]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blobdiff - fs/nfs/pnfs.c
pnfs: serialize LAYOUTGET(openstateid)
[mirror_ubuntu-bionic-kernel.git] / fs / nfs / pnfs.c
index b483026e82aa80298ea260d57d932b797c77e468..59ed68bf79faeb93d722f6c275f5c581726f8d93 100644 (file)
  */
 
 #include <linux/nfs_fs.h>
+#include "internal.h"
 #include "pnfs.h"
 
 #define NFSDBG_FACILITY                NFSDBG_PNFS
 
-/* STUB that returns the equivalent of "no module found" */
+/* Locking:
+ *
+ * pnfs_spinlock:
+ *      protects pnfs_modules_tbl.
+ */
+static DEFINE_SPINLOCK(pnfs_spinlock);
+
+/*
+ * pnfs_modules_tbl holds all pnfs modules
+ */
+static LIST_HEAD(pnfs_modules_tbl);
+
+/* Return the registered pnfs layout driver module matching given id */
+static struct pnfs_layoutdriver_type *
+find_pnfs_driver_locked(u32 id)
+{
+       struct pnfs_layoutdriver_type *local;
+
+       list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
+               if (local->id == id)
+                       goto out;
+       local = NULL;
+out:
+       dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
+       return local;
+}
+
 static struct pnfs_layoutdriver_type *
 find_pnfs_driver(u32 id)
 {
-       return NULL;
+       struct pnfs_layoutdriver_type *local;
+
+       spin_lock(&pnfs_spinlock);
+       local = find_pnfs_driver_locked(id);
+       spin_unlock(&pnfs_spinlock);
+       return local;
 }
 
 void
 unset_pnfs_layoutdriver(struct nfs_server *nfss)
 {
+       if (nfss->pnfs_curr_ld) {
+               nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
+               module_put(nfss->pnfs_curr_ld->owner);
+       }
        nfss->pnfs_curr_ld = NULL;
 }
 
@@ -74,7 +110,18 @@ set_pnfs_layoutdriver(struct nfs_server *server, u32 id)
                        goto out_no_driver;
                }
        }
+       if (!try_module_get(ld_type->owner)) {
+               dprintk("%s: Could not grab reference on module\n", __func__);
+               goto out_no_driver;
+       }
        server->pnfs_curr_ld = ld_type;
+       if (ld_type->set_layoutdriver(server)) {
+               printk(KERN_ERR
+                      "%s: Error initializing mount point for layout driver %u.\n",
+                      __func__, id);
+               module_put(ld_type->owner);
+               goto out_no_driver;
+       }
        dprintk("%s: pNFS module for %u set\n", __func__, id);
        return;
 
@@ -82,3 +129,682 @@ out_no_driver:
        dprintk("%s: Using NFSv4 I/O\n", __func__);
        server->pnfs_curr_ld = NULL;
 }
+
+int
+pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
+{
+       int status = -EINVAL;
+       struct pnfs_layoutdriver_type *tmp;
+
+       if (ld_type->id == 0) {
+               printk(KERN_ERR "%s id 0 is reserved\n", __func__);
+               return status;
+       }
+       if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
+               printk(KERN_ERR "%s Layout driver must provide "
+                      "alloc_lseg and free_lseg.\n", __func__);
+               return status;
+       }
+
+       spin_lock(&pnfs_spinlock);
+       tmp = find_pnfs_driver_locked(ld_type->id);
+       if (!tmp) {
+               list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
+               status = 0;
+               dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
+                       ld_type->name);
+       } else {
+               printk(KERN_ERR "%s Module with id %d already loaded!\n",
+                       __func__, ld_type->id);
+       }
+       spin_unlock(&pnfs_spinlock);
+
+       return status;
+}
+EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
+
+void
+pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
