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 [sun.com URL with a
  20  * copy of GPLv2].
  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(TASK_UNINTERRUPTIBLE);
 546                         cfs_waitlink_init(&waitlink);
 547                         cfs_waitq_add(&page_pools.epp_waitq, &waitlink);
 548
 549                         spin_unlock(&page_pools.epp_lock);
 550                         cfs_schedule();
 551                         spin_lock(&page_pools.epp_lock);
 552
 553                         LASSERT(page_pools.epp_waitqlen > 0);
 554                         page_pools.epp_waitqlen--;
 555                 }
 556
 557                 LASSERT(page_pools.epp_pages_short >= desc->bd_max_iov);
 558                 page_pools.epp_pages_short -= desc->bd_max_iov;
 559
 560                 this_idle = 0;
 561                 goto again;
 562         }
 563
 564         /* record max wait time */
 565         if (unlikely(tick != 0)) {
 566                 tick = cfs_time_current() - tick;
 567                 if (tick > page_pools.epp_st_max_wait)
 568                         page_pools.epp_st_max_wait = tick;
 569         }
 570
 571         /* proceed with rest of allocation */
 572         page_pools.epp_free_pages -= desc->bd_max_iov;
 573
 574         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 575         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 576
 577         for (i = 0; i < desc->bd_max_iov; i++) {
 578                 LASSERT(page_pools.epp_pools[p_idx][g_idx] != NULL);
 579                 desc->bd_enc_pages[i] = page_pools.epp_pools[p_idx][g_idx];
 580                 page_pools.epp_pools[p_idx][g_idx] = NULL;
 581
 582                 if (++g_idx == PAGES_PER_POOL) {
 583                         p_idx++;
 584                         g_idx = 0;
 585                 }
 586         }
 587
 588         if (page_pools.epp_free_pages < page_pools.epp_st_lowfree)
 589                 page_pools.epp_st_lowfree = page_pools.epp_free_pages;
 590
 591         /*
 592          * new idle index = (old * weight + new) / (weight + 1)
 593          */
 594         if (this_idle == -1) {
 595                 this_idle = page_pools.epp_free_pages * IDLE_IDX_MAX /
 596                             page_pools.epp_total_pages;
 597         }
 598         page_pools.epp_idle_idx = (page_pools.epp_idle_idx * IDLE_IDX_WEIGHT +
 599                                    this_idle) /
 600                                   (IDLE_IDX_WEIGHT + 1);
 601
 602         page_pools.epp_last_access = cfs_time_current_sec();
 603
 604         spin_unlock(&page_pools.epp_lock);
 605         return 0;
 606 }
 607 EXPORT_SYMBOL(sptlrpc_enc_pool_get_pages);
 608
 609 void sptlrpc_enc_pool_put_pages(struct ptlrpc_bulk_desc *desc)
 610 {
 611         int     p_idx, g_idx;
 612         int     i;
 613
 614         if (desc->bd_enc_pages == NULL)
 615                 return;
 616         if (desc->bd_max_iov == 0)
 617                 return;
 618
 619         spin_lock(&page_pools.epp_lock);
 620
 621         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 622         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 623
 624         LASSERT(page_pools.epp_free_pages + desc->bd_max_iov <=
 625                 page_pools.epp_total_pages);
 626         LASSERT(page_pools.epp_pools[p_idx]);
 627
 628         for (i = 0; i < desc->bd_max_iov; i++) {
 629                 LASSERT(desc->bd_enc_pages[i] != NULL);
 630                 LASSERT(g_idx != 0 || page_pools.epp_pools[p_idx]);
 631                 LASSERT(page_pools.epp_pools[p_idx][g_idx] == NULL);
 632
 633                 page_pools.epp_pools[p_idx][g_idx] = desc->bd_enc_pages[i];
 634
 635                 if (++g_idx == PAGES_PER_POOL) {
 636                         p_idx++;
 637                         g_idx = 0;
 638                 }
 639         }
 640
 641         page_pools.epp_free_pages += desc->bd_max_iov;
 642
 643         enc_pools_wakeup();
 644
 645         spin_unlock(&page_pools.epp_lock);
 646
 647         OBD_FREE(desc->bd_enc_pages,
 648                  desc->bd_max_iov * sizeof(*desc->bd_enc_pages));
 649         desc->bd_enc_pages = NULL;
 650 }
 651 EXPORT_SYMBOL(sptlrpc_enc_pool_put_pages);
 652
 653 /*
 654  * we don't do much stuff for add_user/del_user anymore, except adding some
 655  * initial pages in add_user() if current pools are empty, rest would be
 656  * handled by the pools's self-adaption.
