lustre/ptlrpc/sec_bulk.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  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 #ifndef EXPORT_SYMTAB
  42 #define EXPORT_SYMTAB
  43 #endif
  44 #define DEBUG_SUBSYSTEM S_SEC
  45
  46 #include <libcfs/libcfs.h>
  47 #ifndef __KERNEL__
  48 #include <liblustre.h>
  49 #include <libcfs/list.h>
  50 #else
  51 #include <linux/crypto.h>
  52 #endif
  53
  54 #include <obd.h>
  55 #include <obd_cksum.h>
  56 #include <obd_class.h>
  57 #include <obd_support.h>
  58 #include <lustre_net.h>
  59 #include <lustre_import.h>
  60 #include <lustre_dlm.h>
  61 #include <lustre_sec.h>
  62
  63 #include "ptlrpc_internal.h"
  64
  65 /****************************************
  66  * bulk encryption page pools           *
  67  ****************************************/
  68
  69 #ifdef __KERNEL__
  70
  71 #define PTRS_PER_PAGE   (CFS_PAGE_SIZE / sizeof(void *))
  72 #define PAGES_PER_POOL  (PTRS_PER_PAGE)
  73
  74 #define IDLE_IDX_MAX            (100)
  75 #define IDLE_IDX_WEIGHT         (3)
  76
  77 #define CACHE_QUIESCENT_PERIOD  (20)
  78
  79 static struct ptlrpc_enc_page_pool {
  80         /*
  81          * constants
  82          */
  83         unsigned long    epp_max_pages;   /* maximum pages can hold, const */
  84         unsigned int     epp_max_pools;   /* number of pools, const */
  85
  86         /*
  87          * wait queue in case of not enough free pages.
  88          */
  89         cfs_waitq_t      epp_waitq;       /* waiting threads */
  90         unsigned int     epp_waitqlen;    /* wait queue length */
  91         unsigned long    epp_pages_short; /* # of pages wanted of in-q users */
  92         unsigned int     epp_growing:1;   /* during adding pages */
  93
  94         /*
  95          * indicating how idle the pools are, from 0 to MAX_IDLE_IDX
  96          * this is counted based on each time when getting pages from
  97          * the pools, not based on time. which means in case that system
  98          * is idled for a while but the idle_idx might still be low if no
  99          * activities happened in the pools.
 100          */
 101         unsigned long    epp_idle_idx;
 102
 103         /* last shrink time due to mem tight */
 104         long             epp_last_shrink;
 105         long             epp_last_access;
 106
 107         /*
 108          * in-pool pages bookkeeping
 109          */
 110         spinlock_t       epp_lock;        /* protect following fields */
 111         unsigned long    epp_total_pages; /* total pages in pools */
 112         unsigned long    epp_free_pages;  /* current pages available */
 113
 114         /*
 115          * statistics
 116          */
 117         unsigned long    epp_st_max_pages;      /* # of pages ever reached */
 118         unsigned int     epp_st_grows;          /* # of grows */
 119         unsigned int     epp_st_grow_fails;     /* # of add pages failures */
 120         unsigned int     epp_st_shrinks;        /* # of shrinks */
 121         unsigned long    epp_st_access;         /* # of access */
 122         unsigned long    epp_st_missings;       /* # of cache missing */
 123         unsigned long    epp_st_lowfree;        /* lowest free pages reached */
 124         unsigned int     epp_st_max_wqlen;      /* highest waitqueue length */
 125         cfs_time_t       epp_st_max_wait;       /* in jeffies */
 126         /*
 127          * pointers to pools
 128          */
 129         cfs_page_t    ***epp_pools;
 130 } page_pools;
 131
 132 /*
 133  * memory shrinker
 134  */
 135 const int pools_shrinker_seeks = DEFAULT_SEEKS;
 136 static struct shrinker *pools_shrinker = NULL;
 137
 138
 139 /*
 140  * /proc/fs/lustre/sptlrpc/encrypt_page_pools
 141  */
 142 int sptlrpc_proc_read_enc_pool(char *page, char **start, off_t off, int count,
 143                                int *eof, void *data)
 144 {
 145         int     rc;
 146
 147         spin_lock(&page_pools.epp_lock);
 148
 149         rc = snprintf(page, count,
 150                       "physical pages:          %lu\n"
 151                       "pages per pool:          %lu\n"
 152                       "max pages:               %lu\n"
 153                       "max pools:               %u\n"
 154                       "total pages:             %lu\n"
 155                       "total free:              %lu\n"
 156                       "idle index:              %lu/100\n"
 157                       "last shrink:             %lds\n"
 158                       "last access:             %lds\n"
 159                       "max pages reached:       %lu\n"
 160                       "grows:                   %u\n"
 161                       "grows failure:           %u\n"
 162                       "shrinks:                 %u\n"
 163                       "cache access:            %lu\n"
 164                       "cache missing:           %lu\n"
 165                       "low free mark:           %lu\n"
 166                       "max waitqueue depth:     %u\n"
 167                       "max wait time:           "CFS_TIME_T"/%u\n"
 168                       ,
 169                       num_physpages,
 170                       PAGES_PER_POOL,
 171                       page_pools.epp_max_pages,
 172                       page_pools.epp_max_pools,
 173                       page_pools.epp_total_pages,
 174                       page_pools.epp_free_pages,
 175                       page_pools.epp_idle_idx,
 176                       cfs_time_current_sec() - page_pools.epp_last_shrink,
 177                       cfs_time_current_sec() - page_pools.epp_last_access,
 178                       page_pools.epp_st_max_pages,
 179                       page_pools.epp_st_grows,
 180                       page_pools.epp_st_grow_fails,
 181                       page_pools.epp_st_shrinks,
 182                       page_pools.epp_st_access,
 183                       page_pools.epp_st_missings,
 184                       page_pools.epp_st_lowfree,
 185                       page_pools.epp_st_max_wqlen,
 186                       page_pools.epp_st_max_wait, HZ
 187                      );
 188
 189         spin_unlock(&page_pools.epp_lock);
 190         return rc;
 191 }
 192
 193 static void enc_pools_release_free_pages(long npages)
 194 {
 195         int     p_idx, g_idx;
 196         int     p_idx_max1, p_idx_max2;
 197
 198         LASSERT(npages > 0);
 199         LASSERT(npages <= page_pools.epp_free_pages);
