libglusterfs,mount/fuse: implement gidcache mechanism in fuse-bridge

This change genericizes the cache mechanism implemented in commit
8efd2845 into libglusterfs/src/gidcache.[ch] and adds fuse-bridge as
a client. The cache mechanism is fundamentally equivalent, with some
minor changes:

  - Change cache key from uid_t to uint64_t.
  - Modify the cache add logic to locate and use an entry with a
    matching ID, should it already exist. This addresses a bug in
    the existing mechanism where an expired entry supercedes a newly
    added entry in lookup, causing repeated adds and flushing of a
    cache bucket.

The fuse group cache is disabled by default. It can be enabled via
the 'gid-timeout' fuse-bridge translator option and accompanying
mount option (i.e., '-o gid-timeout=1' for a 1s entry timeout).

BUG: 800892
Change-Id: I0b34a2263ca48dbb154790a4a44fc70b733e9114
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-on: http://review.gluster.com/3676
Tested-by: Gluster Build System <jenkins@build.gluster.com>
Reviewed-by: Jeff Darcy <jdarcy@redhat.com>
Reviewed-by: Anand Avati <avati@redhat.com>

3 files changed, 32 insertions, 117 deletions

diff --git a/xlators/nfs/server/src/nfs-fops.c b/xlators/nfs/server/src/nfs-fops.c
index 87c511d546f..e2eedf43398 100644
--- a/xlators/nfs/server/src/nfs-fops.c
+++ b/xlators/nfs/server/src/nfs-fops.c
@@ -39,13 +39,6 @@
 #include <libgen.h>
 #include <semaphore.h>
 
-/*
- * We treat this as a very simple set-associative LRU cache, with entries aged
- * out after a configurable interval.  Hardly rocket science, but lots of
- * details to worry about.
- */
-#define BUCKET_START(p,n)       ((p) + ((n) * AUX_GID_CACHE_ASSOC))
-
 void
 nfs_fix_groups (xlator_t *this, call_stack_t *root)
 {
@@ -56,47 +49,23 @@ nfs_fix_groups (xlator_t *this, call_stack_t *root)
         int              ngroups;
         int              i;
         struct nfs_state *priv = this->private;
-        aux_gid_list_t   *agl = NULL;
-        int              bucket = 0;
-        time_t           now = 0;
+        const gid_list_t *agl;
+	gid_list_t gl;
 
         if (!priv->server_aux_gids) {
                 return;
         }
 
-        LOCK(&priv->aux_gid_lock);
-        now = time(NULL);
-        bucket = root->uid % priv->aux_gid_nbuckets;
-        agl = BUCKET_START(priv->aux_gid_cache,bucket);
-        for (i = 0; i < AUX_GID_CACHE_ASSOC; ++i, ++agl) {
-                if (!agl->gid_list) {
-                        continue;
-                }
-                if (agl->uid != root->uid) {
-                        continue;
-                }
-                /*
-                 * We don't put new entries in the cache when expiration=0, but
-                 * there might be entries still in there if expiration was
-                 * changed very recently.  Writing the check this way ensures
-                 * that they're not used.
-                 */
-                if (now < agl->deadline) {
-                        for (ngroups = 0; ngroups < agl->gid_count; ++ngroups) {
-                                root->groups[ngroups] = agl->gid_list[ngroups];
-                        }
-                        UNLOCK(&priv->aux_gid_lock);
-                        root->ngrps = ngroups;
-                        return;
-                }
-                /*
-                 * We're not going to find any more UID matches, and reaping
-                 * is handled further down to maintain LRU order.
-                 */
-                break;
-        }
-        UNLOCK(&priv->aux_gid_lock);
+	agl = gid_cache_lookup(&priv->gid_cache, root->uid);
+	if (agl) {
+		for (ngroups = 0; ngroups < agl->gl_count; ngroups++) 
+			root->groups[ngroups] = agl->gl_list[ngroups];
+		root->ngrps = ngroups;
+		gid_cache_release(&priv->gid_cache, agl);
+		return;
+	}
 
+	/* No cached list found. */
         if (getpwuid_r(root->uid,&mypw,mystrs,sizeof(mystrs),&result) != 0) {
                 gf_log (this->name, GF_LOG_ERROR,
                         "getpwuid_r(%u) failed", root->uid);
@@ -119,51 +88,18 @@ nfs_fix_groups (xlator_t *this, call_stack_t *root)
                 return;
         }
 