 657  */
 658 int sptlrpc_enc_pool_add_user(void)
 659 {
 660         int     need_grow = 0;
 661
 662         spin_lock(&page_pools.epp_lock);
 663         if (page_pools.epp_growing == 0 && page_pools.epp_total_pages == 0) {
 664                 page_pools.epp_growing = 1;
 665                 need_grow = 1;
 666         }
 667         spin_unlock(&page_pools.epp_lock);
 668
 669         if (need_grow) {
 670                 enc_pools_add_pages(PTLRPC_MAX_BRW_PAGES);
 671
 672                 spin_lock(&page_pools.epp_lock);
 673                 page_pools.epp_growing = 0;
 674                 enc_pools_wakeup();
 675                 spin_unlock(&page_pools.epp_lock);
 676         }
 677         return 0;
 678 }
 679 EXPORT_SYMBOL(sptlrpc_enc_pool_add_user);
 680
 681 int sptlrpc_enc_pool_del_user(void)
 682 {
 683         return 0;
 684 }
 685 EXPORT_SYMBOL(sptlrpc_enc_pool_del_user);
 686
 687 static inline void enc_pools_alloc(void)
 688 {
 689         LASSERT(page_pools.epp_max_pools);
 690         /*
 691          * on system with huge memory but small page size, this might lead to
 692          * high-order allocation. but it's not common, and we suppose memory
 693          * be not too much fragmented at module loading time.
 694          */
 695         OBD_ALLOC(page_pools.epp_pools,
 696                   page_pools.epp_max_pools * sizeof(*page_pools.epp_pools));
 697 }
 698
 699 static inline void enc_pools_free(void)
 700 {
 701         LASSERT(page_pools.epp_max_pools);
 702         LASSERT(page_pools.epp_pools);
 703
 704         OBD_FREE(page_pools.epp_pools,
 705                  page_pools.epp_max_pools * sizeof(*page_pools.epp_pools));
 706 }
 707
 708 int sptlrpc_enc_pool_init(void)
 709 {
 710         /*
 711          * maximum capacity is 1/8 of total physical memory.
 712          * is the 1/8 a good number?
 713          */
 714         page_pools.epp_max_pages = num_physpages / 8;
 715         page_pools.epp_max_pools = npages_to_npools(page_pools.epp_max_pages);
 716
 717         cfs_waitq_init(&page_pools.epp_waitq);
 718         page_pools.epp_waitqlen = 0;
 719         page_pools.epp_pages_short = 0;
 720
 721         page_pools.epp_growing = 0;
 722
 723         page_pools.epp_idle_idx = 0;
 724         page_pools.epp_last_shrink = cfs_time_current_sec();
 725         page_pools.epp_last_access = cfs_time_current_sec();
 726
 727         spin_lock_init(&page_pools.epp_lock);
 728         page_pools.epp_total_pages = 0;
 729         page_pools.epp_free_pages = 0;
 730
 731         page_pools.epp_st_max_pages = 0;
 732         page_pools.epp_st_grows = 0;
 733         page_pools.epp_st_grow_fails = 0;
 734         page_pools.epp_st_shrinks = 0;
 735         page_pools.epp_st_access = 0;
 736         page_pools.epp_st_missings = 0;
 737         page_pools.epp_st_lowfree = 0;
 738         page_pools.epp_st_max_wqlen = 0;
 739         page_pools.epp_st_max_wait = 0;
 740
 741         enc_pools_alloc();
 742         if (page_pools.epp_pools == NULL)
 743                 return -ENOMEM;
 744
 745         pools_shrinker = set_shrinker(pools_shrinker_seeks, enc_pools_shrink);
 746         if (pools_shrinker == NULL) {
 747                 enc_pools_free();
 748                 return -ENOMEM;
 749         }
 750
 751         return 0;
 752 }
 753
 754 void sptlrpc_enc_pool_fini(void)
 755 {
 756         unsigned long cleaned, npools;
 757
 758         LASSERT(pools_shrinker);
