drivers/staging/zcache/tmem.c

   1 /*
   2  * In-kernel transcendent memory (generic implementation)
   3  *
   4  * Copyright (c) 2009-2011, Dan Magenheimer, Oracle Corp.
   5  *
   6  * The primary purpose of Transcedent Memory ("tmem") is to map object-oriented
   7  * "handles" (triples containing a pool id, and object id, and an index), to
   8  * pages in a page-accessible memory (PAM).  Tmem references the PAM pages via
   9  * an abstract "pampd" (PAM page-descriptor), which can be operated on by a
  10  * set of functions (pamops).  Each pampd contains some representation of
  11  * PAGE_SIZE bytes worth of data. Tmem must support potentially millions of
  12  * pages and must be able to insert, find, and delete these pages at a
  13  * potential frequency of thousands per second concurrently across many CPUs,
  14  * (and, if used with KVM, across many vcpus across many guests).
  15  * Tmem is tracked with a hierarchy of data structures, organized by
  16  * the elements in a handle-tuple: pool_id, object_id, and page index.
  17  * One or more "clients" (e.g. guests) each provide one or more tmem_pools.
  18  * Each pool, contains a hash table of rb_trees of tmem_objs.  Each
  19  * tmem_obj contains a radix-tree-like tree of pointers, with intermediate
  20  * nodes called tmem_objnodes.  Each leaf pointer in this tree points to
  21  * a pampd, which is accessible only through a small set of callbacks
  22  * registered by the PAM implementation (see tmem_register_pamops). Tmem
  23  * does all memory allocation via a set of callbacks registered by the tmem
  24  * host implementation (e.g. see tmem_register_hostops).
  25  */
  26
  27 #include <linux/list.h>
  28 #include <linux/spinlock.h>
  29 #include <linux/atomic.h>
  30
  31 #include "tmem.h"
  32
  33 /* data structure sentinels used for debugging... see tmem.h */
  34 #define POOL_SENTINEL 0x87658765
  35 #define OBJ_SENTINEL 0x12345678
  36 #define OBJNODE_SENTINEL 0xfedcba09
  37
  38 /*
  39  * A tmem host implementation must use this function to register callbacks
  40  * for memory allocation.
  41  */
  42 static struct tmem_hostops tmem_hostops;
  43
  44 static void tmem_objnode_tree_init(void);
  45
  46 void tmem_register_hostops(struct tmem_hostops *m)
  47 {
  48         tmem_objnode_tree_init();
  49         tmem_hostops = *m;
  50 }
  51
  52 /*
  53  * A tmem host implementation must use this function to register
  54  * callbacks for a page-accessible memory (PAM) implementation
  55  */
  56 static struct tmem_pamops tmem_pamops;
  57
  58 void tmem_register_pamops(struct tmem_pamops *m)
  59 {
  60         tmem_pamops = *m;
  61 }
  62
  63 /*
  64  * Oid's are potentially very sparse and tmem_objs may have an indeterminately
  65  * short life, being added and deleted at a relatively high frequency.
  66  * So an rb_tree is an ideal data structure to manage tmem_objs.  But because
  67  * of the potentially huge number of tmem_objs, each pool manages a hashtable
  68  * of rb_trees to reduce search, insert, delete, and rebalancing time.
  69  * Each hashbucket also has a lock to manage concurrent access.
  70  *
  71  * The following routines manage tmem_objs.  When any tmem_obj is accessed,
  72  * the hashbucket lock must be held.