+{
+       dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
+       spin_lock(&pnfs_spinlock);
+       list_del(&ld_type->pnfs_tblid);
+       spin_unlock(&pnfs_spinlock);
+}
+EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
+
+/*
+ * pNFS client layout cache
+ */
+
+static void
+get_layout_hdr_locked(struct pnfs_layout_hdr *lo)
+{
+       assert_spin_locked(&lo->plh_inode->i_lock);
+       lo->plh_refcount++;
+}
+
+static void
+put_layout_hdr_locked(struct pnfs_layout_hdr *lo)
+{
+       assert_spin_locked(&lo->plh_inode->i_lock);
+       BUG_ON(lo->plh_refcount == 0);
+
+       lo->plh_refcount--;
+       if (!lo->plh_refcount) {
+               dprintk("%s: freeing layout cache %p\n", __func__, lo);
+               BUG_ON(!list_empty(&lo->plh_layouts));
+               NFS_I(lo->plh_inode)->layout = NULL;
+               kfree(lo);
+       }
+}
+
+void
+put_layout_hdr(struct inode *inode)
+{
+       spin_lock(&inode->i_lock);
+       put_layout_hdr_locked(NFS_I(inode)->layout);
+       spin_unlock(&inode->i_lock);
+}
+
+static void
+init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg)
+{
+       INIT_LIST_HEAD(&lseg->pls_list);
+       atomic_set(&lseg->pls_refcount, 1);
+       smp_mb();
+       set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
+       lseg->pls_layout = lo;
+}
+
+static void free_lseg(struct pnfs_layout_segment *lseg)
+{
+       struct inode *ino = lseg->pls_layout->plh_inode;
+
+       NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
+       /* Matched by get_layout_hdr in pnfs_insert_layout */
+       put_layout_hdr(ino);
+}
+
+/* The use of tmp_list is necessary because pnfs_curr_ld->free_lseg
+ * could sleep, so must be called outside of the lock.
+ * Returns 1 if object was removed, otherwise return 0.
+ */
+static int
+put_lseg_locked(struct pnfs_layout_segment *lseg,
+               struct list_head *tmp_list)
+{
+       dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
+               atomic_read(&lseg->pls_refcount),
+               test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
+       if (atomic_dec_and_test(&lseg->pls_refcount)) {
+               struct inode *ino = lseg->pls_layout->plh_inode;
+
+               BUG_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
+               list_del(&lseg->pls_list);
+               if (list_empty(&lseg->pls_layout->plh_segs)) {
+                       struct nfs_client *clp;
+
+                       clp = NFS_SERVER(ino)->nfs_client;
+                       spin_lock(&clp->cl_lock);
+                       /* List does not take a reference, so no need for put here */
+                       list_del_init(&lseg->pls_layout->plh_layouts);
+                       spin_unlock(&clp->cl_lock);
+               }
+               list_add(&lseg->pls_list, tmp_list);
+               return 1;
+       }
+       return 0;
+}
+
+static bool
+should_free_lseg(u32 lseg_iomode, u32 recall_iomode)
+{
+       return (recall_iomode == IOMODE_ANY ||
+               lseg_iomode == recall_iomode);
+}
+
+/* Returns 1 if lseg is removed from list, 0 otherwise */
+static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
+                            struct list_head *tmp_list)
+{
+       int rv = 0;
+
+       if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
+               /* Remove the reference keeping the lseg in the
+                * list.  It will now be removed when all
+                * outstanding io is finished.
+                */
+               rv = put_lseg_locked(lseg, tmp_list);
+       }
+       return rv;
+}
+
+/* Returns count of number of matching invalid lsegs remaining in list
+ * after call.
