Directory xattr znodes hold a reference on their parent

[zfs.git] / module / zfs / dsl_deadlist.c

diff --git a/module/zfs/dsl_deadlist.c b/module/zfs/dsl_deadlist.c
index 064f8ac..1e89a68 100644 (file)
--- a/module/zfs/dsl_deadlist.c
+++ b/module/zfs/dsl_deadlist.c
@@ -20,6 +20,7 @@
  */
 /*
  * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2011 by Delphix. All rights reserved.
  */
 
 #include <sys/dsl_dataset.h>
@@ -29,6 +30,26 @@
 #include <sys/zfs_context.h>
 #include <sys/dsl_pool.h>
 
+/*
+ * Deadlist concurrency:
+ *
+ * Deadlists can only be modified from the syncing thread.
+ *
+ * Except for dsl_deadlist_insert(), it can only be modified with the
+ * dp_config_rwlock held with RW_WRITER.
+ *
+ * The accessors (dsl_deadlist_space() and dsl_deadlist_space_range()) can
+ * be called concurrently, from open context, with the dl_config_rwlock held
+ * with RW_READER.
+ *
+ * Therefore, we only need to provide locking between dsl_deadlist_insert() and
+ * the accessors, protecting:
+ *     dl_phys->dl_used,comp,uncomp
+ *     and protecting the dl_tree from being loaded.
+ * The locking is provided by dl_lock.  Note that locking on the bpobj_t
+ * provides its own locking, and dl_oldfmt is immutable.
+ */
+
 static int
 dsl_deadlist_compare(const void *arg1, const void *arg2)
 {
@@ -59,7 +80,9 @@ dsl_deadlist_load_tree(dsl_deadlist_t *dl)
        for (zap_cursor_init(&zc, dl->dl_os, dl->dl_object);
            zap_cursor_retrieve(&zc, &za) == 0;
            zap_cursor_advance(&zc)) {
-               dsl_deadlist_entry_t *dle = kmem_alloc(sizeof (*dle), KM_SLEEP);
+               dsl_deadlist_entry_t *dle;
+
+               dle = kmem_alloc(sizeof (*dle), KM_PUSHPAGE);
                dle->dle_mintxg = strtonum(za.za_name, NULL);
                VERIFY3U(0, ==, bpobj_open(&dle->dle_bpobj, dl->dl_os,
                    za.za_first_integer));
@@ -194,7 +217,7 @@ dsl_deadlist_add_key(dsl_deadlist_t *dl, uint64_t mintxg, dmu_tx_t *tx)
 
        dsl_deadlist_load_tree(dl);
 
-       dle = kmem_alloc(sizeof (*dle), KM_SLEEP);
+       dle = kmem_alloc(sizeof (*dle), KM_PUSHPAGE);
        dle->dle_mintxg = mintxg;
        obj = bpobj_alloc(dl->dl_os, SPA_MAXBLOCKSIZE, tx);
        VERIFY3U(0, ==, bpobj_open(&dle->dle_bpobj, dl->dl_os, obj));
@@ -309,14 +332,14 @@ dsl_deadlist_space(dsl_deadlist_t *dl,
  * return space used in the range (mintxg, maxtxg].
  * Includes maxtxg, does not include mintxg.
  * mintxg and maxtxg must both be keys in the deadlist (unless maxtxg is
- * UINT64_MAX).
+ * larger than any bp in the deadlist (eg. UINT64_MAX)).
  */
 void
 dsl_deadlist_space_range(dsl_deadlist_t *dl, uint64_t mintxg, uint64_t maxtxg,
     uint64_t *usedp, uint64_t *compp, uint64_t *uncompp)
 {
-       dsl_deadlist_entry_t dle_tofind;
        dsl_deadlist_entry_t *dle;
+       dsl_deadlist_entry_t dle_tofind;
        avl_index_t where;
 
        if (dl->dl_oldfmt) {
@@ -325,9 +348,10 @@ dsl_deadlist_space_range(dsl_deadlist_t *dl, uint64_t mintxg, uint64_t maxtxg,
                return;
        }
 
-       dsl_deadlist_load_tree(dl);
        *usedp = *compp = *uncompp = 0;
 
+       mutex_enter(&dl->dl_lock);
+       dsl_deadlist_load_tree(dl);
        dle_tofind.dle_mintxg = mintxg;
        dle = avl_find(&dl->dl_tree, &dle_tofind, &where);
        /*
@@ -336,6 +360,7 @@ dsl_deadlist_space_range(dsl_deadlist_t *dl, uint64_t mintxg, uint64_t maxtxg,
         */
        ASSERT(dle != NULL ||
            avl_nearest(&dl->dl_tree, where, AVL_AFTER) == NULL);
+
        for (; dle && dle->dle_mintxg < maxtxg;
            dle = AVL_NEXT(&dl->dl_tree, dle)) {
                uint64_t used, comp, uncomp;
@@ -347,6 +372,7 @@ dsl_deadlist_space_range(dsl_deadlist_t *dl, uint64_t mintxg, uint64_t maxtxg,
                *compp += comp;
                *uncompp += uncomp;
        }
+       mutex_exit(&dl->dl_lock);
 }
 
 static void