  73  */
  74
  75 /* searches for object==oid in pool, returns locked object if found */
  76 static struct tmem_obj *tmem_obj_find(struct tmem_hashbucket *hb,
  77                                         struct tmem_oid *oidp)
  78 {
  79         struct rb_node *rbnode;
  80         struct tmem_obj *obj;
  81
  82         rbnode = hb->obj_rb_root.rb_node;
  83         while (rbnode) {
  84                 BUG_ON(RB_EMPTY_NODE(rbnode));
  85                 obj = rb_entry(rbnode, struct tmem_obj, rb_tree_node);
  86                 switch (tmem_oid_compare(oidp, &obj->oid)) {
  87                 case 0: /* equal */
  88                         goto out;
  89                 case -1:
  90                         rbnode = rbnode->rb_left;
  91                         break;
  92                 case 1:
  93                         rbnode = rbnode->rb_right;
  94                         break;
  95                 }
  96         }
  97         obj = NULL;
  98 out:
  99         return obj;
 100 }
 101
 102 static void tmem_pampd_destroy_all_in_obj(struct tmem_obj *);
 103
 104 /* free an object that has no more pampds in it */
 105 static void tmem_obj_free(struct tmem_obj *obj, struct tmem_hashbucket *hb)
 106 {
 107         struct tmem_pool *pool;
 108
 109         BUG_ON(obj == NULL);
 110         ASSERT_SENTINEL(obj, OBJ);
 111         BUG_ON(obj->pampd_count > 0);
 112         pool = obj->pool;
 113         BUG_ON(pool == NULL);
 114         if (obj->objnode_tree_root != NULL) /* may be "stump" with no leaves */
 115                 tmem_pampd_destroy_all_in_obj(obj);
 116         BUG_ON(obj->objnode_tree_root != NULL);
 117         BUG_ON((long)obj->objnode_count != 0);
 118         atomic_dec(&pool->obj_count);
 119         BUG_ON(atomic_read(&pool->obj_count) < 0);
 120         INVERT_SENTINEL(obj, OBJ);
 121         obj->pool = NULL;
 122         tmem_oid_set_invalid(&obj->oid);
 123         rb_erase(&obj->rb_tree_node, &hb->obj_rb_root);
 124 }
 125
 126 /*
 127  * initialize, and insert an tmem_object_root (called only if find failed)
 128  */
 129 static void tmem_obj_init(struct tmem_obj *obj, struct tmem_hashbucket *hb,
 130                                         struct tmem_pool *pool,
 131                                         struct tmem_oid *oidp)
 132 {
 133         struct rb_root *root = &hb->obj_rb_root;
 134         struct rb_node **new = &(root->rb_node), *parent = NULL;
 135         struct tmem_obj *this;
 136
 137         BUG_ON(pool == NULL);
 138         atomic_inc(&pool->obj_count);
 139         obj->objnode_tree_height = 0;
 140         obj->objnode_tree_root = NULL;
 141         obj->pool = pool;
 142         obj->oid = *oidp;
 143         obj->objnode_count = 0;
 144         obj->pampd_count = 0;
 145         SET_SENTINEL(obj, OBJ);
 146         while (*new) {
 147                 BUG_ON(RB_EMPTY_NODE(*new));
 148                 this = rb_entry(*new, struct tmem_obj, rb_tree_node);
 149                 parent = *new;
 150                 switch (tmem_oid_compare(oidp, &this->oid)) {
 151                 case 0:
 152                         BUG(); /* already present; should never happen! */
 153                         break;
 154                 case -1:
 155                         new = &(*new)->rb_left;
 156                         break;
 157                 case 1:
 158                         new = &(*new)->rb_right;
 159                         break;
 160                 }
 161         }
 162         rb_link_node(&obj->rb_tree_node, parent, new);
 163         rb_insert_color(&obj->rb_tree_node, root);
 164 }
 165
 166 /*
 167  * Tmem is managed as a set of tmem_pools with certain attributes, such as
 168  * "ephemeral" vs "persistent".  These attributes apply to all tmem_objs
 169  * and all pampds that belong to a tmem_pool.  A tmem_pool is created
 170  * or deleted relatively rarely (for example, when a filesystem is
 171  * mounted or unmounted.
 172  */
 173
 174 /* flush all data from a pool and, optionally, free it */
 175 static void tmem_pool_flush(struct tmem_pool *pool, bool destroy)
 176 {
 177         struct rb_node *rbnode;
 178         struct tmem_obj *obj;
 179         struct tmem_hashbucket *hb = &pool->hashbucket[0];
 180         int i;
 181
 182         BUG_ON(pool == NULL);
 183         for (i = 0; i < TMEM_HASH_BUCKETS; i++, hb++) {
 184                 spin_lock(&hb->lock);
 185                 rbnode = rb_first(&hb->obj_rb_root);
 186                 while (rbnode != NULL) {
 187                         obj = rb_entry(rbnode, struct tmem_obj, rb_tree_node);
 188                         rbnode = rb_next(rbnode);
 189                         tmem_pampd_destroy_all_in_obj(obj);
 190                         tmem_obj_free(obj, hb);
 191                         (*tmem_hostops.obj_free)(obj, pool);
 192                 }
 193                 spin_unlock(&hb->lock);
 194         }
 195         if (destroy)
 196                 list_del(&pool->pool_list);
 197 }
 198
 199 /*
 200  * A tmem_obj contains a radix-tree-like tree in which the intermediate
 201  * nodes are called tmem_objnodes.  (The kernel lib/radix-tree.c implementation
 202  * is very specialized and tuned for specific uses and is not particularly
 203  * suited for use from this code, though some code from the core algorithms has
 204  * been reused, thus the copyright notices below).  Each tmem_objnode contains
 205  * a set of pointers which point to either a set of intermediate tmem_objnodes
 206  * or a set of of pampds.
