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 <libcfs/libcfs.h>
  44 #ifndef __KERNEL__
  45 #include <liblustre.h>
  46 #include <libcfs/list.h>
  47 #else
  48 #include <linux/crypto.h>
  49 #endif
  50
  51 #include <obd.h>
  52 #include <obd_cksum.h>
  53 #include <obd_class.h>
  54 #include <obd_support.h>
  55 #include <lustre_net.h>
  56 #include <lustre_import.h>
  57 #include <lustre_dlm.h>
  58 #include <lustre_sec.h>
  59
  60 #include "ptlrpc_internal.h"
  61
  62 /****************************************
  63  * bulk encryption page pools           *
  64  ****************************************/
  65
  66 #ifdef __KERNEL__
  67
  68 #define PTRS_PER_PAGE   (PAGE_CACHE_SIZE / sizeof(void *))
  69 #define PAGES_PER_POOL  (PTRS_PER_PAGE)
  70
  71 #define IDLE_IDX_MAX            (100)
  72 #define IDLE_IDX_WEIGHT         (3)
  73
  74 #define CACHE_QUIESCENT_PERIOD  (20)
  75
  76 static struct ptlrpc_enc_page_pool {
  77         /*
  78          * constants
  79          */
  80         unsigned long    epp_max_pages;   /* maximum pages can hold, const */
  81         unsigned int     epp_max_pools;   /* number of pools, const */
  82
  83         /*
  84          * wait queue in case of not enough free pages.
  85          */
  86         wait_queue_head_t    epp_waitq;       /* waiting threads */
  87         unsigned int     epp_waitqlen;    /* wait queue length */
  88         unsigned long    epp_pages_short; /* # of pages wanted of in-q users */
  89         unsigned int     epp_growing:1;   /* during adding pages */
  90
  91         /*
  92          * indicating how idle the pools are, from 0 to MAX_IDLE_IDX
  93          * this is counted based on each time when getting pages from
  94          * the pools, not based on time. which means in case that system
  95          * is idled for a while but the idle_idx might still be low if no
  96          * activities happened in the pools.
  97          */
  98         unsigned long    epp_idle_idx;
  99
 100         /* last shrink time due to mem tight */
 101         long             epp_last_shrink;
 102         long             epp_last_access;
 103
 104         /*
 105          * in-pool pages bookkeeping
 106          */
 107         spinlock_t       epp_lock;         /* protect following fields */
 108         unsigned long    epp_total_pages; /* total pages in pools */
 109         unsigned long    epp_free_pages;  /* current pages available */
 110
 111         /*
 112          * statistics
 113          */
 114         unsigned long    epp_st_max_pages;      /* # of pages ever reached */
 115         unsigned int     epp_st_grows;          /* # of grows */
 116         unsigned int     epp_st_grow_fails;     /* # of add pages failures */
 117         unsigned int     epp_st_shrinks;        /* # of shrinks */
 118         unsigned long    epp_st_access;         /* # of access */
 119         unsigned long    epp_st_missings;       /* # of cache missing */
 120         unsigned long    epp_st_lowfree;        /* lowest free pages reached */
 121         unsigned int     epp_st_max_wqlen;      /* highest waitqueue length */
 122         cfs_time_t       epp_st_max_wait;       /* in jeffies */
 123         /*
 124          * pointers to pools
 125          */
 126         struct page    ***epp_pools;
 127 } page_pools;
 128
 129 /*
 130  * memory shrinker
 131  */
 132 const int pools_shrinker_seeks = DEFAULT_SEEKS;
 133 static struct shrinker *pools_shrinker;
 134
 135
 136 /*
 137  * /proc/fs/lustre/sptlrpc/encrypt_page_pools
 138  */
 139 int sptlrpc_proc_enc_pool_seq_show(struct seq_file *m, void *v)
 140 {
 141         int     rc;
 142
 143         spin_lock(&page_pools.epp_lock);
 144
 145         rc = seq_printf(m,
 146                       "physical pages:          %lu\n"
 147                       "pages per pool:          %lu\n"
 148                       "max pages:               %lu\n"
 149                       "max pools:               %u\n"
 150                       "total pages:             %lu\n"
 151                       "total free:              %lu\n"
 152                       "idle index:              %lu/100\n"
 153                       "last shrink:             %lds\n"
 154                       "last access:             %lds\n"
 155                       "max pages reached:       %lu\n"
 156                       "grows:                   %u\n"