+ */
+static int
+mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
+                           struct list_head *tmp_list,
+                           u32 iomode)
+{
+       struct pnfs_layout_segment *lseg, *next;
+       int invalid = 0, removed = 0;
+
+       dprintk("%s:Begin lo %p\n", __func__, lo);
+
+       list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
+               if (should_free_lseg(lseg->pls_range.iomode, iomode)) {
+                       dprintk("%s: freeing lseg %p iomode %d "
+                               "offset %llu length %llu\n", __func__,
+                               lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
+                               lseg->pls_range.length);
+                       invalid++;
+                       removed += mark_lseg_invalid(lseg, tmp_list);
+               }
+       dprintk("%s:Return %i\n", __func__, invalid - removed);
+       return invalid - removed;
+}
+
+static void
+pnfs_free_lseg_list(struct list_head *free_me)
+{
+       struct pnfs_layout_segment *lseg, *tmp;
+
+       list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
+               list_del(&lseg->pls_list);
+               free_lseg(lseg);
+       }
+}
+
+void
+pnfs_destroy_layout(struct nfs_inode *nfsi)
+{
+       struct pnfs_layout_hdr *lo;
+       LIST_HEAD(tmp_list);
+
+       spin_lock(&nfsi->vfs_inode.i_lock);
+       lo = nfsi->layout;
+       if (lo) {
+               set_bit(NFS_LAYOUT_DESTROYED, &nfsi->layout->plh_flags);
+               mark_matching_lsegs_invalid(lo, &tmp_list, IOMODE_ANY);
+               /* Matched by refcount set to 1 in alloc_init_layout_hdr */
+               put_layout_hdr_locked(lo);
+       }
+       spin_unlock(&nfsi->vfs_inode.i_lock);
+       pnfs_free_lseg_list(&tmp_list);
+}
+
+/*
+ * Called by the state manger to remove all layouts established under an
+ * expired lease.
+ */
+void
+pnfs_destroy_all_layouts(struct nfs_client *clp)
+{
+       struct pnfs_layout_hdr *lo;
+       LIST_HEAD(tmp_list);
+
+       spin_lock(&clp->cl_lock);
+       list_splice_init(&clp->cl_layouts, &tmp_list);
+       spin_unlock(&clp->cl_lock);
+
+       while (!list_empty(&tmp_list)) {
+               lo = list_entry(tmp_list.next, struct pnfs_layout_hdr,
+                               plh_layouts);
+               dprintk("%s freeing layout for inode %lu\n", __func__,
+                       lo->plh_inode->i_ino);
+               pnfs_destroy_layout(NFS_I(lo->plh_inode));
+       }
+}
+
+/* update lo->plh_stateid with new if is more recent */
+static void
+pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo,
+                       const nfs4_stateid *new)
+{
+       u32 oldseq, newseq;
+
+       oldseq = be32_to_cpu(lo->plh_stateid.stateid.seqid);
+       newseq = be32_to_cpu(new->stateid.seqid);
+       if ((int)(newseq - oldseq) > 0)
+               memcpy(&lo->plh_stateid, &new->stateid, sizeof(new->stateid));
+}
+
+/* lget is set to 1 if called from inside send_layoutget call chain */
+static bool
+pnfs_layoutgets_blocked(struct pnfs_layout_hdr *lo, int lget)
+{
+       return (list_empty(&lo->plh_segs) &&
+                (atomic_read(&lo->plh_outstanding) > lget));
+}
+
+int
+pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
+                             struct nfs4_state *open_state)
+{
+       int status = 0;
+
+       dprintk("--> %s\n", __func__);
+       spin_lock(&lo->plh_inode->i_lock);
+       if (pnfs_layoutgets_blocked(lo, 1)) {
+               status = -EAGAIN;
+       } else if (list_empty(&lo->plh_segs)) {
+               int seq;
+
+               do {
+                       seq = read_seqbegin(&open_state->seqlock);
+                       memcpy(dst->data, open_state->stateid.data,
+                              sizeof(open_state->stateid.data));
+               } while (read_seqretry(&open_state->seqlock, seq));
+       } else
+               memcpy(dst->data, lo->plh_stateid.data, sizeof(lo->plh_stateid.data));
+       spin_unlock(&lo->plh_inode->i_lock);
+       dprintk("<-- %s\n", __func__);
+       return status;
+}
+
+/*
+* Get layout from server.