 207  *
 208  * Portions Copyright (C) 2001 Momchil Velikov
 209  * Portions Copyright (C) 2001 Christoph Hellwig
 210  * Portions Copyright (C) 2005 SGI, Christoph Lameter <clameter@sgi.com>
 211  */
 212
 213 struct tmem_objnode_tree_path {
 214         struct tmem_objnode *objnode;
 215         int offset;
 216 };
 217
 218 /* objnode height_to_maxindex translation */
 219 static unsigned long tmem_objnode_tree_h2max[OBJNODE_TREE_MAX_PATH + 1];
 220
 221 static void tmem_objnode_tree_init(void)
 222 {
 223         unsigned int ht, tmp;
 224
 225         for (ht = 0; ht < ARRAY_SIZE(tmem_objnode_tree_h2max); ht++) {
 226                 tmp = ht * OBJNODE_TREE_MAP_SHIFT;
 227                 if (tmp >= OBJNODE_TREE_INDEX_BITS)
 228                         tmem_objnode_tree_h2max[ht] = ~0UL;
 229                 else
 230                         tmem_objnode_tree_h2max[ht] =
 231                             (~0UL >> (OBJNODE_TREE_INDEX_BITS - tmp - 1)) >> 1;
 232         }
 233 }
 234
 235 static struct tmem_objnode *tmem_objnode_alloc(struct tmem_obj *obj)
 236 {
 237         struct tmem_objnode *objnode;
 238
 239         ASSERT_SENTINEL(obj, OBJ);
 240         BUG_ON(obj->pool == NULL);
 241         ASSERT_SENTINEL(obj->pool, POOL);
 242         objnode = (*tmem_hostops.objnode_alloc)(obj->pool);
 243         if (unlikely(objnode == NULL))
 244                 goto out;
 245         objnode->obj = obj;
 246         SET_SENTINEL(objnode, OBJNODE);
 247         memset(&objnode->slots, 0, sizeof(objnode->slots));
 248         objnode->slots_in_use = 0;
 249         obj->objnode_count++;
 250 out:
 251         return objnode;
 252 }
 253
 254 static void tmem_objnode_free(struct tmem_objnode *objnode)
 255 {
 256         struct tmem_pool *pool;
 257         int i;
 258
 259         BUG_ON(objnode == NULL);
 260         for (i = 0; i < OBJNODE_TREE_MAP_SIZE; i++)
 261                 BUG_ON(objnode->slots[i] != NULL);
 262         ASSERT_SENTINEL(objnode, OBJNODE);
 263         INVERT_SENTINEL(objnode, OBJNODE);
 264         BUG_ON(objnode->obj == NULL);
 265         ASSERT_SENTINEL(objnode->obj, OBJ);
 266         pool = objnode->obj->pool;
 267         BUG_ON(pool == NULL);
 268         ASSERT_SENTINEL(pool, POOL);
 269         objnode->obj->objnode_count--;
 270         objnode->obj = NULL;
 271         (*tmem_hostops.objnode_free)(objnode, pool);
 272 }
 273
 274 /*
 275  * lookup index in object and return associated pampd (or NULL if not found)
 276  */
 277 static void *tmem_pampd_lookup_in_obj(struct tmem_obj *obj, uint32_t index)
 278 {
 279         unsigned int height, shift;
 280         struct tmem_objnode **slot = NULL;
 281
 282         BUG_ON(obj == NULL);
 283         ASSERT_SENTINEL(obj, OBJ);
 284         BUG_ON(obj->pool == NULL);
 285         ASSERT_SENTINEL(obj->pool, POOL);
 286
 287         height = obj->objnode_tree_height;
 288         if (index > tmem_objnode_tree_h2max[obj->objnode_tree_height])
 289                 goto out;
 290         if (height == 0 && obj->objnode_tree_root) {
 291                 slot = &obj->objnode_tree_root;
 292                 goto out;
 293         }
 294         shift = (height-1) * OBJNODE_TREE_MAP_SHIFT;
 295         slot = &obj->objnode_tree_root;
 296         while (height > 0) {
 297                 if (*slot == NULL)
 298                         goto out;
 299                 slot = (struct tmem_objnode **)
 300                         ((*slot)->slots +
 301                          ((index >> shift) & OBJNODE_TREE_MAP_MASK));
 302                 shift -= OBJNODE_TREE_MAP_SHIFT;
 303                 height--;
 304         }
 305 out:
 306         return slot != NULL ? *slot : NULL;
 307 }
 308
 309 static int tmem_pampd_add_to_obj(struct tmem_obj *obj, uint32_t index,
 310                                         void *pampd)
 311 {
 312         int ret = 0;
 313         struct tmem_objnode *objnode = NULL, *newnode, *slot;
 314         unsigned int height, shift;
 315         int offset = 0;
 316
 317         /* if necessary, extend the tree to be higher  */