 157                       "grows failure:           %u\n"
 158                       "shrinks:                 %u\n"
 159                       "cache access:            %lu\n"
 160                       "cache missing:           %lu\n"
 161                       "low free mark:           %lu\n"
 162                       "max waitqueue depth:     %u\n"
 163                       "max wait time:           "CFS_TIME_T"/%u\n"
 164                       ,
 165                       totalram_pages,
 166                       PAGES_PER_POOL,
 167                       page_pools.epp_max_pages,
 168                       page_pools.epp_max_pools,
 169                       page_pools.epp_total_pages,
 170                       page_pools.epp_free_pages,
 171                       page_pools.epp_idle_idx,
 172                       cfs_time_current_sec() - page_pools.epp_last_shrink,
 173                       cfs_time_current_sec() - page_pools.epp_last_access,
 174                       page_pools.epp_st_max_pages,
 175                       page_pools.epp_st_grows,
 176                       page_pools.epp_st_grow_fails,
 177                       page_pools.epp_st_shrinks,
 178                       page_pools.epp_st_access,
 179                       page_pools.epp_st_missings,
 180                       page_pools.epp_st_lowfree,
 181                       page_pools.epp_st_max_wqlen,
 182                       page_pools.epp_st_max_wait, HZ
 183                      );
 184
 185         spin_unlock(&page_pools.epp_lock);
 186         return rc;
 187 }
 188
 189 static void enc_pools_release_free_pages(long npages)
 190 {
 191         int     p_idx, g_idx;
 192         int     p_idx_max1, p_idx_max2;
 193
 194         LASSERT(npages > 0);
 195         LASSERT(npages <= page_pools.epp_free_pages);
 196         LASSERT(page_pools.epp_free_pages <= page_pools.epp_total_pages);
 197
 198         /* max pool index before the release */
 199         p_idx_max2 = (page_pools.epp_total_pages - 1) / PAGES_PER_POOL;
 200
 201         page_pools.epp_free_pages -= npages;
 202         page_pools.epp_total_pages -= npages;
 203
 204         /* max pool index after the release */
 205         p_idx_max1 = page_pools.epp_total_pages == 0 ? -1 :
 206                      ((page_pools.epp_total_pages - 1) / PAGES_PER_POOL);
 207
 208         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 209         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 210         LASSERT(page_pools.epp_pools[p_idx]);
 211
 212         while (npages--) {
 213                 LASSERT(page_pools.epp_pools[p_idx]);
 214                 LASSERT(page_pools.epp_pools[p_idx][g_idx] != NULL);
 215
 216                 __free_page(page_pools.epp_pools[p_idx][g_idx]);
 217                 page_pools.epp_pools[p_idx][g_idx] = NULL;
 218
 219                 if (++g_idx == PAGES_PER_POOL) {
 220                         p_idx++;
 221                         g_idx = 0;
 222                 }
 223         }
 224
 225         /* free unused pools */
 226         while (p_idx_max1 < p_idx_max2) {
 227                 LASSERT(page_pools.epp_pools[p_idx_max2]);
 228                 OBD_FREE(page_pools.epp_pools[p_idx_max2], PAGE_CACHE_SIZE);
 229                 page_pools.epp_pools[p_idx_max2] = NULL;
 230                 p_idx_max2--;
 231         }
 232 }
 233
 234 /*
 235  * we try to keep at least PTLRPC_MAX_BRW_PAGES pages in the pool.
 236  */
 237 static unsigned long enc_pools_shrink_count(struct shrinker *s,
 238                                             struct shrink_control *sc)
 239 {
 240         /*
 241          * if no pool access for a long time, we consider it's fully idle.
 242          * a little race here is fine.
 243          */
 244         if (unlikely(cfs_time_current_sec() - page_pools.epp_last_access >
 245                      CACHE_QUIESCENT_PERIOD)) {
 246                 spin_lock(&page_pools.epp_lock);
 247                 page_pools.epp_idle_idx = IDLE_IDX_MAX;
 248                 spin_unlock(&page_pools.epp_lock);
 249         }
 250
 251         LASSERT(page_pools.epp_idle_idx <= IDLE_IDX_MAX);
 252         return max((int)page_pools.epp_free_pages - PTLRPC_MAX_BRW_PAGES, 0) *
 253                 (IDLE_IDX_MAX - page_pools.epp_idle_idx) / IDLE_IDX_MAX;
 254 }
 255
 256 /*
 257  * we try to keep at least PTLRPC_MAX_BRW_PAGES pages in the pool.