+*    for now, assume that whole file layouts are requested.
+*    arg->offset: 0
+*    arg->length: all ones
+*/
+static struct pnfs_layout_segment *
+send_layoutget(struct pnfs_layout_hdr *lo,
+          struct nfs_open_context *ctx,
+          u32 iomode)
+{
+       struct inode *ino = lo->plh_inode;
+       struct nfs_server *server = NFS_SERVER(ino);
+       struct nfs4_layoutget *lgp;
+       struct pnfs_layout_segment *lseg = NULL;
+
+       dprintk("--> %s\n", __func__);
+
+       BUG_ON(ctx == NULL);
+       lgp = kzalloc(sizeof(*lgp), GFP_KERNEL);
+       if (lgp == NULL)
+               return NULL;
+       lgp->args.minlength = NFS4_MAX_UINT64;
+       lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
+       lgp->args.range.iomode = iomode;
+       lgp->args.range.offset = 0;
+       lgp->args.range.length = NFS4_MAX_UINT64;
+       lgp->args.type = server->pnfs_curr_ld->id;
+       lgp->args.inode = ino;
+       lgp->args.ctx = get_nfs_open_context(ctx);
+       lgp->lsegpp = &lseg;
+
+       /* Synchronously retrieve layout information from server and
+        * store in lseg.
+        */
+       nfs4_proc_layoutget(lgp);
+       if (!lseg) {
+               /* remember that LAYOUTGET failed and suspend trying */
+               set_bit(lo_fail_bit(iomode), &lo->plh_flags);
+       }
+       return lseg;
+}
+
+/*
+ * Compare two layout segments for sorting into layout cache.
+ * We want to preferentially return RW over RO layouts, so ensure those
+ * are seen first.
+ */
+static s64
+cmp_layout(u32 iomode1, u32 iomode2)
+{
+       /* read > read/write */
+       return (int)(iomode2 == IOMODE_READ) - (int)(iomode1 == IOMODE_READ);
+}
+
+static void
+pnfs_insert_layout(struct pnfs_layout_hdr *lo,
+                  struct pnfs_layout_segment *lseg)
+{
+       struct pnfs_layout_segment *lp;
+       int found = 0;
+
+       dprintk("%s:Begin\n", __func__);
+
+       assert_spin_locked(&lo->plh_inode->i_lock);
+       if (list_empty(&lo->plh_segs)) {
+               struct nfs_client *clp = NFS_SERVER(lo->plh_inode)->nfs_client;
+
+               spin_lock(&clp->cl_lock);
+               BUG_ON(!list_empty(&lo->plh_layouts));
+               list_add_tail(&lo->plh_layouts, &clp->cl_layouts);
+               spin_unlock(&clp->cl_lock);
+       }
+       list_for_each_entry(lp, &lo->plh_segs, pls_list) {
+               if (cmp_layout(lp->pls_range.iomode, lseg->pls_range.iomode) > 0)
+                       continue;
+               list_add_tail(&lseg->pls_list, &lp->pls_list);
+               dprintk("%s: inserted lseg %p "
+                       "iomode %d offset %llu length %llu before "
+                       "lp %p iomode %d offset %llu length %llu\n",
+                       __func__, lseg, lseg->pls_range.iomode,
+                       lseg->pls_range.offset, lseg->pls_range.length,
+                       lp, lp->pls_range.iomode, lp->pls_range.offset,
+                       lp->pls_range.length);
+               found = 1;
+               break;
