drivers/staging/lustre/lustre/ptlrpc/sec_bulk.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) 2011, 2012, 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/ptlrpc/sec_bulk.c
  37  *
  38  * Author: Eric Mei <ericm@clusterfs.com>
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_SEC
  42
  43 #include <linux/libcfs/libcfs.h>
  44 #include <linux/crypto.h>
  45
  46 #include <obd.h>
  47 #include <obd_cksum.h>
  48 #include <obd_class.h>
  49 #include <obd_support.h>
  50 #include <lustre_net.h>
  51 #include <lustre_import.h>
  52 #include <lustre_dlm.h>
  53 #include <lustre_sec.h>
  54
  55 #include "ptlrpc_internal.h"
  56
  57 /****************************************
  58  * bulk encryption page pools      *
  59  ****************************************/
  60
  61
  62 #define PTRS_PER_PAGE   (PAGE_CACHE_SIZE / sizeof(void *))
  63 #define PAGES_PER_POOL  (PTRS_PER_PAGE)
  64
  65 #define IDLE_IDX_MAX        (100)
  66 #define IDLE_IDX_WEIGHT  (3)
  67
  68 #define CACHE_QUIESCENT_PERIOD  (20)
  69
  70 static struct ptlrpc_enc_page_pool {
  71         /*
  72          * constants
  73          */
  74         unsigned long    epp_max_pages;   /* maximum pages can hold, const */
  75         unsigned int     epp_max_pools;   /* number of pools, const */
  76
  77         /*
  78          * wait queue in case of not enough free pages.
  79          */
  80         wait_queue_head_t      epp_waitq;       /* waiting threads */
  81         unsigned int     epp_waitqlen;    /* wait queue length */
  82         unsigned long    epp_pages_short; /* # of pages wanted of in-q users */
  83         unsigned int     epp_growing:1;   /* during adding pages */
  84
  85         /*
  86          * indicating how idle the pools are, from 0 to MAX_IDLE_IDX
  87          * this is counted based on each time when getting pages from
  88          * the pools, not based on time. which means in case that system
  89          * is idled for a while but the idle_idx might still be low if no
  90          * activities happened in the pools.
  91          */
  92         unsigned long    epp_idle_idx;
  93
  94         /* last shrink time due to mem tight */
  95         long         epp_last_shrink;
  96         long         epp_last_access;
  97
  98         /*
  99          * in-pool pages bookkeeping
 100          */
 101         spinlock_t       epp_lock;         /* protect following fields */
 102         unsigned long    epp_total_pages; /* total pages in pools */
 103         unsigned long    epp_free_pages;  /* current pages available */
 104
 105         /*
 106          * statistics
 107          */
 108         unsigned long    epp_st_max_pages;      /* # of pages ever reached */
 109         unsigned int     epp_st_grows;    /* # of grows */
 110         unsigned int     epp_st_grow_fails;     /* # of add pages failures */
 111         unsigned int     epp_st_shrinks;        /* # of shrinks */
 112         unsigned long    epp_st_access;  /* # of access */
 113         unsigned long    epp_st_missings;       /* # of cache missing */
 114         unsigned long    epp_st_lowfree;        /* lowest free pages reached */
 115         unsigned int     epp_st_max_wqlen;      /* highest waitqueue length */
 116         cfs_time_t       epp_st_max_wait;       /* in jeffies */
 117         /*
 118          * pointers to pools
 119          */
 120         struct page    ***epp_pools;
 121 } page_pools;
 122
 123 /*
 124  * memory shrinker
 125  */
 126 const int pools_shrinker_seeks = DEFAULT_SEEKS;
 127 static struct shrinker *pools_shrinker = NULL;
 128
 129
 130 /*
 131  * /proc/fs/lustre/sptlrpc/encrypt_page_pools
 132  */
 133 int sptlrpc_proc_enc_pool_seq_show(struct seq_file *m, void *v)
 134 {
 135         int     rc;
 136
 137         spin_lock(&page_pools.epp_lock);
 138
 139         rc = seq_printf(m,
 140                       "physical pages:    %lu\n"
 141                       "pages per pool:    %lu\n"
 142                       "max pages:              %lu\n"
 143                       "max pools:              %u\n"
 144                       "total pages:          %lu\n"
 145                       "total free:            %lu\n"
 146                       "idle index:            %lu/100\n"
 147                       "last shrink:          %lds\n"
 148                       "last access:          %lds\n"
 149                       "max pages reached:       %lu\n"
 150                       "grows:              %u\n"
 151                       "grows failure:      %u\n"
 152                       "shrinks:          %u\n"
 153                       "cache access:        %lu\n"