 200         LASSERT(page_pools.epp_free_pages <= page_pools.epp_total_pages);
 201
 202         /* max pool index before the release */
 203         p_idx_max2 = (page_pools.epp_total_pages - 1) / PAGES_PER_POOL;
 204
 205         page_pools.epp_free_pages -= npages;
 206         page_pools.epp_total_pages -= npages;
 207
 208         /* max pool index after the release */
 209         p_idx_max1 = page_pools.epp_total_pages == 0 ? 0 :
 210                      ((page_pools.epp_total_pages - 1) / PAGES_PER_POOL);
 211
 212         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 213         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 214         LASSERT(page_pools.epp_pools[p_idx]);
 215
 216         while (npages--) {
 217                 LASSERT(page_pools.epp_pools[p_idx]);
 218                 LASSERT(page_pools.epp_pools[p_idx][g_idx] != NULL);
 219
 220                 cfs_free_page(page_pools.epp_pools[p_idx][g_idx]);
 221                 page_pools.epp_pools[p_idx][g_idx] = NULL;
 222
 223                 if (++g_idx == PAGES_PER_POOL) {
 224                         p_idx++;
 225                         g_idx = 0;
 226                 }
 227         };
 228
 229         /* free unused pools */
 230         while (p_idx_max1 < p_idx_max2) {
 231                 LASSERT(page_pools.epp_pools[p_idx_max2]);
 232                 OBD_FREE(page_pools.epp_pools[p_idx_max2], CFS_PAGE_SIZE);
 233                 page_pools.epp_pools[p_idx_max2] = NULL;
 234                 p_idx_max2--;
 235         }
 236 }
 237
 238 /*
 239  * could be called frequently for query (@nr_to_scan == 0)
 240  */
 241 static int enc_pools_shrink(int nr_to_scan, unsigned int gfp_mask)
 242 {
 243         unsigned long   ret;
 244
 245         spin_lock(&page_pools.epp_lock);
 246
 247         if (nr_to_scan > page_pools.epp_free_pages)
 248                 nr_to_scan = page_pools.epp_free_pages;
 249
 250         if (nr_to_scan > 0) {
 251                 enc_pools_release_free_pages(nr_to_scan);
 252                 CDEBUG(D_SEC, "released %d pages, %ld left\n",
 253                        nr_to_scan, page_pools.epp_free_pages);
 254
 255                 page_pools.epp_st_shrinks++;
 256                 page_pools.epp_last_shrink = cfs_time_current_sec();
 257         }
 258
 259         /*
 260          * try to keep at least PTLRPC_MAX_BRW_PAGES pages in the pool
 261          */
 262         if (page_pools.epp_free_pages <= PTLRPC_MAX_BRW_PAGES) {
 263                 ret = 0;
 264                 goto out_unlock;
 265         }
 266
 267         /*
 268          * if no pool access for a long time, we consider it's fully idle
 269          */
 270         if (cfs_time_current_sec() - page_pools.epp_last_access >
 271             CACHE_QUIESCENT_PERIOD)
 272                 page_pools.epp_idle_idx = IDLE_IDX_MAX;
 273
 274         LASSERT(page_pools.epp_idle_idx <= IDLE_IDX_MAX);
 275         ret = (page_pools.epp_free_pages * page_pools.epp_idle_idx /
 276                IDLE_IDX_MAX);
 277         if (page_pools.epp_free_pages - ret < PTLRPC_MAX_BRW_PAGES)
 278                 ret = page_pools.epp_free_pages - PTLRPC_MAX_BRW_PAGES;
 279
 280 out_unlock:
 281         spin_unlock(&page_pools.epp_lock);
 282         return ret;
 283 }
 284
 285 static inline
 286 int npages_to_npools(unsigned long npages)
 287 {
 288         return (int) ((npages + PAGES_PER_POOL - 1) / PAGES_PER_POOL);
 289 }
 290
 291 /*
 292  * return how many pages cleaned up.
 293  */
 294 static unsigned long enc_pools_cleanup(cfs_page_t ***pools, int npools)
 295 {
 296         unsigned long cleaned = 0;
 297         int           i, j;
 298
 299         for (i = 0; i < npools; i++) {
 300                 if (pools[i]) {
 301                         for (j = 0; j < PAGES_PER_POOL; j++) {
 302                                 if (pools[i][j]) {
 303                                         cfs_free_page(pools[i][j]);
 304                                         cleaned++;
 305                                 }
 306                         }
 307                         OBD_FREE(pools[i], CFS_PAGE_SIZE);
 308                         pools[i] = NULL;
 309                 }
 310         }
 311
 312         return cleaned;
 313 }
 314
 315 /*
 316  * merge @npools pointed by @pools which contains @npages new pages
 317  * into current pools.
 318  *
 319  * we have options to avoid most memory copy with some tricks. but we choose
 320  * the simplest way to avoid complexity. It's not frequently called.
 321  */
 322 static void enc_pools_insert(cfs_page_t ***pools, int npools, int npages)
 323 {
 324         int     freeslot;
 325         int     op_idx, np_idx, og_idx, ng_idx;
 326         int     cur_npools, end_npools;
 327
 328         LASSERT(npages > 0);
 329         LASSERT(page_pools.epp_total_pages+npages <= page_pools.epp_max_pages);
 330         LASSERT(npages_to_npools(npages) == npools);
 331
 332         spin_lock(&page_pools.epp_lock);
 333
 334         /*
 335          * (1) fill all the free slots of current pools.
 336          */
 337         /* free slots are those left by rent pages, and the extra ones with
 338          * index >= total_pages, locate at the tail of last pool. */
 339         freeslot = page_pools.epp_total_pages % PAGES_PER_POOL;
 340         if (freeslot != 0)
 341                 freeslot = PAGES_PER_POOL - freeslot;
 342         freeslot += page_pools.epp_total_pages - page_pools.epp_free_pages;
 343
 344         op_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 345         og_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 346         np_idx = npools - 1;
 347         ng_idx = (npages - 1) % PAGES_PER_POOL;
 348
 349         while (freeslot) {
 350                 LASSERT(page_pools.epp_pools[op_idx][og_idx] == NULL);
 351                 LASSERT(pools[np_idx][ng_idx] != NULL);
 352
 353                 page_pools.epp_pools[op_idx][og_idx] = pools[np_idx][ng_idx];
 354                 pools[np_idx][ng_idx] = NULL;
 355
 356                 freeslot--;
 357
 358                 if (++og_idx == PAGES_PER_POOL) {
 359                         op_idx++;
 360                         og_idx = 0;
 361                 }
 362                 if (--ng_idx < 0) {
 363                         if (np_idx == 0)
 364                                 break;
 365                         np_idx--;
 366                         ng_idx = PAGES_PER_POOL - 1;
 367                 }
 368         }
 369
 370         /*
 371          * (2) add pools if needed.
