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_read_enc_pool(char *page, char **start, off_t off, int count,
 140                                int *eof, void *data)
 141 {
 142         int     rc;
 143
 144         spin_lock(&page_pools.epp_lock);
 145
 146         rc = snprintf(page, count,
 147                       "physical pages:          %lu\n"
 148                       "pages per pool:          %lu\n"
 149                       "max pages:               %lu\n"
 150                       "max pools:               %u\n"
 151                       "total pages:             %lu\n"
 152                       "total free:              %lu\n"
 153                       "idle index:              %lu/100\n"
 154                       "last shrink:             %lds\n"
 155                       "last access:             %lds\n"
 156                       "max pages reached:       %lu\n"
 157                       "grows:                   %u\n"
 158                       "grows failure:           %u\n"
 159                       "shrinks:                 %u\n"
 160                       "cache access:            %lu\n"
 161                       "cache missing:           %lu\n"
 162                       "low free mark:           %lu\n"
 163                       "max waitqueue depth:     %u\n"
 164                       "max wait time:           "CFS_TIME_T"/%u\n"
 165                       ,
 166                       totalram_pages,
 167                       PAGES_PER_POOL,
 168                       page_pools.epp_max_pages,
 169                       page_pools.epp_max_pools,
 170                       page_pools.epp_total_pages,
 171                       page_pools.epp_free_pages,
 172                       page_pools.epp_idle_idx,
 173                       cfs_time_current_sec() - page_pools.epp_last_shrink,
 174                       cfs_time_current_sec() - page_pools.epp_last_access,
 175                       page_pools.epp_st_max_pages,
 176                       page_pools.epp_st_grows,
 177                       page_pools.epp_st_grow_fails,
 178                       page_pools.epp_st_shrinks,
 179                       page_pools.epp_st_access,
 180                       page_pools.epp_st_missings,
 181                       page_pools.epp_st_lowfree,
 182                       page_pools.epp_st_max_wqlen,
 183                       page_pools.epp_st_max_wait, HZ
 184                      );
 185
 186         spin_unlock(&page_pools.epp_lock);
 187         return rc;
 188 }
 189
 190 static void enc_pools_release_free_pages(long npages)
 191 {
 192         int     p_idx, g_idx;
 193         int     p_idx_max1, p_idx_max2;
 194
 195         LASSERT(npages > 0);
 196         LASSERT(npages <= page_pools.epp_free_pages);
 197         LASSERT(page_pools.epp_free_pages <= page_pools.epp_total_pages);
 198
 199         /* max pool index before the release */
 200         p_idx_max2 = (page_pools.epp_total_pages - 1) / PAGES_PER_POOL;
 201
 202         page_pools.epp_free_pages -= npages;
 203         page_pools.epp_total_pages -= npages;
 204
 205         /* max pool index after the release */
 206         p_idx_max1 = page_pools.epp_total_pages == 0 ? -1 :
 207                      ((page_pools.epp_total_pages - 1) / PAGES_PER_POOL);
 208
 209         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 210         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 211         LASSERT(page_pools.epp_pools[p_idx]);
 212
 213         while (npages--) {
 214                 LASSERT(page_pools.epp_pools[p_idx]);
 215                 LASSERT(page_pools.epp_pools[p_idx][g_idx] != NULL);
 216
 217                 __free_page(page_pools.epp_pools[p_idx][g_idx]);
 218                 page_pools.epp_pools[p_idx][g_idx] = NULL;
 219
 220                 if (++g_idx == PAGES_PER_POOL) {
 221                         p_idx++;
 222                         g_idx = 0;
 223                 }
 224         };
 225
 226         /* free unused pools */
 227         while (p_idx_max1 < p_idx_max2) {
 228                 LASSERT(page_pools.epp_pools[p_idx_max2]);
 229                 OBD_FREE(page_pools.epp_pools[p_idx_max2], PAGE_CACHE_SIZE);
 230                 page_pools.epp_pools[p_idx_max2] = NULL;
 231                 p_idx_max2--;
 232         }
 233 }
 234
 235 /*
 236  * could be called frequently for query (@nr_to_scan == 0).
 237  * we try to keep at least PTLRPC_MAX_BRW_PAGES pages in the pool.