 759         LASSERT(page_pools.epp_pools);
 760         LASSERT(page_pools.epp_total_pages == page_pools.epp_free_pages);
 761
 762         remove_shrinker(pools_shrinker);
 763
 764         npools = npages_to_npools(page_pools.epp_total_pages);
 765         cleaned = enc_pools_cleanup(page_pools.epp_pools, npools);
 766         LASSERT(cleaned == page_pools.epp_total_pages);
 767
 768         enc_pools_free();
 769
 770         if (page_pools.epp_st_access > 0) {
 771                 CWARN("max pages %lu, grows %u, grow fails %u, shrinks %u, "
 772                       "access %lu, missing %lu, max qlen %u, max wait "
 773                       CFS_TIME_T"/%d\n",
 774                       page_pools.epp_st_max_pages, page_pools.epp_st_grows,
 775                       page_pools.epp_st_grow_fails,
 776                       page_pools.epp_st_shrinks, page_pools.epp_st_access,
 777                       page_pools.epp_st_missings, page_pools.epp_st_max_wqlen,
 778                       page_pools.epp_st_max_wait, HZ);
 779         }
 780 }
 781
 782 #else /* !__KERNEL__ */
 783
 784 int sptlrpc_enc_pool_get_pages(struct ptlrpc_bulk_desc *desc)
 785 {
 786         return 0;
 787 }
 788
 789 void sptlrpc_enc_pool_put_pages(struct ptlrpc_bulk_desc *desc)
 790 {
 791 }
 792
 793 int sptlrpc_enc_pool_init(void)
 794 {
 795         return 0;
 796 }
 797
 798 void sptlrpc_enc_pool_fini(void)
 799 {
 800 }
 801 #endif
 802
 803 /****************************************
 804  * Helpers to assist policy modules to  *
 805  * implement checksum funcationality    *
 806  ****************************************/
 807
 808 static struct sptlrpc_hash_type hash_types[] = {
 809         [BULK_HASH_ALG_NULL]    = { "null",     "null",         0 },
 810         [BULK_HASH_ALG_ADLER32] = { "adler32",  "adler32",      4 },
 811         [BULK_HASH_ALG_CRC32]   = { "crc32",    "crc32",        4 },
 812         [BULK_HASH_ALG_MD5]     = { "md5",      "md5",          16 },
 813         [BULK_HASH_ALG_SHA1]    = { "sha1",     "sha1",         20 },
 814         [BULK_HASH_ALG_SHA256]  = { "sha256",   "sha256",       32 },
 815         [BULK_HASH_ALG_SHA384]  = { "sha384",   "sha384",       48 },
 816         [BULK_HASH_ALG_SHA512]  = { "sha512",   "sha512",       64 },
 817         [BULK_HASH_ALG_WP256]   = { "wp256",    "wp256",        32 },
 818         [BULK_HASH_ALG_WP384]   = { "wp384",    "wp384",        48 },
 819         [BULK_HASH_ALG_WP512]   = { "wp512",    "wp512",        64 },
 820 };
 821
 822 const struct sptlrpc_hash_type *sptlrpc_get_hash_type(__u8 hash_alg)
 823 {
 824         struct sptlrpc_hash_type *ht;
 825
 826         if (hash_alg < BULK_HASH_ALG_MAX) {
 827                 ht = &hash_types[hash_alg];
 828                 if (ht->sht_tfm_name)
 829                         return ht;
 830         }
 831         return NULL;
 832 }
 833 EXPORT_SYMBOL(sptlrpc_get_hash_type);
 834
 835 const char * sptlrpc_get_hash_name(__u8 hash_alg)
 836 {
 837         const struct sptlrpc_hash_type *ht;
 838
 839         ht = sptlrpc_get_hash_type(hash_alg);
 840         if (ht)
 841                 return ht->sht_name;
 842         else
 843                 return "unknown";
 844 }
 845 EXPORT_SYMBOL(sptlrpc_get_hash_name);
 846
 847 int bulk_sec_desc_size(__u8 hash_alg, int request, int read)
 848 {
 849         int size = sizeof(struct ptlrpc_bulk_sec_desc);
 850
 851         LASSERT(hash_alg < BULK_HASH_ALG_MAX);