 372          */
 373         cur_npools = (page_pools.epp_total_pages + PAGES_PER_POOL - 1) /
 374                      PAGES_PER_POOL;
 375         end_npools = (page_pools.epp_total_pages + npages + PAGES_PER_POOL -1) /
 376                      PAGES_PER_POOL;
 377         LASSERT(end_npools <= page_pools.epp_max_pools);
 378
 379         np_idx = 0;
 380         while (cur_npools < end_npools) {
 381                 LASSERT(page_pools.epp_pools[cur_npools] == NULL);
 382                 LASSERT(np_idx < npools);
 383                 LASSERT(pools[np_idx] != NULL);
 384
 385                 page_pools.epp_pools[cur_npools++] = pools[np_idx];
 386                 pools[np_idx++] = NULL;
 387         }
 388
 389         page_pools.epp_total_pages += npages;
 390         page_pools.epp_free_pages += npages;
 391         page_pools.epp_st_lowfree = page_pools.epp_free_pages;
 392
 393         if (page_pools.epp_total_pages > page_pools.epp_st_max_pages)
 394                 page_pools.epp_st_max_pages = page_pools.epp_total_pages;
 395
 396         CDEBUG(D_SEC, "add %d pages to total %lu\n", npages,
 397                page_pools.epp_total_pages);
 398
 399         spin_unlock(&page_pools.epp_lock);
 400 }
 401
 402 static int enc_pools_add_pages(int npages)
 403 {
 404         static DECLARE_MUTEX(sem_add_pages);
 405         cfs_page_t   ***pools;
 406         int             npools, alloced = 0;
 407         int             i, j, rc = -ENOMEM;
 408
 409         if (npages < PTLRPC_MAX_BRW_PAGES)
 410                 npages = PTLRPC_MAX_BRW_PAGES;
 411
 412         down(&sem_add_pages);
 413
 414         if (npages + page_pools.epp_total_pages > page_pools.epp_max_pages)
 415                 npages = page_pools.epp_max_pages - page_pools.epp_total_pages;
 416         LASSERT(npages > 0);
 417
 418         page_pools.epp_st_grows++;
 419
 420         npools = npages_to_npools(npages);
 421         OBD_ALLOC(pools, npools * sizeof(*pools));
 422         if (pools == NULL)
 423                 goto out;
 424
 425         for (i = 0; i < npools; i++) {
 426                 OBD_ALLOC(pools[i], CFS_PAGE_SIZE);
 427                 if (pools[i] == NULL)
 428                         goto out_pools;
 429
 430                 for (j = 0; j < PAGES_PER_POOL && alloced < npages; j++) {
 431                         pools[i][j] = cfs_alloc_page(CFS_ALLOC_IO |
 432                                                      CFS_ALLOC_HIGH);
 433                         if (pools[i][j] == NULL)
 434                                 goto out_pools;
 435
 436                         alloced++;
 437                 }
 438         }
 439
 440         enc_pools_insert(pools, npools, npages);
 441         CDEBUG(D_SEC, "added %d pages into pools\n", npages);
 442         rc = 0;
 443
 444 out_pools:
 445         enc_pools_cleanup(pools, npools);
 446         OBD_FREE(pools, npools * sizeof(*pools));
 447 out:
 448         if (rc) {
 449                 page_pools.epp_st_grow_fails++;
 450                 CERROR("Failed to allocate %d enc pages\n", npages);
 451         }
 452
 453         up(&sem_add_pages);
 454         return rc;
 455 }
 456
 457 static inline void enc_pools_wakeup(void)
 458 {
 459         if (unlikely(page_pools.epp_waitqlen)) {
 460                 LASSERT(page_pools.epp_waitqlen > 0);
 461                 LASSERT(cfs_waitq_active(&page_pools.epp_waitq));
 462                 cfs_waitq_broadcast(&page_pools.epp_waitq);
 463         }
 464 }
 465
 466 static int enc_pools_should_grow(int page_needed, long now)
 467 {
 468         /* don't grow if someone else is growing the pools right now,
 469          * or the pools has reached its full capacity
 470          */
 471         if (page_pools.epp_growing ||
 472             page_pools.epp_total_pages == page_pools.epp_max_pages)
 473                 return 0;
 474
 475         /* if total pages is not enough, we need to grow */
 476         if (page_pools.epp_total_pages < page_needed)
 477                 return 1;
 478
 479         /* if we just did a shrink due to memory tight, we'd better
 480          * wait a while to grow again.
 481          */
 482         if (now - page_pools.epp_last_shrink < 2)
 483                 return 0;
 484
 485         /*
 486          * here we perhaps need consider other factors like wait queue
 487          * length, idle index, etc. ?
 488          */
 489
 490         /* grow the pools in any other cases */
 491         return 1;
 492 }
 493
 494 /*
 495  * we allocate the requested pages atomically.
 496  */
 497 int sptlrpc_enc_pool_get_pages(struct ptlrpc_bulk_desc *desc)
 498 {
 499         cfs_waitlink_t  waitlink;
 500         unsigned long   this_idle = -1;
 501         cfs_time_t      tick = 0;
 502         long            now;
 503         int             p_idx, g_idx;
 504         int             i;
 505
 506         LASSERT(desc->bd_max_iov > 0);
 507         LASSERT(desc->bd_max_iov <= page_pools.epp_max_pages);
 508
 509         /* resent bulk, enc pages might have been allocated previously */
 510         if (desc->bd_enc_pages != NULL)
 511                 return 0;
 512
 513         OBD_ALLOC(desc->bd_enc_pages,
 514                   desc->bd_max_iov * sizeof(*desc->bd_enc_pages));
 515         if (desc->bd_enc_pages == NULL)
 516                 return -ENOMEM;
 517
 518         spin_lock(&page_pools.epp_lock);
 519
 520         page_pools.epp_st_access++;
 521 again:
 522         if (unlikely(page_pools.epp_free_pages < desc->bd_max_iov)) {
 523                 if (tick == 0)
 524                         tick = cfs_time_current();
 525
 526                 now = cfs_time_current_sec();
 527
 528                 page_pools.epp_st_missings++;
 529                 page_pools.epp_pages_short += desc->bd_max_iov;
 530
 531                 if (enc_pools_should_grow(desc->bd_max_iov, now)) {
 532                         page_pools.epp_growing = 1;
 533
 534                         spin_unlock(&page_pools.epp_lock);
 535                         enc_pools_add_pages(page_pools.epp_pages_short / 2);
 536                         spin_lock(&page_pools.epp_lock);
 537
 538                         page_pools.epp_growing = 0;
 539                 } else {
 540                         if (++page_pools.epp_waitqlen >
 541                             page_pools.epp_st_max_wqlen)
 542                                 page_pools.epp_st_max_wqlen =