 238  */
 239 static int enc_pools_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
 240 {
 241         if (unlikely(shrink_param(sc, nr_to_scan) != 0)) {
 242                 spin_lock(&page_pools.epp_lock);
 243                 shrink_param(sc, nr_to_scan) = min_t(unsigned long,
 244                                                    shrink_param(sc, nr_to_scan),
 245                                                    page_pools.epp_free_pages -
 246                                                    PTLRPC_MAX_BRW_PAGES);
 247                 if (shrink_param(sc, nr_to_scan) > 0) {
 248                         enc_pools_release_free_pages(shrink_param(sc,
 249                                                                   nr_to_scan));
 250                         CDEBUG(D_SEC, "released %ld pages, %ld left\n",
 251                                (long)shrink_param(sc, nr_to_scan),
 252                                page_pools.epp_free_pages);
 253
 254                         page_pools.epp_st_shrinks++;
 255                         page_pools.epp_last_shrink = cfs_time_current_sec();
 256                 }
 257                 spin_unlock(&page_pools.epp_lock);
 258         }
 259
 260         /*
 261          * if no pool access for a long time, we consider it's fully idle.
 262          * a little race here is fine.
 263          */
 264         if (unlikely(cfs_time_current_sec() - page_pools.epp_last_access >
 265                      CACHE_QUIESCENT_PERIOD)) {
 266                 spin_lock(&page_pools.epp_lock);
 267                 page_pools.epp_idle_idx = IDLE_IDX_MAX;
 268                 spin_unlock(&page_pools.epp_lock);
 269         }
 270
 271         LASSERT(page_pools.epp_idle_idx <= IDLE_IDX_MAX);
 272         return max((int)page_pools.epp_free_pages - PTLRPC_MAX_BRW_PAGES, 0) *
 273                 (IDLE_IDX_MAX - page_pools.epp_idle_idx) / IDLE_IDX_MAX;
 274 }
 275
 276 static inline
 277 int npages_to_npools(unsigned long npages)
 278 {
 279         return (int) ((npages + PAGES_PER_POOL - 1) / PAGES_PER_POOL);
 280 }
 281
 282 /*
 283  * return how many pages cleaned up.
 284  */
 285 static unsigned long enc_pools_cleanup(struct page ***pools, int npools)
 286 {
 287         unsigned long cleaned = 0;
 288         int           i, j;
 289
 290         for (i = 0; i < npools; i++) {
 291                 if (pools[i]) {
 292                         for (j = 0; j < PAGES_PER_POOL; j++) {
 293                                 if (pools[i][j]) {
 294                                         __free_page(pools[i][j]);
 295                                         cleaned++;
 296                                 }
 297                         }
 298                         OBD_FREE(pools[i], PAGE_CACHE_SIZE);
 299                         pools[i] = NULL;
 300                 }
 301         }
 302
 303         return cleaned;
 304 }
 305
 306 /*
 307  * merge @npools pointed by @pools which contains @npages new pages
 308  * into current pools.
 309  *
 310  * we have options to avoid most memory copy with some tricks. but we choose
 311  * the simplest way to avoid complexity. It's not frequently called.
 312  */
 313 static void enc_pools_insert(struct page ***pools, int npools, int npages)
 314 {
 315         int     freeslot;
 316         int     op_idx, np_idx, og_idx, ng_idx;
 317         int     cur_npools, end_npools;
 318
 319         LASSERT(npages > 0);
 320         LASSERT(page_pools.epp_total_pages+npages <= page_pools.epp_max_pages);
 321         LASSERT(npages_to_npools(npages) == npools);
 322         LASSERT(page_pools.epp_growing);
 323
 324         spin_lock(&page_pools.epp_lock);
 325
 326         /*
 327          * (1) fill all the free slots of current pools.
 328          */
 329         /* free slots are those left by rent pages, and the extra ones with
 330          * index >= total_pages, locate at the tail of last pool. */
 331         freeslot = page_pools.epp_total_pages % PAGES_PER_POOL;
 332         if (freeslot != 0)
 333                 freeslot = PAGES_PER_POOL - freeslot;
 334         freeslot += page_pools.epp_total_pages - page_pools.epp_free_pages;
 335
 336         op_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 337         og_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 338         np_idx = npools - 1;
 339         ng_idx = (npages - 1) % PAGES_PER_POOL;
 340
 341         while (freeslot) {
 342                 LASSERT(page_pools.epp_pools[op_idx][og_idx] == NULL);
 343                 LASSERT(pools[np_idx][ng_idx] != NULL);
 344
 345                 page_pools.epp_pools[op_idx][og_idx] = pools[np_idx][ng_idx];
 346                 pools[np_idx][ng_idx] = NULL;
 347
 348                 freeslot--;
 349
 350                 if (++og_idx == PAGES_PER_POOL) {
 351                         op_idx++;
 352                         og_idx = 0;
 353                 }
 354                 if (--ng_idx < 0) {
 355                         if (np_idx == 0)
 356                                 break;
 357                         np_idx--;
 358                         ng_idx = PAGES_PER_POOL - 1;
 359                 }
 360         }
 361
 362         /*
 363          * (2) add pools if needed.