 852
 853         /* read request don't need extra data */
 854         if (!(read && request))
 855                 size += hash_types[hash_alg].sht_size;
 856
 857         return size;
 858 }
 859 EXPORT_SYMBOL(bulk_sec_desc_size);
 860
 861 int bulk_sec_desc_unpack(struct lustre_msg *msg, int offset)
 862 {
 863         struct ptlrpc_bulk_sec_desc *bsd;
 864         int    size = msg->lm_buflens[offset];
 865
 866         bsd = lustre_msg_buf(msg, offset, sizeof(*bsd));
 867         if (bsd == NULL) {
 868                 CERROR("Invalid bulk sec desc: size %d\n", size);
 869                 return -EINVAL;
 870         }
 871
 872         /* nothing to swab */
 873
 874         if (unlikely(bsd->bsd_version != 0)) {
 875                 CERROR("Unexpected version %u\n", bsd->bsd_version);
 876                 return -EPROTO;
 877         }
 878
 879         if (unlikely(bsd->bsd_flags != 0)) {
 880                 CERROR("Unexpected flags %x\n", bsd->bsd_flags);
 881                 return -EPROTO;
 882         }
 883
 884         if (unlikely(!sptlrpc_get_hash_type(bsd->bsd_hash_alg))) {
 885                 CERROR("Unsupported checksum algorithm %u\n",
 886                        bsd->bsd_hash_alg);
 887                 return -EINVAL;
 888         }
 889
 890         if (unlikely(!sptlrpc_get_ciph_type(bsd->bsd_ciph_alg))) {
 891                 CERROR("Unsupported cipher algorithm %u\n",
 892                        bsd->bsd_ciph_alg);
 893                 return -EINVAL;
 894         }
 895
 896         if (unlikely(size > sizeof(*bsd)) &&
 897             size < sizeof(*bsd) + hash_types[bsd->bsd_hash_alg].sht_size) {
 898                 CERROR("Mal-formed checksum data: csum alg %u, size %d\n",
 899                        bsd->bsd_hash_alg, size);
 900                 return -EINVAL;
 901         }
 902
 903         return 0;
 904 }
 905 EXPORT_SYMBOL(bulk_sec_desc_unpack);
 906
 907 #ifdef __KERNEL__
 908
 909 #ifdef HAVE_ADLER
 910 static int do_bulk_checksum_adler32(struct ptlrpc_bulk_desc *desc, void *buf)
 911 {
 912         struct page    *page;
 913         int             off;
 914         char           *ptr;
 915         __u32           adler32 = 1;
 916         int             len, i;
 917
 918         for (i = 0; i < desc->bd_iov_count; i++) {
 919                 page = desc->bd_iov[i].kiov_page;
 920                 off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK;
 921                 ptr = cfs_kmap(page) + off;
 922                 len = desc->bd_iov[i].kiov_len;
 923
 924                 adler32 = adler32(adler32, ptr, len);
 925
 926                 cfs_kunmap(page);
 927         }
 928
 929         adler32 = cpu_to_le32(adler32);
 930         memcpy(buf, &adler32, sizeof(adler32));
 931         return 0;
 932 }
 933 #endif
 934
 935 static int do_bulk_checksum_crc32(struct ptlrpc_bulk_desc *desc, void *buf)
 936 {
 937         struct page    *page;
 938         int             off;
 939         char           *ptr;
 940         __u32           crc32 = ~0;
 941         int             len, i;
 942
 943         for (i = 0; i < desc->bd_iov_count; i++) {
 944                 page = desc->bd_iov[i].kiov_page;
 945                 off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK;
 946                 ptr = cfs_kmap(page) + off;
 947                 len = desc->bd_iov[i].kiov_len;
 948
 949                 crc32 = crc32_le(crc32, ptr, len);
 950
