drivers/staging/lustre/lustre/ptlrpc/sec_bulk.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/ptlrpc/sec_bulk.c
  37  *
  38  * Author: Eric Mei <ericm@clusterfs.com>
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_SEC
  42
  43 #include <linux/libcfs/libcfs.h>
  44 #include <linux/crypto.h>
  45
  46 #include <obd.h>
  47 #include <obd_cksum.h>
  48 #include <obd_class.h>
  49 #include <obd_support.h>
  50 #include <lustre_net.h>
  51 #include <lustre_import.h>
  52 #include <lustre_dlm.h>
  53 #include <lustre_sec.h>
  54
  55 #include "ptlrpc_internal.h"
  56
  57 /****************************************
  58  * bulk encryption page pools      *
  59  ****************************************/
  60
  61
  62 #define PTRS_PER_PAGE   (PAGE_CACHE_SIZE / sizeof(void *))
  63 #define PAGES_PER_POOL  (PTRS_PER_PAGE)
  64
  65 #define IDLE_IDX_MAX        (100)
  66 #define IDLE_IDX_WEIGHT  (3)
  67
  68 #define CACHE_QUIESCENT_PERIOD  (20)
  69
  70 static struct ptlrpc_enc_page_pool {
  71         /*
  72          * constants
  73          */
  74         unsigned long    epp_max_pages;   /* maximum pages can hold, const */
  75         unsigned int     epp_max_pools;   /* number of pools, const */
  76
  77         /*
  78          * wait queue in case of not enough free pages.
  79          */
  80         wait_queue_head_t      epp_waitq;       /* waiting threads */
  81         unsigned int     epp_waitqlen;    /* wait queue length */
  82         unsigned long    epp_pages_short; /* # of pages wanted of in-q users */
  83         unsigned int     epp_growing:1;   /* during adding pages */
  84
  85         /*
  86          * indicating how idle the pools are, from 0 to MAX_IDLE_IDX
  87          * this is counted based on each time when getting pages from
  88          * the pools, not based on time. which means in case that system
  89          * is idled for a while but the idle_idx might still be low if no
  90          * activities happened in the pools.
  91          */
  92         unsigned long    epp_idle_idx;
  93
  94         /* last shrink time due to mem tight */
  95         long         epp_last_shrink;
  96         long         epp_last_access;
  97
  98         /*
  99          * in-pool pages bookkeeping
 100          */
 101         spinlock_t       epp_lock;         /* protect following fields */
 102         unsigned long    epp_total_pages; /* total pages in pools */
 103         unsigned long    epp_free_pages;  /* current pages available */
 104
 105         /*
 106          * statistics
 107          */
 108         unsigned long    epp_st_max_pages;      /* # of pages ever reached */
 109         unsigned int     epp_st_grows;    /* # of grows */
 110         unsigned int     epp_st_grow_fails;     /* # of add pages failures */
 111         unsigned int     epp_st_shrinks;        /* # of shrinks */
 112         unsigned long    epp_st_access;  /* # of access */
 113         unsigned long    epp_st_missings;       /* # of cache missing */
 114         unsigned long    epp_st_lowfree;        /* lowest free pages reached */
 115         unsigned int     epp_st_max_wqlen;      /* highest waitqueue length */
 116         cfs_time_t       epp_st_max_wait;       /* in jeffies */
 117         /*
 118          * pointers to pools
 119          */
 120         struct page    ***epp_pools;
 121 } page_pools;
 122
 123 /*
 124  * memory shrinker
 125  */
 126 const int pools_shrinker_seeks = DEFAULT_SEEKS;
 127 static struct shrinker *pools_shrinker = NULL;
 128
 129
 130 /*
 131  * /proc/fs/lustre/sptlrpc/encrypt_page_pools
 132  */
 133 int sptlrpc_proc_read_enc_pool(char *page, char **start, off_t off, int count,
 134                                int *eof, void *data)
 135 {
 136         int     rc;
 137
 138         spin_lock(&page_pools.epp_lock);
 139
 140         rc = snprintf(page, count,
 141                       "physical pages:    %lu\n"
 142                       "pages per pool:    %lu\n"
 143                       "max pages:              %lu\n"
 144                       "max pools:              %u\n"
 145                       "total pages:          %lu\n"
 146                       "total free:            %lu\n"
 147                       "idle index:            %lu/100\n"
 148                       "last shrink:          %lds\n"
 149                       "last access:          %lds\n"
 150                       "max pages reached:       %lu\n"
 151                       "grows:              %u\n"
 152                       "grows failure:      %u\n"
 153                       "shrinks:          %u\n"
 154                       "cache access:        %lu\n"
