drivers/staging/lustre/lustre/fld/fld_cache.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2012, 2013, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/fld/fld_cache.c
  37  *
  38  * FLD (Fids Location Database)
  39  *
  40  * Author: Pravin Shelar <pravin.shelar@sun.com>
  41  * Author: Yury Umanets <umka@clusterfs.com>
  42  */
  43
  44 #define DEBUG_SUBSYSTEM S_FLD
  45
  46 #include "../../include/linux/libcfs/libcfs.h"
  47 #include <linux/module.h>
  48 #include <asm/div64.h>
  49
  50 #include "../include/obd.h"
  51 #include "../include/obd_class.h"
  52 #include "../include/lustre_ver.h"
  53 #include "../include/obd_support.h"
  54 #include "../include/lprocfs_status.h"
  55
  56 #include "../include/lustre_req_layout.h"
  57 #include "../include/lustre_fld.h"
  58 #include "fld_internal.h"
  59
  60 /**
  61  * create fld cache.
  62  */
  63 struct fld_cache *fld_cache_init(const char *name,
  64                                  int cache_size, int cache_threshold)
  65 {
  66         struct fld_cache *cache;
  67
  68         LASSERT(name);
  69         LASSERT(cache_threshold < cache_size);
  70
  71         cache = kzalloc(sizeof(*cache), GFP_NOFS);
  72         if (!cache)
  73                 return ERR_PTR(-ENOMEM);
  74
  75         INIT_LIST_HEAD(&cache->fci_entries_head);
  76         INIT_LIST_HEAD(&cache->fci_lru);
  77
  78         cache->fci_cache_count = 0;
  79         rwlock_init(&cache->fci_lock);
  80
  81         strlcpy(cache->fci_name, name,
  82                 sizeof(cache->fci_name));
  83
  84         cache->fci_cache_size = cache_size;
  85         cache->fci_threshold = cache_threshold;
  86
  87         /* Init fld cache info. */
  88         memset(&cache->fci_stat, 0, sizeof(cache->fci_stat));
  89
  90         CDEBUG(D_INFO, "%s: FLD cache - Size: %d, Threshold: %d\n",
  91                cache->fci_name, cache_size, cache_threshold);
  92
  93         return cache;
  94 }
  95
  96 /**
  97  * destroy fld cache.
  98  */
  99 void fld_cache_fini(struct fld_cache *cache)
 100 {
 101         __u64 pct;
 102
 103         LASSERT(cache);
 104         fld_cache_flush(cache);
 105
 106         if (cache->fci_stat.fst_count > 0) {
 107                 pct = cache->fci_stat.fst_cache * 100;
 108                 do_div(pct, cache->fci_stat.fst_count);
 109         } else {
 110                 pct = 0;
 111         }
 112
 113         CDEBUG(D_INFO, "FLD cache statistics (%s):\n", cache->fci_name);
 114         CDEBUG(D_INFO, "  Total reqs: %llu\n", cache->fci_stat.fst_count);
 115         CDEBUG(D_INFO, "  Cache reqs: %llu\n", cache->fci_stat.fst_cache);
 116         CDEBUG(D_INFO, "  Cache hits: %llu%%\n", pct);
 117
 118         kfree(cache);
 119 }
 120
 121 /**
 122  * delete given node from list.
 123  */
 124 static void fld_cache_entry_delete(struct fld_cache *cache,
 125                                    struct fld_cache_entry *node)
 126 {
 127         list_del(&node->fce_list);
 128         list_del(&node->fce_lru);
 129         cache->fci_cache_count--;
 130         kfree(node);
 131 }
 132
 133 /**
 134  * fix list by checking new entry with NEXT entry in order.