 154                       "cache missing:      %lu\n"
 155                       "low free mark:      %lu\n"
 156                       "max waitqueue depth:     %u\n"
 157                       "max wait time:      "CFS_TIME_T"/%u\n"
 158                       ,
 159                       totalram_pages,
 160                       PAGES_PER_POOL,
 161                       page_pools.epp_max_pages,
 162                       page_pools.epp_max_pools,
 163                       page_pools.epp_total_pages,
 164                       page_pools.epp_free_pages,
 165                       page_pools.epp_idle_idx,
 166                       cfs_time_current_sec() - page_pools.epp_last_shrink,
 167                       cfs_time_current_sec() - page_pools.epp_last_access,
 168                       page_pools.epp_st_max_pages,
 169                       page_pools.epp_st_grows,
 170                       page_pools.epp_st_grow_fails,
 171                       page_pools.epp_st_shrinks,
 172                       page_pools.epp_st_access,
 173                       page_pools.epp_st_missings,
 174                       page_pools.epp_st_lowfree,
 175                       page_pools.epp_st_max_wqlen,
 176                       page_pools.epp_st_max_wait, HZ
 177                      );
 178
 179         spin_unlock(&page_pools.epp_lock);
 180         return rc;
 181 }
 182
 183 static void enc_pools_release_free_pages(long npages)
 184 {
 185         int     p_idx, g_idx;
 186         int     p_idx_max1, p_idx_max2;
 187
 188         LASSERT(npages > 0);
 189         LASSERT(npages <= page_pools.epp_free_pages);
 190         LASSERT(page_pools.epp_free_pages <= page_pools.epp_total_pages);
 191
 192         /* max pool index before the release */
 193         p_idx_max2 = (page_pools.epp_total_pages - 1) / PAGES_PER_POOL;
 194
 195         page_pools.epp_free_pages -= npages;
 196         page_pools.epp_total_pages -= npages;
 197
 198         /* max pool index after the release */
 199         p_idx_max1 = page_pools.epp_total_pages == 0 ? -1 :
 200                      ((page_pools.epp_total_pages - 1) / PAGES_PER_POOL);
 201
 202         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 203         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 204         LASSERT(page_pools.epp_pools[p_idx]);
 205
 206         while (npages--) {
 207                 LASSERT(page_pools.epp_pools[p_idx]);
 208                 LASSERT(page_pools.epp_pools[p_idx][g_idx] != NULL);
 209
 210                 __free_page(page_pools.epp_pools[p_idx][g_idx]);
 211                 page_pools.epp_pools[p_idx][g_idx] = NULL;
 212
 213                 if (++g_idx == PAGES_PER_POOL) {
 214                         p_idx++;
 215                         g_idx = 0;
 216                 }
 217         };
 218
 219         /* free unused pools */
 220         while (p_idx_max1 < p_idx_max2) {
 221                 LASSERT(page_pools.epp_pools[p_idx_max2]);
 222                 OBD_FREE(page_pools.epp_pools[p_idx_max2], PAGE_CACHE_SIZE);
 223                 page_pools.epp_pools[p_idx_max2] = NULL;
 224                 p_idx_max2--;
 225         }
 226 }
 227
 228 /*
 229  * could be called frequently for query (@nr_to_scan == 0).
 230  * we try to keep at least PTLRPC_MAX_BRW_PAGES pages in the pool.
 231  */
 232 static int enc_pools_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
 233 {
 234         if (unlikely(shrink_param(sc, nr_to_scan) != 0)) {
 235                 spin_lock(&page_pools.epp_lock);
 236                 shrink_param(sc, nr_to_scan) = min_t(unsigned long,
 237                                                    shrink_param(sc, nr_to_scan),
 238                                                    page_pools.epp_free_pages -
 239                                                    PTLRPC_MAX_BRW_PAGES);
 240                 if (shrink_param(sc, nr_to_scan) > 0) {
 241                         enc_pools_release_free_pages(shrink_param(sc,
 242                                                                   nr_to_scan));
 243                         CDEBUG(D_SEC, "released %ld pages, %ld left\n",
 244                                (long)shrink_param(sc, nr_to_scan),
 245                                page_pools.epp_free_pages);
 246
 247                         page_pools.epp_st_shrinks++;
 248                         page_pools.epp_last_shrink = cfs_time_current_sec();
 249                 }
 250                 spin_unlock(&page_pools.epp_lock);
 251         }
 252
 253         /*
 254          * if no pool access for a long time, we consider it's fully idle.