 155                       "cache missing:      %lu\n"
 156                       "low free mark:      %lu\n"
 157                       "max waitqueue depth:     %u\n"
 158                       "max wait time:      "CFS_TIME_T"/%u\n"
 159                       ,
 160                       num_physpages,
 161                       PAGES_PER_POOL,
 162                       page_pools.epp_max_pages,
 163                       page_pools.epp_max_pools,
 164                       page_pools.epp_total_pages,
 165                       page_pools.epp_free_pages,
 166                       page_pools.epp_idle_idx,
 167                       cfs_time_current_sec() - page_pools.epp_last_shrink,
 168                       cfs_time_current_sec() - page_pools.epp_last_access,
 169                       page_pools.epp_st_max_pages,
 170                       page_pools.epp_st_grows,
 171                       page_pools.epp_st_grow_fails,
 172                       page_pools.epp_st_shrinks,
 173                       page_pools.epp_st_access,
 174                       page_pools.epp_st_missings,
 175                       page_pools.epp_st_lowfree,
 176                       page_pools.epp_st_max_wqlen,
 177                       page_pools.epp_st_max_wait, HZ
 178                      );
 179
 180         spin_unlock(&page_pools.epp_lock);
 181         return rc;
 182 }
 183
 184 static void enc_pools_release_free_pages(long npages)
 185 {
 186         int     p_idx, g_idx;
 187         int     p_idx_max1, p_idx_max2;
 188
 189         LASSERT(npages > 0);
 190         LASSERT(npages <= page_pools.epp_free_pages);
 191         LASSERT(page_pools.epp_free_pages <= page_pools.epp_total_pages);
 192
 193         /* max pool index before the release */
 194         p_idx_max2 = (page_pools.epp_total_pages - 1) / PAGES_PER_POOL;
 195
 196         page_pools.epp_free_pages -= npages;
 197         page_pools.epp_total_pages -= npages;
 198
 199         /* max pool index after the release */
 200         p_idx_max1 = page_pools.epp_total_pages == 0 ? -1 :
 201                      ((page_pools.epp_total_pages - 1) / PAGES_PER_POOL);
 202
 203         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 204         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 205         LASSERT(page_pools.epp_pools[p_idx]);
 206
 207         while (npages--) {
 208                 LASSERT(page_pools.epp_pools[p_idx]);
 209                 LASSERT(page_pools.epp_pools[p_idx][g_idx] != NULL);
 210
 211                 __free_page(page_pools.epp_pools[p_idx][g_idx]);
 212                 page_pools.epp_pools[p_idx][g_idx] = NULL;
 213
 214                 if (++g_idx == PAGES_PER_POOL) {
 215                         p_idx++;
 216                         g_idx = 0;
 217                 }
 218         };
 219
 220         /* free unused pools */
 221         while (p_idx_max1 < p_idx_max2) {
 222                 LASSERT(page_pools.epp_pools[p_idx_max2]);
 223                 OBD_FREE(page_pools.epp_pools[p_idx_max2], PAGE_CACHE_SIZE);
 224                 page_pools.epp_pools[p_idx_max2] = NULL;
 225                 p_idx_max2--;
 226         }
 227 }
 228
 229 /*
 230  * could be called frequently for query (@nr_to_scan == 0).
 231  * we try to keep at least PTLRPC_MAX_BRW_PAGES pages in the pool.
 232  */
 233 static int enc_pools_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
 234 {
 235         if (unlikely(shrink_param(sc, nr_to_scan) != 0)) {
 236                 spin_lock(&page_pools.epp_lock);
 237                 shrink_param(sc, nr_to_scan) = min_t(unsigned long,
 238                                                    shrink_param(sc, nr_to_scan),
 239                                                    page_pools.epp_free_pages -
 240                                                    PTLRPC_MAX_BRW_PAGES);
 241                 if (shrink_param(sc, nr_to_scan) > 0) {
 242                         enc_pools_release_free_pages(shrink_param(sc,
 243                                                                   nr_to_scan));
 244                         CDEBUG(D_SEC, "released %ld pages, %ld left\n",
 245                                (long)shrink_param(sc, nr_to_scan),
 246                                page_pools.epp_free_pages);
 247
 248                         page_pools.epp_st_shrinks++;
 249                         page_pools.epp_last_shrink = cfs_time_current_sec();
 250                 }
 251                 spin_unlock(&page_pools.epp_lock);
 252         }
 253
 254         /*
 255          * if no pool access for a long time, we consider it's fully idle.
