETTE_DELAY);
+
+       if (ctl == NFSD_FILE_LAUNDRETTE_NOFLUSH)
+               return;
+
+       /* ...and don't delay flushing if we're out of control */
+       if (count >= NFSD_FILE_LRU_LIMIT)
+               flush_delayed_work(&nfsd_filecache_laundrette);
+}
+
+static void
+nfsd_file_slab_free(struct rcu_head *rcu)
+{
+       struct nfsd_file *nf = container_of(rcu, struct nfsd_file, nf_rcu);
+
+       put_cred(nf->nf_cred);
+       kmem_cache_free(nfsd_file_slab, nf);
+}
+
+static void
+nfsd_file_mark_free(struct fsnotify_mark *mark)
+{
+       struct nfsd_file_mark *nfm = container_of(mark, struct nfsd_file_mark,
+                                                 nfm_mark);
+
+       kmem_cache_free(nfsd_file_mark_slab, nfm);
+}
+
+static struct nfsd_file_mark *
+nfsd_file_mark_get(struct nfsd_file_mark *nfm)
+{
+       if (!atomic_inc_not_zero(&nfm->nfm_ref))
+               return NULL;
+       return nfm;
+}
+
+static void
+nfsd_file_mark_put(struct nfsd_file_mark *nfm)
+{
+       if (atomic_dec_and_test(&nfm->nfm_ref)) {
+
+               fsnotify_destroy_mark(&nfm->nfm_mark, nfsd_file_fsnotify_group);
+               fsnotify_put_mark(&nfm->nfm_mark);
+       }
+}
+
+static struct nfsd_file_mark *
+nfsd_file_mark_find_or_create(struct nfsd_file *nf)
+{
+       int                     err;
+       struct fsnotify_mark    *mark;
+       struct nfsd_file_mark   *nfm = NULL, *new;
+       struct inode *inode = nf->nf_inode;
+
+       do {
+               mutex_lock(&nfsd_file_fsnotify_group->mark_mutex);
+               mark = fsnotify_find_mark(&inode->i_fsnotify_marks,
+                               nfsd_file_fsnotify_group);
+               if (mark) {
+                       nfm = nfsd_file_mark_get(container_of(mark,
+                                                struct nfsd_file_mark,
+                                                nfm_mark));
+                       mutex_unlock(&nfsd_file_fsnotify_group->mark_mutex);
+                       fsnotify_put_mark(mark);
+                       if (likely(nfm))
+                               break;
+               } else
+                       mutex_unlock(&nfsd_file_fsnotify_group->mark_mutex);
+
+               /* allocate a new nfm */
+               new = kmem_cache_alloc(nfsd_file_mark_slab, GFP_KERNEL);
+               if (!new)
+                       return NULL;
+               fsnotify_init_mark(&new->nfm_mark, nfsd_file_fsnotify_group);
+               new->nfm_mark.mask = FS_ATTRIB|FS_DELETE_SELF;
+               atomic_set(&new->nfm_ref, 1);
+
+               err = fsnotify_add_inode_mark(&new->nfm_mark, inode, 0);
+
+               /*
+                * If the add was successful, then return the object.
+                * Otherwise, we need to put the reference we hold on the
+                * nfm_mark. The fsnotify code will take a reference and put
+                * it on failure, so we can't just free it directly. It's also
+                * not safe to call fsnotify_destroy_mark on it as the
+                * mark->group will be NULL. Thus, we can't let the nfm_ref
+                * counter drive the destruction at this point.
+                */
+               if (likely(!err))
+                       nfm = new;
+               else
+                       fsnotify_put_mark(&new->nfm_mark);
+       } while (unlikely(err == -EEXIST));
+
+       return nfm;
+}
+
+static struct nfsd_file *
+nfsd_file_alloc(struct inode *inode, unsigned int may, unsigned int hashval)
+{
+       struct nfsd_file *nf;
+
+       nf = kmem_cache_alloc(nfsd_file_slab, GFP_KERNEL);
+       if (nf) {
+               INIT_HLIST_NODE(&nf->nf_node);
+               INIT_LIST_HEAD(&nf->nf_lru);
+               nf->nf_file = NULL;
+               nf->nf_cred = get_current_cred();
+               nf->nf_flags = 0;
+               nf->nf_inode = inode;
+               nf->nf_hashval = hashval;
+               atomic_set(&nf->nf_ref, 1);
+               nf->nf_may = may & NFSD_FILE_MAY_MASK;
+               if (may & NFSD_MAY_NOT_BREAK_LEASE) {
+                       if (may & NFSD_MAY_WRITE)
+                               __set_bit(NFSD_FILE_BREAK_WRITE, &nf->nf_flags);
+                       if (may & NFSD_MAY_READ)
+                               __set_bit(NFSD_FILE_BREAK_READ, &nf->nf_flags);
+               }
+               nf->nf_mark = NULL;
+               trace_nfsd_file_alloc(nf);
+       }
+       return nf;
+}
+
+static bool
+nfsd_file_free(struct nfsd_file *nf)
+{
+       bool flush = false;
+
+       trace_nfsd_file_put_final(nf);
+       if (nf->nf_mark)
+               nfsd_file_mark_put(nf->nf_mark);
+       if (nf->nf_file) {
+               get_file(nf->nf_file);
+               filp_close(nf->nf_file, NULL);
+               fput(nf->nf_file);
+               flush = true;
+       }
+       call_rcu(&nf->nf_rcu, nfsd_file_slab_free);
+       return flush;
+}
+
+static void
+nfsd_file_do_unhash(struct nfsd_file *nf)
+{
+       lockdep_assert_held(&nfsd_file_hashtbl[nf->nf_hashval].nfb_lock);
+
+       trace_nfsd_file_unhash(nf);
+
+       --nfsd_file_hashtbl[nf->nf_hashval].nfb_count;
+       hlist_del_rcu(&nf->nf_node);
+       if (!list_empty(&nf->nf_lru))
+               list_lru_del(&nfsd_file_lru, &nf->nf_lru);
+       atomic_long_dec(&nfsd_filecache_count);
+}
+
+static bool
+nfsd_file_unhash(struct nfsd_file *nf)
+{
+       if (test_and_clear_bit(NFSD_FILE_HASHED, &nf->nf_flags)) {
+               nfsd_file_do_unhash(nf);
+               return true;
+       }
+       return false;
+}
+
+/*
+ * Return true if the file was unhashed.
+ */
+static bool
+nfsd_file_unhash_and_release_locked(struct nfsd_file *nf, struct list_head *dispose)
+{
+       lockdep_assert_held(&nfsd_file_hashtbl[nf->nf_hashval].nfb_lock);
+
+       trace_nfsd_file_unhash_and_release_locked(nf);
+       if (!nfsd_file_unhash(nf))
+               return false;
+       /* keep final reference for nfsd_file_lru_dispose */