+       }
+       if (!found) {
+               list_add_tail(&lseg->pls_list, &lo->plh_segs);
+               dprintk("%s: inserted lseg %p "
+                       "iomode %d offset %llu length %llu at tail\n",
+                       __func__, lseg, lseg->pls_range.iomode,
+                       lseg->pls_range.offset, lseg->pls_range.length);
+       }
+       get_layout_hdr_locked(lo);
+
+       dprintk("%s:Return\n", __func__);
+}
+
+static struct pnfs_layout_hdr *
+alloc_init_layout_hdr(struct inode *ino)
+{
+       struct pnfs_layout_hdr *lo;
+
+       lo = kzalloc(sizeof(struct pnfs_layout_hdr), GFP_KERNEL);
+       if (!lo)
+               return NULL;
+       lo->plh_refcount = 1;
+       INIT_LIST_HEAD(&lo->plh_layouts);
+       INIT_LIST_HEAD(&lo->plh_segs);
+       lo->plh_inode = ino;
+       return lo;
+}
+
+static struct pnfs_layout_hdr *
+pnfs_find_alloc_layout(struct inode *ino)
+{
+       struct nfs_inode *nfsi = NFS_I(ino);
+       struct pnfs_layout_hdr *new = NULL;
+
+       dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
+
+       assert_spin_locked(&ino->i_lock);
+       if (nfsi->layout) {
+               if (test_bit(NFS_LAYOUT_DESTROYED, &nfsi->layout->plh_flags))
+                       return NULL;
+               else
+                       return nfsi->layout;
+       }
+       spin_unlock(&ino->i_lock);
+       new = alloc_init_layout_hdr(ino);
+       spin_lock(&ino->i_lock);
+
+       if (likely(nfsi->layout == NULL))       /* Won the race? */
+               nfsi->layout = new;
+       else
+               kfree(new);
+       return nfsi->layout;
+}
+
+/*
+ * iomode matching rules:
+ * iomode      lseg    match
+ * -----       -----   -----
+ * ANY         READ    true
+ * ANY         RW      true
+ * RW          READ    false
+ * RW          RW      true
+ * READ                READ    true
+ * READ                RW      true
+ */
+static int
+is_matching_lseg(struct pnfs_layout_segment *lseg, u32 iomode)
+{
+       return (iomode != IOMODE_RW || lseg->pls_range.iomode == IOMODE_RW);
+}
+
+/*
+ * lookup range in layout
+ */
+static struct pnfs_layout_segment *
+pnfs_has_layout(struct pnfs_layout_hdr *lo, u32 iomode)
+{
+       struct pnfs_layout_segment *lseg, *ret = NULL;
+
+       dprintk("%s:Begin\n", __func__);
+
+       assert_spin_locked(&lo->plh_inode->i_lock);
+       list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
+               if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
+                   is_matching_lseg(lseg, iomode)) {
+                       ret = lseg;
+                       break;
+               }
+               if (cmp_layout(iomode, lseg->pls_range.iomode) > 0)
+                       break;
+       }
+
+       dprintk("%s:Return lseg %p ref %d\n",
+               __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
+       return ret;
+}
+
+/*
+ * Layout segment is retreived from the server if not cached.
+ * The appropriate layout segment is referenced and returned to the caller.