 318         if (index > tmem_objnode_tree_h2max[obj->objnode_tree_height]) {
 319                 height = obj->objnode_tree_height + 1;
 320                 if (index > tmem_objnode_tree_h2max[height])
 321                         while (index > tmem_objnode_tree_h2max[height])
 322                                 height++;
 323                 if (obj->objnode_tree_root == NULL) {
 324                         obj->objnode_tree_height = height;
 325                         goto insert;
 326                 }
 327                 do {
 328                         newnode = tmem_objnode_alloc(obj);
 329                         if (!newnode) {
 330                                 ret = -ENOMEM;
 331                                 goto out;
 332                         }
 333                         newnode->slots[0] = obj->objnode_tree_root;
 334                         newnode->slots_in_use = 1;
 335                         obj->objnode_tree_root = newnode;
 336                         obj->objnode_tree_height++;
 337                 } while (height > obj->objnode_tree_height);
 338         }
 339 insert:
 340         slot = obj->objnode_tree_root;
 341         height = obj->objnode_tree_height;
 342         shift = (height-1) * OBJNODE_TREE_MAP_SHIFT;
 343         while (height > 0) {
 344                 if (slot == NULL) {
 345                         /* add a child objnode.  */
 346                         slot = tmem_objnode_alloc(obj);
 347                         if (!slot) {
 348                                 ret = -ENOMEM;
 349                                 goto out;
 350                         }
 351                         if (objnode) {
 352
 353                                 objnode->slots[offset] = slot;
 354                                 objnode->slots_in_use++;
 355                         } else
 356                                 obj->objnode_tree_root = slot;
 357                 }
 358                 /* go down a level */
 359                 offset = (index >> shift) & OBJNODE_TREE_MAP_MASK;
 360                 objnode = slot;
 361                 slot = objnode->slots[offset];
 362                 shift -= OBJNODE_TREE_MAP_SHIFT;
 363                 height--;
 364         }
 365         BUG_ON(slot != NULL);
 366         if (objnode) {
 367                 objnode->slots_in_use++;
 368                 objnode->slots[offset] = pampd;
 369         } else
 370                 obj->objnode_tree_root = pampd;
 371         obj->pampd_count++;
 372 out:
 373         return ret;
 374 }
 375
 376 static void *tmem_pampd_delete_from_obj(struct tmem_obj *obj, uint32_t index)
 377 {
 378         struct tmem_objnode_tree_path path[OBJNODE_TREE_MAX_PATH + 1];
 379         struct tmem_objnode_tree_path *pathp = path;
 380         struct tmem_objnode *slot = NULL;
 381         unsigned int height, shift;
 382         int offset;
 383
 384         BUG_ON(obj == NULL);
 385         ASSERT_SENTINEL(obj, OBJ);
 386         BUG_ON(obj->pool == NULL);
 387         ASSERT_SENTINEL(obj->pool, POOL);
 388         height = obj->objnode_tree_height;
 389         if (index > tmem_objnode_tree_h2max[height])
 390                 goto out;
 391         slot = obj->objnode_tree_root;
 392         if (height == 0 && obj->objnode_tree_root) {
 393                 obj->objnode_tree_root = NULL;
 394                 goto out;
 395         }
 396         shift = (height - 1) * OBJNODE_TREE_MAP_SHIFT;
 397         pathp->objnode = NULL;
 398         do {
 399                 if (slot == NULL)
 400                         goto out;
 401                 pathp++;
 402                 offset = (index >> shift) & OBJNODE_TREE_MAP_MASK;
 403                 pathp->offset = offset;
 404                 pathp->objnode = slot;
 405                 slot = slot->slots[offset];
 406                 shift -= OBJNODE_TREE_MAP_SHIFT;
 407                 height--;
 408         } while (height > 0);
 409         if (slot == NULL)
 410                 goto out;
 411         while (pathp->objnode) {
 412                 pathp->objnode->slots[pathp->offset] = NULL;
 413                 pathp->objnode->slots_in_use--;