 255          * a little race here is fine.
 256          */
 257         if (unlikely(cfs_time_current_sec() - page_pools.epp_last_access >
 258                      CACHE_QUIESCENT_PERIOD)) {
 259                 spin_lock(&page_pools.epp_lock);
 260                 page_pools.epp_idle_idx = IDLE_IDX_MAX;
 261                 spin_unlock(&page_pools.epp_lock);
 262         }
 263
 264         LASSERT(page_pools.epp_idle_idx <= IDLE_IDX_MAX);
 265         return max((int)page_pools.epp_free_pages - PTLRPC_MAX_BRW_PAGES, 0) *
 266                 (IDLE_IDX_MAX - page_pools.epp_idle_idx) / IDLE_IDX_MAX;
 267 }
 268
 269 static inline
 270 int npages_to_npools(unsigned long npages)
 271 {
 272         return (int) ((npages + PAGES_PER_POOL - 1) / PAGES_PER_POOL);
 273 }
 274
 275 /*
 276  * return how many pages cleaned up.
 277  */
 278 static unsigned long enc_pools_cleanup(struct page ***pools, int npools)
 279 {
 280         unsigned long cleaned = 0;
 281         int        i, j;
 282
 283         for (i = 0; i < npools; i++) {
 284                 if (pools[i]) {
 285                         for (j = 0; j < PAGES_PER_POOL; j++) {
 286                                 if (pools[i][j]) {
 287                                         __free_page(pools[i][j]);
 288                                         cleaned++;
 289                                 }
 290                         }
 291                         OBD_FREE(pools[i], PAGE_CACHE_SIZE);
 292                         pools[i] = NULL;
 293                 }
 294         }
 295
 296         return cleaned;
 297 }
 298
 299 /*
 300  * merge @npools pointed by @pools which contains @npages new pages
 301  * into current pools.
 302  *
 303  * we have options to avoid most memory copy with some tricks. but we choose
 304  * the simplest way to avoid complexity. It's not frequently called.
 305  */
 306 static void enc_pools_insert(struct page ***pools, int npools, int npages)
 307 {
 308         int     freeslot;
 309         int     op_idx, np_idx, og_idx, ng_idx;
 310         int     cur_npools, end_npools;
 311
 312         LASSERT(npages > 0);
 313         LASSERT(page_pools.epp_total_pages+npages <= page_pools.epp_max_pages);
 314         LASSERT(npages_to_npools(npages) == npools);
 315         LASSERT(page_pools.epp_growing);
 316
 317         spin_lock(&page_pools.epp_lock);
 318
 319         /*
 320          * (1) fill all the free slots of current pools.
 321          */
 322         /* free slots are those left by rent pages, and the extra ones with
 323          * index >= total_pages, locate at the tail of last pool. */
 324         freeslot = page_pools.epp_total_pages % PAGES_PER_POOL;
 325         if (freeslot != 0)
 326                 freeslot = PAGES_PER_POOL - freeslot;
 327         freeslot += page_pools.epp_total_pages - page_pools.epp_free_pages;
 328
 329         op_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 330         og_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 331         np_idx = npools - 1;
 332         ng_idx = (npages - 1) % PAGES_PER_POOL;
 333
 334         while (freeslot) {
 335                 LASSERT(page_pools.epp_pools[op_idx][og_idx] == NULL);
 336                 LASSERT(pools[np_idx][ng_idx] != NULL);
 337
 338                 page_pools.epp_pools[op_idx][og_idx] = pools[np_idx][ng_idx];
 339                 pools[np_idx][ng_idx] = NULL;
 340
 341                 freeslot--;
 342
 343                 if (++og_idx == PAGES_PER_POOL) {
 344                         op_idx++;
 345                         og_idx = 0;
 346                 }
 347                 if (--ng_idx < 0) {
 348                         if (np_idx == 0)
 349                                 break;
 350                         np_idx--;
 351                         ng_idx = PAGES_PER_POOL - 1;
 352                 }
 353         }
 354
 355         /*
 356          * (2) add pools if needed.