 951                 cfs_kunmap(page);
 952         }
 953
 954         crc32 = cpu_to_le32(crc32);
 955         memcpy(buf, &crc32, sizeof(crc32));
 956         return 0;
 957 }
 958
 959 static int do_bulk_checksum(struct ptlrpc_bulk_desc *desc, __u32 alg, void *buf)
 960 {
 961         struct hash_desc    hdesc;
 962         struct scatterlist *sl;
 963         int i, rc = 0, bytes = 0;
 964
 965         LASSERT(alg > BULK_HASH_ALG_NULL &&
 966                 alg < BULK_HASH_ALG_MAX);
 967
 968         switch (alg) {
 969         case BULK_HASH_ALG_ADLER32:
 970 #ifdef HAVE_ADLER
 971                 return do_bulk_checksum_adler32(desc, buf);
 972 #else
 973                 CERROR("Adler32 not supported\n");
 974                 return -EINVAL;
 975 #endif
 976         case BULK_HASH_ALG_CRC32:
 977                 return do_bulk_checksum_crc32(desc, buf);
 978         }
 979
 980         hdesc.tfm = ll_crypto_alloc_hash(hash_types[alg].sht_tfm_name, 0, 0);
 981         if (hdesc.tfm == NULL) {
 982                 CERROR("Unable to allocate TFM %s\n", hash_types[alg].sht_name);
 983                 return -ENOMEM;
 984         }
 985         hdesc.flags = 0;
 986
 987         OBD_ALLOC(sl, sizeof(*sl) * desc->bd_iov_count);
 988         if (sl == NULL) {
 989                 rc = -ENOMEM;
 990                 goto out_tfm;
 991         }
 992
 993         for (i = 0; i < desc->bd_iov_count; i++) {
 994                 sl[i].page = desc->bd_iov[i].kiov_page;
 995                 sl[i].offset = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK;
 996                 sl[i].length = desc->bd_iov[i].kiov_len;
 997                 bytes += desc->bd_iov[i].kiov_len;
 998         }
 999
1000         ll_crypto_hash_init(&hdesc);
1001         ll_crypto_hash_update(&hdesc, sl, bytes);
1002         ll_crypto_hash_final(&hdesc, buf);
1003
1004         OBD_FREE(sl, sizeof(*sl) * desc->bd_iov_count);
1005
1006 out_tfm:
1007         ll_crypto_free_hash(hdesc.tfm);
1008         return rc;
1009 }
1010
1011 #else /* !__KERNEL__ */
1012
1013 static int do_bulk_checksum(struct ptlrpc_bulk_desc *desc, __u32 alg, void *buf)
1014 {
1015         __u32   csum32;
1016         int     i;
1017
1018         LASSERT(alg == BULK_HASH_ALG_ADLER32 || alg == BULK_HASH_ALG_CRC32);
1019
1020         if (alg == BULK_HASH_ALG_ADLER32)
1021                 csum32 = 1;
1022         else
1023                 csum32 = ~0;
1024
1025         for (i = 0; i < desc->bd_iov_count; i++) {
1026                 unsigned char *ptr = desc->bd_iov[i].iov_base;
1027                 int len = desc->bd_iov[i].iov_len;
1028
1029                 switch (alg) {
1030                 case BULK_HASH_ALG_ADLER32:
1031 #ifdef HAVE_ADLER
1032                         csum32 = adler32(csum32, ptr, len);
1033 #else
1034                         CERROR("Adler32 not supported\n");
1035                         return -EINVAL;
1036 #endif
1037                         break;
1038                 case BULK_HASH_ALG_CRC32:
1039                         csum32 = crc32_le(csum32, ptr, len);
1040                         break;
1041                 }
1042         }
1043
1044         csum32 = cpu_to_le32(csum32);
1045         memcpy(buf, &csum32, sizeof(csum32));
1046         return 0;
1047 }
1048
1049 #endif /* __KERNEL__ */
1050
1051 /*
1052  * perform algorithm @alg checksum on @desc, store result in @buf.
1053  * if anything goes wrong, leave 'alg' be BULK_HASH_ALG_NULL.