 256          * a little race here is fine.
 257          */
 258         if (unlikely(cfs_time_current_sec() - page_pools.epp_last_access >
 259                      CACHE_QUIESCENT_PERIOD)) {
 260                 spin_lock(&page_pools.epp_lock);
 261                 page_pools.epp_idle_idx = IDLE_IDX_MAX;
 262                 spin_unlock(&page_pools.epp_lock);
 263         }
 264
 265         LASSERT(page_pools.epp_idle_idx <= IDLE_IDX_MAX);
 266         return max((int)page_pools.epp_free_pages - PTLRPC_MAX_BRW_PAGES, 0) *
 267                 (IDLE_IDX_MAX - page_pools.epp_idle_idx) / IDLE_IDX_MAX;
 268 }
 269
 270 static inline
 271 int npages_to_npools(unsigned long npages)
 272 {
 273         return (int) ((npages + PAGES_PER_POOL - 1) / PAGES_PER_POOL);
 274 }
 275
 276 /*
 277  * return how many pages cleaned up.
 278  */
 279 static unsigned long enc_pools_cleanup(struct page ***pools, int npools)
 280 {
 281         unsigned long cleaned = 0;
 282         int        i, j;
 283
 284         for (i = 0; i < npools; i++) {
 285                 if (pools[i]) {
 286                         for (j = 0; j < PAGES_PER_POOL; j++) {
 287                                 if (pools[i][j]) {
 288                                         __free_page(pools[i][j]);
 289                                         cleaned++;
 290                                 }
 291                         }
 292                         OBD_FREE(pools[i], PAGE_CACHE_SIZE);
 293                         pools[i] = NULL;
 294                 }
 295         }
 296
 297         return cleaned;
 298 }
 299
 300 /*
 301  * merge @npools pointed by @pools which contains @npages new pages
 302  * into current pools.
 303  *
 304  * we have options to avoid most memory copy with some tricks. but we choose
 305  * the simplest way to avoid complexity. It's not frequently called.
 306  */
 307 static void enc_pools_insert(struct page ***pools, int npools, int npages)
 308 {
 309         int     freeslot;
 310         int     op_idx, np_idx, og_idx, ng_idx;
 311         int     cur_npools, end_npools;
 312
 313         LASSERT(npages > 0);
 314         LASSERT(page_pools.epp_total_pages+npages <= page_pools.epp_max_pages);
 315         LASSERT(npages_to_npools(npages) == npools);
 316         LASSERT(page_pools.epp_growing);
 317
 318         spin_lock(&page_pools.epp_lock);
 319
 320         /*
 321          * (1) fill all the free slots of current pools.
 322          */
 323         /* free slots are those left by rent pages, and the extra ones with
 324          * index >= total_pages, locate at the tail of last pool. */
 325         freeslot = page_pools.epp_total_pages % PAGES_PER_POOL;
 326         if (freeslot != 0)
 327                 freeslot = PAGES_PER_POOL - freeslot;
 328         freeslot += page_pools.epp_total_pages - page_pools.epp_free_pages;
 329
 330         op_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 331         og_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 332         np_idx = npools - 1;
 333         ng_idx = (npages - 1) % PAGES_PER_POOL;
 334
 335         while (freeslot) {
 336                 LASSERT(page_pools.epp_pools[op_idx][og_idx] == NULL);
 337                 LASSERT(pools[np_idx][ng_idx] != NULL);
 338
 339                 page_pools.epp_pools[op_idx][og_idx] = pools[np_idx][ng_idx];
 340                 pools[np_idx][ng_idx] = NULL;
 341
 342                 freeslot--;
 343
 344                 if (++og_idx == PAGES_PER_POOL) {
 345                         op_idx++;
 346                         og_idx = 0;
 347                 }
 348                 if (--ng_idx < 0) {
 349                         if (np_idx == 0)
 350                                 break;
 351                         np_idx--;
 352                         ng_idx = PAGES_PER_POOL - 1;
 353                 }
 354         }
 355
 356         /*
 357          * (2) add pools if needed.