 414                 if (pathp->objnode->slots_in_use) {
 415                         if (pathp->objnode == obj->objnode_tree_root) {
 416                                 while (obj->objnode_tree_height > 0 &&
 417                                   obj->objnode_tree_root->slots_in_use == 1 &&
 418                                   obj->objnode_tree_root->slots[0]) {
 419                                         struct tmem_objnode *to_free =
 420                                                 obj->objnode_tree_root;
 421
 422                                         obj->objnode_tree_root =
 423                                                         to_free->slots[0];
 424                                         obj->objnode_tree_height--;
 425                                         to_free->slots[0] = NULL;
 426                                         to_free->slots_in_use = 0;
 427                                         tmem_objnode_free(to_free);
 428                                 }
 429                         }
 430                         goto out;
 431                 }
 432                 tmem_objnode_free(pathp->objnode); /* 0 slots used, free it */
 433                 pathp--;
 434         }
 435         obj->objnode_tree_height = 0;
 436         obj->objnode_tree_root = NULL;
 437
 438 out:
 439         if (slot != NULL)
 440                 obj->pampd_count--;
 441         BUG_ON(obj->pampd_count < 0);
 442         return slot;
 443 }
 444
 445 /* recursively walk the objnode_tree destroying pampds and objnodes */
 446 static void tmem_objnode_node_destroy(struct tmem_obj *obj,
 447                                         struct tmem_objnode *objnode,
 448                                         unsigned int ht)
 449 {
 450         int i;
 451
 452         if (ht == 0)
 453                 return;
 454         for (i = 0; i < OBJNODE_TREE_MAP_SIZE; i++) {
 455                 if (objnode->slots[i]) {
 456                         if (ht == 1) {
 457                                 obj->pampd_count--;
 458                                 (*tmem_pamops.free)(objnode->slots[i],
 459                                                                 obj->pool);
 460                                 objnode->slots[i] = NULL;
 461                                 continue;
 462                         }
 463                         tmem_objnode_node_destroy(obj, objnode->slots[i], ht-1);
 464                         tmem_objnode_free(objnode->slots[i]);
 465                         objnode->slots[i] = NULL;
 466                 }
 467         }
 468 }
 469
 470 static void tmem_pampd_destroy_all_in_obj(struct tmem_obj *obj)
 471 {
 472         if (obj->objnode_tree_root == NULL)
 473                 return;
 474         if (obj->objnode_tree_height == 0) {
 475                 obj->pampd_count--;
 476                 (*tmem_pamops.free)(obj->objnode_tree_root, obj->pool);
 477         } else {
 478                 tmem_objnode_node_destroy(obj, obj->objnode_tree_root,
 479                                         obj->objnode_tree_height);
 480                 tmem_objnode_free(obj->objnode_tree_root);
 481                 obj->objnode_tree_height = 0;
 482         }
 483         obj->objnode_tree_root = NULL;
 484 }
 485
 486 /*
 487  * Tmem is operated on by a set of well-defined actions:
 488  * "put", "get", "flush", "flush_object", "new pool" and "destroy pool".
 489  * (The tmem ABI allows for subpages and exchanges but these operations
 490  * are not included in this implementation.)
 491  *
 492  * These "tmem core" operations are implemented in the following functions.
 493  */
 494
 495 /*
 496  * "Put" a page, e.g. copy a page from the kernel into newly allocated
 497  * PAM space (if such space is available).  Tmem_put is complicated by
 498  * a corner case: What if a page with matching handle already exists in
 499  * tmem?  To guarantee coherency, one of two actions is necessary: Either
 500  * the data for the page must be overwritten, or the page must be
 501  * "flushed" so that the data is not accessible to a subsequent "get".