 135  */
 136 static void fld_fix_new_list(struct fld_cache *cache)
 137 {
 138         struct fld_cache_entry *f_curr;
 139         struct fld_cache_entry *f_next;
 140         struct lu_seq_range *c_range;
 141         struct lu_seq_range *n_range;
 142         struct list_head *head = &cache->fci_entries_head;
 143
 144 restart_fixup:
 145
 146         list_for_each_entry_safe(f_curr, f_next, head, fce_list) {
 147                 c_range = &f_curr->fce_range;
 148                 n_range = &f_next->fce_range;
 149
 150                 LASSERT(range_is_sane(c_range));
 151                 if (&f_next->fce_list == head)
 152                         break;
 153
 154                 if (c_range->lsr_flags != n_range->lsr_flags)
 155                         continue;
 156
 157                 LASSERTF(c_range->lsr_start <= n_range->lsr_start,
 158                          "cur lsr_start "DRANGE" next lsr_start "DRANGE"\n",
 159                          PRANGE(c_range), PRANGE(n_range));
 160
 161                 /* check merge possibility with next range */
 162                 if (c_range->lsr_end == n_range->lsr_start) {
 163                         if (c_range->lsr_index != n_range->lsr_index)
 164                                 continue;
 165                         n_range->lsr_start = c_range->lsr_start;
 166                         fld_cache_entry_delete(cache, f_curr);
 167                         continue;
 168                 }
 169
 170                 /* check if current range overlaps with next range. */
 171                 if (n_range->lsr_start < c_range->lsr_end) {
 172                         if (c_range->lsr_index == n_range->lsr_index) {
 173                                 n_range->lsr_start = c_range->lsr_start;
 174                                 n_range->lsr_end = max(c_range->lsr_end,
 175                                                        n_range->lsr_end);
 176                                 fld_cache_entry_delete(cache, f_curr);
 177                         } else {
 178                                 if (n_range->lsr_end <= c_range->lsr_end) {
 179                                         *n_range = *c_range;
 180                                         fld_cache_entry_delete(cache, f_curr);
 181                                 } else
 182                                         n_range->lsr_start = c_range->lsr_end;
 183                         }
 184
 185                         /* we could have overlap over next
 186                          * range too. better restart.
 187                          */
 188                         goto restart_fixup;
 189                 }
 190
 191                 /* kill duplicates */
 192                 if (c_range->lsr_start == n_range->lsr_start &&
 193                     c_range->lsr_end == n_range->lsr_end)
 194                         fld_cache_entry_delete(cache, f_curr);
 195         }
 196 }
 197
 198 /**
 199  * add node to fld cache
 200  */
 201 static inline void fld_cache_entry_add(struct fld_cache *cache,
 202                                        struct fld_cache_entry *f_new,
 203                                        struct list_head *pos)
 204 {
 205         list_add(&f_new->fce_list, pos);
 206         list_add(&f_new->fce_lru, &cache->fci_lru);
 207
 208         cache->fci_cache_count++;
 209         fld_fix_new_list(cache);
 210 }
 211
 212 /**
 213  * Check if cache needs to be shrunk. If so - do it.
 214  * Remove one entry in list and so on until cache is shrunk enough.
 215  */
 216 static int fld_cache_shrink(struct fld_cache *cache)
 217 {
 218         struct fld_cache_entry *flde;
 219         struct list_head *curr;
 220         int num = 0;
 221
 222         if (cache->fci_cache_count < cache->fci_cache_size)
 223                 return 0;
 224
 225         curr = cache->fci_lru.prev;
 226
 227         while (cache->fci_cache_count + cache->fci_threshold >
 228                cache->fci_cache_size && curr != &cache->fci_lru) {
 229                 flde = list_entry(curr, struct fld_cache_entry, fce_lru);
 230                 curr = curr->prev;
 231                 fld_cache_entry_delete(cache, flde);
 232                 num++;
 233         }
 234
 235         CDEBUG(D_INFO, "%s: FLD cache - Shrunk by %d entries\n",
 236                cache->fci_name, num);
 237
 238         return 0;
 239 }
 240
 241 /**
 242  * kill all fld cache entries.
 243  */
 244 void fld_cache_flush(struct fld_cache *cache)
 245 {
 246         write_lock(&cache->fci_lock);
 247         cache->fci_cache_size = 0;
 248         fld_cache_shrink(cache);
 249         write_unlock(&cache->fci_lock);
 250 }
 251
 252 /**
 253  * punch hole in existing range. divide this range and add new
 254  * entry accordingly.
 255  */
 256
 257 static void fld_cache_punch_hole(struct fld_cache *cache,
 258                                  struct fld_cache_entry *f_curr,
 259                                  struct fld_cache_entry *f_new)
 260 {
 261         const struct lu_seq_range *range = &f_new->fce_range;
 262         const u64 new_start  = range->lsr_start;
 263         const u64 new_end  = range->lsr_end;
 264         struct fld_cache_entry *fldt;
 265
 266         fldt = kzalloc(sizeof(*fldt), GFP_ATOMIC);
 267         if (!fldt) {
 268                 kfree(f_new);
 269                 /* overlap is not allowed, so dont mess up list. */
 270                 return;
 271         }
 272         /*  break f_curr RANGE into three RANGES:
 273          *      f_curr, f_new , fldt
 274          */
 275
 276         /* f_new = *range */
 277
 278         /* fldt */
 279         fldt->fce_range.lsr_start = new_end;
 280         fldt->fce_range.lsr_end = f_curr->fce_range.lsr_end;
 281         fldt->fce_range.lsr_index = f_curr->fce_range.lsr_index;
 282
 283         /* f_curr */
 284         f_curr->fce_range.lsr_end = new_start;
 285
 286         /* add these two entries to list */
 287         fld_cache_entry_add(cache, f_new, &f_curr->fce_list);
 288         fld_cache_entry_add(cache, fldt, &f_new->fce_list);
 289
 290         /* no need to fixup */
 291 }
 292
 293 /**
 294  * handle range overlap in fld cache.