+ */
+struct pnfs_layout_segment *
+pnfs_update_layout(struct inode *ino,
+                  struct nfs_open_context *ctx,
+                  enum pnfs_iomode iomode)
+{
+       struct nfs_inode *nfsi = NFS_I(ino);
+       struct pnfs_layout_hdr *lo;
+       struct pnfs_layout_segment *lseg = NULL;
+
+       if (!pnfs_enabled_sb(NFS_SERVER(ino)))
+               return NULL;
+       spin_lock(&ino->i_lock);
+       lo = pnfs_find_alloc_layout(ino);
+       if (lo == NULL) {
+               dprintk("%s ERROR: can't get pnfs_layout_hdr\n", __func__);
+               goto out_unlock;
+       }
+
+       /* Check to see if the layout for the given range already exists */
+       lseg = pnfs_has_layout(lo, iomode);
+       if (lseg) {
+               dprintk("%s: Using cached lseg %p for iomode %d)\n",
+                       __func__, lseg, iomode);
+               goto out_unlock;
+       }
+
+       /* if LAYOUTGET already failed once we don't try again */
+       if (test_bit(lo_fail_bit(iomode), &nfsi->layout->plh_flags))
+               goto out_unlock;
+
+       if (pnfs_layoutgets_blocked(lo, 0))
+               goto out_unlock;
+       atomic_inc(&lo->plh_outstanding);
+
+       get_layout_hdr_locked(lo);
+       spin_unlock(&ino->i_lock);
+
+       lseg = send_layoutget(lo, ctx, iomode);
+       atomic_dec(&lo->plh_outstanding);
+       put_layout_hdr(ino);
+out:
+       dprintk("%s end, state 0x%lx lseg %p\n", __func__,
+               nfsi->layout->plh_flags, lseg);
+       return lseg;
+out_unlock:
+       spin_unlock(&ino->i_lock);
+       goto out;
+}
+
+int
+pnfs_layout_process(struct nfs4_layoutget *lgp)
+{
+       struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
+       struct nfs4_layoutget_res *res = &lgp->res;
+       struct pnfs_layout_segment *lseg;
+       struct inode *ino = lo->plh_inode;
+       int status = 0;
+
+       /* Inject layout blob into I/O device driver */
+       lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res);
+       if (!lseg || IS_ERR(lseg)) {
+               if (!lseg)
+                       status = -ENOMEM;
+               else
+                       status = PTR_ERR(lseg);
+               dprintk("%s: Could not allocate layout: error %d\n",
+                      __func__, status);
+               goto out;
+       }
+
+       spin_lock(&ino->i_lock);
+       init_lseg(lo, lseg);
+       lseg->pls_range = res->range;
+       *lgp->lsegpp = lseg;
+       pnfs_insert_layout(lo, lseg);
+
+       /* Done processing layoutget. Set the layout stateid */
+       pnfs_set_layout_stateid(lo, &res->stateid);
+       spin_unlock(&ino->i_lock);
+out:
+       return status;
+}
+
+/*
+ * Device ID cache. Currently supports one layout type per struct nfs_client.
+ * Add layout type to the lookup key to expand to support multiple types.
+ */
+int
+pnfs_alloc_init_deviceid_cache(struct nfs_client *clp,
+                        void (*free_callback)(struct pnfs_deviceid_node *))
+{
+       struct pnfs_deviceid_cache *c;
+
+       c = kzalloc(sizeof(struct pnfs_deviceid_cache), GFP_KERNEL);
+       if (!c)
+               return -ENOMEM;
+       spin_lock(&clp->cl_lock);
+       if (clp->cl_devid_cache != NULL) {
+               atomic_inc(&clp->cl_devid_cache->dc_ref);
+               dprintk("%s [kref [%d]]\n", __func__,
+                       atomic_read(&clp->cl_devid_cache->dc_ref));
+               kfree(c);
+       } else {
+               /* kzalloc initializes hlists */
+               spin_lock_init(&c->dc_lock);
+               atomic_set(&c->dc_ref, 1);
+               c->dc_free_callback = free_callback;
+               clp->cl_devid_cache = c;