+       if (atomic_add_unless(&nf->nf_ref, -1, 1))
+               return true;
+
+       list_add(&nf->nf_lru, dispose);
+       return true;
+}
+
+static int
+nfsd_file_put_noref(struct nfsd_file *nf)
+{
+       int count;
+       trace_nfsd_file_put(nf);
+
+       count = atomic_dec_return(&nf->nf_ref);
+       if (!count) {
+               WARN_ON(test_bit(NFSD_FILE_HASHED, &nf->nf_flags));
+               nfsd_file_free(nf);
+       }
+       return count;
+}
+
+void
+nfsd_file_put(struct nfsd_file *nf)
+{
+       bool is_hashed = test_bit(NFSD_FILE_HASHED, &nf->nf_flags) != 0;
+
+       set_bit(NFSD_FILE_REFERENCED, &nf->nf_flags);
+       if (nfsd_file_put_noref(nf) == 1 && is_hashed)
+               nfsd_file_schedule_laundrette(NFSD_FILE_LAUNDRETTE_MAY_FLUSH);
+}
+
+struct nfsd_file *
+nfsd_file_get(struct nfsd_file *nf)
+{
+       if (likely(atomic_inc_not_zero(&nf->nf_ref)))
+               return nf;
+       return NULL;
+}
+
+static void
+nfsd_file_dispose_list(struct list_head *dispose)
+{
+       struct nfsd_file *nf;
+
+       while(!list_empty(dispose)) {
+               nf = list_first_entry(dispose, struct nfsd_file, nf_lru);
+               list_del(&nf->nf_lru);
+               nfsd_file_put_noref(nf);
+       }
+}
+
+static void
+nfsd_file_dispose_list_sync(struct list_head *dispose)
+{
+       bool flush = false;
+       struct nfsd_file *nf;
+
+       while(!list_empty(dispose)) {
+               nf = list_first_entry(dispose, struct nfsd_file, nf_lru);
+               list_del(&nf->nf_lru);
+               if (!atomic_dec_and_test(&nf->nf_ref))
+                       continue;
+               if (nfsd_file_free(nf))
+                       flush = true;
+       }
+       if (flush)
+               flush_delayed_fput();
+}
+
+/*
+ * Note this can deadlock with nfsd_file_cache_purge.
+ */
+static enum lru_status
+nfsd_file_lru_cb(struct list_head *item, struct list_lru_one *lru,
+                spinlock_t *lock, void *arg)
+       __releases(lock)
+       __acquires(lock)
+{
+       struct list_head *head = arg;
+       struct nfsd_file *nf = list_entry(item, struct nfsd_file, nf_lru);
+
+       /*
+        * Do a lockless refcount check. The hashtable holds one reference, so
+        * we look to see if anything else has a reference, or if any have
+        * been put since the shrinker last ran. Those don't get unhashed and
+        * released.
+        *
+        * Note that in the put path, we set the flag and then decrement the
+        * counter. Here we check the counter and then test and clear the flag.
+        * That order is deliberate to ensure that we can do this locklessly.
+        */
+       if (atomic_read(&nf->nf_ref) > 1)
+               goto out_skip;
+       if (test_and_clear_bit(NFSD_FILE_REFERENCED, &nf->nf_flags))
+               goto out_rescan;
+
+       if (!test_and_clear_bit(NFSD_FILE_HASHED, &nf->nf_flags))
+               goto out_skip;
+
+       list_lru_isolate_move(lru, &nf->nf_lru, head);
+       return LRU_REMOVED;
+out_rescan:
+       set_bit(NFSD_FILE_LRU_RESCAN, &nfsd_file_lru_flags);
+out_skip:
+       return LRU_SKIP;
+}
+
+static void
+nfsd_file_lru_dispose(struct list_head *head)
+{
+       while(!list_empty(head)) {
+               struct nfsd_file *nf = list_first_entry(head,
+                               struct nfsd_file, nf_lru);
+               list_del_init(&nf->nf_lru);
+               spin_lock(&nfsd_file_hashtbl[nf->nf_hashval].nfb_lock);
+               nfsd_file_do_unhash(nf);
+               spin_unlock(&nfsd_file_hashtbl[nf->nf_hashval].nfb_lock);
+               nfsd_file_put_noref(nf);
+       }
+}
+
+static unsigned long
+nfsd_file_lru_count(struct shrinker *s, struct shrink_control *sc)
+{
+       return list_lru_count(&nfsd_file_lru);
+}
+
+static unsigned long
+nfsd_file_lru_scan(struct shrinker *s, struct shrink_control *sc)
+{
+       LIST_HEAD(head);
+       unsigned long ret;
+
+       ret = list_lru_shrink_walk(&nfsd_file_lru, sc, nfsd_file_lru_cb, &head);
+       nfsd_file_lru_dispose(&head);
+       return ret;
+}
+
+static struct shrinker nfsd_file_shrinker = {
+       .scan_objects = nfsd_file_lru_scan,
+       .count_objects = nfsd_file_lru_count,
+       .seeks = 1,
+};
+
+static void
+__nfsd_file_close_inode(struct inode *inode, unsigned int hashval,
+                       struct list_head *dispose)
+{
+       struct nfsd_file        *nf;
+       struct hlist_node       *tmp;
+
+       spin_lock(&nfsd_file_hashtbl[hashval].nfb_lock);
+       hlist_for_each_entry_safe(nf, tmp, &nfsd_file_hashtbl[hashval].nfb_head, nf_node) {
+               if (inode == nf->nf_inode)
+                       nfsd_file_unhash_and_release_locked(nf, dispose);
+       }
+       spin_unlock(&nfsd_file_hashtbl[hashval].nfb_lock);
+}
+
+/**
+ * nfsd_file_close_inode_sync - attempt to forcibly close a nfsd_file
+ * @inode: inode of the file to attempt to remove
+ *
+ * Walk the whole hash bucket, looking for any files that correspond to "inode".
+ * If any do, then unhash them and put the hashtable reference to them and
+ * destroy any that had their last reference put. Also ensure that any of the
+ * fputs also have their final __fput done as well.
+ */
+void
+nfsd_file_close_inode_sync(struct inode *inode)
+{
+       unsigned int            hashval = (unsigned int)hash_long(inode->i_ino,
+                                               NFSD_FILE_HASH_BITS);
+       LIST_HEAD(dispose);
+
+       __nfsd_file_close_inode(inode, hashval, &dispose);
+       trace_nfsd_file_close_inode_sync(inode, hashval, !list_empty(&dispose));
+       nfsd_file_dispose_list_sync(&dispose);
+}
+
+/**
+ * nfsd_file_close_inode_sync - attempt to forcibly close a nfsd_file
+ * @inode: inode of the file to attempt to remove
+ *
+ * Walk the whole hash bucket, looking for any files that correspond to "inode".
+ * If any do, then unhash them and put the hashtable reference to them and