1054  */
1055 static
1056 int generate_bulk_csum(struct ptlrpc_bulk_desc *desc, __u32 alg,
1057                        struct ptlrpc_bulk_sec_desc *bsd, int bsdsize)
1058 {
1059         int rc;
1060
1061         LASSERT(bsd);
1062         LASSERT(alg < BULK_HASH_ALG_MAX);
1063
1064         bsd->bsd_hash_alg = BULK_HASH_ALG_NULL;
1065
1066         if (alg == BULK_HASH_ALG_NULL)
1067                 return 0;
1068
1069         LASSERT(bsdsize >= sizeof(*bsd) + hash_types[alg].sht_size);
1070
1071         rc = do_bulk_checksum(desc, alg, bsd->bsd_csum);
1072         if (rc == 0)
1073                 bsd->bsd_hash_alg = alg;
1074
1075         return rc;
1076 }
1077
1078 static
1079 int verify_bulk_csum(struct ptlrpc_bulk_desc *desc, int read,
1080                      struct ptlrpc_bulk_sec_desc *bsdv, int bsdvsize,
1081                      struct ptlrpc_bulk_sec_desc *bsdr, int bsdrsize)
1082 {
1083         char *csum_p;
1084         char *buf = NULL;
1085         int   csum_size, rc = 0;
1086
1087         LASSERT(bsdv);
1088         LASSERT(bsdv->bsd_hash_alg < BULK_HASH_ALG_MAX);
1089
1090         if (bsdr)
1091                 bsdr->bsd_hash_alg = BULK_HASH_ALG_NULL;
1092
1093         if (bsdv->bsd_hash_alg == BULK_HASH_ALG_NULL)
1094                 return 0;
1095
1096         /* for all supported algorithms */
1097         csum_size = hash_types[bsdv->bsd_hash_alg].sht_size;
1098
1099         if (bsdvsize < sizeof(*bsdv) + csum_size) {
1100                 CERROR("verifier size %d too small, require %d\n",
1101                        bsdvsize, (int) sizeof(*bsdv) + csum_size);
1102                 return -EINVAL;
1103         }
1104
1105         if (bsdr) {
1106                 LASSERT(bsdrsize >= sizeof(*bsdr) + csum_size);
1107                 csum_p = (char *) bsdr->bsd_csum;
1108         } else {
1109                 OBD_ALLOC(buf, csum_size);
1110                 if (buf == NULL)
1111                         return -EINVAL;
1112                 csum_p = buf;
1113         }
1114
1115         rc = do_bulk_checksum(desc, bsdv->bsd_hash_alg, csum_p);
1116
1117         if (memcmp(bsdv->bsd_csum, csum_p, csum_size)) {
1118                 CERROR("BAD %s CHECKSUM (%s), data mutated during "
1119                        "transfer!\n", read ? "READ" : "WRITE",
1120                        hash_types[bsdv->bsd_hash_alg].sht_name);
1121                 rc = -EINVAL;
1122         } else {
1123                 CDEBUG(D_SEC, "bulk %s checksum (%s) verified\n",
1124                       read ? "read" : "write",
1125                       hash_types[bsdv->bsd_hash_alg].sht_name);
1126         }
1127
1128         if (bsdr) {
1129                 bsdr->bsd_hash_alg = bsdv->bsd_hash_alg;
1130                 memcpy(bsdr->bsd_csum, csum_p, csum_size);
1131         } else {
1132                 LASSERT(buf);
1133                 OBD_FREE(buf, csum_size);
1134         }
1135
1136         return rc;
1137 }
1138
1139 int bulk_csum_cli_request(struct ptlrpc_bulk_desc *desc, int read,
1140                           __u32 alg, struct lustre_msg *rmsg, int roff)
1141 {
1142         struct ptlrpc_bulk_sec_desc *bsdr;
1143         int    rsize, rc = 0;
1144
1145         rsize = rmsg->lm_buflens[roff];
1146         bsdr = lustre_msg_buf(rmsg, roff, sizeof(*bsdr));
1147
1148         LASSERT(bsdr);
1149         LASSERT(rsize >= sizeof(*bsdr));
1150         LASSERT(alg < BULK_HASH_ALG_MAX);
1151
1152         if (read) {
1153                 bsdr->bsd_hash_alg = alg;
1154         } else {
1155                 rc = generate_bulk_csum(desc, alg, bsdr, rsize);
1156                 if (rc)
1157                         CERROR("bulk write: client failed to compute "
1158                                "checksum: %d\n", rc);
1159
1160                 /* For sending we only compute the wrong checksum instead
1161                  * of corrupting the data so it is still correct on a redo */
1162                 if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND) &&
1163                     bsdr->bsd_hash_alg != BULK_HASH_ALG_NULL)
1164                         bsdr->bsd_csum[0] ^= 0x1;
1165         }
1166