 543                                                 page_pools.epp_waitqlen;
 544
 545                         set_current_state(CFS_TASK_UNINT);
 546                         cfs_waitlink_init(&waitlink);
 547                         cfs_waitq_add(&page_pools.epp_waitq, &waitlink);
 548
 549                         spin_unlock(&page_pools.epp_lock);
 550                         cfs_waitq_wait(&waitlink, CFS_TASK_UNINT);
 551                         cfs_waitq_del(&page_pools.epp_waitq, &waitlink);
 552                         spin_lock(&page_pools.epp_lock);
 553
 554                         LASSERT(page_pools.epp_waitqlen > 0);
 555                         page_pools.epp_waitqlen--;
 556                 }
 557
 558                 LASSERT(page_pools.epp_pages_short >= desc->bd_max_iov);
 559                 page_pools.epp_pages_short -= desc->bd_max_iov;
 560
 561                 this_idle = 0;
 562                 goto again;
 563         }
 564
 565         /* record max wait time */
 566         if (unlikely(tick != 0)) {
 567                 tick = cfs_time_current() - tick;
 568                 if (tick > page_pools.epp_st_max_wait)
 569                         page_pools.epp_st_max_wait = tick;
 570         }
 571
 572         /* proceed with rest of allocation */
 573         page_pools.epp_free_pages -= desc->bd_max_iov;
 574
 575         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 576         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 577
 578         for (i = 0; i < desc->bd_max_iov; i++) {
 579                 LASSERT(page_pools.epp_pools[p_idx][g_idx] != NULL);
 580                 desc->bd_enc_pages[i] = page_pools.epp_pools[p_idx][g_idx];
 581                 page_pools.epp_pools[p_idx][g_idx] = NULL;
 582
 583                 if (++g_idx == PAGES_PER_POOL) {
 584                         p_idx++;
 585                         g_idx = 0;
 586                 }
 587         }
 588
 589         if (page_pools.epp_free_pages < page_pools.epp_st_lowfree)
 590                 page_pools.epp_st_lowfree = page_pools.epp_free_pages;
 591
 592         /*
 593          * new idle index = (old * weight + new) / (weight + 1)
 594          */
 595         if (this_idle == -1) {
 596                 this_idle = page_pools.epp_free_pages * IDLE_IDX_MAX /
 597                             page_pools.epp_total_pages;
 598         }
 599         page_pools.epp_idle_idx = (page_pools.epp_idle_idx * IDLE_IDX_WEIGHT +
 600                                    this_idle) /
 601                                   (IDLE_IDX_WEIGHT + 1);
 602
 603         page_pools.epp_last_access = cfs_time_current_sec();
 604
 605         spin_unlock(&page_pools.epp_lock);
 606         return 0;
 607 }
 608 EXPORT_SYMBOL(sptlrpc_enc_pool_get_pages);
 609
 610 void sptlrpc_enc_pool_put_pages(struct ptlrpc_bulk_desc *desc)
 611 {
 612         int     p_idx, g_idx;
 613         int     i;
 614
 615         if (desc->bd_enc_pages == NULL)
 616                 return;
 617         if (desc->bd_max_iov == 0)
 618                 return;
 619
 620         spin_lock(&page_pools.epp_lock);
 621
 622         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 623         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 624
 625         LASSERT(page_pools.epp_free_pages + desc->bd_max_iov <=
 626                 page_pools.epp_total_pages);
 627         LASSERT(page_pools.epp_pools[p_idx]);
 628
 629         for (i = 0; i < desc->bd_max_iov; i++) {
 630                 LASSERT(desc->bd_enc_pages[i] != NULL);
 631                 LASSERT(g_idx != 0 || page_pools.epp_pools[p_idx]);
 632                 LASSERT(page_pools.epp_pools[p_idx][g_idx] == NULL);
 633
 634                 page_pools.epp_pools[p_idx][g_idx] = desc->bd_enc_pages[i];
 635
 636                 if (++g_idx == PAGES_PER_POOL) {
 637                         p_idx++;
 638                         g_idx = 0;
 639                 }
 640         }
 641
 642         page_pools.epp_free_pages += desc->bd_max_iov;
 643
 644         enc_pools_wakeup();
 645
 646         spin_unlock(&page_pools.epp_lock);
 647
 648         OBD_FREE(desc->bd_enc_pages,
 649                  desc->bd_max_iov * sizeof(*desc->bd_enc_pages));
 650         desc->bd_enc_pages = NULL;
 651 }
 652 EXPORT_SYMBOL(sptlrpc_enc_pool_put_pages);
 653
 654 /*
 655  * we don't do much stuff for add_user/del_user anymore, except adding some
 656  * initial pages in add_user() if current pools are empty, rest would be
 657  * handled by the pools's self-adaption.
 658  */
 659 int sptlrpc_enc_pool_add_user(void)
 660 {
 661         int     need_grow = 0;
 662
 663         spin_lock(&page_pools.epp_lock);
 664         if (page_pools.epp_growing == 0 && page_pools.epp_total_pages == 0) {
 665                 page_pools.epp_growing = 1;
 666                 need_grow = 1;
 667         }
 668         spin_unlock(&page_pools.epp_lock);
 669
 670         if (need_grow) {
 671                 enc_pools_add_pages(PTLRPC_MAX_BRW_PAGES);
 672
 673                 spin_lock(&page_pools.epp_lock);
 674                 page_pools.epp_growing = 0;
 675                 enc_pools_wakeup();
 676                 spin_unlock(&page_pools.epp_lock);
 677         }
 678         return 0;
 679 }
 680 EXPORT_SYMBOL(sptlrpc_enc_pool_add_user);
 681
 682 int sptlrpc_enc_pool_del_user(void)
 683 {
 684         return 0;
 685 }
 686 EXPORT_SYMBOL(sptlrpc_enc_pool_del_user);
 687
 688 static inline void enc_pools_alloc(void)
 689 {
 690         LASSERT(page_pools.epp_max_pools);
 691         /*
 692          * on system with huge memory but small page size, this might lead to
 693          * high-order allocation. but it's not common, and we suppose memory
 694          * be not too much fragmented at module loading time.
 695          */
 696         OBD_ALLOC(page_pools.epp_pools,
 697                   page_pools.epp_max_pools * sizeof(*page_pools.epp_pools));
 698 }
 699
 700 static inline void enc_pools_free(void)
 701 {
 702         LASSERT(page_pools.epp_max_pools);
 703         LASSERT(page_pools.epp_pools);
 704
 705         OBD_FREE(page_pools.epp_pools,
 706                  page_pools.epp_max_pools * sizeof(*page_pools.epp_pools));
 707 }
 708
 709 int sptlrpc_enc_pool_init(void)
 710 {
 711         /*
 712          * maximum capacity is 1/8 of total physical memory.
 713          * is the 1/8 a good number?