+ * destroy any that had their last reference put.
+ */
+static void
+nfsd_file_close_inode(struct inode *inode)
+{
+       unsigned int            hashval = (unsigned int)hash_long(inode->i_ino,
+                                               NFSD_FILE_HASH_BITS);
+       LIST_HEAD(dispose);
+
+       __nfsd_file_close_inode(inode, hashval, &dispose);
+       trace_nfsd_file_close_inode(inode, hashval, !list_empty(&dispose));
+       nfsd_file_dispose_list(&dispose);
+}
+
+/**
+ * nfsd_file_delayed_close - close unused nfsd_files
+ * @work: dummy
+ *
+ * Walk the LRU list and close any entries that have not been used since
+ * the last scan.
+ *
+ * Note this can deadlock with nfsd_file_cache_purge.
+ */
+static void
+nfsd_file_delayed_close(struct work_struct *work)
+{
+       LIST_HEAD(head);
+
+       list_lru_walk(&nfsd_file_lru, nfsd_file_lru_cb, &head, LONG_MAX);
+
+       if (test_and_clear_bit(NFSD_FILE_LRU_RESCAN, &nfsd_file_lru_flags))
+               nfsd_file_schedule_laundrette(NFSD_FILE_LAUNDRETTE_NOFLUSH);
+
+       if (!list_empty(&head)) {
+               nfsd_file_lru_dispose(&head);
+               flush_delayed_fput();
+       }
+}
+
+static int
+nfsd_file_lease_notifier_call(struct notifier_block *nb, unsigned long arg,
+                           void *data)
+{
+       struct file_lock *fl = data;
+
+       /* Only close files for F_SETLEASE leases */
+       if (fl->fl_flags & FL_LEASE)
+               nfsd_file_close_inode_sync(file_inode(fl->fl_file));
+       return 0;
+}
+
+static struct notifier_block nfsd_file_lease_notifier = {
+       .notifier_call = nfsd_file_lease_notifier_call,
+};
+
+static int
+nfsd_file_fsnotify_handle_event(struct fsnotify_group *group,
+                               struct inode *inode,
+                               u32 mask, const void *data, int data_type,
+                               const struct qstr *file_name, u32 cookie,
+                               struct fsnotify_iter_info *iter_info)
+{
+       trace_nfsd_file_fsnotify_handle_event(inode, mask);
+
+       /* Should be no marks on non-regular files */
+       if (!S_ISREG(inode->i_mode)) {
+               WARN_ON_ONCE(1);
+               return 0;
+       }
+
+       /* don't close files if this was not the last link */
+       if (mask & FS_ATTRIB) {
+               if (inode->i_nlink)
+                       return 0;
+       }
+
+       nfsd_file_close_inode(inode);
+       return 0;
+}
+
+
+static const struct fsnotify_ops nfsd_file_fsnotify_ops = {
+       .handle_event = nfsd_file_fsnotify_handle_event,
+       .free_mark = nfsd_file_mark_free,
+};
+
+int
+nfsd_file_cache_init(void)
+{
+       int             ret = -ENOMEM;
+       unsigned int    i;
+
+       clear_bit(NFSD_FILE_SHUTDOWN, &nfsd_file_lru_flags);
+
+       if (nfsd_file_hashtbl)
+               return 0;
+
+       nfsd_file_hashtbl = kcalloc(NFSD_FILE_HASH_SIZE,
+                               sizeof(*nfsd_file_hashtbl), GFP_KERNEL);
+       if (!nfsd_file_hashtbl) {
+               pr_err("nfsd: unable to allocate nfsd_file_hashtbl\n");
+               goto out_err;
+       }
+
+       nfsd_file_slab = kmem_cache_create("nfsd_file",
+                               sizeof(struct nfsd_file), 0, 0, NULL);
+       if (!nfsd_file_slab) {
+               pr_err("nfsd: unable to create nfsd_file_slab\n");
+               goto out_err;
+       }
+
+       nfsd_file_mark_slab = kmem_cache_create("nfsd_file_mark",
+                                       sizeof(struct nfsd_file_mark), 0, 0, NULL);
+       if (!nfsd_file_mark_slab) {
+               pr_err("nfsd: unable to create nfsd_file_mark_slab\n");
+               goto out_err;
+       }
+
+
+       ret = list_lru_init(&nfsd_file_lru);
+       if (ret) {
+               pr_err("nfsd: failed to init nfsd_file_lru: %d\n", ret);
+               goto out_err;
+       }
+
+       ret = register_shrinker(&nfsd_file_shrinker);
+       if (ret) {
+               pr_err("nfsd: failed to register nfsd_file_shrinker: %d\n", ret);
+               goto out_lru;
+       }
+
+       ret = lease_register_notifier(&nfsd_file_lease_notifier);
+       if (ret) {
+               pr_err("nfsd: unable to register lease notifier: %d\n", ret);
+               goto out_shrinker;
+       }
+
+       nfsd_file_fsnotify_group = fsnotify_alloc_group(&nfsd_file_fsnotify_ops);
+       if (IS_ERR(nfsd_file_fsnotify_group)) {
+               pr_err("nfsd: unable to create fsnotify group: %ld\n",
+                       PTR_ERR(nfsd_file_fsnotify_group));
+               nfsd_file_fsnotify_group = NULL;
+               goto out_notifier;
+       }
+
+       for (i = 0; i < NFSD_FILE_HASH_SIZE; i++) {
+               INIT_HLIST_HEAD(&nfsd_file_hashtbl[i].nfb_head);
+               spin_lock_init(&nfsd_file_hashtbl[i].nfb_lock);
+       }
+
+       INIT_DELAYED_WORK(&nfsd_filecache_laundrette, nfsd_file_delayed_close);
+out:
+       return ret;
+out_notifier:
+       lease_unregister_notifier(&nfsd_file_lease_notifier);
+out_shrinker:
+       unregister_shrinker(&nfsd_file_shrinker);
+out_lru:
+       list_lru_destroy(&nfsd_file_lru);
+out_err:
+       kmem_cache_destroy(nfsd_file_slab);
+       nfsd_file_slab = NULL;
+       kmem_cache_destroy(nfsd_file_mark_slab);
+       nfsd_file_mark_slab = NULL;
+       kfree(nfsd_file_hashtbl);
+       nfsd_file_hashtbl = NULL;
+       goto out;
+}
+
+/*
+ * Note this can deadlock with nfsd_file_lru_cb.
+ */
+void
+nfsd_file_cache_purge(void)
+{
+       unsigned int            i;
+       struct nfsd_file        *nf;
+       LIST_HEAD(dispose);
+       bool del;
+
+       if (!nfsd_file_hashtbl)
+               return;
+
+       for (i = 0; i < NFSD_FILE_HASH_SIZE; i++) {
+               spin_lock(&nfsd_file_hashtbl[i].nfb_lock);
+               while(!hlist_empty(&nfsd_file_hashtbl[i].nfb_head)) {
+                       nf = hlist_entry(nfsd_file_hashtbl[i].nfb_head.first,
+                                        struct nfsd_file, nf_node);