1167         return rc;
1168 }
1169 EXPORT_SYMBOL(bulk_csum_cli_request);
1170
1171 int bulk_csum_cli_reply(struct ptlrpc_bulk_desc *desc, int read,
1172                         struct lustre_msg *rmsg, int roff,
1173                         struct lustre_msg *vmsg, int voff)
1174 {
1175         struct ptlrpc_bulk_sec_desc *bsdv, *bsdr;
1176         int    rsize, vsize;
1177
1178         rsize = rmsg->lm_buflens[roff];
1179         vsize = vmsg->lm_buflens[voff];
1180         bsdr = lustre_msg_buf(rmsg, roff, 0);
1181         bsdv = lustre_msg_buf(vmsg, voff, 0);
1182
1183         if (bsdv == NULL || vsize < sizeof(*bsdv)) {
1184                 CERROR("Invalid checksum verifier from server: size %d\n",
1185                        vsize);
1186                 return -EINVAL;
1187         }
1188
1189         LASSERT(bsdr);
1190         LASSERT(rsize >= sizeof(*bsdr));
1191         LASSERT(vsize >= sizeof(*bsdv));
1192
1193         if (bsdr->bsd_hash_alg != bsdv->bsd_hash_alg) {
1194                 CERROR("bulk %s: checksum algorithm mismatch: client request "
1195                        "%s but server reply with %s. try to use the new one "
1196                        "for checksum verification\n",
1197                        read ? "read" : "write",
1198                        hash_types[bsdr->bsd_hash_alg].sht_name,
1199                        hash_types[bsdv->bsd_hash_alg].sht_name);
1200         }
1201
1202         if (read)
1203                 return verify_bulk_csum(desc, 1, bsdv, vsize, NULL, 0);
1204         else {
1205                 char *cli, *srv, *new = NULL;
1206                 int csum_size = hash_types[bsdr->bsd_hash_alg].sht_size;
1207
1208                 LASSERT(bsdr->bsd_hash_alg < BULK_HASH_ALG_MAX);
1209                 if (bsdr->bsd_hash_alg == BULK_HASH_ALG_NULL)
1210                         return 0;
1211
1212                 if (vsize < sizeof(*bsdv) + csum_size) {
1213                         CERROR("verifier size %d too small, require %d\n",
1214                                vsize, (int) sizeof(*bsdv) + csum_size);
1215                         return -EINVAL;
1216                 }
1217
1218                 cli = (char *) (bsdr + 1);
1219                 srv = (char *) (bsdv + 1);
1220
1221                 if (!memcmp(cli, srv, csum_size)) {
1222                         /* checksum confirmed */
1223                         CDEBUG(D_SEC, "bulk write checksum (%s) confirmed\n",
1224                                hash_types[bsdr->bsd_hash_alg].sht_name);
1225                         return 0;
1226                 }
1227
1228                 /* checksum mismatch, re-compute a new one and compare with
1229                  * others, give out proper warnings. */
1230                 OBD_ALLOC(new, csum_size);
1231                 if (new == NULL)
1232                         return -ENOMEM;
1233
1234                 do_bulk_checksum(desc, bsdr->bsd_hash_alg, new);
1235
1236                 if (!memcmp(new, srv, csum_size)) {
1237                         CERROR("BAD WRITE CHECKSUM (%s): pages were mutated "
1238                                "on the client after we checksummed them\n",
1239                                hash_types[bsdr->bsd_hash_alg].sht_name);
1240                 } else if (!memcmp(new, cli, csum_size)) {
1241                         CERROR("BAD WRITE CHECKSUM (%s): pages were mutated "
1242                                "in transit\n",
1243                                hash_types[bsdr->bsd_hash_alg].sht_name);
1244                 } else {
1245                         CERROR("BAD WRITE CHECKSUM (%s): pages were mutated "
1246                                "in transit, and the current page contents "
1247                                "don't match the originals and what the server "
1248                                "received\n",
1249                                hash_types[bsdr->bsd_hash_alg].sht_name);
1250                 }
1251                 OBD_FREE(new, csum_size);
1252
1253                 return -EINVAL;
1254         }
1255 }
1256 EXPORT_SYMBOL(bulk_csum_cli_reply);
1257
1258 #ifdef __KERNEL__
