[PATCH] ext2: fsid for statvfs
[linux-2.6.git] / fs / gfs2 / glock.c
index 87b37fe..55f5333 100644 (file)
 #include <linux/delay.h>
 #include <linux/sort.h>
 #include <linux/jhash.h>
-#include <linux/kref.h>
 #include <linux/kallsyms.h>
 #include <linux/gfs2_ondisk.h>
+#include <linux/list.h>
+#include <linux/lm_interface.h>
 #include <asm/uaccess.h>
 
 #include "gfs2.h"
-#include "lm_interface.h"
 #include "incore.h"
 #include "glock.h"
 #include "glops.h"
 #include "super.h"
 #include "util.h"
 
-/*  Must be kept in sync with the beginning of struct gfs2_glock  */
-struct glock_plug {
-       struct list_head gl_list;
-       unsigned long gl_flags;
-};
-
 struct greedy {
        struct gfs2_holder gr_gh;
-       struct work_struct gr_work;
+       struct delayed_work gr_work;
+};
+
+struct gfs2_gl_hash_bucket {
+        struct hlist_head hb_list;
 };
 
 typedef void (*glock_examiner) (struct gfs2_glock * gl);
 
 static int gfs2_dump_lockstate(struct gfs2_sbd *sdp);
 static int dump_glock(struct gfs2_glock *gl);
+static int dump_inode(struct gfs2_inode *ip);
+
+#define GFS2_GL_HASH_SHIFT      15
+#define GFS2_GL_HASH_SIZE       (1 << GFS2_GL_HASH_SHIFT)
+#define GFS2_GL_HASH_MASK       (GFS2_GL_HASH_SIZE - 1)
 
 static struct gfs2_gl_hash_bucket gl_hash_table[GFS2_GL_HASH_SIZE];
 
+/*
+ * Despite what you might think, the numbers below are not arbitrary :-)
+ * They are taken from the ipv4 routing hash code, which is well tested
+ * and thus should be nearly optimal. Later on we might tweek the numbers
+ * but for now this should be fine.
+ *
+ * The reason for putting the locks in a separate array from the list heads
+ * is that we can have fewer locks than list heads and save memory. We use
+ * the same hash function for both, but with a different hash mask.
+ */
+#if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_SPINLOCK) || \
+       defined(CONFIG_PROVE_LOCKING)
+
+#ifdef CONFIG_LOCKDEP
+# define GL_HASH_LOCK_SZ        256
+#else
+# if NR_CPUS >= 32
+#  define GL_HASH_LOCK_SZ       4096
+# elif NR_CPUS >= 16
+#  define GL_HASH_LOCK_SZ       2048
+# elif NR_CPUS >= 8
+#  define GL_HASH_LOCK_SZ       1024
+# elif NR_CPUS >= 4
+#  define GL_HASH_LOCK_SZ       512
+# else
+#  define GL_HASH_LOCK_SZ       256
+# endif
+#endif
+
+/* We never want more locks than chains */
+#if GFS2_GL_HASH_SIZE < GL_HASH_LOCK_SZ
+# undef GL_HASH_LOCK_SZ
+# define GL_HASH_LOCK_SZ GFS2_GL_HASH_SIZE
+#endif
+
+static rwlock_t gl_hash_locks[GL_HASH_LOCK_SZ];
+
+static inline rwlock_t *gl_lock_addr(unsigned int x)
+{
+       return &gl_hash_locks[x & (GL_HASH_LOCK_SZ-1)];
+}
+#else /* not SMP, so no spinlocks required */
+static inline rwlock_t *gl_lock_addr(x)
+{
+       return NULL;
+}
+#endif
+
 /**
  * relaxed_state_ok - is a requested lock compatible with the current lock mode?
  * @actual: the current state of the lock
@@ -127,23 +178,7 @@ static void glock_free(struct gfs2_glock *gl)
 
 void gfs2_glock_hold(struct gfs2_glock *gl)
 {
-       kref_get(&gl->gl_ref);
-}
-
-/* All work is done after the return from kref_put() so we
-   can release the write_lock before the free. */
-
-static void kill_glock(struct kref *kref)
-{
-       struct gfs2_glock *gl = container_of(kref, struct gfs2_glock, gl_ref);
-       struct gfs2_sbd *sdp = gl->gl_sbd;
-
-       gfs2_assert(sdp, gl->gl_state == LM_ST_UNLOCKED);
-       gfs2_assert(sdp, list_empty(&gl->gl_reclaim));
-       gfs2_assert(sdp, list_empty(&gl->gl_holders));
-       gfs2_assert(sdp, list_empty(&gl->gl_waiters1));
-       gfs2_assert(sdp, list_empty(&gl->gl_waiters2));
-       gfs2_assert(sdp, list_empty(&gl->gl_waiters3));
+       atomic_inc(&gl->gl_ref);
 }
 