+                       del = nfsd_file_unhash_and_release_locked(nf, &dispose);
+
+                       /*
+                        * Deadlock detected! Something marked this entry as
+                        * unhased, but hasn't removed it from the hash list.
+                        */
+                       WARN_ON_ONCE(!del);
+               }
+               spin_unlock(&nfsd_file_hashtbl[i].nfb_lock);
+               nfsd_file_dispose_list(&dispose);
+       }
+}
+
+void
+nfsd_file_cache_shutdown(void)
+{
+       LIST_HEAD(dispose);
+
+       set_bit(NFSD_FILE_SHUTDOWN, &nfsd_file_lru_flags);
+
+       lease_unregister_notifier(&nfsd_file_lease_notifier);
+       unregister_shrinker(&nfsd_file_shrinker);
+       /*
+        * make sure all callers of nfsd_file_lru_cb are done before
+        * calling nfsd_file_cache_purge
+        */
+       cancel_delayed_work_sync(&nfsd_filecache_laundrette);
+       nfsd_file_cache_purge();
+       list_lru_destroy(&nfsd_file_lru);
+       rcu_barrier();
+       fsnotify_put_group(nfsd_file_fsnotify_group);
+       nfsd_file_fsnotify_group = NULL;
+       kmem_cache_destroy(nfsd_file_slab);
+       nfsd_file_slab = NULL;
+       fsnotify_wait_marks_destroyed();
+       kmem_cache_destroy(nfsd_file_mark_slab);
+       nfsd_file_mark_slab = NULL;
+       kfree(nfsd_file_hashtbl);
+       nfsd_file_hashtbl = NULL;
+}
+
+static bool
+nfsd_match_cred(const struct cred *c1, const struct cred *c2)
+{
+       int i;
+
+       if (!uid_eq(c1->fsuid, c2->fsuid))
+               return false;
+       if (!gid_eq(c1->fsgid, c2->fsgid))
+               return false;
+       if (c1->group_info == NULL || c2->group_info == NULL)
+               return c1->group_info == c2->group_info;
+       if (c1->group_info->ngroups != c2->group_info->ngroups)
+               return false;
+       for (i = 0; i < c1->group_info->ngroups; i++) {
+               if (!gid_eq(c1->group_info->gid[i], c2->group_info->gid[i]))
+                       return false;
+       }
+       return true;
+}
+
+static struct nfsd_file *
+nfsd_file_find_locked(struct inode *inode, unsigned int may_flags,
+                       unsigned int hashval)
+{
+       struct nfsd_file *nf;
+       unsigned char need = may_flags & NFSD_FILE_MAY_MASK;
+
+       hlist_for_each_entry_rcu(nf, &nfsd_file_hashtbl[hashval].nfb_head,
+                                nf_node) {
+               if ((need & nf->nf_may) != need)
+                       continue;
+               if (nf->nf_inode != inode)
+                       continue;
+               if (!nfsd_match_cred(nf->nf_cred, current_cred()))
+                       continue;
+               if (nfsd_file_get(nf) != NULL)
+                       return nf;
+       }
+       return NULL;
+}
+
+/**
+ * nfsd_file_is_cached - are there any cached open files for this fh?
+ * @inode: inode of the file to check
+ *
+ * Scan the hashtable for open files that match this fh. Returns true if there
+ * are any, and false if not.
+ */
+bool
+nfsd_file_is_cached(struct inode *inode)
+{
+       bool                    ret = false;
+       struct nfsd_file        *nf;
+       unsigned int            hashval;
+
+        hashval = (unsigned int)hash_long(inode->i_ino, NFSD_FILE_HASH_BITS);
+
+       rcu_read_lock();
+       hlist_for_each_entry_rcu(nf, &nfsd_file_hashtbl[hashval].nfb_head,
+                                nf_node) {
+               if (inode == nf->nf_inode) {
+                       ret = true;
+                       break;
+               }
+       }
+       rcu_read_unlock();
+       trace_nfsd_file_is_cached(inode, hashval, (int)ret);
+       return ret;
+}
+
+__be32
+nfsd_file_acquire(struct svc_rqst *rqstp, struct svc_fh *fhp,
+                 unsigned int may_flags, struct nfsd_file **pnf)
+{
+       __be32  status;
+       struct nfsd_file *nf, *new;
+       struct inode *inode;
+       unsigned int hashval;
+
+       /* FIXME: skip this if fh_dentry is already set? */
+       status = fh_verify(rqstp, fhp, S_IFREG,
+                               may_flags|NFSD_MAY_OWNER_OVERRIDE);
+       if (status != nfs_ok)
+               return status;
+
+       inode = d_inode(fhp->fh_dentry);
+       hashval = (unsigned int)hash_long(inode->i_ino, NFSD_FILE_HASH_BITS);
+retry:
+       rcu_read_lock();
+       nf = nfsd_file_find_locked(inode, may_flags, hashval);
+       rcu_read_unlock();
+       if (nf)
+               goto wait_for_construction;
+
+       new = nfsd_file_alloc(inode, may_flags, hashval);
+       if (!new) {
+               trace_nfsd_file_acquire(rqstp, hashval, inode, may_flags,
+                                       NULL, nfserr_jukebox);
+               return nfserr_jukebox;
+       }
+
+       spin_lock(&nfsd_file_hashtbl[hashval].nfb_lock);
+       nf = nfsd_file_find_locked(inode, may_flags, hashval);
+       if (nf == NULL)
+               goto open_file;
+       spin_unlock(&nfsd_file_hashtbl[hashval].nfb_lock);
+       nfsd_file_slab_free(&new->nf_rcu);
+
+wait_for_construction:
+       wait_on_bit(&nf->nf_flags, NFSD_FILE_PENDING, TASK_UNINTERRUPTIBLE);
+
+       /* Did construction of this file fail? */
+       if (!test_bit(NFSD_FILE_HASHED, &nf->nf_flags)) {
+               nfsd_file_put_noref(nf);
+               goto retry;
+       }
+
+       this_cpu_inc(nfsd_file_cache_hits);
+
+       if (!(may_flags & NFSD_MAY_NOT_BREAK_LEASE)) {
+               bool write = (may_flags & NFSD_MAY_WRITE);
+
+               if (test_bit(NFSD_FILE_BREAK_READ, &nf->nf_flags) ||
+                   (test_bit(NFSD_FILE_BREAK_WRITE, &nf->nf_flags) && write)) {
+                       status = nfserrno(nfsd_open_break_lease(
+                                       file_inode(nf->nf_file), may_flags));
+                       if (status == nfs_ok) {
+                               clear_bit(NFSD_FILE_BREAK_READ, &nf->nf_flags);
+                               if (write)
+                                       clear_bit(NFSD_FILE_BREAK_WRITE,