 714          */
 715         page_pools.epp_max_pages = num_physpages / 8;
 716         page_pools.epp_max_pools = npages_to_npools(page_pools.epp_max_pages);
 717
 718         cfs_waitq_init(&page_pools.epp_waitq);
 719         page_pools.epp_waitqlen = 0;
 720         page_pools.epp_pages_short = 0;
 721
 722         page_pools.epp_growing = 0;
 723
 724         page_pools.epp_idle_idx = 0;
 725         page_pools.epp_last_shrink = cfs_time_current_sec();
 726         page_pools.epp_last_access = cfs_time_current_sec();
 727
 728         spin_lock_init(&page_pools.epp_lock);
 729         page_pools.epp_total_pages = 0;
 730         page_pools.epp_free_pages = 0;
 731
 732         page_pools.epp_st_max_pages = 0;
 733         page_pools.epp_st_grows = 0;
 734         page_pools.epp_st_grow_fails = 0;
 735         page_pools.epp_st_shrinks = 0;
 736         page_pools.epp_st_access = 0;
 737         page_pools.epp_st_missings = 0;
 738         page_pools.epp_st_lowfree = 0;
 739         page_pools.epp_st_max_wqlen = 0;
 740         page_pools.epp_st_max_wait = 0;
 741
 742         enc_pools_alloc();
 743         if (page_pools.epp_pools == NULL)
 744                 return -ENOMEM;
 745
 746         pools_shrinker = set_shrinker(pools_shrinker_seeks, enc_pools_shrink);
 747         if (pools_shrinker == NULL) {
 748                 enc_pools_free();
 749                 return -ENOMEM;
 750         }
 751
 752         return 0;
 753 }
 754
 755 void sptlrpc_enc_pool_fini(void)
 756 {
 757         unsigned long cleaned, npools;
 758
 759         LASSERT(pools_shrinker);
 760         LASSERT(page_pools.epp_pools);
 761         LASSERT(page_pools.epp_total_pages == page_pools.epp_free_pages);
 762
 763         remove_shrinker(pools_shrinker);
 764
 765         npools = npages_to_npools(page_pools.epp_total_pages);
 766         cleaned = enc_pools_cleanup(page_pools.epp_pools, npools);
 767         LASSERT(cleaned == page_pools.epp_total_pages);
 768
 769         enc_pools_free();
 770
 771         if (page_pools.epp_st_access > 0) {
 772                 CWARN("max pages %lu, grows %u, grow fails %u, shrinks %u, "
 773                       "access %lu, missing %lu, max qlen %u, max wait "
 774                       CFS_TIME_T"/%d\n",
 775                       page_pools.epp_st_max_pages, page_pools.epp_st_grows,
 776                       page_pools.epp_st_grow_fails,
 777                       page_pools.epp_st_shrinks, page_pools.epp_st_access,
 778                       page_pools.epp_st_missings, page_pools.epp_st_max_wqlen,
 779                       page_pools.epp_st_max_wait, HZ);
 780         }
 781 }
 782
 783 #else /* !__KERNEL__ */
 784
 785 int sptlrpc_enc_pool_get_pages(struct ptlrpc_bulk_desc *desc)
 786 {
 787         return 0;
 788 }
 789
 790 void sptlrpc_enc_pool_put_pages(struct ptlrpc_bulk_desc *desc)
 791 {
 792 }
 793
 794 int sptlrpc_enc_pool_init(void)
 795 {
 796         return 0;
 797 }
 798
 799 void sptlrpc_enc_pool_fini(void)
 800 {
 801 }
 802 #endif
 803
 804 /****************************************
 805  * Helpers to assist policy modules to  *
 806  * implement checksum funcationality    *
 807  ****************************************/
 808
 809 static struct sptlrpc_hash_type hash_types[] = {
 810         [BULK_HASH_ALG_NULL]    = { "null",     "null",         0 },
 811         [BULK_HASH_ALG_ADLER32] = { "adler32",  "adler32",      4 },
 812         [BULK_HASH_ALG_CRC32]   = { "crc32",    "crc32",        4 },
 813         [BULK_HASH_ALG_MD5]     = { "md5",      "md5",          16 },
 814         [BULK_HASH_ALG_SHA1]    = { "sha1",     "sha1",         20 },
 815         [BULK_HASH_ALG_SHA256]  = { "sha256",   "sha256",       32 },
 816         [BULK_HASH_ALG_SHA384]  = { "sha384",   "sha384",       48 },
 817         [BULK_HASH_ALG_SHA512]  = { "sha512",   "sha512",       64 },
 818         [BULK_HASH_ALG_WP256]   = { "wp256",    "wp256",        32 },
 819         [BULK_HASH_ALG_WP384]   = { "wp384",    "wp384",        48 },
 820         [BULK_HASH_ALG_WP512]   = { "wp512",    "wp512",        64 },
 821 };
 822
 823 const struct sptlrpc_hash_type *sptlrpc_get_hash_type(__u8 hash_alg)
 824 {
 825         struct sptlrpc_hash_type *ht;
 826
 827         if (hash_alg < BULK_HASH_ALG_MAX) {
 828                 ht = &hash_types[hash_alg];
 829                 if (ht->sht_tfm_name)
 830                         return ht;
 831         }
 832         return NULL;
 833 }
 834 EXPORT_SYMBOL(sptlrpc_get_hash_type);
 835
 836 const char * sptlrpc_get_hash_name(__u8 hash_alg)
 837 {
 838         const struct sptlrpc_hash_type *ht;
 839
 840         ht = sptlrpc_get_hash_type(hash_alg);
 841         if (ht)
 842                 return ht->sht_name;
 843         else
 844                 return "unknown";
 845 }
 846 EXPORT_SYMBOL(sptlrpc_get_hash_name);
 847
 848 int bulk_sec_desc_size(__u8 hash_alg, int request, int read)
 849 {
 850         int size = sizeof(struct ptlrpc_bulk_sec_desc);
 851
 852         LASSERT(hash_alg < BULK_HASH_ALG_MAX);
 853
 854         /* read request don't need extra data */
 855         if (!(read && request))
 856                 size += hash_types[hash_alg].sht_size;
 857
 858         return size;
 859 }
 860 EXPORT_SYMBOL(bulk_sec_desc_size);
 861
 862 int bulk_sec_desc_unpack(struct lustre_msg *msg, int offset)
 863 {
 864         struct ptlrpc_bulk_sec_desc *bsd;
 865         int    size = msg->lm_buflens[offset];
 866
 867         bsd = lustre_msg_buf(msg, offset, sizeof(*bsd));
 868         if (bsd == NULL) {
 869                 CERROR("Invalid bulk sec desc: size %d\n", size);
 870                 return -EINVAL;
 871         }
 872
 873         /* nothing to swab */
 874
 875         if (unlikely(bsd->bsd_version != 0)) {
 876                 CERROR("Unexpected version %u\n", bsd->bsd_version);
 877                 return -EPROTO;
 878         }
 879
 880         if (unlikely(bsd->bsd_flags != 0)) {