 258  */
 259 static unsigned long enc_pools_shrink_scan(struct shrinker *s,
 260                                            struct shrink_control *sc)
 261 {
 262         spin_lock(&page_pools.epp_lock);
 263         sc->nr_to_scan = min_t(unsigned long, sc->nr_to_scan,
 264                               page_pools.epp_free_pages - PTLRPC_MAX_BRW_PAGES);
 265         if (sc->nr_to_scan > 0) {
 266                 enc_pools_release_free_pages(sc->nr_to_scan);
 267                 CDEBUG(D_SEC, "released %ld pages, %ld left\n",
 268                        (long)sc->nr_to_scan, page_pools.epp_free_pages);
 269
 270                 page_pools.epp_st_shrinks++;
 271                 page_pools.epp_last_shrink = cfs_time_current_sec();
 272         }
 273         spin_unlock(&page_pools.epp_lock);
 274
 275         /*
 276          * if no pool access for a long time, we consider it's fully idle.
 277          * a little race here is fine.
 278          */
 279         if (unlikely(cfs_time_current_sec() - page_pools.epp_last_access >
 280                      CACHE_QUIESCENT_PERIOD)) {
 281                 spin_lock(&page_pools.epp_lock);
 282                 page_pools.epp_idle_idx = IDLE_IDX_MAX;
 283                 spin_unlock(&page_pools.epp_lock);
 284         }
 285
 286         LASSERT(page_pools.epp_idle_idx <= IDLE_IDX_MAX);
 287         return sc->nr_to_scan;
 288 }
 289
 290 #ifndef HAVE_SHRINKER_COUNT
 291 /*
 292  * could be called frequently for query (@nr_to_scan == 0).
 293  * we try to keep at least PTLRPC_MAX_BRW_PAGES pages in the pool.
 294  */
 295 static int enc_pools_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
 296 {
 297         struct shrink_control scv = {
 298                 .nr_to_scan = shrink_param(sc, nr_to_scan),
 299                 .gfp_mask   = shrink_param(sc, gfp_mask)
 300         };
 301 #if !defined(HAVE_SHRINKER_WANT_SHRINK_PTR) && !defined(HAVE_SHRINK_CONTROL)
 302         struct shrinker* shrinker = NULL;
 303 #endif
 304
 305         enc_pools_shrink_scan(shrinker, &scv);
 306
 307         return enc_pools_shrink_count(shrinker, &scv);
 308 }
 309
 310 #endif /* HAVE_SHRINKER_COUNT */
 311
 312 static inline
 313 int npages_to_npools(unsigned long npages)
 314 {
 315         return (int) ((npages + PAGES_PER_POOL - 1) / PAGES_PER_POOL);
 316 }
 317
 318 /*
 319  * return how many pages cleaned up.
 320  */
 321 static unsigned long enc_pools_cleanup(struct page ***pools, int npools)
 322 {
 323         unsigned long cleaned = 0;
 324         int           i, j;
 325
 326         for (i = 0; i < npools; i++) {
 327                 if (pools[i]) {
 328                         for (j = 0; j < PAGES_PER_POOL; j++) {
 329                                 if (pools[i][j]) {
 330                                         __free_page(pools[i][j]);
 331                                         cleaned++;
 332                                 }
 333                         }
 334                         OBD_FREE(pools[i], PAGE_CACHE_SIZE);
 335                         pools[i] = NULL;
 336                 }
 337         }
 338
 339         return cleaned;
 340 }
 341
 342 /*
 343  * merge @npools pointed by @pools which contains @npages new pages
 344  * into current pools.
 345  *
 346  * we have options to avoid most memory copy with some tricks. but we choose
 347  * the simplest way to avoid complexity. It's not frequently called.
 348  */
 349 static void enc_pools_insert(struct page ***pools, int npools, int npages)
 350 {
 351         int     freeslot;
 352         int     op_idx, np_idx, og_idx, ng_idx;
 353         int     cur_npools, end_npools;
 354
 355         LASSERT(npages > 0);
 356         LASSERT(page_pools.epp_total_pages+npages <= page_pools.epp_max_pages);
 357         LASSERT(npages_to_npools(npages) == npools);
 358         LASSERT(page_pools.epp_growing);
 359
 360         spin_lock(&page_pools.epp_lock);
 361
 362         /*
 363          * (1) fill all the free slots of current pools.
 364          */
 365         /* free slots are those left by rent pages, and the extra ones with
 366          * index >= total_pages, locate at the tail of last pool. */
 367         freeslot = page_pools.epp_total_pages % PAGES_PER_POOL;
 368         if (freeslot != 0)
 369                 freeslot = PAGES_PER_POOL - freeslot;
 370         freeslot += page_pools.epp_total_pages - page_pools.epp_free_pages;
 371
 372         op_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 373         og_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 374         np_idx = npools - 1;
 375         ng_idx = (npages - 1) % PAGES_PER_POOL;
 376
 377         while (freeslot) {
 378                 LASSERT(page_pools.epp_pools[op_idx][og_idx] == NULL);
 379                 LASSERT(pools[np_idx][ng_idx] != NULL);
 380
 381                 page_pools.epp_pools[op_idx][og_idx] = pools[np_idx][ng_idx];
 382                 pools[np_idx][ng_idx] = NULL;
 383
 384                 freeslot--;
 385
 386                 if (++og_idx == PAGES_PER_POOL) {
 387                         op_idx++;
 388                         og_idx = 0;
 389                 }
 390                 if (--ng_idx < 0) {
 391                         if (np_idx == 0)
 392                                 break;
 393                         np_idx--;
 394                         ng_idx = PAGES_PER_POOL - 1;
 395                 }
 396         }
 397
 398         /*
 399          * (2) add pools if needed.