 /**
@@ -154,19 +189,25 @@ static void kill_glock(struct kref *kref)
 
 int gfs2_glock_put(struct gfs2_glock *gl)
 {
-       struct gfs2_gl_hash_bucket *bucket = gl->gl_bucket;
        int rv = 0;
+       struct gfs2_sbd *sdp = gl->gl_sbd;
 
-       write_lock(&bucket->hb_lock);
-       if (kref_put(&gl->gl_ref, kill_glock)) {
-               list_del_init(&gl->gl_list);
-               write_unlock(&bucket->hb_lock);
+       write_lock(gl_lock_addr(gl->gl_hash));
+       if (atomic_dec_and_test(&gl->gl_ref)) {
+               hlist_del(&gl->gl_list);
+               write_unlock(gl_lock_addr(gl->gl_hash));
                BUG_ON(spin_is_locked(&gl->gl_spin));
+               gfs2_assert(sdp, gl->gl_state == LM_ST_UNLOCKED);
+               gfs2_assert(sdp, list_empty(&gl->gl_reclaim));
+               gfs2_assert(sdp, list_empty(&gl->gl_holders));
+               gfs2_assert(sdp, list_empty(&gl->gl_waiters1));
+               gfs2_assert(sdp, list_empty(&gl->gl_waiters2));
+               gfs2_assert(sdp, list_empty(&gl->gl_waiters3));
                glock_free(gl);
                rv = 1;
                goto out;
        }
-       write_unlock(&bucket->hb_lock);
+       write_unlock(gl_lock_addr(gl->gl_hash));
 out:
        return rv;
 }
@@ -203,21 +244,20 @@ static inline int queue_empty(struct gfs2_glock *gl, struct list_head *head)
  * Returns: NULL, or the struct gfs2_glock with the requested number
  */
 
-static struct gfs2_glock *search_bucket(struct gfs2_gl_hash_bucket *bucket,
+static struct gfs2_glock *search_bucket(unsigned int hash,
                                        const struct gfs2_sbd *sdp,
                                        const struct lm_lockname *name)
 {
        struct gfs2_glock *gl;
+       struct hlist_node *h;
 
-       list_for_each_entry(gl, &bucket->hb_list, gl_list) {
-               if (test_bit(GLF_PLUG, &gl->gl_flags))
-                       continue;
+       hlist_for_each_entry(gl, h, &gl_hash_table[hash].hb_list, gl_list) {
                if (!lm_name_equal(&gl->gl_name, name))
                        continue;
                if (gl->gl_sbd != sdp)
                        continue;
 
-               kref_get(&gl->gl_ref);
+               atomic_inc(&gl->gl_ref);
 
                return gl;
        }
@@ -236,12 +276,12 @@ static struct gfs2_glock *search_bucket(struct gfs2_gl_hash_bucket *bucket,
 static struct gfs2_glock *gfs2_glock_find(const struct gfs2_sbd *sdp,
                                          const struct lm_lockname *name)
 {
-       struct gfs2_gl_hash_bucket *bucket = &gl_hash_table[gl_hash(sdp, name)];
+       unsigned int hash = gl_hash(sdp, name);
        struct gfs2_glock *gl;
 
