sys/cddl/contrib/opensolaris/uts/common/fs/zfs/sys/dmu_impl.h

   1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  22  * Copyright 2010 Sun Microsystems, Inc.  All rights reserved.
  23  * Use is subject to license terms.
  24  */
  25
  26 #ifndef _SYS_DMU_IMPL_H
  27 #define _SYS_DMU_IMPL_H
  28
  29 #include <sys/txg_impl.h>
  30 #include <sys/zio.h>
  31 #include <sys/dnode.h>
  32 #include <sys/kstat.h>
  33 #include <sys/zfs_context.h>
  34
  35 #ifdef  __cplusplus
  36 extern "C" {
  37 #endif
  38
  39 /*
  40  * This is the locking strategy for the DMU.  Numbers in parenthesis are
  41  * cases that use that lock order, referenced below:
  42  *
  43  * ARC is self-contained
  44  * bplist is self-contained
  45  * refcount is self-contained
  46  * txg is self-contained (hopefully!)
  47  * zst_lock
  48  * zf_rwlock
  49  *
  50  * XXX try to improve evicting path?
  51  *
  52  * dp_config_rwlock > os_obj_lock > dn_struct_rwlock >
  53  *      dn_dbufs_mtx > hash_mutexes > db_mtx > dd_lock > leafs
  54  *
  55  * dp_config_rwlock
  56  *    must be held before: everything
  57  *    protects dd namespace changes
  58  *    protects property changes globally
  59  *    held from:
  60  *      dsl_dir_open/r:
  61  *      dsl_dir_create_sync/w:
  62  *      dsl_dir_sync_destroy/w:
  63  *      dsl_dir_rename_sync/w:
  64  *      dsl_prop_changed_notify/r:
  65  *
  66  * os_obj_lock
  67  *   must be held before:
  68  *      everything except dp_config_rwlock
  69  *   protects os_obj_next
  70  *   held from:
  71  *      dmu_object_alloc: dn_dbufs_mtx, db_mtx, hash_mutexes, dn_struct_rwlock
  72  *
  73  * dn_struct_rwlock
  74  *   must be held before:
  75  *      everything except dp_config_rwlock and os_obj_lock
  76  *   protects structure of dnode (eg. nlevels)
  77  *      db_blkptr can change when syncing out change to nlevels
  78  *      dn_maxblkid
  79  *      dn_nlevels
  80  *      dn_*blksz*
  81  *      phys nlevels, maxblkid, physical blkptr_t's (?)
  82  *   held from:
  83  *      callers of dbuf_read_impl, dbuf_hold[_impl], dbuf_prefetch
  84  *      dmu_object_info_from_dnode: dn_dirty_mtx (dn_datablksz)
  85  *      dmu_tx_count_free:
  86  *      dbuf_read_impl: db_mtx, dmu_zfetch()
  87  *      dmu_zfetch: zf_rwlock/r, zst_lock, dbuf_prefetch()
  88  *      dbuf_new_size: db_mtx
  89  *      dbuf_dirty: db_mtx
  90  *      dbuf_findbp: (callers, phys? - the real need)
  91  *      dbuf_create: dn_dbufs_mtx, hash_mutexes, db_mtx (phys?)
  92  *      dbuf_prefetch: dn_dirty_mtx, hash_mutexes, db_mtx, dn_dbufs_mtx
  93  *      dbuf_hold_impl: hash_mutexes, db_mtx, dn_dbufs_mtx, dbuf_findbp()
  94  *      dnode_sync/w (increase_indirection): db_mtx (phys)
  95  *      dnode_set_blksz/w: dn_dbufs_mtx (dn_*blksz*)
  96  *      dnode_new_blkid/w: (dn_maxblkid)
  97  *      dnode_free_range/w: dn_dirty_mtx (dn_maxblkid)
  98  *      dnode_next_offset: (phys)
  99  *
 100  * dn_dbufs_mtx
 101  *    must be held before:
 102  *      db_mtx, hash_mutexes
 103  *    protects:
 104  *      dn_dbufs
 105  *      dn_evicted
 106  *    held from:
 107  *      dmu_evict_user: db_mtx (dn_dbufs)
 108  *      dbuf_free_range: db_mtx (dn_dbufs)
 109  *      dbuf_remove_ref: db_mtx, callees:
 110  *              dbuf_hash_remove: hash_mutexes, db_mtx
 111  *      dbuf_create: hash_mutexes, db_mtx (dn_dbufs)
 112  *      dnode_set_blksz: (dn_dbufs)
 113  *
 114  * hash_mutexes (global)
 115  *   must be held before:
 116  *      db_mtx
 117  *   protects dbuf_hash_table (global) and db_hash_next
 118  *   held from:
 119  *      dbuf_find: db_mtx
 120  *      dbuf_hash_insert: db_mtx
 121  *      dbuf_hash_remove: db_mtx
 122  *
 123  * db_mtx (meta-leaf)
 124  *   must be held before:
 125  *      dn_mtx, dn_dirty_mtx, dd_lock (leaf mutexes)
 126  *   protects:
 127  *      db_state
 128  *      db_holds
 129  *      db_buf
 130  *      db_changed
 131  *      db_data_pending
 132  *      db_dirtied
 133  *      db_link
 134  *      db_dirty_node (??)