+                                                 &nf->nf_flags);
+                       }
+               }
+       }
+out:
+       if (status == nfs_ok) {
+               *pnf = nf;
+       } else {
+               nfsd_file_put(nf);
+               nf = NULL;
+       }
+
+       trace_nfsd_file_acquire(rqstp, hashval, inode, may_flags, nf, status);
+       return status;
+open_file:
+       nf = new;
+       /* Take reference for the hashtable */
+       atomic_inc(&nf->nf_ref);
+       __set_bit(NFSD_FILE_HASHED, &nf->nf_flags);
+       __set_bit(NFSD_FILE_PENDING, &nf->nf_flags);
+       list_lru_add(&nfsd_file_lru, &nf->nf_lru);
+       hlist_add_head_rcu(&nf->nf_node, &nfsd_file_hashtbl[hashval].nfb_head);
+       ++nfsd_file_hashtbl[hashval].nfb_count;
+       nfsd_file_hashtbl[hashval].nfb_maxcount = max(nfsd_file_hashtbl[hashval].nfb_maxcount,
+                       nfsd_file_hashtbl[hashval].nfb_count);
+       spin_unlock(&nfsd_file_hashtbl[hashval].nfb_lock);
+       atomic_long_inc(&nfsd_filecache_count);
+
+       nf->nf_mark = nfsd_file_mark_find_or_create(nf);
+       if (nf->nf_mark)
+               status = nfsd_open_verified(rqstp, fhp, S_IFREG,
+                               may_flags, &nf->nf_file);
+       else
+               status = nfserr_jukebox;
+       /*
+        * If construction failed, or we raced with a call to unlink()
+        * then unhash.
+        */
+       if (status != nfs_ok || inode->i_nlink == 0) {
+               bool do_free;
+               spin_lock(&nfsd_file_hashtbl[hashval].nfb_lock);
+               do_free = nfsd_file_unhash(nf);
+               spin_unlock(&nfsd_file_hashtbl[hashval].nfb_lock);
+               if (do_free)
+                       nfsd_file_put_noref(nf);
+       }
+       clear_bit_unlock(NFSD_FILE_PENDING, &nf->nf_flags);
+       smp_mb__after_atomic();
+       wake_up_bit(&nf->nf_flags, NFSD_FILE_PENDING);
+       goto out;
+}
+
+/*
+ * Note that fields may be added, removed or reordered in the future. Programs
+ * scraping this file for info should test the labels to ensure they're
+ * getting the correct field.
+ */
+static int nfsd_file_cache_stats_show(struct seq_file *m, void *v)
+{
+       unsigned int i, count = 0, longest = 0;
+       unsigned long hits = 0;
+
+       /*
+        * No need for spinlocks here since we're not terribly interested in
+        * accuracy. We do take the nfsd_mutex simply to ensure that we
+        * don't end up racing with server shutdown
+        */
+       mutex_lock(&nfsd_mutex);
+       if (nfsd_file_hashtbl) {
+               for (i = 0; i < NFSD_FILE_HASH_SIZE; i++) {
+                       count += nfsd_file_hashtbl[i].nfb_count;
+                       longest = max(longest, nfsd_file_hashtbl[i].nfb_count);
+               }
+       }
+       mutex_unlock(&nfsd_mutex);
+
+       for_each_possible_cpu(i)
+               hits += per_cpu(nfsd_file_cache_hits, i);
+
+       seq_printf(m, "total entries: %u\n", count);
+       seq_printf(m, "longest chain: %u\n", longest);
+       seq_printf(m, "cache hits:    %lu\n", hits);
+       return 0;
+}
+
+int nfsd_file_cache_stats_open(struct inode *inode, struct file *file)
+{
+       return single_open(file, nfsd_file_cache_stats_show, NULL);
+}
diff --git a/fs/nfsd/filecache.h b/fs/nfsd/filecache.h
new file mode 100644 (file)
index 0000000..0c0c671
--- /dev/null
@@ -0,0 +1,60 @@
+#ifndef _FS_NFSD_FILECACHE_H
+#define _FS_NFSD_FILECACHE_H
+
+#include <linux/fsnotify_backend.h>
+
+/*
+ * This is the fsnotify_mark container that nfsd attaches to the files that it
+ * is holding open. Note that we have a separate refcount here aside from the
+ * one in the fsnotify_mark. We only want a single fsnotify_mark attached to
+ * the inode, and for each nfsd_file to hold a reference to it.
+ *
+ * The fsnotify_mark is itself refcounted, but that's not sufficient to tell us
+ * how to put that reference. If there are still outstanding nfsd_files that
+ * reference the mark, then we would want to call fsnotify_put_mark on it.
+ * If there were not, then we'd need to call fsnotify_destroy_mark. Since we
+ * can't really tell the difference, we use the nfm_mark to keep track of how
+ * many nfsd_files hold references to the mark. When that counter goes to zero
+ * then we know to call fsnotify_destroy_mark on it.
+ */
+struct nfsd_file_mark {
+       struct fsnotify_mark    nfm_mark;
+       atomic_t                nfm_ref;
+};
+
+/*
+ * 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.
+ */
+struct nfsd_file {
+       struct hlist_node       nf_node;
+       struct list_head        nf_lru;
+       struct rcu_head         nf_rcu;
+       struct file             *nf_file;
+       const struct cred       *nf_cred;
+#define NFSD_FILE_HASHED       (0)
+#define NFSD_FILE_PENDING      (1)
+#define NFSD_FILE_BREAK_READ   (2)
+#define NFSD_FILE_BREAK_WRITE  (3)
+#define NFSD_FILE_REFERENCED   (4)
+       unsigned long           nf_flags;
+       struct inode            *nf_inode;
+       unsigned int            nf_hashval;
+       atomic_t                nf_ref;
+       unsigned char           nf_may;
+       struct nfsd_file_mark   *nf_mark;
+};
+
+int nfsd_file_cache_init(void);
+void nfsd_file_cache_purge(void);
+void nfsd_file_cache_shutdown(void);
+void nfsd_file_put(struct nfsd_file *nf);
+struct nfsd_file *nfsd_file_get(struct nfsd_file *nf);
+void nfsd_file_close_inode_sync(struct inode *inode);
+bool nfsd_file_is_cached(struct inode *inode);
+__be32 nfsd_file_acquire(struct svc_rqst *rqstp, struct svc_fh *fhp,
+                 unsigned int may_flags, struct nfsd_file **nfp);
+int    nfsd_file_cache_stats_open(struct inode *, struct file *);
+#endif /* _FS_NFSD_FILECACHE_H */
index 18d94ea984ba4add43d1af0ca90ed155fc67c700..a6b1eab7b722cc37c2d0b6439e2d2ce2c1ed251d 100644 (file)
@@ -27,6 +27,7 @@
 #include "cache.h"
 #include "vfs.h"
 #include "netns.h"
+#include "filecache.h"
 