 358          */
 359         cur_npools = (page_pools.epp_total_pages + PAGES_PER_POOL - 1) /
 360                      PAGES_PER_POOL;
 361         end_npools = (page_pools.epp_total_pages + npages + PAGES_PER_POOL -1) /
 362                      PAGES_PER_POOL;
 363         LASSERT(end_npools <= page_pools.epp_max_pools);
 364
 365         np_idx = 0;
 366         while (cur_npools < end_npools) {
 367                 LASSERT(page_pools.epp_pools[cur_npools] == NULL);
 368                 LASSERT(np_idx < npools);
 369                 LASSERT(pools[np_idx] != NULL);
 370
 371                 page_pools.epp_pools[cur_npools++] = pools[np_idx];
 372                 pools[np_idx++] = NULL;
 373         }
 374
 375         page_pools.epp_total_pages += npages;
 376         page_pools.epp_free_pages += npages;
 377         page_pools.epp_st_lowfree = page_pools.epp_free_pages;
 378
 379         if (page_pools.epp_total_pages > page_pools.epp_st_max_pages)
 380                 page_pools.epp_st_max_pages = page_pools.epp_total_pages;
 381
 382         CDEBUG(D_SEC, "add %d pages to total %lu\n", npages,
 383                page_pools.epp_total_pages);
 384
 385         spin_unlock(&page_pools.epp_lock);
 386 }
 387
 388 static int enc_pools_add_pages(int npages)
 389 {
 390         static DEFINE_MUTEX(add_pages_mutex);
 391         struct page   ***pools;
 392         int          npools, alloced = 0;
 393         int          i, j, rc = -ENOMEM;
 394
 395         if (npages < PTLRPC_MAX_BRW_PAGES)
 396                 npages = PTLRPC_MAX_BRW_PAGES;
 397
 398         mutex_lock(&add_pages_mutex);
 399
 400         if (npages + page_pools.epp_total_pages > page_pools.epp_max_pages)
 401                 npages = page_pools.epp_max_pages - page_pools.epp_total_pages;
 402         LASSERT(npages > 0);
 403
 404         page_pools.epp_st_grows++;
 405
 406         npools = npages_to_npools(npages);
 407         OBD_ALLOC(pools, npools * sizeof(*pools));
 408         if (pools == NULL)
 409                 goto out;
 410
 411         for (i = 0; i < npools; i++) {
 412                 OBD_ALLOC(pools[i], PAGE_CACHE_SIZE);
 413                 if (pools[i] == NULL)
 414                         goto out_pools;
 415
 416                 for (j = 0; j < PAGES_PER_POOL && alloced < npages; j++) {
 417                         pools[i][j] = alloc_page(__GFP_IO |
 418                                                      __GFP_HIGHMEM);
 419                         if (pools[i][j] == NULL)
 420                                 goto out_pools;
 421
 422                         alloced++;
 423                 }
 424         }
 425         LASSERT(alloced == npages);
 426
 427         enc_pools_insert(pools, npools, npages);
 428         CDEBUG(D_SEC, "added %d pages into pools\n", npages);
 429         rc = 0;
 430
 431 out_pools:
 432         enc_pools_cleanup(pools, npools);
 433         OBD_FREE(pools, npools * sizeof(*pools));
 434 out:
 435         if (rc) {
 436                 page_pools.epp_st_grow_fails++;
 437                 CERROR("Failed to allocate %d enc pages\n", npages);
 438         }
 439
 440         mutex_unlock(&add_pages_mutex);
 441         return rc;
 442 }
 443
 444 static inline void enc_pools_wakeup(void)
 445 {
 446         LASSERT(spin_is_locked(&page_pools.epp_lock));
 447         LASSERT(page_pools.epp_waitqlen >= 0);
 448
 449         if (unlikely(page_pools.epp_waitqlen)) {
 450                 LASSERT(waitqueue_active(&page_pools.epp_waitq));
 451                 wake_up_all(&page_pools.epp_waitq);
 452         }
 453 }
 454
 455 static int enc_pools_should_grow(int page_needed, long now)
 456 {
 457         /* don't grow if someone else is growing the pools right now,
 458          * or the pools has reached its full capacity
 459          */
 460         if (page_pools.epp_growing ||
 461             page_pools.epp_total_pages == page_pools.epp_max_pages)
 462                 return 0;
 463
 464         /* if total pages is not enough, we need to grow */
 465         if (page_pools.epp_total_pages < page_needed)
 466                 return 1;
 467
 468         /*
 469          * we wanted to return 0 here if there was a shrink just happened
 470          * moment ago, but this may cause deadlock if both client and ost
 471          * live on single node.