 400          */
 401         cur_npools = (page_pools.epp_total_pages + PAGES_PER_POOL - 1) /
 402                      PAGES_PER_POOL;
 403         end_npools = (page_pools.epp_total_pages + npages + PAGES_PER_POOL -1) /
 404                      PAGES_PER_POOL;
 405         LASSERT(end_npools <= page_pools.epp_max_pools);
 406
 407         np_idx = 0;
 408         while (cur_npools < end_npools) {
 409                 LASSERT(page_pools.epp_pools[cur_npools] == NULL);
 410                 LASSERT(np_idx < npools);
 411                 LASSERT(pools[np_idx] != NULL);
 412
 413                 page_pools.epp_pools[cur_npools++] = pools[np_idx];
 414                 pools[np_idx++] = NULL;
 415         }
 416
 417         page_pools.epp_total_pages += npages;
 418         page_pools.epp_free_pages += npages;
 419         page_pools.epp_st_lowfree = page_pools.epp_free_pages;
 420
 421         if (page_pools.epp_total_pages > page_pools.epp_st_max_pages)
 422                 page_pools.epp_st_max_pages = page_pools.epp_total_pages;
 423
 424         CDEBUG(D_SEC, "add %d pages to total %lu\n", npages,
 425                page_pools.epp_total_pages);
 426
 427         spin_unlock(&page_pools.epp_lock);
 428 }
 429
 430 static int enc_pools_add_pages(int npages)
 431 {
 432         static DEFINE_MUTEX(add_pages_mutex);
 433         struct page   ***pools;
 434         int             npools, alloced = 0;
 435         int             i, j, rc = -ENOMEM;
 436
 437         if (npages < PTLRPC_MAX_BRW_PAGES)
 438                 npages = PTLRPC_MAX_BRW_PAGES;
 439
 440         mutex_lock(&add_pages_mutex);
 441
 442         if (npages + page_pools.epp_total_pages > page_pools.epp_max_pages)
 443                 npages = page_pools.epp_max_pages - page_pools.epp_total_pages;
 444         LASSERT(npages > 0);
 445
 446         page_pools.epp_st_grows++;
 447
 448         npools = npages_to_npools(npages);
 449         OBD_ALLOC(pools, npools * sizeof(*pools));
 450         if (pools == NULL)
 451                 goto out;
 452
 453         for (i = 0; i < npools; i++) {
 454                 OBD_ALLOC(pools[i], PAGE_CACHE_SIZE);
 455                 if (pools[i] == NULL)
 456                         goto out_pools;
 457
 458                 for (j = 0; j < PAGES_PER_POOL && alloced < npages; j++) {
 459                         pools[i][j] = alloc_page(GFP_NOFS |
 460                                                  __GFP_HIGHMEM);
 461                         if (pools[i][j] == NULL)
 462                                 goto out_pools;
 463
 464                         alloced++;
 465                 }
 466         }
 467         LASSERT(alloced == npages);
 468
 469         enc_pools_insert(pools, npools, npages);
 470         CDEBUG(D_SEC, "added %d pages into pools\n", npages);
 471         rc = 0;
 472
 473 out_pools:
 474         enc_pools_cleanup(pools, npools);
 475         OBD_FREE(pools, npools * sizeof(*pools));
 476 out:
 477         if (rc) {
 478                 page_pools.epp_st_grow_fails++;
 479                 CERROR("Failed to allocate %d enc pages\n", npages);
 480         }
 481
 482         mutex_unlock(&add_pages_mutex);
 483         return rc;
 484 }
 485
 486 static inline void enc_pools_wakeup(void)
 487 {
 488         assert_spin_locked(&page_pools.epp_lock);
 489
 490         if (unlikely(page_pools.epp_waitqlen)) {
 491                 LASSERT(waitqueue_active(&page_pools.epp_waitq));
 492                 wake_up_all(&page_pools.epp_waitq);
 493         }
 494 }
 495
 496 static int enc_pools_should_grow(int page_needed, long now)
 497 {
 498         /* don't grow if someone else is growing the pools right now,
 499          * or the pools has reached its full capacity
 500          */
 501         if (page_pools.epp_growing ||
 502             page_pools.epp_total_pages == page_pools.epp_max_pages)
 503                 return 0;
 504
 505         /* if total pages is not enough, we need to grow */
 506         if (page_pools.epp_total_pages < page_needed)
 507                 return 1;
 508
 509         /*
 510          * we wanted to return 0 here if there was a shrink just happened
 511          * moment ago, but this may cause deadlock if both client and ost
 512          * live on single node.