-       read_lock(&bucket->hb_lock);
-       gl = search_bucket(bucket, sdp, name);
-       read_unlock(&bucket->hb_lock);
+       read_lock(gl_lock_addr(hash));
+       gl = search_bucket(hash, sdp, name);
+       read_unlock(gl_lock_addr(hash));
 
        return gl;
 }
@@ -263,18 +303,14 @@ int gfs2_glock_get(struct gfs2_sbd *sdp, u64 number,
                   const struct gfs2_glock_operations *glops, int create,
                   struct gfs2_glock **glp)
 {
-       struct lm_lockname name;
+       struct lm_lockname name = { .ln_number = number, .ln_type = glops->go_type };
        struct gfs2_glock *gl, *tmp;
-       struct gfs2_gl_hash_bucket *bucket;
+       unsigned int hash = gl_hash(sdp, &name);
        int error;
 
-       name.ln_number = number;
-       name.ln_type = glops->go_type;
-       bucket = &gl_hash_table[gl_hash(sdp, &name)];
-
-       read_lock(&bucket->hb_lock);
-       gl = search_bucket(bucket, sdp, &name);
-       read_unlock(&bucket->hb_lock);
+       read_lock(gl_lock_addr(hash));
+       gl = search_bucket(hash, sdp, &name);
+       read_unlock(gl_lock_addr(hash));
 
        if (gl || !create) {
                *glp = gl;
@@ -287,8 +323,9 @@ int gfs2_glock_get(struct gfs2_sbd *sdp, u64 number,
 
        gl->gl_flags = 0;
        gl->gl_name = name;
-       kref_init(&gl->gl_ref);
+       atomic_set(&gl->gl_ref, 1);
        gl->gl_state = LM_ST_UNLOCKED;
+       gl->gl_hash = hash;
        gl->gl_owner = NULL;
        gl->gl_ip = 0;
        gl->gl_ops = glops;
@@ -297,7 +334,6 @@ int gfs2_glock_get(struct gfs2_sbd *sdp, u64 number,
        gl->gl_vn = 0;
        gl->gl_stamp = jiffies;
        gl->gl_object = NULL;
-       gl->gl_bucket = bucket;
        gl->gl_sbd = sdp;
        gl->gl_aspace = NULL;
        lops_init_le(&gl->gl_le, &gfs2_glock_lops);
@@ -316,15 +352,15 @@ int gfs2_glock_get(struct gfs2_sbd *sdp, u64 number,
        if (error)
                goto fail_aspace;
 
-       write_lock(&bucket->hb_lock);
-       tmp = search_bucket(bucket, sdp, &name);
+       write_lock(gl_lock_addr(hash));
+       tmp = search_bucket(hash, sdp, &name);
        if (tmp) {
-               write_unlock(&bucket->hb_lock);
+               write_unlock(gl_lock_addr(hash));
                glock_free(gl);
                gl = tmp;
        } else {
-               list_add_tail(&gl->gl_list, &bucket->hb_list);
-               write_unlock(&bucket->hb_lock);
+               hlist_add_head(&gl->gl_list, &gl_hash_table[hash].hb_list);
+               write_unlock(gl_lock_addr(hash));
        }
 
        *glp = gl;
@@ -335,7 +371,7 @@ fail_aspace:
        if (gl->gl_aspace)
                gfs2_aspace_put(gl->gl_aspace);
 fail:
-       kmem_cache_free(gfs2_glock_cachep, gl); 
+       kmem_cache_free(gfs2_glock_cachep, gl);
        return error;
 }
 
@@ -578,7 +614,7 @@ static int rq_greedy(struct gfs2_holder *gh)
        gfs2_holder_uninit(gh);
        kfree(container_of(gh, struct greedy, gr_gh));
 
-       spin_lock(&gl->gl_spin);                
+       spin_lock(&gl->gl_spin);
 
        return 0;
 }
@@ -1148,11 +1184,11 @@ static void add_to_queue(struct gfs2_holder *gh)
        if (existing) {
                print_symbol(KERN_WARNING "original: %s\n", existing->gh_ip);
                printk(KERN_INFO "pid : %d\n", existing->gh_owner->pid);
-               printk(KERN_INFO "lock type : %d lock state : %d\n", 
+               printk(KERN_INFO "lock type : %d lock state : %d\n",
                                existing->gh_gl->gl_name.ln_type, existing->gh_gl->gl_state);
                print_symbol(KERN_WARNING "new: %s\n", gh->gh_ip);
                printk(KERN_INFO "pid : %d\n", gh->gh_owner->pid);
-               printk(KERN_INFO "lock type : %d lock state : %d\n", 
+               printk(KERN_INFO "lock type : %d lock state : %d\n",
                                gl->gl_name.ln_type, gl->gl_state);
                BUG();
        }
@@ -1167,7 +1203,7 @@ static void add_to_queue(struct gfs2_holder *gh)
        if (gh->gh_flags & LM_FLAG_PRIORITY)
                list_add(&gh->gh_list, &gl->gl_waiters3);
        else
-               list_add_tail(&gh->gh_list, &gl->gl_waiters3);  
+               list_add_tail(&gh->gh_list, &gl->gl_waiters3);
 }
 