 357          */
 358         cur_npools = (page_pools.epp_total_pages + PAGES_PER_POOL - 1) /
 359                      PAGES_PER_POOL;
 360         end_npools = (page_pools.epp_total_pages + npages + PAGES_PER_POOL -1) /
 361                      PAGES_PER_POOL;
 362         LASSERT(end_npools <= page_pools.epp_max_pools);
 363
 364         np_idx = 0;
 365         while (cur_npools < end_npools) {
 366                 LASSERT(page_pools.epp_pools[cur_npools] == NULL);
 367                 LASSERT(np_idx < npools);
 368                 LASSERT(pools[np_idx] != NULL);
 369
 370                 page_pools.epp_pools[cur_npools++] = pools[np_idx];
 371                 pools[np_idx++] = NULL;
 372         }
 373
 374         page_pools.epp_total_pages += npages;
 375         page_pools.epp_free_pages += npages;
 376         page_pools.epp_st_lowfree = page_pools.epp_free_pages;
 377
 378         if (page_pools.epp_total_pages > page_pools.epp_st_max_pages)
 379                 page_pools.epp_st_max_pages = page_pools.epp_total_pages;
 380
 381         CDEBUG(D_SEC, "add %d pages to total %lu\n", npages,
 382                page_pools.epp_total_pages);
 383
 384         spin_unlock(&page_pools.epp_lock);
 385 }
 386
 387 static int enc_pools_add_pages(int npages)
 388 {
 389         static DEFINE_MUTEX(add_pages_mutex);
 390         struct page   ***pools;
 391         int          npools, alloced = 0;
 392         int          i, j, rc = -ENOMEM;
 393
 394         if (npages < PTLRPC_MAX_BRW_PAGES)
 395                 npages = PTLRPC_MAX_BRW_PAGES;
 396
 397         mutex_lock(&add_pages_mutex);
 398
 399         if (npages + page_pools.epp_total_pages > page_pools.epp_max_pages)
 400                 npages = page_pools.epp_max_pages - page_pools.epp_total_pages;
 401         LASSERT(npages > 0);
 402
 403         page_pools.epp_st_grows++;
 404
 405         npools = npages_to_npools(npages);
 406         OBD_ALLOC(pools, npools * sizeof(*pools));
 407         if (pools == NULL)
 408                 goto out;
 409
 410         for (i = 0; i < npools; i++) {
 411                 OBD_ALLOC(pools[i], PAGE_CACHE_SIZE);
 412                 if (pools[i] == NULL)
 413                         goto out_pools;
 414
 415                 for (j = 0; j < PAGES_PER_POOL && alloced < npages; j++) {
 416                         pools[i][j] = alloc_page(__GFP_IO |
 417                                                      __GFP_HIGHMEM);
 418                         if (pools[i][j] == NULL)
 419                                 goto out_pools;
 420
 421                         alloced++;
 422                 }
 423         }
 424         LASSERT(alloced == npages);
 425
 426         enc_pools_insert(pools, npools, npages);
 427         CDEBUG(D_SEC, "added %d pages into pools\n", npages);
 428         rc = 0;
 429
 430 out_pools:
 431         enc_pools_cleanup(pools, npools);
 432         OBD_FREE(pools, npools * sizeof(*pools));
 433 out:
 434         if (rc) {
 435                 page_pools.epp_st_grow_fails++;
 436                 CERROR("Failed to allocate %d enc pages\n", npages);
 437         }
 438
 439         mutex_unlock(&add_pages_mutex);
 440         return rc;
 441 }
 442
 443 static inline void enc_pools_wakeup(void)
 444 {
 445         LASSERT(spin_is_locked(&page_pools.epp_lock));
 446         LASSERT(page_pools.epp_waitqlen >= 0);
 447
 448         if (unlikely(page_pools.epp_waitqlen)) {
 449                 LASSERT(waitqueue_active(&page_pools.epp_waitq));
 450                 wake_up_all(&page_pools.epp_waitq);
 451         }
 452 }
 453
 454 static int enc_pools_should_grow(int page_needed, long now)
 455 {
 456         /* don't grow if someone else is growing the pools right now,
 457          * or the pools has reached its full capacity
 458          */
 459         if (page_pools.epp_growing ||
 460             page_pools.epp_total_pages == page_pools.epp_max_pages)
 461                 return 0;
 462
 463         /* if total pages is not enough, we need to grow */
 464         if (page_pools.epp_total_pages < page_needed)
 465                 return 1;
 466
 467         /*
 468          * we wanted to return 0 here if there was a shrink just happened
 469          * moment ago, but this may cause deadlock if both client and ost
 470          * live on single node.