 513          */
 514 #if 0
 515         if (now - page_pools.epp_last_shrink < 2)
 516                 return 0;
 517 #endif
 518
 519         /*
 520          * here we perhaps need consider other factors like wait queue
 521          * length, idle index, etc. ?
 522          */
 523
 524         /* grow the pools in any other cases */
 525         return 1;
 526 }
 527
 528 /*
 529  * we allocate the requested pages atomically.
 530  */
 531 int sptlrpc_enc_pool_get_pages(struct ptlrpc_bulk_desc *desc)
 532 {
 533         wait_queue_t  waitlink;
 534         unsigned long   this_idle = -1;
 535         cfs_time_t      tick = 0;
 536         long            now;
 537         int             p_idx, g_idx;
 538         int             i;
 539
 540         LASSERT(desc->bd_iov_count > 0);
 541         LASSERT(desc->bd_iov_count <= page_pools.epp_max_pages);
 542
 543         /* resent bulk, enc iov might have been allocated previously */
 544         if (desc->bd_enc_iov != NULL)
 545                 return 0;
 546
 547         OBD_ALLOC(desc->bd_enc_iov,
 548                   desc->bd_iov_count * sizeof(*desc->bd_enc_iov));
 549         if (desc->bd_enc_iov == NULL)
 550                 return -ENOMEM;
 551
 552         spin_lock(&page_pools.epp_lock);
 553
 554         page_pools.epp_st_access++;
 555 again:
 556         if (unlikely(page_pools.epp_free_pages < desc->bd_iov_count)) {
 557                 if (tick == 0)
 558                         tick = cfs_time_current();
 559
 560                 now = cfs_time_current_sec();
 561
 562                 page_pools.epp_st_missings++;
 563                 page_pools.epp_pages_short += desc->bd_iov_count;
 564
 565                 if (enc_pools_should_grow(desc->bd_iov_count, now)) {
 566                         page_pools.epp_growing = 1;
 567
 568                         spin_unlock(&page_pools.epp_lock);
 569                         enc_pools_add_pages(page_pools.epp_pages_short / 2);
 570                         spin_lock(&page_pools.epp_lock);
 571
 572                         page_pools.epp_growing = 0;
 573
 574                         enc_pools_wakeup();
 575                 } else {
 576                         if (++page_pools.epp_waitqlen >
 577                             page_pools.epp_st_max_wqlen)
 578                                 page_pools.epp_st_max_wqlen =
 579                                                 page_pools.epp_waitqlen;
 580
 581                         set_current_state(TASK_UNINTERRUPTIBLE);
 582                         init_waitqueue_entry_current(&waitlink);
 583                         add_wait_queue(&page_pools.epp_waitq, &waitlink);
 584
 585                         spin_unlock(&page_pools.epp_lock);
 586                         waitq_wait(&waitlink, TASK_UNINTERRUPTIBLE);
 587                         remove_wait_queue(&page_pools.epp_waitq, &waitlink);
 588                         LASSERT(page_pools.epp_waitqlen > 0);
 589                         spin_lock(&page_pools.epp_lock);
 590                         page_pools.epp_waitqlen--;
 591                 }
 592
 593                 LASSERT(page_pools.epp_pages_short >= desc->bd_iov_count);
 594                 page_pools.epp_pages_short -= desc->bd_iov_count;
 595
 596                 this_idle = 0;
 597                 goto again;
 598         }
 599
 600         /* record max wait time */
 601         if (unlikely(tick != 0)) {
 602                 tick = cfs_time_current() - tick;
 603                 if (tick > page_pools.epp_st_max_wait)
 604                         page_pools.epp_st_max_wait = tick;
 605         }
 606
 607         /* proceed with rest of allocation */
 608         page_pools.epp_free_pages -= desc->bd_iov_count;
 609
 610         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 611         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 612
 613         for (i = 0; i < desc->bd_iov_count; i++) {
 614                 LASSERT(page_pools.epp_pools[p_idx][g_idx] != NULL);
 615                 desc->bd_enc_iov[i].kiov_page =
 616                                         page_pools.epp_pools[p_idx][g_idx];
 617                 page_pools.epp_pools[p_idx][g_idx] = NULL;
 618
 619                 if (++g_idx == PAGES_PER_POOL) {
 620                         p_idx++;
 621                         g_idx = 0;
 622                 }
 623         }
 624
 625         if (page_pools.epp_free_pages < page_pools.epp_st_lowfree)