 /**
@@ -1332,9 +1368,9 @@ static void gfs2_glock_prefetch(struct gfs2_glock *gl, unsigned int state,
        glops->go_xmote_th(gl, state, flags);
 }
 
-static void greedy_work(void *data)
+static void greedy_work(struct work_struct *work)
 {
-       struct greedy *gr = data;
+       struct greedy *gr = container_of(work, struct greedy, gr_work.work);
        struct gfs2_holder *gh = &gr->gr_gh;
        struct gfs2_glock *gl = gh->gh_gl;
        const struct gfs2_glock_operations *glops = gl->gl_ops;
@@ -1386,7 +1422,7 @@ int gfs2_glock_be_greedy(struct gfs2_glock *gl, unsigned int time)
 
        gfs2_holder_init(gl, 0, 0, gh);
        set_bit(HIF_GREEDY, &gh->gh_iflags);
-       INIT_WORK(&gr->gr_work, greedy_work, gr);
+       INIT_DELAYED_WORK(&gr->gr_work, greedy_work);
 
        set_bit(GLF_SKIP_WAITERS2, &gl->gl_flags);
        schedule_delayed_work(&gr->gr_work, time);
@@ -1443,26 +1479,20 @@ int gfs2_glock_nq_num(struct gfs2_sbd *sdp, u64 number,
 
 static int glock_compare(const void *arg_a, const void *arg_b)
 {
-       struct gfs2_holder *gh_a = *(struct gfs2_holder **)arg_a;
-       struct gfs2_holder *gh_b = *(struct gfs2_holder **)arg_b;
-       struct lm_lockname *a = &gh_a->gh_gl->gl_name;
-       struct lm_lockname *b = &gh_b->gh_gl->gl_name;
-       int ret = 0;
+       const struct gfs2_holder *gh_a = *(const struct gfs2_holder **)arg_a;
+       const struct gfs2_holder *gh_b = *(const struct gfs2_holder **)arg_b;
+       const struct lm_lockname *a = &gh_a->gh_gl->gl_name;
+       const struct lm_lockname *b = &gh_b->gh_gl->gl_name;
 
        if (a->ln_number > b->ln_number)
-               ret = 1;
-       else if (a->ln_number < b->ln_number)
-               ret = -1;
-       else {
-               if (gh_a->gh_state == LM_ST_SHARED &&
-                   gh_b->gh_state == LM_ST_EXCLUSIVE)
-                       ret = 1;
-               else if (!(gh_a->gh_flags & GL_LOCAL_EXCL) &&
-                        (gh_b->gh_flags & GL_LOCAL_EXCL))
-                       ret = 1;
-       }
-
-       return ret;
+               return 1;
+       if (a->ln_number < b->ln_number)
+               return -1;
+       if (gh_a->gh_state == LM_ST_SHARED && gh_b->gh_state == LM_ST_EXCLUSIVE)
+               return 1;
+       if (!(gh_a->gh_flags & GL_LOCAL_EXCL) && (gh_b->gh_flags & GL_LOCAL_EXCL))
+               return 1;
+       return 0;
 }
 
 /**
@@ -1713,8 +1743,9 @@ static void blocking_cb(struct gfs2_sbd *sdp, struct lm_lockname *name,
  * a journal from another client needs to be recovered.
  */
 
-void gfs2_glock_cb(struct gfs2_sbd *sdp, unsigned int type, void *data)
+void gfs2_glock_cb(void *cb_data, unsigned int type, void *data)
 {
+       struct gfs2_sbd *sdp = cb_data;
 
        switch (type) {
        case LM_CB_NEED_E:
@@ -1760,22 +1791,6 @@ void gfs2_glock_cb(struct gfs2_sbd *sdp, unsigned int type, void *data)
 }
 