 881                 CERROR("Unexpected flags %x\n", bsd->bsd_flags);
 882                 return -EPROTO;
 883         }
 884
 885         if (unlikely(!sptlrpc_get_hash_type(bsd->bsd_hash_alg))) {
 886                 CERROR("Unsupported checksum algorithm %u\n",
 887                        bsd->bsd_hash_alg);
 888                 return -EINVAL;
 889         }
 890
 891         if (unlikely(!sptlrpc_get_ciph_type(bsd->bsd_ciph_alg))) {
 892                 CERROR("Unsupported cipher algorithm %u\n",
 893                        bsd->bsd_ciph_alg);
 894                 return -EINVAL;
 895         }
 896
 897         if (unlikely(size > sizeof(*bsd)) &&
 898             size < sizeof(*bsd) + hash_types[bsd->bsd_hash_alg].sht_size) {
 899                 CERROR("Mal-formed checksum data: csum alg %u, size %d\n",
 900                        bsd->bsd_hash_alg, size);
 901                 return -EINVAL;
 902         }
 903
 904         return 0;
 905 }
 906 EXPORT_SYMBOL(bulk_sec_desc_unpack);
 907
 908 #ifdef __KERNEL__
 909
 910 #ifdef HAVE_ADLER
 911 static int do_bulk_checksum_adler32(struct ptlrpc_bulk_desc *desc, void *buf)
 912 {
 913         struct page    *page;
 914         int             off;
 915         char           *ptr;
 916         __u32           adler32 = 1;
 917         int             len, i;
 918
 919         for (i = 0; i < desc->bd_iov_count; i++) {
 920                 page = desc->bd_iov[i].kiov_page;
 921                 off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK;
 922                 ptr = cfs_kmap(page) + off;
 923                 len = desc->bd_iov[i].kiov_len;
 924
 925                 adler32 = adler32(adler32, ptr, len);
 926
 927                 cfs_kunmap(page);
 928         }
 929
 930         adler32 = cpu_to_le32(adler32);
 931         memcpy(buf, &adler32, sizeof(adler32));
 932         return 0;
 933 }
 934 #endif
 935
 936 static int do_bulk_checksum_crc32(struct ptlrpc_bulk_desc *desc, void *buf)
 937 {
 938         struct page    *page;
 939         int             off;
 940         char           *ptr;
 941         __u32           crc32 = ~0;
 942         int             len, i;
 943
 944         for (i = 0; i < desc->bd_iov_count; i++) {
 945                 page = desc->bd_iov[i].kiov_page;
 946                 off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK;
 947                 ptr = cfs_kmap(page) + off;
 948                 len = desc->bd_iov[i].kiov_len;
 949
 950                 crc32 = crc32_le(crc32, ptr, len);
 951
 952                 cfs_kunmap(page);
 953         }
 954
 955         crc32 = cpu_to_le32(crc32);
 956         memcpy(buf, &crc32, sizeof(crc32));
 957         return 0;
 958 }
 959
 960 static int do_bulk_checksum(struct ptlrpc_bulk_desc *desc, __u32 alg, void *buf)
 961 {
 962         struct hash_desc    hdesc;
 963         struct scatterlist *sl;
 964         int i, rc = 0, bytes = 0;
 965
 966         LASSERT(alg > BULK_HASH_ALG_NULL &&
 967                 alg < BULK_HASH_ALG_MAX);
 968
 969         switch (alg) {
 970         case BULK_HASH_ALG_ADLER32:
 971 #ifdef HAVE_ADLER
 972                 return do_bulk_checksum_adler32(desc, buf);
 973 #else
 974                 CERROR("Adler32 not supported\n");
 975                 return -EINVAL;
 976 #endif
 977         case BULK_HASH_ALG_CRC32:
 978                 return do_bulk_checksum_crc32(desc, buf);
 979         }
 980
 981         hdesc.tfm = ll_crypto_alloc_hash(hash_types[alg].sht_tfm_name, 0, 0);
 982         if (hdesc.tfm == NULL) {
 983                 CERROR("Unable to allocate TFM %s\n", hash_types[alg].sht_name);
 984                 return -ENOMEM;
 985         }
 986         hdesc.flags = 0;
 987
 988         OBD_ALLOC(sl, sizeof(*sl) * desc->bd_iov_count);
 989         if (sl == NULL) {
 990                 rc = -ENOMEM;
 991                 goto out_tfm;
 992         }
 993
 994         for (i = 0; i < desc->bd_iov_count; i++) {
 995                 sl[i].page = desc->bd_iov[i].kiov_page;
 996                 sl[i].offset = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK;
 997                 sl[i].length = desc->bd_iov[i].kiov_len;
 998                 bytes += desc->bd_iov[i].kiov_len;
 999         }
1000
1001         ll_crypto_hash_init(&hdesc);
1002         ll_crypto_hash_update(&hdesc, sl, bytes);
1003         ll_crypto_hash_final(&hdesc, buf);
1004
1005         OBD_FREE(sl, sizeof(*sl) * desc->bd_iov_count);
1006
1007 out_tfm:
1008         ll_crypto_free_hash(hdesc.tfm);
1009         return rc;
1010 }
1011
1012 #else /* !__KERNEL__ */
1013
1014 static int do_bulk_checksum(struct ptlrpc_bulk_desc *desc, __u32 alg, void *buf)
1015 {
1016         __u32   csum32;
1017         int     i;
1018
1019         LASSERT(alg == BULK_HASH_ALG_ADLER32 || alg == BULK_HASH_ALG_CRC32);
1020
1021         if (alg == BULK_HASH_ALG_ADLER32)
1022                 csum32 = 1;
1023         else
1024                 csum32 = ~0;
1025
1026         for (i = 0; i < desc->bd_iov_count; i++) {
1027                 unsigned char *ptr = desc->bd_iov[i].iov_base;
1028                 int len = desc->bd_iov[i].iov_len;
1029
1030                 switch (alg) {
1031                 case BULK_HASH_ALG_ADLER32:
1032 #ifdef HAVE_ADLER
1033                         csum32 = adler32(csum32, ptr, len);
1034 #else
1035                         CERROR("Adler32 not supported\n");
1036                         return -EINVAL;
1037 #endif
1038                         break;
1039                 case BULK_HASH_ALG_CRC32:
1040                         csum32 = crc32_le(csum32, ptr, len);
1041                         break;
1042                 }
1043         }
1044
1045         csum32 = cpu_to_le32(csum32);
1046         memcpy(buf, &csum32, sizeof(csum32));
1047         return 0;
1048 }
1049
1050 #endif /* __KERNEL__ */
1051
1052 /*
1053  * perform algorithm @alg checksum on @desc, store result in @buf.
1054  * if anything goes wrong, leave 'alg' be BULK_HASH_ALG_NULL.