+               dprintk("%s [new]\n", __func__);
+       }
+       spin_unlock(&clp->cl_lock);
+       return 0;
+}
+EXPORT_SYMBOL_GPL(pnfs_alloc_init_deviceid_cache);
+
+/*
+ * Called from pnfs_layoutdriver_type->free_lseg
+ * last layout segment reference frees deviceid
+ */
+void
+pnfs_put_deviceid(struct pnfs_deviceid_cache *c,
+                 struct pnfs_deviceid_node *devid)
+{
+       struct nfs4_deviceid *id = &devid->de_id;
+       struct pnfs_deviceid_node *d;
+       struct hlist_node *n;
+       long h = nfs4_deviceid_hash(id);
+
+       dprintk("%s [%d]\n", __func__, atomic_read(&devid->de_ref));
+       if (!atomic_dec_and_lock(&devid->de_ref, &c->dc_lock))
+               return;
+
+       hlist_for_each_entry_rcu(d, n, &c->dc_deviceids[h], de_node)
+               if (!memcmp(&d->de_id, id, sizeof(*id))) {
+                       hlist_del_rcu(&d->de_node);
+                       spin_unlock(&c->dc_lock);
+                       synchronize_rcu();
+                       c->dc_free_callback(devid);
+                       return;
+               }
+       spin_unlock(&c->dc_lock);
+       /* Why wasn't it found in  the list? */
+       BUG();
+}
+EXPORT_SYMBOL_GPL(pnfs_put_deviceid);
+
+/* Find and reference a deviceid */
+struct pnfs_deviceid_node *
+pnfs_find_get_deviceid(struct pnfs_deviceid_cache *c, struct nfs4_deviceid *id)
+{
+       struct pnfs_deviceid_node *d;
+       struct hlist_node *n;
+       long hash = nfs4_deviceid_hash(id);
+
+       dprintk("--> %s hash %ld\n", __func__, hash);
+       rcu_read_lock();
+       hlist_for_each_entry_rcu(d, n, &c->dc_deviceids[hash], de_node) {
+               if (!memcmp(&d->de_id, id, sizeof(*id))) {
+                       if (!atomic_inc_not_zero(&d->de_ref)) {
+                               goto fail;
+                       } else {
+                               rcu_read_unlock();
+                               return d;
+                       }
+               }
+       }
+fail:
+       rcu_read_unlock();
+       return NULL;
+}
+EXPORT_SYMBOL_GPL(pnfs_find_get_deviceid);
+
+/*
+ * Add a deviceid to the cache.
+ * GETDEVICEINFOs for same deviceid can race. If deviceid is found, discard new
+ */
+struct pnfs_deviceid_node *
+pnfs_add_deviceid(struct pnfs_deviceid_cache *c, struct pnfs_deviceid_node *new)
+{
+       struct pnfs_deviceid_node *d;
+       long hash = nfs4_deviceid_hash(&new->de_id);
+
+       dprintk("--> %s hash %ld\n", __func__, hash);
+       spin_lock(&c->dc_lock);
+       d = pnfs_find_get_deviceid(c, &new->de_id);
+       if (d) {
+               spin_unlock(&c->dc_lock);
+               dprintk("%s [discard]\n", __func__);
+               c->dc_free_callback(new);
+               return d;
+       }
+       INIT_HLIST_NODE(&new->de_node);
+       atomic_set(&new->de_ref, 1);
+       hlist_add_head_rcu(&new->de_node, &c->dc_deviceids[hash]);
+       spin_unlock(&c->dc_lock);
+       dprintk("%s [new]\n", __func__);
+       return new;
+}
+EXPORT_SYMBOL_GPL(pnfs_add_deviceid);
+
+void
+pnfs_put_deviceid_cache(struct nfs_client *clp)
+{
+       struct pnfs_deviceid_cache *local = clp->cl_devid_cache;
+
+       dprintk("--> %s cl_devid_cache %p\n", __func__, clp->cl_devid_cache);
+       if (atomic_dec_and_lock(&local->dc_ref, &clp->cl_lock)) {
+               int i;
+               /* Verify cache is empty */
+               for (i = 0; i < NFS4_DEVICE_ID_HASH_SIZE; i++)
+                       BUG_ON(!hlist_empty(&local->dc_deviceids[i]));
+               clp->cl_devid_cache = NULL;
+               spin_unlock(&clp->cl_lock);
+               kfree(local);
+       }
+}
+EXPORT_SYMBOL_GPL(pnfs_put_deviceid_cache);