 471          */
 472 #if 0
 473         if (now - page_pools.epp_last_shrink < 2)
 474                 return 0;
 475 #endif
 476
 477         /*
 478          * here we perhaps need consider other factors like wait queue
 479          * length, idle index, etc. ?
 480          */
 481
 482         /* grow the pools in any other cases */
 483         return 1;
 484 }
 485
 486 /*
 487  * we allocate the requested pages atomically.
 488  */
 489 int sptlrpc_enc_pool_get_pages(struct ptlrpc_bulk_desc *desc)
 490 {
 491         wait_queue_t  waitlink;
 492         unsigned long   this_idle = -1;
 493         cfs_time_t      tick = 0;
 494         long        now;
 495         int          p_idx, g_idx;
 496         int          i;
 497
 498         LASSERT(desc->bd_iov_count > 0);
 499         LASSERT(desc->bd_iov_count <= page_pools.epp_max_pages);
 500
 501         /* resent bulk, enc iov might have been allocated previously */
 502         if (desc->bd_enc_iov != NULL)
 503                 return 0;
 504
 505         OBD_ALLOC(desc->bd_enc_iov,
 506                   desc->bd_iov_count * sizeof(*desc->bd_enc_iov));
 507         if (desc->bd_enc_iov == NULL)
 508                 return -ENOMEM;
 509
 510         spin_lock(&page_pools.epp_lock);
 511
 512         page_pools.epp_st_access++;
 513 again:
 514         if (unlikely(page_pools.epp_free_pages < desc->bd_iov_count)) {
 515                 if (tick == 0)
 516                         tick = cfs_time_current();
 517
 518                 now = cfs_time_current_sec();
 519
 520                 page_pools.epp_st_missings++;
 521                 page_pools.epp_pages_short += desc->bd_iov_count;
 522
 523                 if (enc_pools_should_grow(desc->bd_iov_count, now)) {
 524                         page_pools.epp_growing = 1;
 525
 526                         spin_unlock(&page_pools.epp_lock);
 527                         enc_pools_add_pages(page_pools.epp_pages_short / 2);
 528                         spin_lock(&page_pools.epp_lock);
 529
 530                         page_pools.epp_growing = 0;
 531
 532                         enc_pools_wakeup();
 533                 } else {
 534                         if (++page_pools.epp_waitqlen >
 535                             page_pools.epp_st_max_wqlen)
 536                                 page_pools.epp_st_max_wqlen =
 537                                                 page_pools.epp_waitqlen;
 538
 539                         set_current_state(TASK_UNINTERRUPTIBLE);
 540                         init_waitqueue_entry_current(&waitlink);
 541                         add_wait_queue(&page_pools.epp_waitq, &waitlink);
 542
 543                         spin_unlock(&page_pools.epp_lock);
 544                         waitq_wait(&waitlink, TASK_UNINTERRUPTIBLE);
 545                         remove_wait_queue(&page_pools.epp_waitq, &waitlink);
 546                         LASSERT(page_pools.epp_waitqlen > 0);
 547                         spin_lock(&page_pools.epp_lock);
 548                         page_pools.epp_waitqlen--;
 549                 }
 550
 551                 LASSERT(page_pools.epp_pages_short >= desc->bd_iov_count);
 552                 page_pools.epp_pages_short -= desc->bd_iov_count;
 553
 554                 this_idle = 0;
 555                 goto again;
 556         }
 557
 558         /* record max wait time */
 559         if (unlikely(tick != 0)) {
 560                 tick = cfs_time_current() - tick;
 561                 if (tick > page_pools.epp_st_max_wait)
 562                         page_pools.epp_st_max_wait = tick;
 563         }
 564
 565         /* proceed with rest of allocation */
 566         page_pools.epp_free_pages -= desc->bd_iov_count;
 567
 568         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 569         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 570
 571         for (i = 0; i < desc->bd_iov_count; i++) {
 572                 LASSERT(page_pools.epp_pools[p_idx][g_idx] != NULL);
 573                 desc->bd_enc_iov[i].kiov_page =
 574                                         page_pools.epp_pools[p_idx][g_idx];
 575                 page_pools.epp_pools[p_idx][g_idx] = NULL;
 576
 577                 if (++g_idx == PAGES_PER_POOL) {
 578                         p_idx++;
 579                         g_idx = 0;
 580                 }