 295  */
 296 static void fld_cache_overlap_handle(struct fld_cache *cache,
 297                                      struct fld_cache_entry *f_curr,
 298                                      struct fld_cache_entry *f_new)
 299 {
 300         const struct lu_seq_range *range = &f_new->fce_range;
 301         const u64 new_start  = range->lsr_start;
 302         const u64 new_end  = range->lsr_end;
 303         const u32 mdt = range->lsr_index;
 304
 305         /* this is overlap case, these case are checking overlapping with
 306          * prev range only. fixup will handle overlapping with next range.
 307          */
 308
 309         if (f_curr->fce_range.lsr_index == mdt) {
 310                 f_curr->fce_range.lsr_start = min(f_curr->fce_range.lsr_start,
 311                                                   new_start);
 312
 313                 f_curr->fce_range.lsr_end = max(f_curr->fce_range.lsr_end,
 314                                                 new_end);
 315
 316                 kfree(f_new);
 317                 fld_fix_new_list(cache);
 318
 319         } else if (new_start <= f_curr->fce_range.lsr_start &&
 320                         f_curr->fce_range.lsr_end <= new_end) {
 321                 /* case 1: new range completely overshadowed existing range.
 322                  *       e.g. whole range migrated. update fld cache entry
 323                  */
 324
 325                 f_curr->fce_range = *range;
 326                 kfree(f_new);
 327                 fld_fix_new_list(cache);
 328
 329         } else if (f_curr->fce_range.lsr_start < new_start &&
 330                         new_end < f_curr->fce_range.lsr_end) {
 331                 /* case 2: new range fit within existing range. */
 332
 333                 fld_cache_punch_hole(cache, f_curr, f_new);
 334
 335         } else  if (new_end <= f_curr->fce_range.lsr_end) {
 336                 /* case 3: overlap:
 337                  *       [new_start [c_start  new_end)  c_end)
 338                  */
 339
 340                 LASSERT(new_start <= f_curr->fce_range.lsr_start);
 341
 342                 f_curr->fce_range.lsr_start = new_end;
 343                 fld_cache_entry_add(cache, f_new, f_curr->fce_list.prev);
 344
 345         } else if (f_curr->fce_range.lsr_start <= new_start) {
 346                 /* case 4: overlap:
 347                  *       [c_start [new_start c_end) new_end)
 348                  */
 349
 350                 LASSERT(f_curr->fce_range.lsr_end <= new_end);
 351
 352                 f_curr->fce_range.lsr_end = new_start;
 353                 fld_cache_entry_add(cache, f_new, &f_curr->fce_list);
 354         } else
 355                 CERROR("NEW range ="DRANGE" curr = "DRANGE"\n",
 356                        PRANGE(range), PRANGE(&f_curr->fce_range));
 357 }
 358
 359 struct fld_cache_entry
 360 *fld_cache_entry_create(const struct lu_seq_range *range)
 361 {
 362         struct fld_cache_entry *f_new;
 363
 364         LASSERT(range_is_sane(range));
 365
 366         f_new = kzalloc(sizeof(*f_new), GFP_NOFS);
 367         if (!f_new)
 368                 return ERR_PTR(-ENOMEM);
 369
 370         f_new->fce_range = *range;
 371         return f_new;
 372 }
 373
 374 /**
 375  * Insert FLD entry in FLD cache.
 376  *
 377  * This function handles all cases of merging and breaking up of
 378  * ranges.
 379  */
 380 static int fld_cache_insert_nolock(struct fld_cache *cache,
 381                                    struct fld_cache_entry *f_new)
 382 {
 383         struct fld_cache_entry *f_curr;
 384         struct fld_cache_entry *n;
 385         struct list_head *head;
 386         struct list_head *prev = NULL;
 387         const u64 new_start  = f_new->fce_range.lsr_start;
 388         const u64 new_end  = f_new->fce_range.lsr_end;
 389         __u32 new_flags  = f_new->fce_range.lsr_flags;
 390
 391         /*
 392          * Duplicate entries are eliminated in insert op.