 626                 page_pools.epp_st_lowfree = page_pools.epp_free_pages;
 627
 628         /*
 629          * new idle index = (old * weight + new) / (weight + 1)
 630          */
 631         if (this_idle == -1) {
 632                 this_idle = page_pools.epp_free_pages * IDLE_IDX_MAX /
 633                             page_pools.epp_total_pages;
 634         }
 635         page_pools.epp_idle_idx = (page_pools.epp_idle_idx * IDLE_IDX_WEIGHT +
 636                                    this_idle) /
 637                                   (IDLE_IDX_WEIGHT + 1);
 638
 639         page_pools.epp_last_access = cfs_time_current_sec();
 640
 641         spin_unlock(&page_pools.epp_lock);
 642         return 0;
 643 }
 644 EXPORT_SYMBOL(sptlrpc_enc_pool_get_pages);
 645
 646 void sptlrpc_enc_pool_put_pages(struct ptlrpc_bulk_desc *desc)
 647 {
 648         int     p_idx, g_idx;
 649         int     i;
 650
 651         if (desc->bd_enc_iov == NULL)
 652                 return;
 653
 654         LASSERT(desc->bd_iov_count > 0);
 655
 656         spin_lock(&page_pools.epp_lock);
 657
 658         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 659         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 660
 661         LASSERT(page_pools.epp_free_pages + desc->bd_iov_count <=
 662                 page_pools.epp_total_pages);
 663         LASSERT(page_pools.epp_pools[p_idx]);
 664
 665         for (i = 0; i < desc->bd_iov_count; i++) {
 666                 LASSERT(desc->bd_enc_iov[i].kiov_page != NULL);
 667                 LASSERT(g_idx != 0 || page_pools.epp_pools[p_idx]);
 668                 LASSERT(page_pools.epp_pools[p_idx][g_idx] == NULL);
 669
 670                 page_pools.epp_pools[p_idx][g_idx] =
 671                                         desc->bd_enc_iov[i].kiov_page;
 672
 673                 if (++g_idx == PAGES_PER_POOL) {
 674                         p_idx++;
 675                         g_idx = 0;
 676                 }
 677         }
 678
 679         page_pools.epp_free_pages += desc->bd_iov_count;
 680
 681         enc_pools_wakeup();
 682
 683         spin_unlock(&page_pools.epp_lock);
 684
 685         OBD_FREE(desc->bd_enc_iov,
 686                  desc->bd_iov_count * sizeof(*desc->bd_enc_iov));
 687         desc->bd_enc_iov = NULL;
 688 }
 689 EXPORT_SYMBOL(sptlrpc_enc_pool_put_pages);
 690
 691 /*
 692  * we don't do much stuff for add_user/del_user anymore, except adding some
 693  * initial pages in add_user() if current pools are empty, rest would be
 694  * handled by the pools's self-adaption.
 695  */
 696 int sptlrpc_enc_pool_add_user(void)
 697 {
 698         int     need_grow = 0;
 699
 700         spin_lock(&page_pools.epp_lock);
 701         if (page_pools.epp_growing == 0 && page_pools.epp_total_pages == 0) {
 702                 page_pools.epp_growing = 1;
 703                 need_grow = 1;
 704         }
 705         spin_unlock(&page_pools.epp_lock);
 706
 707         if (need_grow) {
 708                 enc_pools_add_pages(PTLRPC_MAX_BRW_PAGES +
 709                                     PTLRPC_MAX_BRW_PAGES);
 710
 711                 spin_lock(&page_pools.epp_lock);
 712                 page_pools.epp_growing = 0;
 713                 enc_pools_wakeup();
 714                 spin_unlock(&page_pools.epp_lock);
 715         }
 716         return 0;
 717 }
 718 EXPORT_SYMBOL(sptlrpc_enc_pool_add_user);
 719
 720 int sptlrpc_enc_pool_del_user(void)
 721 {
 722         return 0;
 723 }
 724 EXPORT_SYMBOL(sptlrpc_enc_pool_del_user);
 725
 726 static inline void enc_pools_alloc(void)
 727 {
 728         LASSERT(page_pools.epp_max_pools);
 729         OBD_ALLOC_LARGE(page_pools.epp_pools,
 730                         page_pools.epp_max_pools *
 731                         sizeof(*page_pools.epp_pools));
 732 }
 733
 734 static inline void enc_pools_free(void)
 735 {
 736         LASSERT(page_pools.epp_max_pools);
 737         LASSERT(page_pools.epp_pools);
 738
 739         OBD_FREE_LARGE(page_pools.epp_pools,
 740                        page_pools.epp_max_pools *
 741                        sizeof(*page_pools.epp_pools));
 742 }
 743
 744 int sptlrpc_enc_pool_init(void)
 745 {
 746         DEF_SHRINKER_VAR(shvar, enc_pools_shrink,
 747                          enc_pools_shrink_count, enc_pools_shrink_scan);
 748         /*
 749          * maximum capacity is 1/8 of total physical memory.