 581         }
 582
 583         if (page_pools.epp_free_pages < page_pools.epp_st_lowfree)
 584                 page_pools.epp_st_lowfree = page_pools.epp_free_pages;
 585
 586         /*
 587          * new idle index = (old * weight + new) / (weight + 1)
 588          */
 589         if (this_idle == -1) {
 590                 this_idle = page_pools.epp_free_pages * IDLE_IDX_MAX /
 591                             page_pools.epp_total_pages;
 592         }
 593         page_pools.epp_idle_idx = (page_pools.epp_idle_idx * IDLE_IDX_WEIGHT +
 594                                    this_idle) /
 595                                   (IDLE_IDX_WEIGHT + 1);
 596
 597         page_pools.epp_last_access = cfs_time_current_sec();
 598
 599         spin_unlock(&page_pools.epp_lock);
 600         return 0;
 601 }
 602 EXPORT_SYMBOL(sptlrpc_enc_pool_get_pages);
 603
 604 void sptlrpc_enc_pool_put_pages(struct ptlrpc_bulk_desc *desc)
 605 {
 606         int     p_idx, g_idx;
 607         int     i;
 608
 609         if (desc->bd_enc_iov == NULL)
 610                 return;
 611
 612         LASSERT(desc->bd_iov_count > 0);
 613
 614         spin_lock(&page_pools.epp_lock);
 615
 616         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 617         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 618
 619         LASSERT(page_pools.epp_free_pages + desc->bd_iov_count <=
 620                 page_pools.epp_total_pages);
 621         LASSERT(page_pools.epp_pools[p_idx]);
 622
 623         for (i = 0; i < desc->bd_iov_count; i++) {
 624                 LASSERT(desc->bd_enc_iov[i].kiov_page != NULL);
 625                 LASSERT(g_idx != 0 || page_pools.epp_pools[p_idx]);
 626                 LASSERT(page_pools.epp_pools[p_idx][g_idx] == NULL);
 627
 628                 page_pools.epp_pools[p_idx][g_idx] =
 629                                         desc->bd_enc_iov[i].kiov_page;
 630
 631                 if (++g_idx == PAGES_PER_POOL) {
 632                         p_idx++;
 633                         g_idx = 0;
 634                 }
 635         }
 636
 637         page_pools.epp_free_pages += desc->bd_iov_count;
 638
 639         enc_pools_wakeup();
 640
 641         spin_unlock(&page_pools.epp_lock);
 642
 643         OBD_FREE(desc->bd_enc_iov,
 644                  desc->bd_iov_count * sizeof(*desc->bd_enc_iov));
 645         desc->bd_enc_iov = NULL;
 646 }
 647 EXPORT_SYMBOL(sptlrpc_enc_pool_put_pages);
 648
 649 /*
 650  * we don't do much stuff for add_user/del_user anymore, except adding some
 651  * initial pages in add_user() if current pools are empty, rest would be
 652  * handled by the pools's self-adaption.
 653  */
 654 int sptlrpc_enc_pool_add_user(void)
 655 {
 656         int     need_grow = 0;
 657
 658         spin_lock(&page_pools.epp_lock);
 659         if (page_pools.epp_growing == 0 && page_pools.epp_total_pages == 0) {
 660                 page_pools.epp_growing = 1;
 661                 need_grow = 1;
 662         }
 663         spin_unlock(&page_pools.epp_lock);
 664
 665         if (need_grow) {
 666                 enc_pools_add_pages(PTLRPC_MAX_BRW_PAGES +
 667                                     PTLRPC_MAX_BRW_PAGES);
 668
 669                 spin_lock(&page_pools.epp_lock);
 670                 page_pools.epp_growing = 0;
 671                 enc_pools_wakeup();
 672                 spin_unlock(&page_pools.epp_lock);
 673         }
 674         return 0;
 675 }
 676 EXPORT_SYMBOL(sptlrpc_enc_pool_add_user);
 677
 678 int sptlrpc_enc_pool_del_user(void)
 679 {
 680         return 0;
 681 }
 682 EXPORT_SYMBOL(sptlrpc_enc_pool_del_user);
 683
 684 static inline void enc_pools_alloc(void)
 685 {
 686         LASSERT(page_pools.epp_max_pools);
 687         OBD_ALLOC_LARGE(page_pools.epp_pools,
 688                         page_pools.epp_max_pools *
 689                         sizeof(*page_pools.epp_pools));
 690 }
 691
 692 static inline void enc_pools_free(void)
 693 {
 694         LASSERT(page_pools.epp_max_pools);
 695         LASSERT(page_pools.epp_pools);
 696
 697         OBD_FREE_LARGE(page_pools.epp_pools,
 698                        page_pools.epp_max_pools *
 699                        sizeof(*page_pools.epp_pools));
 700 }
 701
 702 int sptlrpc_enc_pool_init(void)
 703 {
 704         /*
 705          * maximum capacity is 1/8 of total physical memory.