 #define NFSDDBG_FACILITY       NFSDDBG_SVC
 
@@ -313,6 +314,9 @@ static int nfsd_startup_generic(int nrservs)
        if (nfsd_users++)
                return 0;
 
+       ret = nfsd_file_cache_init();
+       if (ret)
+               goto dec_users;
        /*
         * Readahead param cache - will no-op if it already exists.
         * (Note therefore results will be suboptimal if number of
@@ -320,7 +324,7 @@ static int nfsd_startup_generic(int nrservs)
         */
        ret = nfsd_racache_init(2*nrservs);
        if (ret)
-               goto dec_users;
+               goto out_file_cache;
 
        ret = nfs4_state_start();
        if (ret)
@@ -329,6 +333,8 @@ static int nfsd_startup_generic(int nrservs)
 
 out_racache:
        nfsd_racache_shutdown();
+out_file_cache:
+       nfsd_file_cache_shutdown();
 dec_users:
        nfsd_users--;
        return ret;
@@ -340,6 +346,7 @@ static void nfsd_shutdown_generic(void)
                return;
 
        nfs4_state_shutdown();
+       nfsd_file_cache_shutdown();
        nfsd_racache_shutdown();
 }
 
index 80933e4334d84b9d80884527d6b029ee8f6aa867..ffc78a0e28b24e59c2976400e4309f861dddfc8f 100644 (file)
@@ -126,6 +126,8 @@ DEFINE_NFSD_ERR_EVENT(read_err);
 DEFINE_NFSD_ERR_EVENT(write_err);
 