 393          * So we don't need to search new entry before starting
 394          * insertion loop.
 395          */
 396
 397         if (!cache->fci_no_shrink)
 398                 fld_cache_shrink(cache);
 399
 400         head = &cache->fci_entries_head;
 401
 402         list_for_each_entry_safe(f_curr, n, head, fce_list) {
 403                 /* add list if next is end of list */
 404                 if (new_end < f_curr->fce_range.lsr_start ||
 405                     (new_end == f_curr->fce_range.lsr_start &&
 406                      new_flags != f_curr->fce_range.lsr_flags))
 407                         break;
 408
 409                 prev = &f_curr->fce_list;
 410                 /* check if this range is to left of new range. */
 411                 if (new_start < f_curr->fce_range.lsr_end &&
 412                     new_flags == f_curr->fce_range.lsr_flags) {
 413                         fld_cache_overlap_handle(cache, f_curr, f_new);
 414                         goto out;
 415                 }
 416         }
 417
 418         if (!prev)
 419                 prev = head;
 420
 421         CDEBUG(D_INFO, "insert range "DRANGE"\n", PRANGE(&f_new->fce_range));
 422         /* Add new entry to cache and lru list. */
 423         fld_cache_entry_add(cache, f_new, prev);
 424 out:
 425         return 0;
 426 }
 427
 428 int fld_cache_insert(struct fld_cache *cache,
 429                      const struct lu_seq_range *range)
 430 {
 431         struct fld_cache_entry  *flde;
 432         int rc;
 433
 434         flde = fld_cache_entry_create(range);
 435         if (IS_ERR(flde))
 436                 return PTR_ERR(flde);
 437
 438         write_lock(&cache->fci_lock);
 439         rc = fld_cache_insert_nolock(cache, flde);
 440         write_unlock(&cache->fci_lock);
 441         if (rc)
 442                 kfree(flde);
 443
 444         return rc;
 445 }
 446
 447 /**
 448  * Delete FLD entry in FLD cache.
 449  *
 450  */
 451
 452 struct fld_cache_entry
 453 *fld_cache_entry_lookup_nolock(struct fld_cache *cache,
 454                               struct lu_seq_range *range)
 455 {
 456         struct fld_cache_entry *flde;
 457         struct fld_cache_entry *got = NULL;
 458         struct list_head *head;
 459
 460         head = &cache->fci_entries_head;
 461         list_for_each_entry(flde, head, fce_list) {
 462                 if (range->lsr_start == flde->fce_range.lsr_start ||
 463                     (range->lsr_end == flde->fce_range.lsr_end &&
 464                      range->lsr_flags == flde->fce_range.lsr_flags)) {
 465                         got = flde;
 466                         break;
 467                 }
 468         }
 469
 470         return got;
 471 }
 472
 473 /**
 474  * lookup \a seq sequence for range in fld cache.
 475  */
 476 struct fld_cache_entry
 477 *fld_cache_entry_lookup(struct fld_cache *cache, struct lu_seq_range *range)
 478 {
 479         struct fld_cache_entry *got = NULL;
 480
 481         read_lock(&cache->fci_lock);
 482         got = fld_cache_entry_lookup_nolock(cache, range);
 483         read_unlock(&cache->fci_lock);
 484         return got;
 485 }
 486
 487 /**
 488  * lookup \a seq sequence for range in fld cache.
 489  */
 490 int fld_cache_lookup(struct fld_cache *cache,
 491                      const u64 seq, struct lu_seq_range *range)
 492 {
 493         struct fld_cache_entry *flde;
 494         struct fld_cache_entry *prev = NULL;
 495         struct list_head *head;
 496
 497         read_lock(&cache->fci_lock);
 498         head = &cache->fci_entries_head;
 499
 500         cache->fci_stat.fst_count++;
 501         list_for_each_entry(flde, head, fce_list) {
 502                 if (flde->fce_range.lsr_start > seq) {
 503                         if (prev)
 504                                 *range = prev->fce_range;
 505                         break;
 506                 }
 507
 508                 prev = flde;
 509                 if (range_within(&flde->fce_range, seq)) {
 510                         *range = flde->fce_range;
 511
 512                         cache->fci_stat.fst_cache++;
 513                         read_unlock(&cache->fci_lock);
 514                         return 0;
 515                 }
 516         }
 517         read_unlock(&cache->fci_lock);
 518         return -ENOENT;
 519 }