 472          */
 473 #if 0
 474         if (now - page_pools.epp_last_shrink < 2)
 475                 return 0;
 476 #endif
 477
 478         /*
 479          * here we perhaps need consider other factors like wait queue
 480          * length, idle index, etc. ?
 481          */
 482
 483         /* grow the pools in any other cases */
 484         return 1;
 485 }
 486
 487 /*
 488  * we allocate the requested pages atomically.
 489  */
 490 int sptlrpc_enc_pool_get_pages(struct ptlrpc_bulk_desc *desc)
 491 {
 492         wait_queue_t  waitlink;
 493         unsigned long   this_idle = -1;
 494         cfs_time_t      tick = 0;
 495         long        now;
 496         int          p_idx, g_idx;
 497         int          i;
 498
 499         LASSERT(desc->bd_iov_count > 0);
 500         LASSERT(desc->bd_iov_count <= page_pools.epp_max_pages);
 501
 502         /* resent bulk, enc iov might have been allocated previously */
 503         if (desc->bd_enc_iov != NULL)
 504                 return 0;
 505
 506         OBD_ALLOC(desc->bd_enc_iov,
 507                   desc->bd_iov_count * sizeof(*desc->bd_enc_iov));
 508         if (desc->bd_enc_iov == NULL)
 509                 return -ENOMEM;
 510
 511         spin_lock(&page_pools.epp_lock);
 512
 513         page_pools.epp_st_access++;
 514 again:
 515         if (unlikely(page_pools.epp_free_pages < desc->bd_iov_count)) {
 516                 if (tick == 0)
 517                         tick = cfs_time_current();
 518
 519                 now = cfs_time_current_sec();
 520
 521                 page_pools.epp_st_missings++;
 522                 page_pools.epp_pages_short += desc->bd_iov_count;
 523
 524                 if (enc_pools_should_grow(desc->bd_iov_count, now)) {
 525                         page_pools.epp_growing = 1;
 526
 527                         spin_unlock(&page_pools.epp_lock);
 528                         enc_pools_add_pages(page_pools.epp_pages_short / 2);
 529                         spin_lock(&page_pools.epp_lock);
 530
 531                         page_pools.epp_growing = 0;
 532
 533                         enc_pools_wakeup();
 534                 } else {
 535                         if (++page_pools.epp_waitqlen >
 536                             page_pools.epp_st_max_wqlen)
 537                                 page_pools.epp_st_max_wqlen =
 538                                                 page_pools.epp_waitqlen;
 539
 540                         set_current_state(TASK_UNINTERRUPTIBLE);
 541                         init_waitqueue_entry_current(&waitlink);
 542                         add_wait_queue(&page_pools.epp_waitq, &waitlink);
 543
 544                         spin_unlock(&page_pools.epp_lock);
 545                         waitq_wait(&waitlink, TASK_UNINTERRUPTIBLE);
 546                         remove_wait_queue(&page_pools.epp_waitq, &waitlink);
 547                         LASSERT(page_pools.epp_waitqlen > 0);
 548                         spin_lock(&page_pools.epp_lock);
 549                         page_pools.epp_waitqlen--;
 550                 }
 551
 552                 LASSERT(page_pools.epp_pages_short >= desc->bd_iov_count);
 553                 page_pools.epp_pages_short -= desc->bd_iov_count;
 554
 555                 this_idle = 0;
 556                 goto again;
 557         }
 558
 559         /* record max wait time */
 560         if (unlikely(tick != 0)) {
 561                 tick = cfs_time_current() - tick;
 562                 if (tick > page_pools.epp_st_max_wait)
 563                         page_pools.epp_st_max_wait = tick;
 564         }
 565
 566         /* proceed with rest of allocation */
 567         page_pools.epp_free_pages -= desc->bd_iov_count;
 568
 569         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 570         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 571
 572         for (i = 0; i < desc->bd_iov_count; i++) {
 573                 LASSERT(page_pools.epp_pools[p_idx][g_idx] != NULL);
 574                 desc->bd_enc_iov[i].kiov_page =
 575                                         page_pools.epp_pools[p_idx][g_idx];
 576                 page_pools.epp_pools[p_idx][g_idx] = NULL;
 577
 578                 if (++g_idx == PAGES_PER_POOL) {
 579                         p_idx++;
 580                         g_idx = 0;
 581                 }