1259 static void corrupt_bulk_data(struct ptlrpc_bulk_desc *desc)
1260 {
1261         char           *ptr;
1262         unsigned int    off, i;
1263
1264         for (i = 0; i < desc->bd_iov_count; i++) {
1265                 if (desc->bd_iov[i].kiov_len == 0)
1266                         continue;
1267
1268                 ptr = cfs_kmap(desc->bd_iov[i].kiov_page);
1269                 off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK;
1270                 ptr[off] ^= 0x1;
1271                 cfs_kunmap(desc->bd_iov[i].kiov_page);
1272                 return;
1273         }
1274 }
1275 #else
1276 static void corrupt_bulk_data(struct ptlrpc_bulk_desc *desc)
1277 {
1278 }
1279 #endif /* __KERNEL__ */
1280
1281 int bulk_csum_svc(struct ptlrpc_bulk_desc *desc, int read,
1282                   struct ptlrpc_bulk_sec_desc *bsdv, int vsize,
1283                   struct ptlrpc_bulk_sec_desc *bsdr, int rsize)
1284 {
1285         int    rc;
1286
1287         LASSERT(vsize >= sizeof(*bsdv));
1288         LASSERT(rsize >= sizeof(*bsdr));
1289         LASSERT(bsdv && bsdr);
1290
1291         if (read) {
1292                 rc = generate_bulk_csum(desc, bsdv->bsd_hash_alg, bsdr, rsize);
1293                 if (rc)
1294                         CERROR("bulk read: server failed to generate %s "
1295                                "checksum: %d\n",
1296                                hash_types[bsdv->bsd_hash_alg].sht_name, rc);
1297
1298                 /* corrupt the data after we compute the checksum, to
1299                  * simulate an OST->client data error */
1300                 if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE))
1301                         corrupt_bulk_data(desc);
1302         } else {
1303                 rc = verify_bulk_csum(desc, 0, bsdv, vsize, bsdr, rsize);
1304         }
1305
1306         return rc;
1307 }
1308 EXPORT_SYMBOL(bulk_csum_svc);
1309
1310 /****************************************
1311  * Helpers to assist policy modules to  *
1312  * implement encryption funcationality  *
1313  ****************************************/
1314
1315 /* FIXME */
1316 #ifndef __KERNEL__
1317 #define CRYPTO_TFM_MODE_ECB     (0)
1318 #define CRYPTO_TFM_MODE_CBC     (1)
1319 #endif
1320
1321 static struct sptlrpc_ciph_type cipher_types[] = {
1322         [BULK_CIPH_ALG_NULL]    = {
1323                 "null",         "null",       0,                   0,  0
1324         },
1325         [BULK_CIPH_ALG_ARC4]    = {
1326                 "arc4",         "ecb(arc4)",       0, 0,  16
1327         },
1328         [BULK_CIPH_ALG_AES128]  = {
1329                 "aes128",       "cbc(aes)",        0, 16, 16
1330         },
1331         [BULK_CIPH_ALG_AES192]  = {
1332                 "aes192",       "cbc(aes)",        0, 16, 24
1333         },
1334         [BULK_CIPH_ALG_AES256]  = {
1335                 "aes256",       "cbc(aes)",        0, 16, 32
1336         },
1337         [BULK_CIPH_ALG_CAST128] = {
1338                 "cast128",      "cbc(cast5)",      0, 8,  16
1339         },
1340         [BULK_CIPH_ALG_CAST256] = {
1341                 "cast256",      "cbc(cast6)",      0, 16, 32
1342         },
1343         [BULK_CIPH_ALG_TWOFISH128] = {
1344                 "twofish128",   "cbc(twofish)",    0, 16, 16
1345         },
1346         [BULK_CIPH_ALG_TWOFISH256] = {
1347                 "twofish256",   "cbc(twofish)",    0, 16, 32
1348         },
1349 };
1350
1351 const struct sptlrpc_ciph_type *sptlrpc_get_ciph_type(__u8 ciph_alg)
1352 {
1353         struct sptlrpc_ciph_type *ct;
1354
1355         if (ciph_alg < BULK_CIPH_ALG_MAX) {
1356                 ct = &cipher_types[ciph_alg];
1357                 if (ct->sct_tfm_name)
1358                         return ct;
1359         }
1360         return NULL;
1361 }
1362 EXPORT_SYMBOL(sptlrpc_get_ciph_type);
1363
1364 const char *sptlrpc_get_ciph_name(__u8 ciph_alg)
1365 {
1366         const struct sptlrpc_ciph_type *ct;
1367
1368         ct = sptlrpc_get_ciph_type(ciph_alg);
1369         if (ct)
1370                 return ct->sct_name;
1371         else
1372                 return "unknown";
1373 }
1374 EXPORT_SYMBOL(sptlrpc_get_ciph_name);