 #include "state.h"
+#include "filecache.h"
+#include "vfs.h"
 
 DECLARE_EVENT_CLASS(nfsd_stateid_class,
        TP_PROTO(stateid_t *stp),
@@ -164,6 +166,144 @@ DEFINE_STATEID_EVENT(layout_recall_done);
 DEFINE_STATEID_EVENT(layout_recall_fail);
 DEFINE_STATEID_EVENT(layout_recall_release);
 
+#define show_nf_flags(val)                                             \
+       __print_flags(val, "|",                                         \
+               { 1 << NFSD_FILE_HASHED,        "HASHED" },             \
+               { 1 << NFSD_FILE_PENDING,       "PENDING" },            \
+               { 1 << NFSD_FILE_BREAK_READ,    "BREAK_READ" },         \
+               { 1 << NFSD_FILE_BREAK_WRITE,   "BREAK_WRITE" },        \
+               { 1 << NFSD_FILE_REFERENCED,    "REFERENCED"})
+
+/* FIXME: This should probably be fleshed out in the future. */
+#define show_nf_may(val)                                               \
+       __print_flags(val, "|",                                         \
+               { NFSD_MAY_READ,                "READ" },               \
+               { NFSD_MAY_WRITE,               "WRITE" },              \
+               { NFSD_MAY_NOT_BREAK_LEASE,     "NOT_BREAK_LEASE" })
+
+DECLARE_EVENT_CLASS(nfsd_file_class,
+       TP_PROTO(struct nfsd_file *nf),
+       TP_ARGS(nf),
+       TP_STRUCT__entry(
+               __field(unsigned int, nf_hashval)
+               __field(void *, nf_inode)
+               __field(int, nf_ref)
+               __field(unsigned long, nf_flags)
+               __field(unsigned char, nf_may)
+               __field(struct file *, nf_file)
+       ),
+       TP_fast_assign(
+               __entry->nf_hashval = nf->nf_hashval;
+               __entry->nf_inode = nf->nf_inode;
+               __entry->nf_ref = atomic_read(&nf->nf_ref);
+               __entry->nf_flags = nf->nf_flags;
+               __entry->nf_may = nf->nf_may;
+               __entry->nf_file = nf->nf_file;
+       ),
+       TP_printk("hash=0x%x inode=0x%p ref=%d flags=%s may=%s file=%p",
+               __entry->nf_hashval,
+               __entry->nf_inode,
+               __entry->nf_ref,
+               show_nf_flags(__entry->nf_flags),
+               show_nf_may(__entry->nf_may),
+               __entry->nf_file)
+)
+
+#define DEFINE_NFSD_FILE_EVENT(name) \
+DEFINE_EVENT(nfsd_file_class, name, \
+       TP_PROTO(struct nfsd_file *nf), \
+       TP_ARGS(nf))
+
+DEFINE_NFSD_FILE_EVENT(nfsd_file_alloc);
+DEFINE_NFSD_FILE_EVENT(nfsd_file_put_final);
+DEFINE_NFSD_FILE_EVENT(nfsd_file_unhash);
+DEFINE_NFSD_FILE_EVENT(nfsd_file_put);
+DEFINE_NFSD_FILE_EVENT(nfsd_file_unhash_and_release_locked);
+
+TRACE_EVENT(nfsd_file_acquire,
+       TP_PROTO(struct svc_rqst *rqstp, unsigned int hash,
+                struct inode *inode, unsigned int may_flags,
+                struct nfsd_file *nf, __be32 status),
+
+       TP_ARGS(rqstp, hash, inode, may_flags, nf, status),
+
+       TP_STRUCT__entry(
+               __field(__be32, xid)
+               __field(unsigned int, hash)
+               __field(void *, inode)
+               __field(unsigned int, may_flags)
+               __field(int, nf_ref)
+               __field(unsigned long, nf_flags)
+               __field(unsigned char, nf_may)
+               __field(struct file *, nf_file)
+               __field(__be32, status)
+       ),
+
+       TP_fast_assign(
+               __entry->xid = rqstp->rq_xid;
+               __entry->hash = hash;
+               __entry->inode = inode;
+               __entry->may_flags = may_flags;
+               __entry->nf_ref = nf ? atomic_read(&nf->nf_ref) : 0;
+               __entry->nf_flags = nf ? nf->nf_flags : 0;
+               __entry->nf_may = nf ? nf->nf_may : 0;
+               __entry->nf_file = nf ? nf->nf_file : NULL;
+               __entry->status = status;
+       ),
+
+       TP_printk("xid=0x%x hash=0x%x inode=0x%p may_flags=%s ref=%d nf_flags=%s nf_may=%s nf_file=0x%p status=%u",
+                       be32_to_cpu(__entry->xid), __entry->hash, __entry->inode,
+                       show_nf_may(__entry->may_flags), __entry->nf_ref,
+                       show_nf_flags(__entry->nf_flags),
+                       show_nf_may(__entry->nf_may), __entry->nf_file,
+                       be32_to_cpu(__entry->status))
+);
+
+DECLARE_EVENT_CLASS(nfsd_file_search_class,
+       TP_PROTO(struct inode *inode, unsigned int hash, int found),
+       TP_ARGS(inode, hash, found),
+       TP_STRUCT__entry(
+               __field(struct inode *, inode)
+               __field(unsigned int, hash)
+               __field(int, found)
+       ),
+       TP_fast_assign(
+               __entry->inode = inode;
+               __entry->hash = hash;
+               __entry->found = found;
+       ),
+       TP_printk("hash=0x%x inode=0x%p found=%d", __entry->hash,
+                       __entry->inode, __entry->found)
+);
+
+#define DEFINE_NFSD_FILE_SEARCH_EVENT(name)                            \
+DEFINE_EVENT(nfsd_file_search_class, name,                             \
+       TP_PROTO(struct inode *inode, unsigned int hash, int found),    \
+       TP_ARGS(inode, hash, found))
+
+DEFINE_NFSD_FILE_SEARCH_EVENT(nfsd_file_close_inode_sync);
+DEFINE_NFSD_FILE_SEARCH_EVENT(nfsd_file_close_inode);
+DEFINE_NFSD_FILE_SEARCH_EVENT(nfsd_file_is_cached);
+
+TRACE_EVENT(nfsd_file_fsnotify_handle_event,
+       TP_PROTO(struct inode *inode, u32 mask),
+       TP_ARGS(inode, mask),
+       TP_STRUCT__entry(
+               __field(struct inode *, inode)
+               __field(unsigned int, nlink)
+               __field(umode_t, mode)
+               __field(u32, mask)
+       ),
+       TP_fast_assign(
+               __entry->inode = inode;
+               __entry->nlink = inode->i_nlink;
+               __entry->mode = inode->i_mode;
+               __entry->mask = mask;
+       ),
+       TP_printk("inode=0x%p nlink=%u mode=0%ho mask=0x%x", __entry->inode,
+                       __entry->nlink, __entry->mode, __entry->mask)
+);
+
 #endif /* _NFSD_TRACE_H */
 