 /**
- * gfs2_iopen_go_callback - Try to kick the inode/vnode associated with an
- *                          iopen glock from memory
- * @io_gl: the iopen glock
- * @state: the state into which the glock should be put
- *
- */
-
-void gfs2_iopen_go_callback(struct gfs2_glock *io_gl, unsigned int state)
-{
-
-       if (state != LM_ST_UNLOCKED)
-               return;
-       /* FIXME: remove this? */
-}
-
-/**
  * demote_ok - Check to see if it's ok to unlock a glock
  * @gl: the glock
  *
@@ -1867,53 +1882,37 @@ void gfs2_reclaim_glock(struct gfs2_sbd *sdp)
  */
 
 static int examine_bucket(glock_examiner examiner, struct gfs2_sbd *sdp,
-                         struct gfs2_gl_hash_bucket *bucket)
+                         unsigned int hash)
 {
-       struct glock_plug plug;
-       struct list_head *tmp;
-       struct gfs2_glock *gl;
-       int entries;
-
-       /* Add "plug" to end of bucket list, work back up list from there */
-       memset(&plug.gl_flags, 0, sizeof(unsigned long));
-       set_bit(GLF_PLUG, &plug.gl_flags);
-
-       write_lock(&bucket->hb_lock);
-       list_add(&plug.gl_list, &bucket->hb_list);
-       write_unlock(&bucket->hb_lock);
-
-       for (;;) {
-               write_lock(&bucket->hb_lock);
-
-               for (;;) {
-                       tmp = plug.gl_list.next;
-
-                       if (tmp == &bucket->hb_list) {
-                               list_del(&plug.gl_list);
-                               entries = !list_empty(&bucket->hb_list);
-                               write_unlock(&bucket->hb_lock);
-                               return entries;
-                       }
-                       gl = list_entry(tmp, struct gfs2_glock, gl_list);
+       struct gfs2_glock *gl, *prev = NULL;
+       int has_entries = 0;
+       struct hlist_head *head = &gl_hash_table[hash].hb_list;
 
-                       /* Move plug up list */
-                       list_move(&plug.gl_list, &gl->gl_list);
-
-                       if (test_bit(GLF_PLUG, &gl->gl_flags))
-                               continue;
-                       if (gl->gl_sbd != sdp)
-                               continue;
-
-                       /* examiner() must glock_put() */
+       read_lock(gl_lock_addr(hash));
+       /* Can't use hlist_for_each_entry - don't want prefetch here */
+       if (hlist_empty(head))
+               goto out;
+       gl = list_entry(head->first, struct gfs2_glock, gl_list);
+       while(1) {
+               if (gl->gl_sbd == sdp) {
                        gfs2_glock_hold(gl);
-
-                       break;
+                       read_unlock(gl_lock_addr(hash));
+                       if (prev)
+                               gfs2_glock_put(prev);
+                       prev = gl;
+                       examiner(gl);
+                       has_entries = 1;
+                       read_lock(gl_lock_addr(hash));
                }
-
-               write_unlock(&bucket->hb_lock);
-
-               examiner(gl);
+               if (gl->gl_list.next == NULL)
+                       break;
+               gl = list_entry(gl->gl_list.next, struct gfs2_glock, gl_list);
        }
+out:
+       read_unlock(gl_lock_addr(hash));
+       if (prev)
+               gfs2_glock_put(prev);
+       return has_entries;
 }
 
 /**
@@ -1925,23 +1924,19 @@ static int examine_bucket(glock_examiner examiner, struct gfs2_sbd *sdp,
 static void scan_glock(struct gfs2_glock *gl)
 {
        if (gl->gl_ops == &gfs2_inode_glops)
-               goto out;
+               return;
 
        if (gfs2_glmutex_trylock(gl)) {
                if (queue_empty(gl, &gl->gl_holders) &&
-                   gl->gl_state != LM_ST_UNLOCKED &&
-                   demote_ok(gl))
+                   gl->gl_state != LM_ST_UNLOCKED && demote_ok(gl))
                        goto out_schedule;
                gfs2_glmutex_unlock(gl);
        }
-out:
-       gfs2_glock_put(gl);
        return;
 
 out_schedule:
        gfs2_glmutex_unlock(gl);
        gfs2_glock_schedule_for_reclaim(gl);
-       gfs2_glock_put(gl);
 }
 