 750          * is the 1/8 a good number?
 751          */
 752         page_pools.epp_max_pages = totalram_pages / 8;
 753         page_pools.epp_max_pools = npages_to_npools(page_pools.epp_max_pages);
 754
 755         init_waitqueue_head(&page_pools.epp_waitq);
 756         page_pools.epp_waitqlen = 0;
 757         page_pools.epp_pages_short = 0;
 758
 759         page_pools.epp_growing = 0;
 760
 761         page_pools.epp_idle_idx = 0;
 762         page_pools.epp_last_shrink = cfs_time_current_sec();
 763         page_pools.epp_last_access = cfs_time_current_sec();
 764
 765         spin_lock_init(&page_pools.epp_lock);
 766         page_pools.epp_total_pages = 0;
 767         page_pools.epp_free_pages = 0;
 768
 769         page_pools.epp_st_max_pages = 0;
 770         page_pools.epp_st_grows = 0;
 771         page_pools.epp_st_grow_fails = 0;
 772         page_pools.epp_st_shrinks = 0;
 773         page_pools.epp_st_access = 0;
 774         page_pools.epp_st_missings = 0;
 775         page_pools.epp_st_lowfree = 0;
 776         page_pools.epp_st_max_wqlen = 0;
 777         page_pools.epp_st_max_wait = 0;
 778
 779         enc_pools_alloc();
 780         if (page_pools.epp_pools == NULL)
 781                 return -ENOMEM;
 782
 783         pools_shrinker = set_shrinker(pools_shrinker_seeks, &shvar);
 784         if (pools_shrinker == NULL) {
 785                 enc_pools_free();
 786                 return -ENOMEM;
 787         }
 788
 789         return 0;
 790 }
 791
 792 void sptlrpc_enc_pool_fini(void)
 793 {
 794         unsigned long cleaned, npools;
 795
 796         LASSERT(pools_shrinker);
 797         LASSERT(page_pools.epp_pools);
 798         LASSERT(page_pools.epp_total_pages == page_pools.epp_free_pages);
 799
 800         remove_shrinker(pools_shrinker);
 801
 802         npools = npages_to_npools(page_pools.epp_total_pages);
 803         cleaned = enc_pools_cleanup(page_pools.epp_pools, npools);
 804         LASSERT(cleaned == page_pools.epp_total_pages);
 805
 806         enc_pools_free();
 807
 808         if (page_pools.epp_st_access > 0) {
 809                 CDEBUG(D_SEC,
 810                        "max pages %lu, grows %u, grow fails %u, shrinks %u, "
 811                        "access %lu, missing %lu, max qlen %u, max wait "
 812                        CFS_TIME_T"/%d\n",
 813                        page_pools.epp_st_max_pages, page_pools.epp_st_grows,
 814                        page_pools.epp_st_grow_fails,
 815                        page_pools.epp_st_shrinks, page_pools.epp_st_access,
 816                        page_pools.epp_st_missings, page_pools.epp_st_max_wqlen,
 817                        page_pools.epp_st_max_wait, HZ);
 818         }
 819 }
 820
 821 #else /* !__KERNEL__ */
 822
 823 int sptlrpc_enc_pool_get_pages(struct ptlrpc_bulk_desc *desc)
 824 {
 825         return 0;
 826 }
 827
 828 void sptlrpc_enc_pool_put_pages(struct ptlrpc_bulk_desc *desc)
 829 {
 830 }
 831
 832 int sptlrpc_enc_pool_init(void)
 833 {
 834         return 0;
 835 }
 836
 837 void sptlrpc_enc_pool_fini(void)
 838 {
 839 }
 840 #endif
 841
 842 static int cfs_hash_alg_id[] = {
 843         [BULK_HASH_ALG_NULL]    = CFS_HASH_ALG_NULL,
 844         [BULK_HASH_ALG_ADLER32] = CFS_HASH_ALG_ADLER32,
 845         [BULK_HASH_ALG_CRC32]   = CFS_HASH_ALG_CRC32,
 846         [BULK_HASH_ALG_MD5]     = CFS_HASH_ALG_MD5,
 847         [BULK_HASH_ALG_SHA1]    = CFS_HASH_ALG_SHA1,
 848         [BULK_HASH_ALG_SHA256]  = CFS_HASH_ALG_SHA256,
 849         [BULK_HASH_ALG_SHA384]  = CFS_HASH_ALG_SHA384,
 850         [BULK_HASH_ALG_SHA512]  = CFS_HASH_ALG_SHA512,
 851 };
 852 const char * sptlrpc_get_hash_name(__u8 hash_alg)
 853 {