 706          * is the 1/8 a good number?
 707          */
 708         page_pools.epp_max_pages = totalram_pages / 8;
 709         page_pools.epp_max_pools = npages_to_npools(page_pools.epp_max_pages);
 710
 711         init_waitqueue_head(&page_pools.epp_waitq);
 712         page_pools.epp_waitqlen = 0;
 713         page_pools.epp_pages_short = 0;
 714
 715         page_pools.epp_growing = 0;
 716
 717         page_pools.epp_idle_idx = 0;
 718         page_pools.epp_last_shrink = cfs_time_current_sec();
 719         page_pools.epp_last_access = cfs_time_current_sec();
 720
 721         spin_lock_init(&page_pools.epp_lock);
 722         page_pools.epp_total_pages = 0;
 723         page_pools.epp_free_pages = 0;
 724
 725         page_pools.epp_st_max_pages = 0;
 726         page_pools.epp_st_grows = 0;
 727         page_pools.epp_st_grow_fails = 0;
 728         page_pools.epp_st_shrinks = 0;
 729         page_pools.epp_st_access = 0;
 730         page_pools.epp_st_missings = 0;
 731         page_pools.epp_st_lowfree = 0;
 732         page_pools.epp_st_max_wqlen = 0;
 733         page_pools.epp_st_max_wait = 0;
 734
 735         enc_pools_alloc();
 736         if (page_pools.epp_pools == NULL)
 737                 return -ENOMEM;
 738
 739         pools_shrinker = set_shrinker(pools_shrinker_seeks,
 740                                           enc_pools_shrink);
 741         if (pools_shrinker == NULL) {
 742                 enc_pools_free();
 743                 return -ENOMEM;
 744         }
 745
 746         return 0;
 747 }
 748
 749 void sptlrpc_enc_pool_fini(void)
 750 {
 751         unsigned long cleaned, npools;
 752
 753         LASSERT(pools_shrinker);
 754         LASSERT(page_pools.epp_pools);
 755         LASSERT(page_pools.epp_total_pages == page_pools.epp_free_pages);
 756
 757         remove_shrinker(pools_shrinker);
 758
 759         npools = npages_to_npools(page_pools.epp_total_pages);
 760         cleaned = enc_pools_cleanup(page_pools.epp_pools, npools);
 761         LASSERT(cleaned == page_pools.epp_total_pages);
 762
 763         enc_pools_free();
 764
 765         if (page_pools.epp_st_access > 0) {
 766                 CDEBUG(D_SEC,
 767                        "max pages %lu, grows %u, grow fails %u, shrinks %u, "
 768                        "access %lu, missing %lu, max qlen %u, max wait "
 769                        CFS_TIME_T"/%d\n",
 770                        page_pools.epp_st_max_pages, page_pools.epp_st_grows,
 771                        page_pools.epp_st_grow_fails,
 772                        page_pools.epp_st_shrinks, page_pools.epp_st_access,
 773                        page_pools.epp_st_missings, page_pools.epp_st_max_wqlen,
 774                        page_pools.epp_st_max_wait, HZ);
 775         }
 776 }
 777
 778
 779 static int cfs_hash_alg_id[] = {
 780         [BULK_HASH_ALG_NULL]    = CFS_HASH_ALG_NULL,
 781         [BULK_HASH_ALG_ADLER32] = CFS_HASH_ALG_ADLER32,
 782         [BULK_HASH_ALG_CRC32]   = CFS_HASH_ALG_CRC32,
 783         [BULK_HASH_ALG_MD5]     = CFS_HASH_ALG_MD5,
 784         [BULK_HASH_ALG_SHA1]    = CFS_HASH_ALG_SHA1,
 785         [BULK_HASH_ALG_SHA256]  = CFS_HASH_ALG_SHA256,
 786         [BULK_HASH_ALG_SHA384]  = CFS_HASH_ALG_SHA384,
 787         [BULK_HASH_ALG_SHA512]  = CFS_HASH_ALG_SHA512,
 788 };
 789 const char * sptlrpc_get_hash_name(__u8 hash_alg)
 790 {
 791         return cfs_crypto_hash_name(cfs_hash_alg_id[hash_alg]);
 792 }