 582         }
 583
 584         if (page_pools.epp_free_pages < page_pools.epp_st_lowfree)
 585                 page_pools.epp_st_lowfree = page_pools.epp_free_pages;
 586
 587         /*
 588          * new idle index = (old * weight + new) / (weight + 1)
 589          */
 590         if (this_idle == -1) {
 591                 this_idle = page_pools.epp_free_pages * IDLE_IDX_MAX /
 592                             page_pools.epp_total_pages;
 593         }
 594         page_pools.epp_idle_idx = (page_pools.epp_idle_idx * IDLE_IDX_WEIGHT +
 595                                    this_idle) /
 596                                   (IDLE_IDX_WEIGHT + 1);
 597
 598         page_pools.epp_last_access = cfs_time_current_sec();
 599
 600         spin_unlock(&page_pools.epp_lock);
 601         return 0;
 602 }
 603 EXPORT_SYMBOL(sptlrpc_enc_pool_get_pages);
 604
 605 void sptlrpc_enc_pool_put_pages(struct ptlrpc_bulk_desc *desc)
 606 {
 607         int     p_idx, g_idx;
 608         int     i;
 609
 610         if (desc->bd_enc_iov == NULL)
 611                 return;
 612
 613         LASSERT(desc->bd_iov_count > 0);
 614
 615         spin_lock(&page_pools.epp_lock);
 616
 617         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 618         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 619
 620         LASSERT(page_pools.epp_free_pages + desc->bd_iov_count <=
 621                 page_pools.epp_total_pages);
 622         LASSERT(page_pools.epp_pools[p_idx]);
 623
 624         for (i = 0; i < desc->bd_iov_count; i++) {
 625                 LASSERT(desc->bd_enc_iov[i].kiov_page != NULL);
 626                 LASSERT(g_idx != 0 || page_pools.epp_pools[p_idx]);
 627                 LASSERT(page_pools.epp_pools[p_idx][g_idx] == NULL);
 628
 629                 page_pools.epp_pools[p_idx][g_idx] =
 630                                         desc->bd_enc_iov[i].kiov_page;
 631
 632                 if (++g_idx == PAGES_PER_POOL) {
 633                         p_idx++;
 634                         g_idx = 0;
 635                 }
 636         }
 637
 638         page_pools.epp_free_pages += desc->bd_iov_count;
 639
 640         enc_pools_wakeup();
 641
 642         spin_unlock(&page_pools.epp_lock);
 643
 644         OBD_FREE(desc->bd_enc_iov,
 645                  desc->bd_iov_count * sizeof(*desc->bd_enc_iov));
 646         desc->bd_enc_iov = NULL;
 647 }
 648 EXPORT_SYMBOL(sptlrpc_enc_pool_put_pages);
 649
 650 /*
 651  * we don't do much stuff for add_user/del_user anymore, except adding some
 652  * initial pages in add_user() if current pools are empty, rest would be
 653  * handled by the pools's self-adaption.
 654  */
 655 int sptlrpc_enc_pool_add_user(void)
 656 {
 657         int     need_grow = 0;
 658
 659         spin_lock(&page_pools.epp_lock);
 660         if (page_pools.epp_growing == 0 && page_pools.epp_total_pages == 0) {
 661                 page_pools.epp_growing = 1;
 662                 need_grow = 1;
 663         }
 664         spin_unlock(&page_pools.epp_lock);
 665
 666         if (need_grow) {
 667                 enc_pools_add_pages(PTLRPC_MAX_BRW_PAGES +
 668                                     PTLRPC_MAX_BRW_PAGES);
 669
 670                 spin_lock(&page_pools.epp_lock);
 671                 page_pools.epp_growing = 0;
 672                 enc_pools_wakeup();
 673                 spin_unlock(&page_pools.epp_lock);
 674         }
 675         return 0;
 676 }
 677 EXPORT_SYMBOL(sptlrpc_enc_pool_add_user);
 678
 679 int sptlrpc_enc_pool_del_user(void)
 680 {
 681         return 0;
 682 }
 683 EXPORT_SYMBOL(sptlrpc_enc_pool_del_user);
 684
 685 static inline void enc_pools_alloc(void)
 686 {
 687         LASSERT(page_pools.epp_max_pools);
 688         OBD_ALLOC_LARGE(page_pools.epp_pools,
 689                         page_pools.epp_max_pools *
 690                         sizeof(*page_pools.epp_pools));
 691 }
 692
 693 static inline void enc_pools_free(void)
 694 {
 695         LASSERT(page_pools.epp_max_pools);
 696         LASSERT(page_pools.epp_pools);
 697
 698         OBD_FREE_LARGE(page_pools.epp_pools,
 699                        page_pools.epp_max_pools *
 700                        sizeof(*page_pools.epp_pools));
 701 }
 702
 703 int sptlrpc_enc_pool_init(void)
 704 {
 705         /*
 706          * maximum capacity is 1/8 of total physical memory.