 502  * Since these "duplicate puts" are relatively rare, this implementation
 503  * always flushes for simplicity.
 504  */
 505 int tmem_put(struct tmem_pool *pool, struct tmem_oid *oidp, uint32_t index,
 506                 struct page *page)
 507 {
 508         struct tmem_obj *obj = NULL, *objfound = NULL, *objnew = NULL;
 509         void *pampd = NULL, *pampd_del = NULL;
 510         int ret = -ENOMEM;
 511         bool ephemeral;
 512         struct tmem_hashbucket *hb;
 513
 514         ephemeral = is_ephemeral(pool);
 515         hb = &pool->hashbucket[tmem_oid_hash(oidp)];
 516         spin_lock(&hb->lock);
 517         obj = objfound = tmem_obj_find(hb, oidp);
 518         if (obj != NULL) {
 519                 pampd = tmem_pampd_lookup_in_obj(objfound, index);
 520                 if (pampd != NULL) {
 521                         /* if found, is a dup put, flush the old one */
 522                         pampd_del = tmem_pampd_delete_from_obj(obj, index);
 523                         BUG_ON(pampd_del != pampd);
 524                         (*tmem_pamops.free)(pampd, pool);
 525                         if (obj->pampd_count == 0) {
 526                                 objnew = obj;
 527                                 objfound = NULL;
 528                         }
 529                         pampd = NULL;
 530                 }
 531         } else {
 532                 obj = objnew = (*tmem_hostops.obj_alloc)(pool);
 533                 if (unlikely(obj == NULL)) {
 534                         ret = -ENOMEM;
 535                         goto out;
 536                 }
 537                 tmem_obj_init(obj, hb, pool, oidp);
 538         }
 539         BUG_ON(obj == NULL);
 540         BUG_ON(((objnew != obj) && (objfound != obj)) || (objnew == objfound));
 541         pampd = (*tmem_pamops.create)(obj->pool, &obj->oid, index, page);
 542         if (unlikely(pampd == NULL))
 543                 goto free;
 544         ret = tmem_pampd_add_to_obj(obj, index, pampd);
 545         if (unlikely(ret == -ENOMEM))
 546                 /* may have partially built objnode tree ("stump") */
 547                 goto delete_and_free;
 548         goto out;
 549
 550 delete_and_free:
 551         (void)tmem_pampd_delete_from_obj(obj, index);
 552 free:
 553         if (pampd)
 554                 (*tmem_pamops.free)(pampd, pool);
 555         if (objnew) {
 556                 tmem_obj_free(objnew, hb);
 557                 (*tmem_hostops.obj_free)(objnew, pool);
 558         }
 559 out:
 560         spin_unlock(&hb->lock);
 561         return ret;
 562 }
 563
 564 /*
 565  * "Get" a page, e.g. if one can be found, copy the tmem page with the
 566  * matching handle from PAM space to the kernel.  By tmem definition,
 567  * when a "get" is successful on an ephemeral page, the page is "flushed",
 568  * and when a "get" is successful on a persistent page, the page is retained
 569  * in tmem.  Note that to preserve
 570  * coherency, "get" can never be skipped if tmem contains the data.
 571  * That is, if a get is done with a certain handle and fails, any
 572  * subsequent "get" must also fail (unless of course there is a
 573  * "put" done with the same handle).