-        if (priv->aux_gid_max_age) {
-                LOCK(&priv->aux_gid_lock);
-                /* Bucket should still be valid from before. */
-                agl = BUCKET_START(priv->aux_gid_cache,bucket);
-                for (i = 0; i < AUX_GID_CACHE_ASSOC; ++i, ++agl) {
-                        if (!agl->gid_list) {
-                                break;
-                        }
-                }
-                /*
-                 * The way we allocate free entries naturally places the newest
-                 * ones at the highest indices, so evicting the lowest makes
-                 * sense, but that also means we can't just replace it with the
-                 * one that caused the eviction.  That would cause us to thrash
-                 * the first entry while others remain idle.  Therefore, we
-                 * need to slide the other entries down and add the new one at
-                 * the end just as if the *last* slot had been free.
-                 *
-                 * Deadline expiration is also handled here, since the oldest
-                 * expired entry will be in the first position.  This does mean
-                 * the bucket can stay full of expired entries if we're idle
-                 * but, if the small amount of extra memory or scan time before
-                 * we decide to evict someone ever become issues, we could
-                 * easily add a reaper thread.
-                 */
-                if (i >= AUX_GID_CACHE_ASSOC) {
-                        agl = BUCKET_START(priv->aux_gid_cache,bucket);
-                        GF_FREE(agl->gid_list);
-                        for (i = 1; i < AUX_GID_CACHE_ASSOC; ++i) {
-                                agl[0] = agl[1];
-                                ++agl;
-                        }
-                }
-                agl->gid_list = GF_CALLOC(ngroups,sizeof(gid_t),
-                                          gf_nfs_mt_aux_gids);
-                if (agl->gid_list) {
-                        /* It's not fatal if the alloc failed. */
-                        agl->uid = root->uid;
-                        agl->gid_count = ngroups;
-                        memcpy(agl->gid_list,mygroups,sizeof(gid_t)*ngroups);
-                        agl->deadline = now + priv->aux_gid_max_age;
-                }
-                UNLOCK(&priv->aux_gid_lock);
-        }
-
+	/* Add the group data to the cache. */
+	gl.gl_list = GF_CALLOC(ngroups, sizeof(gid_t), gf_nfs_mt_aux_gids);
+	if (gl.gl_list) {
+		/* It's not fatal if the alloc failed. */
+		gl.gl_id = root->uid;
+		gl.gl_count = ngroups;
+		memcpy(gl.gl_list, mygroups, sizeof(gid_t) * ngroups);
+		if (gid_cache_add(&priv->gid_cache, &gl) != 1)
+			GF_FREE(gl.gl_list);
+	}
+
+	/* Copy data to the frame. */
         for (i = 0; i < ngroups; ++i) {
                 gf_log (this->name, GF_LOG_TRACE,
                         "%s is in group %u", result->pw_name, mygroups[i]);
diff --git a/xlators/nfs/server/src/nfs.c b/xlators/nfs/server/src/nfs.c
index ba63bcd7a8c..0a5a9d1e34a 100644
--- a/xlators/nfs/server/src/nfs.c
+++ b/xlators/nfs/server/src/nfs.c
@@ -736,9 +736,14 @@ nfs_init_state (xlator_t *this)
 
         GF_OPTION_INIT (OPT_SERVER_AUX_GIDS, nfs->server_aux_gids,
                         bool, free_foppool);
-        GF_OPTION_INIT (OPT_SERVER_GID_CACHE_TIMEOUT,nfs->aux_gid_max_age,
+        GF_OPTION_INIT (OPT_SERVER_GID_CACHE_TIMEOUT, nfs->server_aux_gids_max_age,
                         uint32, free_foppool);
 
+	if (gid_cache_init(&nfs->gid_cache, nfs->server_aux_gids_max_age) < 0) {
+		gf_log(GF_NFS, GF_LOG_ERROR, "Failed to initialize group cache.");
+		goto free_foppool;
+	}
+
         if (stat("/sbin/rpc.statd", &stbuf) == -1) {
                 gf_log (GF_NFS, GF_LOG_WARNING, "/sbin/rpc.statd not found. "
                         "Disabling NLM");
@@ -827,9 +832,6 @@ init (xlator_t *this) {
                 goto err;
         }
 
-        LOCK_INIT(&nfs->aux_gid_lock);
-        nfs->aux_gid_nbuckets = AUX_GID_CACHE_BUCKETS;
-
         gf_log (GF_NFS, GF_LOG_INFO, "NFS service started");
 err:
 
diff --git a/xlators/nfs/server/src/nfs.h b/xlators/nfs/server/src/nfs.h
index d2a0c134318..c3deba00a02 100644
--- a/xlators/nfs/server/src/nfs.h
+++ b/xlators/nfs/server/src/nfs.h
@@ -29,6 +29,7 @@
 #include "dict.h"
 #include "xlator.h"
 #include "lkowner.h"
+#include "gidcache.h"
 
 #define GF_NFS                  "nfs"
 
@@ -65,28 +66,6 @@ struct nfs_initer_list {
         rpcsvc_program_t        *program;
 };
 
-/*
- * TBD: make the cache size tunable
- *
- * The current size represents a pretty trivial amount of memory, and should
- * provide good hit rates even for quite busy systems.  If we ever want to
- * support really large cache sizes, we'll need to do dynamic allocation
- * instead of just defining an array within nfs_state.  It doesn't make a
- * whole lot of sense to change the associativity, because it won't improve
- * hit rates all that much and will increase the maintenance cost as we have
- * to scan more entries with every lookup/update.
- */
-#define AUX_GID_CACHE_ASSOC     4
-#define AUX_GID_CACHE_BUCKETS   256
-#define AUX_GID_CACHE_SIZE      (AUX_GID_CACHE_ASSOC * AUX_GID_CACHE_BUCKETS)
-
-typedef struct {
-        uid_t                   uid;
-        int                     gid_count;
-        gid_t                   *gid_list;
-        time_t                  deadline;
-} aux_gid_list_t;
-
 struct nfs_state {
         rpcsvc_t                *rpcsvc;
         struct list_head        versions;
@@ -110,10 +89,8 @@ struct nfs_state {
         int                     mount_udp;
         struct rpc_clnt         *rpc_clnt;
         gf_boolean_t            server_aux_gids;
-        gf_lock_t               aux_gid_lock;
-        uint32_t                aux_gid_max_age;
-        unsigned int            aux_gid_nbuckets;
-        aux_gid_list_t          aux_gid_cache[AUX_GID_CACHE_SIZE];
+	uint32_t		server_aux_gids_max_age;
+	gid_cache_t		gid_cache;
 };
 
 #define gf_nfs_dvm_on(nfsstt)   (((struct nfs_state *)nfsstt)->dynamicvolumes == GF_NFS_DVM_ON)