 /**
@@ -1954,10 +1949,8 @@ void gfs2_scand_internal(struct gfs2_sbd *sdp)
 {
        unsigned int x;
 
-       for (x = 0; x < GFS2_GL_HASH_SIZE; x++) {
-               examine_bucket(scan_glock, sdp, &gl_hash_table[x]);
-               cond_resched();
-       }
+       for (x = 0; x < GFS2_GL_HASH_SIZE; x++)
+               examine_bucket(scan_glock, sdp, x);
 }
 
 /**
@@ -1986,11 +1979,8 @@ static void clear_glock(struct gfs2_glock *gl)
                if (queue_empty(gl, &gl->gl_holders) &&
                    gl->gl_state != LM_ST_UNLOCKED)
                        handle_callback(gl, LM_ST_UNLOCKED);
-
                gfs2_glmutex_unlock(gl);
        }
-
-       gfs2_glock_put(gl);
 }
 
 /**
@@ -2012,10 +2002,10 @@ void gfs2_gl_hash_clear(struct gfs2_sbd *sdp, int wait)
 
        for (;;) {
                cont = 0;
-
-               for (x = 0; x < GFS2_GL_HASH_SIZE; x++)
-                       if (examine_bucket(clear_glock, sdp, &gl_hash_table[x]))
+               for (x = 0; x < GFS2_GL_HASH_SIZE; x++) {
+                       if (examine_bucket(clear_glock, sdp, x))
                                cont = 1;
+               }
 
                if (!wait || !cont)
                        break;
@@ -2124,7 +2114,7 @@ static int dump_glock(struct gfs2_glock *gl)
                        printk(" %u", x);
        }
        printk(" \n");
-       printk(KERN_INFO "  gl_ref = %d\n", atomic_read(&gl->gl_ref.refcount));
+       printk(KERN_INFO "  gl_ref = %d\n", atomic_read(&gl->gl_ref));
        printk(KERN_INFO "  gl_state = %u\n", gl->gl_state);
        printk(KERN_INFO "  gl_owner = %s\n", gl->gl_owner->comm);
        print_symbol(KERN_INFO "  gl_ip = %s\n", gl->gl_ip);
@@ -2197,19 +2187,16 @@ out:
 
 static int gfs2_dump_lockstate(struct gfs2_sbd *sdp)
 {
-       struct gfs2_gl_hash_bucket *bucket;
        struct gfs2_glock *gl;
+       struct hlist_node *h;
        unsigned int x;
        int error = 0;
 
        for (x = 0; x < GFS2_GL_HASH_SIZE; x++) {
-               bucket = &gl_hash_table[x];
 
-               read_lock(&bucket->hb_lock);
+               read_lock(gl_lock_addr(x));
 
-               list_for_each_entry(gl, &bucket->hb_list, gl_list) {
-                       if (test_bit(GLF_PLUG, &gl->gl_flags))
-                               continue;
+               hlist_for_each_entry(gl, h, &gl_hash_table[x].hb_list, gl_list) {
                        if (gl->gl_sbd != sdp)
                                continue;
 
@@ -2218,7 +2205,7 @@ static int gfs2_dump_lockstate(struct gfs2_sbd *sdp)
                                break;
                }
 
-               read_unlock(&bucket->hb_lock);
+               read_unlock(gl_lock_addr(x));
 
                if (error)
                        break;
@@ -2232,10 +2219,13 @@ int __init gfs2_glock_init(void)
 {
        unsigned i;
        for(i = 0; i < GFS2_GL_HASH_SIZE; i++) {
-               struct gfs2_gl_hash_bucket *hb = &gl_hash_table[i];
-               rwlock_init(&hb->hb_lock);
-               INIT_LIST_HEAD(&hb->hb_list);
+               INIT_HLIST_HEAD(&gl_hash_table[i].hb_list);
+       }
+#ifdef GL_HASH_LOCK_SZ
+       for(i = 0; i < GL_HASH_LOCK_SZ; i++) {
+               rwlock_init(&gl_hash_locks[i]);
        }
+#endif
        return 0;
 }