 574
 575  */
 576 int tmem_get(struct tmem_pool *pool, struct tmem_oid *oidp,
 577                                 uint32_t index, struct page *page)
 578 {
 579         struct tmem_obj *obj;
 580         void *pampd;
 581         bool ephemeral = is_ephemeral(pool);
 582         uint32_t ret = -1;
 583         struct tmem_hashbucket *hb;
 584
 585         hb = &pool->hashbucket[tmem_oid_hash(oidp)];
 586         spin_lock(&hb->lock);
 587         obj = tmem_obj_find(hb, oidp);
 588         if (obj == NULL)
 589                 goto out;
 590         ephemeral = is_ephemeral(pool);
 591         if (ephemeral)
 592                 pampd = tmem_pampd_delete_from_obj(obj, index);
 593         else
 594                 pampd = tmem_pampd_lookup_in_obj(obj, index);
 595         if (pampd == NULL)
 596                 goto out;
 597         ret = (*tmem_pamops.get_data)(page, pampd, pool);
 598         if (ret < 0)
 599                 goto out;
 600         if (ephemeral) {
 601                 (*tmem_pamops.free)(pampd, pool);
 602                 if (obj->pampd_count == 0) {
 603                         tmem_obj_free(obj, hb);
 604                         (*tmem_hostops.obj_free)(obj, pool);
 605                         obj = NULL;
 606                 }
 607         }
 608         ret = 0;
 609 out:
 610         spin_unlock(&hb->lock);
 611         return ret;
 612 }
 613
 614 /*
 615  * If a page in tmem matches the handle, "flush" this page from tmem such
 616  * that any subsequent "get" does not succeed (unless, of course, there
 617  * was another "put" with the same handle).
 618  */
 619 int tmem_flush_page(struct tmem_pool *pool,
 620                                 struct tmem_oid *oidp, uint32_t index)
 621 {
 622         struct tmem_obj *obj;
 623         void *pampd;
 624         int ret = -1;
 625         struct tmem_hashbucket *hb;
 626
 627         hb = &pool->hashbucket[tmem_oid_hash(oidp)];
 628         spin_lock(&hb->lock);
 629         obj = tmem_obj_find(hb, oidp);
 630         if (obj == NULL)
 631                 goto out;
 632         pampd = tmem_pampd_delete_from_obj(obj, index);
 633         if (pampd == NULL)
 634                 goto out;
 635         (*tmem_pamops.free)(pampd, pool);
 636         if (obj->pampd_count == 0) {
 637                 tmem_obj_free(obj, hb);
 638                 (*tmem_hostops.obj_free)(obj, pool);
 639         }
 640         ret = 0;
 641
 642 out:
 643         spin_unlock(&hb->lock);
 644         return ret;
 645 }
 646
 647 /*
 648  * "Flush" all pages in tmem matching this oid.
 649  */
 650 int tmem_flush_object(struct tmem_pool *pool, struct tmem_oid *oidp)
 651 {
 652         struct tmem_obj *obj;
 653         struct tmem_hashbucket *hb;
 654         int ret = -1;
 655
 656         hb = &pool->hashbucket[tmem_oid_hash(oidp)];
 657         spin_lock(&hb->lock);
 658         obj = tmem_obj_find(hb, oidp);
 659         if (obj == NULL)
 660                 goto out;
 661         tmem_pampd_destroy_all_in_obj(obj);
 662         tmem_obj_free(obj, hb);
 663         (*tmem_hostops.obj_free)(obj, pool);
 664         ret = 0;
 665
 666 out:
 667         spin_unlock(&hb->lock);
 668         return ret;
 669 }
 670
 671 /*
 672  * "Flush" all pages (and tmem_objs) from this tmem_pool and disable
 673  * all subsequent access to this tmem_pool.
 674  */
 675 int tmem_destroy_pool(struct tmem_pool *pool)
 676 {
 677         int ret = -1;
 678
 679         if (pool == NULL)
 680                 goto out;
 681         tmem_pool_flush(pool, 1);
 682         ret = 0;
 683 out:
 684         return ret;
 685 }
 686
 687 static LIST_HEAD(tmem_global_pool_list);
 688
 689 /*
 690  * Create a new tmem_pool with the provided flag and return
 691  * a pool id provided by the tmem host implementation.
 692  */
 693 void tmem_new_pool(struct tmem_pool *pool, uint32_t flags)
 694 {
 695         int persistent = flags & TMEM_POOL_PERSIST;
 696         int shared = flags & TMEM_POOL_SHARED;
 697         struct tmem_hashbucket *hb = &pool->hashbucket[0];
 698         int i;
 699
 700         for (i = 0; i < TMEM_HASH_BUCKETS; i++, hb++) {
 701                 hb->obj_rb_root = RB_ROOT;
 702                 spin_lock_init(&hb->lock);
 703         }
 704         INIT_LIST_HEAD(&pool->pool_list);
 705         atomic_set(&pool->obj_count, 0);
 706         SET_SENTINEL(pool, POOL);
 707         list_add_tail(&pool->pool_list, &tmem_global_pool_list);
 708         pool->persistent = persistent;
 709         pool->shared = shared;
 710 }