1055  */
1056 static
1057 int generate_bulk_csum(struct ptlrpc_bulk_desc *desc, __u32 alg,
1058                        struct ptlrpc_bulk_sec_desc *bsd, int bsdsize)
1059 {
1060         int rc;
1061
1062         LASSERT(bsd);
1063         LASSERT(alg < BULK_HASH_ALG_MAX);
1064
1065         bsd->bsd_hash_alg = BULK_HASH_ALG_NULL;
1066
1067         if (alg == BULK_HASH_ALG_NULL)
1068                 return 0;
1069
1070         LASSERT(bsdsize >= sizeof(*bsd) + hash_types[alg].sht_size);
1071
1072         rc = do_bulk_checksum(desc, alg, bsd->bsd_csum);
1073         if (rc == 0)
1074                 bsd->bsd_hash_alg = alg;
1075
1076         return rc;
1077 }
1078
1079 static
1080 int verify_bulk_csum(struct ptlrpc_bulk_desc *desc, int read,
1081                      struct ptlrpc_bulk_sec_desc *bsdv, int bsdvsize,
1082                      struct ptlrpc_bulk_sec_desc *bsdr, int bsdrsize)
1083 {
1084         char *csum_p;
1085         char *buf = NULL;
1086         int   csum_size, rc = 0;
1087
1088         LASSERT(bsdv);
1089         LASSERT(bsdv->bsd_hash_alg < BULK_HASH_ALG_MAX);
1090
1091         if (bsdr)
1092                 bsdr->bsd_hash_alg = BULK_HASH_ALG_NULL;
1093
1094         if (bsdv->bsd_hash_alg == BULK_HASH_ALG_NULL)
1095                 return 0;
1096
1097         /* for all supported algorithms */
1098         csum_size = hash_types[bsdv->bsd_hash_alg].sht_size;
1099
1100         if (bsdvsize < sizeof(*bsdv) + csum_size) {
1101                 CERROR("verifier size %d too small, require %d\n",
1102                        bsdvsize, (int) sizeof(*bsdv) + csum_size);
1103                 return -EINVAL;
1104         }
1105
1106         if (bsdr) {
1107                 LASSERT(bsdrsize >= sizeof(*bsdr) + csum_size);
1108                 csum_p = (char *) bsdr->bsd_csum;
1109         } else {
1110                 OBD_ALLOC(buf, csum_size);
1111                 if (buf == NULL)
1112                         return -EINVAL;
1113                 csum_p = buf;
1114         }
1115
1116         rc = do_bulk_checksum(desc, bsdv->bsd_hash_alg, csum_p);
1117
1118         if (memcmp(bsdv->bsd_csum, csum_p, csum_size)) {
1119                 CERROR("BAD %s CHECKSUM (%s), data mutated during "
1120                        "transfer!\n", read ? "READ" : "WRITE",
1121                        hash_types[bsdv->bsd_hash_alg].sht_name);
1122                 rc = -EINVAL;
1123         } else {
1124                 CDEBUG(D_SEC, "bulk %s checksum (%s) verified\n",
1125                       read ? "read" : "write",
1126                       hash_types[bsdv->bsd_hash_alg].sht_name);
1127         }
1128
1129         if (bsdr) {
1130                 bsdr->bsd_hash_alg = bsdv->bsd_hash_alg;
1131                 memcpy(bsdr->bsd_csum, csum_p, csum_size);
1132         } else {
1133                 LASSERT(buf);
1134                 OBD_FREE(buf, csum_size);
1135         }
1136
1137         return rc;
1138 }
1139
1140 int bulk_csum_cli_request(struct ptlrpc_bulk_desc *desc, int read,
1141                           __u32 alg, struct lustre_msg *rmsg, int roff)
1142 {
1143         struct ptlrpc_bulk_sec_desc *bsdr;
1144         int    rsize, rc = 0;
1145
1146         rsize = rmsg->lm_buflens[roff];
1147         bsdr = lustre_msg_buf(rmsg, roff, sizeof(*bsdr));
1148
1149         LASSERT(bsdr);
1150         LASSERT(rsize >= sizeof(*bsdr));
1151         LASSERT(alg < BULK_HASH_ALG_MAX);
1152
1153         if (read) {
1154                 bsdr->bsd_hash_alg = alg;
1155         } else {
1156                 rc = generate_bulk_csum(desc, alg, bsdr, rsize);
1157                 if (rc)
1158                         CERROR("bulk write: client failed to compute "
1159                                "checksum: %d\n", rc);
1160
1161                 /* For sending we only compute the wrong checksum instead
1162                  * of corrupting the data so it is still correct on a redo */
1163                 if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND) &&
1164                     bsdr->bsd_hash_alg != BULK_HASH_ALG_NULL)
1165                         bsdr->bsd_csum[0] ^= 0x1;
1166         }
1167
1168         return rc;
1169 }
1170 EXPORT_SYMBOL(bulk_csum_cli_request);
1171
1172 int bulk_csum_cli_reply(struct ptlrpc_bulk_desc *desc, int read,
1173                         struct lustre_msg *rmsg, int roff,
1174                         struct lustre_msg *vmsg, int voff)
1175 {
1176         struct ptlrpc_bulk_sec_desc *bsdv, *bsdr;
1177         int    rsize, vsize;
1178
1179         rsize = rmsg->lm_buflens[roff];
1180         vsize = vmsg->lm_buflens[voff];
1181         bsdr = lustre_msg_buf(rmsg, roff, 0);
1182         bsdv = lustre_msg_buf(vmsg, voff, 0);
1183
1184         if (bsdv == NULL || vsize < sizeof(*bsdv)) {
1185                 CERROR("Invalid checksum verifier from server: size %d\n",
1186                        vsize);
1187                 return -EINVAL;
1188         }
1189
1190         LASSERT(bsdr);
1191         LASSERT(rsize >= sizeof(*bsdr));
1192         LASSERT(vsize >= sizeof(*bsdv));
1193
1194         if (bsdr->bsd_hash_alg != bsdv->bsd_hash_alg) {
1195                 CERROR("bulk %s: checksum algorithm mismatch: client request "
1196                        "%s but server reply with %s. try to use the new one "
1197                        "for checksum verification\n",
1198                        read ? "read" : "write",
1199                        hash_types[bsdr->bsd_hash_alg].sht_name,
1200                        hash_types[bsdv->bsd_hash_alg].sht_name);
1201         }
1202
1203         if (read)
1204                 return verify_bulk_csum(desc, 1, bsdv, vsize, NULL, 0);
1205         else {
1206                 char *cli, *srv, *new = NULL;
1207                 int csum_size = hash_types[bsdr->bsd_hash_alg].sht_size;
1208
1209                 LASSERT(bsdr->bsd_hash_alg < BULK_HASH_ALG_MAX);
1210                 if (bsdr->bsd_hash_alg == BULK_HASH_ALG_NULL)
1211                         return 0;
1212
1213                 if (vsize < sizeof(*bsdv) + csum_size) {
1214                         CERROR("verifier size %d too small, require %d\n",
1215                                vsize, (int) sizeof(*bsdv) + csum_size);