 707          * is the 1/8 a good number?
 708          */
 709         page_pools.epp_max_pages = num_physpages / 8;
 710         page_pools.epp_max_pools = npages_to_npools(page_pools.epp_max_pages);
 711
 712         init_waitqueue_head(&page_pools.epp_waitq);
 713         page_pools.epp_waitqlen = 0;
 714         page_pools.epp_pages_short = 0;
 715
 716         page_pools.epp_growing = 0;
 717
 718         page_pools.epp_idle_idx = 0;
 719         page_pools.epp_last_shrink = cfs_time_current_sec();
 720         page_pools.epp_last_access = cfs_time_current_sec();
 721
 722         spin_lock_init(&page_pools.epp_lock);
 723         page_pools.epp_total_pages = 0;
 724         page_pools.epp_free_pages = 0;
 725
 726         page_pools.epp_st_max_pages = 0;
 727         page_pools.epp_st_grows = 0;
 728         page_pools.epp_st_grow_fails = 0;
 729         page_pools.epp_st_shrinks = 0;
 730         page_pools.epp_st_access = 0;
 731         page_pools.epp_st_missings = 0;
 732         page_pools.epp_st_lowfree = 0;
 733         page_pools.epp_st_max_wqlen = 0;
 734         page_pools.epp_st_max_wait = 0;
 735
 736         enc_pools_alloc();
 737         if (page_pools.epp_pools == NULL)
 738                 return -ENOMEM;
 739
 740         pools_shrinker = set_shrinker(pools_shrinker_seeks,
 741                                           enc_pools_shrink);
 742         if (pools_shrinker == NULL) {
 743                 enc_pools_free();
 744                 return -ENOMEM;
 745         }
 746
 747         return 0;
 748 }
 749
 750 void sptlrpc_enc_pool_fini(void)
 751 {
 752         unsigned long cleaned, npools;
 753
 754         LASSERT(pools_shrinker);
 755         LASSERT(page_pools.epp_pools);
 756         LASSERT(page_pools.epp_total_pages == page_pools.epp_free_pages);
 757
 758         remove_shrinker(pools_shrinker);
 759
 760         npools = npages_to_npools(page_pools.epp_total_pages);
 761         cleaned = enc_pools_cleanup(page_pools.epp_pools, npools);
 762         LASSERT(cleaned == page_pools.epp_total_pages);
 763
 764         enc_pools_free();
 765
 766         if (page_pools.epp_st_access > 0) {
 767                 CDEBUG(D_SEC,
 768                        "max pages %lu, grows %u, grow fails %u, shrinks %u, "
 769                        "access %lu, missing %lu, max qlen %u, max wait "
 770                        CFS_TIME_T"/%d\n",
 771                        page_pools.epp_st_max_pages, page_pools.epp_st_grows,
 772                        page_pools.epp_st_grow_fails,
 773                        page_pools.epp_st_shrinks, page_pools.epp_st_access,
 774                        page_pools.epp_st_missings, page_pools.epp_st_max_wqlen,
 775                        page_pools.epp_st_max_wait, HZ);
 776         }
 777 }
 778
 779
 780 static int cfs_hash_alg_id[] = {
 781         [BULK_HASH_ALG_NULL]    = CFS_HASH_ALG_NULL,
 782         [BULK_HASH_ALG_ADLER32] = CFS_HASH_ALG_ADLER32,
 783         [BULK_HASH_ALG_CRC32]   = CFS_HASH_ALG_CRC32,
 784         [BULK_HASH_ALG_MD5]     = CFS_HASH_ALG_MD5,
 785         [BULK_HASH_ALG_SHA1]    = CFS_HASH_ALG_SHA1,
 786         [BULK_HASH_ALG_SHA256]  = CFS_HASH_ALG_SHA256,
 787         [BULK_HASH_ALG_SHA384]  = CFS_HASH_ALG_SHA384,
 788         [BULK_HASH_ALG_SHA512]  = CFS_HASH_ALG_SHA512,
 789 };
 790 const char * sptlrpc_get_hash_name(__u8 hash_alg)
 791 {
 792         return cfs_crypto_hash_name(cfs_hash_alg_id[hash_alg]);
 793 }