 135  *      db_dirtycnt
 136  *      db_d.*
 137  *      db.*
 138  *   held from:
 139  *      dbuf_dirty: dn_mtx, dn_dirty_mtx
 140  *      dbuf_dirty->dsl_dir_willuse_space: dd_lock
 141  *      dbuf_dirty->dbuf_new_block->dsl_dataset_block_freeable: dd_lock
 142  *      dbuf_undirty: dn_dirty_mtx (db_d)
 143  *      dbuf_write_done: dn_dirty_mtx (db_state)
 144  *      dbuf_*
 145  *      dmu_buf_update_user: none (db_d)
 146  *      dmu_evict_user: none (db_d) (maybe can eliminate)
 147  *      dbuf_find: none (db_holds)
 148  *      dbuf_hash_insert: none (db_holds)
 149  *      dmu_buf_read_array_impl: none (db_state, db_changed)
 150  *      dmu_sync: none (db_dirty_node, db_d)
 151  *      dnode_reallocate: none (db)
 152  *
 153  * dn_mtx (leaf)
 154  *   protects:
 155  *      dn_dirty_dbufs
 156  *      dn_ranges
 157  *      phys accounting
 158  *      dn_allocated_txg
 159  *      dn_free_txg
 160  *      dn_assigned_txg
 161  *      dd_assigned_tx
 162  *      dn_notxholds
 163  *      dn_dirtyctx
 164  *      dn_dirtyctx_firstset
 165  *      (dn_phys copy fields?)
 166  *      (dn_phys contents?)
 167  *   held from:
 168  *      dnode_*
 169  *      dbuf_dirty: none
 170  *      dbuf_sync: none (phys accounting)
 171  *      dbuf_undirty: none (dn_ranges, dn_dirty_dbufs)
 172  *      dbuf_write_done: none (phys accounting)
 173  *      dmu_object_info_from_dnode: none (accounting)
 174  *      dmu_tx_commit: none
 175  *      dmu_tx_hold_object_impl: none
 176  *      dmu_tx_try_assign: dn_notxholds(cv)
 177  *      dmu_tx_unassign: none
 178  *
 179  * dd_lock
 180  *    must be held before:
 181  *      ds_lock
 182  *      ancestors' dd_lock
 183  *    protects:
 184  *      dd_prop_cbs
 185  *      dd_sync_*
 186  *      dd_used_bytes
 187  *      dd_tempreserved
 188  *      dd_space_towrite
 189  *      dd_myname
 190  *      dd_phys accounting?
 191  *    held from:
 192  *      dsl_dir_*
 193  *      dsl_prop_changed_notify: none (dd_prop_cbs)
 194  *      dsl_prop_register: none (dd_prop_cbs)
 195  *      dsl_prop_unregister: none (dd_prop_cbs)
 196  *      dsl_dataset_block_freeable: none (dd_sync_*)
 197  *
 198  * os_lock (leaf)
 199  *   protects:
 200  *      os_dirty_dnodes
 201  *      os_free_dnodes
 202  *      os_dnodes
 203  *      os_downgraded_dbufs
 204  *      dn_dirtyblksz
 205  *      dn_dirty_link
 206  *   held from:
 207  *      dnode_create: none (os_dnodes)
 208  *      dnode_destroy: none (os_dnodes)
 209  *      dnode_setdirty: none (dn_dirtyblksz, os_*_dnodes)
 210  *      dnode_free: none (dn_dirtyblksz, os_*_dnodes)
 211  *
 212  * ds_lock
 213  *    protects:
 214  *      ds_objset
 215  *      ds_open_refcount
 216  *      ds_snapname
 217  *      ds_phys accounting
 218  *      ds_phys userrefs zapobj
 219  *      ds_reserved
 220  *    held from:
 221  *      dsl_dataset_*
 222  *
 223  * dr_mtx (leaf)
 224  *    protects:
 225  *      dr_children
 226  *    held from:
 227  *      dbuf_dirty
 228  *      dbuf_undirty
 229  *      dbuf_sync_indirect
 230  *      dnode_new_blkid
 231  */
 232
 233 struct objset;
 234 struct dmu_pool;
 235
 236 typedef struct dmu_xuio {
 237         int next;
 238         int cnt;
 239         struct arc_buf **bufs;
 240         iovec_t *iovp;
 241 } dmu_xuio_t;
 242
 243 typedef struct xuio_stats {
 244         /* loaned yet not returned arc_buf */
 245         kstat_named_t xuiostat_onloan_rbuf;
 246         kstat_named_t xuiostat_onloan_wbuf;
 247         /* whether a copy is made when loaning out a read buffer */
 248         kstat_named_t xuiostat_rbuf_copied;
 249         kstat_named_t xuiostat_rbuf_nocopy;
 250         /* whether a copy is made when assigning a write buffer */
 251         kstat_named_t xuiostat_wbuf_copied;
 252         kstat_named_t xuiostat_wbuf_nocopy;
 253 } xuio_stats_t;
 254
 255 static xuio_stats_t xuio_stats = {
 256         { "onloan_read_buf",    KSTAT_DATA_UINT64 },
 257         { "onloan_write_buf",   KSTAT_DATA_UINT64 },
 258         { "read_buf_copied",    KSTAT_DATA_UINT64 },
 259         { "read_buf_nocopy",    KSTAT_DATA_UINT64 },
 260         { "write_buf_copied",   KSTAT_DATA_UINT64 },
 261         { "write_buf_nocopy",   KSTAT_DATA_UINT64 }
 262 };
 263
 264 #define XUIOSTAT_INCR(stat, val)        \
 265         atomic_add_64(&xuio_stats.stat.value.ui64, (val))
 266 #define XUIOSTAT_BUMP(stat)     XUIOSTAT_INCR(stat, 1)
 267
 268
 269 #ifdef  __cplusplus
 270 }
 271 #endif
 272
 273 #endif  /* _SYS_DMU_IMPL_H */