 #undef TRACE_INCLUDE_PATH
index c85783e536d595de5816584b3dadac4a73fc4b83..5983206ab0366e33db6fcd79d7cac5378e2d235d 100644 (file)
@@ -699,7 +699,7 @@ nfsd_access(struct svc_rqst *rqstp, struct svc_fh *fhp, u32 *access, u32 *suppor
 }
 #endif /* CONFIG_NFSD_V3 */
 
-static int nfsd_open_break_lease(struct inode *inode, int access)
+int nfsd_open_break_lease(struct inode *inode, int access)
 {
        unsigned int mode;
 
@@ -715,8 +715,8 @@ static int nfsd_open_break_lease(struct inode *inode, int access)
  * and additional flags.
  * N.B. After this call fhp needs an fh_put
  */
-__be32
-nfsd_open(struct svc_rqst *rqstp, struct svc_fh *fhp, umode_t type,
+static __be32
+__nfsd_open(struct svc_rqst *rqstp, struct svc_fh *fhp, umode_t type,
                        int may_flags, struct file **filp)
 {
        struct path     path;
@@ -726,25 +726,6 @@ nfsd_open(struct svc_rqst *rqstp, struct svc_fh *fhp, umode_t type,
        __be32          err;
        int             host_err = 0;
 
-       validate_process_creds();
-
-       /*
-        * If we get here, then the client has already done an "open",
-        * and (hopefully) checked permission - so allow OWNER_OVERRIDE
-        * in case a chmod has now revoked permission.
-        *
-        * Arguably we should also allow the owner override for
-        * directories, but we never have and it doesn't seem to have
-        * caused anyone a problem.  If we were to change this, note
-        * also that our filldir callbacks would need a variant of
-        * lookup_one_len that doesn't check permissions.
-        */
-       if (type == S_IFREG)
-               may_flags |= NFSD_MAY_OWNER_OVERRIDE;
-       err = fh_verify(rqstp, fhp, type, may_flags);
-       if (err)
-               goto out;
-
        path.mnt = fhp->fh_export->ex_path.mnt;
        path.dentry = fhp->fh_dentry;
        inode = d_inode(path.dentry);
@@ -798,10 +779,50 @@ nfsd_open(struct svc_rqst *rqstp, struct svc_fh *fhp, umode_t type,
 out_nfserr:
        err = nfserrno(host_err);
 out:
+       return err;
+}
+
+__be32
+nfsd_open(struct svc_rqst *rqstp, struct svc_fh *fhp, umode_t type,
+               int may_flags, struct file **filp)
+{
+       __be32 err;
+
+       validate_process_creds();
+       /*
+        * If we get here, then the client has already done an "open",
+        * and (hopefully) checked permission - so allow OWNER_OVERRIDE
+        * in case a chmod has now revoked permission.
+        *
+        * Arguably we should also allow the owner override for
+        * directories, but we never have and it doesn't seem to have
+        * caused anyone a problem.  If we were to change this, note
+        * also that our filldir callbacks would need a variant of
+        * lookup_one_len that doesn't check permissions.
+        */
+       if (type == S_IFREG)
+               may_flags |= NFSD_MAY_OWNER_OVERRIDE;
+       err = fh_verify(rqstp, fhp, type, may_flags);
+       if (!err)
+               err = __nfsd_open(rqstp, fhp, type, may_flags, filp);
+       validate_process_creds();
+       return err;
+}
+
+__be32
+nfsd_open_verified(struct svc_rqst *rqstp, struct svc_fh *fhp, umode_t type,
+               int may_flags, struct file **filp)
+{
+       __be32 err;
+
+       validate_process_creds();
+       err = __nfsd_open(rqstp, fhp, type, may_flags, filp);
        validate_process_creds();
        return err;
 }
 
+
+
 struct raparms *
 nfsd_init_raparms(struct file *file)
 {
index db351247892d05155e9cc2d2517229b071e807a1..31fdae34e028f43410fa4e5da791005a8b4d7fcf 100644 (file)
@@ -75,8 +75,11 @@ __be32               do_nfsd_create(struct svc_rqst *, struct svc_fh *,
 __be32         nfsd_commit(struct svc_rqst *, struct svc_fh *,
                                loff_t, unsigned long);
 #endif /* CONFIG_NFSD_V3 */
+int            nfsd_open_break_lease(struct inode *, int);
 __be32         nfsd_open(struct svc_rqst *, struct svc_fh *, umode_t,
                                int, struct file **);
+__be32         nfsd_open_verified(struct svc_rqst *, struct svc_fh *, umode_t,
+                               int, struct file **);
 struct raparms;
 __be32         nfsd_splice_read(struct svc_rqst *rqstp, struct svc_fh *fhp,
                                struct file *file, loff_t offset,