1216                         return -EINVAL;
1217                 }
1218
1219                 cli = (char *) (bsdr + 1);
1220                 srv = (char *) (bsdv + 1);
1221
1222                 if (!memcmp(cli, srv, csum_size)) {
1223                         /* checksum confirmed */
1224                         CDEBUG(D_SEC, "bulk write checksum (%s) confirmed\n",
1225                                hash_types[bsdr->bsd_hash_alg].sht_name);
1226                         return 0;
1227                 }
1228
1229                 /* checksum mismatch, re-compute a new one and compare with
1230                  * others, give out proper warnings. */
1231                 OBD_ALLOC(new, csum_size);
1232                 if (new == NULL)
1233                         return -ENOMEM;
1234
1235                 do_bulk_checksum(desc, bsdr->bsd_hash_alg, new);
1236
1237                 if (!memcmp(new, srv, csum_size)) {
1238                         CERROR("BAD WRITE CHECKSUM (%s): pages were mutated "
1239                                "on the client after we checksummed them\n",
1240                                hash_types[bsdr->bsd_hash_alg].sht_name);
1241                 } else if (!memcmp(new, cli, csum_size)) {
1242                         CERROR("BAD WRITE CHECKSUM (%s): pages were mutated "
1243                                "in transit\n",
1244                                hash_types[bsdr->bsd_hash_alg].sht_name);
1245                 } else {
1246                         CERROR("BAD WRITE CHECKSUM (%s): pages were mutated "
1247                                "in transit, and the current page contents "
1248                                "don't match the originals and what the server "
1249                                "received\n",
1250                                hash_types[bsdr->bsd_hash_alg].sht_name);
1251                 }
1252                 OBD_FREE(new, csum_size);
1253
1254                 return -EINVAL;
1255         }
1256 }
1257 EXPORT_SYMBOL(bulk_csum_cli_reply);
1258
1259 #ifdef __KERNEL__
1260 static void corrupt_bulk_data(struct ptlrpc_bulk_desc *desc)
1261 {
1262         char           *ptr;
1263         unsigned int    off, i;
1264
1265         for (i = 0; i < desc->bd_iov_count; i++) {
1266                 if (desc->bd_iov[i].kiov_len == 0)
1267                         continue;
1268
1269                 ptr = cfs_kmap(desc->bd_iov[i].kiov_page);
1270                 off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK;
1271                 ptr[off] ^= 0x1;
1272                 cfs_kunmap(desc->bd_iov[i].kiov_page);
1273                 return;
1274         }
1275 }
1276 #else
1277 static void corrupt_bulk_data(struct ptlrpc_bulk_desc *desc)
1278 {
1279 }
1280 #endif /* __KERNEL__ */
1281
1282 int bulk_csum_svc(struct ptlrpc_bulk_desc *desc, int read,
1283                   struct ptlrpc_bulk_sec_desc *bsdv, int vsize,
1284                   struct ptlrpc_bulk_sec_desc *bsdr, int rsize)
1285 {
1286         int    rc;
1287
1288         LASSERT(vsize >= sizeof(*bsdv));
1289         LASSERT(rsize >= sizeof(*bsdr));
1290         LASSERT(bsdv && bsdr);
1291
1292         if (read) {
1293                 rc = generate_bulk_csum(desc, bsdv->bsd_hash_alg, bsdr, rsize);
1294                 if (rc)
1295                         CERROR("bulk read: server failed to generate %s "
1296                                "checksum: %d\n",
1297                                hash_types[bsdv->bsd_hash_alg].sht_name, rc);
1298
1299                 /* corrupt the data after we compute the checksum, to
1300                  * simulate an OST->client data error */
1301                 if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE))
1302                         corrupt_bulk_data(desc);
1303         } else {
1304                 rc = verify_bulk_csum(desc, 0, bsdv, vsize, bsdr, rsize);
1305         }
1306
1307         return rc;
1308 }
1309 EXPORT_SYMBOL(bulk_csum_svc);
1310
1311 /****************************************
1312  * Helpers to assist policy modules to  *
1313  * implement encryption funcationality  *
1314  ****************************************/
1315
1316 /* FIXME */
1317 #ifndef __KERNEL__
1318 #define CRYPTO_TFM_MODE_ECB     (0)
1319 #define CRYPTO_TFM_MODE_CBC     (1)
1320 #endif
1321
1322 static struct sptlrpc_ciph_type cipher_types[] = {
1323         [BULK_CIPH_ALG_NULL]    = {
1324                 "null",         "null",       0,                   0,  0
1325         },
1326         [BULK_CIPH_ALG_ARC4]    = {
1327                 "arc4",         "ecb(arc4)",       0, 0,  16
1328         },
1329         [BULK_CIPH_ALG_AES128]  = {
1330                 "aes128",       "cbc(aes)",        0, 16, 16
1331         },
1332         [BULK_CIPH_ALG_AES192]  = {
1333                 "aes192",       "cbc(aes)",        0, 16, 24
1334         },
1335         [BULK_CIPH_ALG_AES256]  = {
1336                 "aes256",       "cbc(aes)",        0, 16, 32
1337         },
1338         [BULK_CIPH_ALG_CAST128] = {
1339                 "cast128",      "cbc(cast5)",      0, 8,  16
1340         },
1341         [BULK_CIPH_ALG_CAST256] = {
1342                 "cast256",      "cbc(cast6)",      0, 16, 32
1343         },
1344         [BULK_CIPH_ALG_TWOFISH128] = {
1345                 "twofish128",   "cbc(twofish)",    0, 16, 16
1346         },
1347         [BULK_CIPH_ALG_TWOFISH256] = {
1348                 "twofish256",   "cbc(twofish)",    0, 16, 32
1349         },
1350 };
1351
1352 const struct sptlrpc_ciph_type *sptlrpc_get_ciph_type(__u8 ciph_alg)
1353 {
1354         struct sptlrpc_ciph_type *ct;
1355
1356         if (ciph_alg < BULK_CIPH_ALG_MAX) {
1357                 ct = &cipher_types[ciph_alg];
1358                 if (ct->sct_tfm_name)
1359                         return ct;
1360         }
1361         return NULL;
1362 }
1363 EXPORT_SYMBOL(sptlrpc_get_ciph_type);
1364
1365 const char *sptlrpc_get_ciph_name(__u8 ciph_alg)
1366 {
1367         const struct sptlrpc_ciph_type *ct;
1368
1369         ct = sptlrpc_get_ciph_type(ciph_alg);
1370         if (ct)
1371                 return ct->sct_name;
1372         else
1373                 return "unknown";
1374 }
1375 EXPORT_SYMBOL(sptlrpc_get_ciph_name);