 854         return cfs_crypto_hash_name(cfs_hash_alg_id[hash_alg]);
 855 }
 856 EXPORT_SYMBOL(sptlrpc_get_hash_name);
 857
 858 __u8 sptlrpc_get_hash_alg(const char *algname)
 859 {
 860         return cfs_crypto_hash_alg(algname);
 861 }
 862 EXPORT_SYMBOL(sptlrpc_get_hash_alg);
 863
 864 int bulk_sec_desc_unpack(struct lustre_msg *msg, int offset, int swabbed)
 865 {
 866         struct ptlrpc_bulk_sec_desc *bsd;
 867         int                          size = msg->lm_buflens[offset];
 868
 869         bsd = lustre_msg_buf(msg, offset, sizeof(*bsd));
 870         if (bsd == NULL) {
 871                 CERROR("Invalid bulk sec desc: size %d\n", size);
 872                 return -EINVAL;
 873         }
 874
 875         if (swabbed) {
 876                 __swab32s(&bsd->bsd_nob);
 877         }
 878
 879         if (unlikely(bsd->bsd_version != 0)) {
 880                 CERROR("Unexpected version %u\n", bsd->bsd_version);
 881                 return -EPROTO;
 882         }
 883
 884         if (unlikely(bsd->bsd_type >= SPTLRPC_BULK_MAX)) {
 885                 CERROR("Invalid type %u\n", bsd->bsd_type);
 886                 return -EPROTO;
 887         }
 888
 889         /* FIXME more sanity check here */
 890
 891         if (unlikely(bsd->bsd_svc != SPTLRPC_BULK_SVC_NULL &&
 892                      bsd->bsd_svc != SPTLRPC_BULK_SVC_INTG &&
 893                      bsd->bsd_svc != SPTLRPC_BULK_SVC_PRIV)) {
 894                 CERROR("Invalid svc %u\n", bsd->bsd_svc);
 895                 return -EPROTO;
 896         }
 897
 898         return 0;
 899 }
 900 EXPORT_SYMBOL(bulk_sec_desc_unpack);
 901
 902 int sptlrpc_get_bulk_checksum(struct ptlrpc_bulk_desc *desc, __u8 alg,
 903                               void *buf, int buflen)
 904 {
 905         struct cfs_crypto_hash_desc     *hdesc;
 906         int                             hashsize;
 907         char                            hashbuf[64];
 908         unsigned int                    bufsize;
 909         int                             i, err;
 910
 911         LASSERT(alg > BULK_HASH_ALG_NULL && alg < BULK_HASH_ALG_MAX);
 912         LASSERT(buflen >= 4);
 913
 914         hdesc = cfs_crypto_hash_init(cfs_hash_alg_id[alg], NULL, 0);
 915         if (IS_ERR(hdesc)) {
 916                 CERROR("Unable to initialize checksum hash %s\n",
 917                        cfs_crypto_hash_name(cfs_hash_alg_id[alg]));
 918                 return PTR_ERR(hdesc);
 919         }
 920
 921         hashsize = cfs_crypto_hash_digestsize(cfs_hash_alg_id[alg]);
 922
 923         for (i = 0; i < desc->bd_iov_count; i++) {
 924 #ifdef __KERNEL__
 925                 cfs_crypto_hash_update_page(hdesc, desc->bd_iov[i].kiov_page,
 926                                   desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK,
 927                                   desc->bd_iov[i].kiov_len);
 928 #else
 929                 cfs_crypto_hash_update(hdesc, desc->bd_iov[i].iov_base,
 930                                   desc->bd_iov[i].iov_len);
 931 #endif
 932         }
 933         if (hashsize > buflen) {
 934                 bufsize = sizeof(hashbuf);
 935                 err = cfs_crypto_hash_final(hdesc, (unsigned char *)hashbuf,
 936                                             &bufsize);
 937                 memcpy(buf, hashbuf, buflen);
 938         } else {
 939                 bufsize = buflen;
 940                 err = cfs_crypto_hash_final(hdesc, (unsigned char *)buf,
 941                                             &bufsize);
 942         }
 943
 944         if (err)
 945                 cfs_crypto_hash_final(hdesc, NULL, NULL);
 946         return err;
 947 }
 948 EXPORT_SYMBOL(sptlrpc_get_bulk_checksum);
 949
 950