 794 EXPORT_SYMBOL(sptlrpc_get_hash_name);
 795
 796 __u8 sptlrpc_get_hash_alg(const char *algname)
 797 {
 798         return cfs_crypto_hash_alg(algname);
 799 }
 800 EXPORT_SYMBOL(sptlrpc_get_hash_alg);
 801
 802 int bulk_sec_desc_unpack(struct lustre_msg *msg, int offset, int swabbed)
 803 {
 804         struct ptlrpc_bulk_sec_desc *bsd;
 805         int                       size = msg->lm_buflens[offset];
 806
 807         bsd = lustre_msg_buf(msg, offset, sizeof(*bsd));
 808         if (bsd == NULL) {
 809                 CERROR("Invalid bulk sec desc: size %d\n", size);
 810                 return -EINVAL;
 811         }
 812
 813         if (swabbed) {
 814                 __swab32s(&bsd->bsd_nob);
 815         }
 816
 817         if (unlikely(bsd->bsd_version != 0)) {
 818                 CERROR("Unexpected version %u\n", bsd->bsd_version);
 819                 return -EPROTO;
 820         }
 821
 822         if (unlikely(bsd->bsd_type >= SPTLRPC_BULK_MAX)) {
 823                 CERROR("Invalid type %u\n", bsd->bsd_type);
 824                 return -EPROTO;
 825         }
 826
 827         /* FIXME more sanity check here */
 828
 829         if (unlikely(bsd->bsd_svc != SPTLRPC_BULK_SVC_NULL &&
 830                      bsd->bsd_svc != SPTLRPC_BULK_SVC_INTG &&
 831                      bsd->bsd_svc != SPTLRPC_BULK_SVC_PRIV)) {
 832                 CERROR("Invalid svc %u\n", bsd->bsd_svc);
 833                 return -EPROTO;
 834         }
 835
 836         return 0;
 837 }
 838 EXPORT_SYMBOL(bulk_sec_desc_unpack);
 839
 840 int sptlrpc_get_bulk_checksum(struct ptlrpc_bulk_desc *desc, __u8 alg,
 841                               void *buf, int buflen)
 842 {
 843         struct cfs_crypto_hash_desc     *hdesc;
 844         int                             hashsize;
 845         char                            hashbuf[64];
 846         unsigned int                    bufsize;
 847         int                             i, err;
 848
 849         LASSERT(alg > BULK_HASH_ALG_NULL && alg < BULK_HASH_ALG_MAX);
 850         LASSERT(buflen >= 4);
 851
 852         hdesc = cfs_crypto_hash_init(cfs_hash_alg_id[alg], NULL, 0);
 853         if (IS_ERR(hdesc)) {
 854                 CERROR("Unable to initialize checksum hash %s\n",
 855                        cfs_crypto_hash_name(cfs_hash_alg_id[alg]));
 856                 return PTR_ERR(hdesc);
 857         }
 858
 859         hashsize = cfs_crypto_hash_digestsize(cfs_hash_alg_id[alg]);
 860
 861         for (i = 0; i < desc->bd_iov_count; i++) {
 862                 cfs_crypto_hash_update_page(hdesc, desc->bd_iov[i].kiov_page,
 863                                   desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK,
 864                                   desc->bd_iov[i].kiov_len);
 865         }
 866         if (hashsize > buflen) {
 867                 bufsize = sizeof(hashbuf);
 868                 err = cfs_crypto_hash_final(hdesc, (unsigned char *)hashbuf,
 869                                             &bufsize);
 870                 memcpy(buf, hashbuf, buflen);
 871         } else {
 872                 bufsize = buflen;
 873                 err = cfs_crypto_hash_final(hdesc, (unsigned char *)buf,
 874                                             &bufsize);
 875         }
 876
 877         if (err)
 878                 cfs_crypto_hash_final(hdesc, NULL, NULL);
 879         return err;
 880 }
 881 EXPORT_SYMBOL(sptlrpc_get_bulk_checksum);