 793 EXPORT_SYMBOL(sptlrpc_get_hash_name);
 794
 795 __u8 sptlrpc_get_hash_alg(const char *algname)
 796 {
 797         return cfs_crypto_hash_alg(algname);
 798 }
 799 EXPORT_SYMBOL(sptlrpc_get_hash_alg);
 800
 801 int bulk_sec_desc_unpack(struct lustre_msg *msg, int offset, int swabbed)
 802 {
 803         struct ptlrpc_bulk_sec_desc *bsd;
 804         int                       size = msg->lm_buflens[offset];
 805
 806         bsd = lustre_msg_buf(msg, offset, sizeof(*bsd));
 807         if (bsd == NULL) {
 808                 CERROR("Invalid bulk sec desc: size %d\n", size);
 809                 return -EINVAL;
 810         }
 811
 812         if (swabbed) {
 813                 __swab32s(&bsd->bsd_nob);
 814         }
 815
 816         if (unlikely(bsd->bsd_version != 0)) {
 817                 CERROR("Unexpected version %u\n", bsd->bsd_version);
 818                 return -EPROTO;
 819         }
 820
 821         if (unlikely(bsd->bsd_type >= SPTLRPC_BULK_MAX)) {
 822                 CERROR("Invalid type %u\n", bsd->bsd_type);
 823                 return -EPROTO;
 824         }
 825
 826         /* FIXME more sanity check here */
 827
 828         if (unlikely(bsd->bsd_svc != SPTLRPC_BULK_SVC_NULL &&
 829                      bsd->bsd_svc != SPTLRPC_BULK_SVC_INTG &&
 830                      bsd->bsd_svc != SPTLRPC_BULK_SVC_PRIV)) {
 831                 CERROR("Invalid svc %u\n", bsd->bsd_svc);
 832                 return -EPROTO;
 833         }
 834
 835         return 0;
 836 }
 837 EXPORT_SYMBOL(bulk_sec_desc_unpack);
 838
 839 int sptlrpc_get_bulk_checksum(struct ptlrpc_bulk_desc *desc, __u8 alg,
 840                               void *buf, int buflen)
 841 {
 842         struct cfs_crypto_hash_desc     *hdesc;
 843         int                             hashsize;
 844         char                            hashbuf[64];
 845         unsigned int                    bufsize;
 846         int                             i, err;
 847
 848         LASSERT(alg > BULK_HASH_ALG_NULL && alg < BULK_HASH_ALG_MAX);
 849         LASSERT(buflen >= 4);
 850
 851         hdesc = cfs_crypto_hash_init(cfs_hash_alg_id[alg], NULL, 0);
 852         if (IS_ERR(hdesc)) {
 853                 CERROR("Unable to initialize checksum hash %s\n",
 854                        cfs_crypto_hash_name(cfs_hash_alg_id[alg]));
 855                 return PTR_ERR(hdesc);
 856         }
 857
 858         hashsize = cfs_crypto_hash_digestsize(cfs_hash_alg_id[alg]);
 859
 860         for (i = 0; i < desc->bd_iov_count; i++) {
 861                 cfs_crypto_hash_update_page(hdesc, desc->bd_iov[i].kiov_page,
 862                                   desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK,
 863                                   desc->bd_iov[i].kiov_len);
 864         }
 865         if (hashsize > buflen) {
 866                 bufsize = sizeof(hashbuf);
 867                 err = cfs_crypto_hash_final(hdesc, (unsigned char *)hashbuf,
 868                                             &bufsize);
 869                 memcpy(buf, hashbuf, buflen);
 870         } else {
 871                 bufsize = buflen;
 872                 err = cfs_crypto_hash_final(hdesc, (unsigned char *)buf,
 873                                             &bufsize);
 874         }
 875
 876         if (err)
 877                 cfs_crypto_hash_final(hdesc, NULL, NULL);
 878         return err;
 879 }
 880 EXPORT_SYMBOL(sptlrpc_get_bulk_checksum);