 364          */
 365         cur_npools = (page_pools.epp_total_pages + PAGES_PER_POOL - 1) /
 366                      PAGES_PER_POOL;
 367         end_npools = (page_pools.epp_total_pages + npages + PAGES_PER_POOL -1) /
 368                      PAGES_PER_POOL;
 369         LASSERT(end_npools <= page_pools.epp_max_pools);
 370
 371         np_idx = 0;
 372         while (cur_npools < end_npools) {
 373                 LASSERT(page_pools.epp_pools[cur_npools] == NULL);
 374                 LASSERT(np_idx < npools);
 375                 LASSERT(pools[np_idx] != NULL);
 376
 377                 page_pools.epp_pools[cur_npools++] = pools[np_idx];
 378                 pools[np_idx++] = NULL;
 379         }
 380
 381         page_pools.epp_total_pages += npages;
 382         page_pools.epp_free_pages += npages;
 383         page_pools.epp_st_lowfree = page_pools.epp_free_pages;
 384
 385         if (page_pools.epp_total_pages > page_pools.epp_st_max_pages)
 386                 page_pools.epp_st_max_pages = page_pools.epp_total_pages;
 387
 388         CDEBUG(D_SEC, "add %d pages to total %lu\n", npages,
 389                page_pools.epp_total_pages);
 390
 391         spin_unlock(&page_pools.epp_lock);
 392 }
 393
 394 static int enc_pools_add_pages(int npages)
 395 {
 396         static DEFINE_MUTEX(add_pages_mutex);
 397         struct page   ***pools;
 398         int             npools, alloced = 0;
 399         int             i, j, rc = -ENOMEM;
 400
 401         if (npages < PTLRPC_MAX_BRW_PAGES)
 402                 npages = PTLRPC_MAX_BRW_PAGES;
 403
 404         mutex_lock(&add_pages_mutex);
 405
 406         if (npages + page_pools.epp_total_pages > page_pools.epp_max_pages)
 407                 npages = page_pools.epp_max_pages - page_pools.epp_total_pages;
 408         LASSERT(npages > 0);
 409
 410         page_pools.epp_st_grows++;
 411
 412         npools = npages_to_npools(npages);
 413         OBD_ALLOC(pools, npools * sizeof(*pools));
 414         if (pools == NULL)
 415                 goto out;
 416
 417         for (i = 0; i < npools; i++) {
 418                 OBD_ALLOC(pools[i], PAGE_CACHE_SIZE);
 419                 if (pools[i] == NULL)
 420                         goto out_pools;
 421
 422                 for (j = 0; j < PAGES_PER_POOL && alloced < npages; j++) {
 423                         pools[i][j] = alloc_page(__GFP_IO |
 424                                                      __GFP_HIGHMEM);
 425                         if (pools[i][j] == NULL)
 426                                 goto out_pools;
 427
 428                         alloced++;
 429                 }
 430         }
 431         LASSERT(alloced == npages);
 432
 433         enc_pools_insert(pools, npools, npages);
 434         CDEBUG(D_SEC, "added %d pages into pools\n", npages);
 435         rc = 0;
 436
 437 out_pools:
 438         enc_pools_cleanup(pools, npools);
 439         OBD_FREE(pools, npools * sizeof(*pools));
 440 out:
 441         if (rc) {
 442                 page_pools.epp_st_grow_fails++;
 443                 CERROR("Failed to allocate %d enc pages\n", npages);
 444         }
 445
 446         mutex_unlock(&add_pages_mutex);
 447         return rc;
 448 }
 449
 450 static inline void enc_pools_wakeup(void)
 451 {
 452         LASSERT(spin_is_locked(&page_pools.epp_lock));
 453
 454         if (unlikely(page_pools.epp_waitqlen)) {
 455                 LASSERT(waitqueue_active(&page_pools.epp_waitq));
 456                 wake_up_all(&page_pools.epp_waitq);
 457         }
 458 }
 459
 460 static int enc_pools_should_grow(int page_needed, long now)
 461 {
 462         /* don't grow if someone else is growing the pools right now,
 463          * or the pools has reached its full capacity
 464          */
 465         if (page_pools.epp_growing ||
 466             page_pools.epp_total_pages == page_pools.epp_max_pages)
 467                 return 0;
 468
 469         /* if total pages is not enough, we need to grow */
 470         if (page_pools.epp_total_pages < page_needed)
 471                 return 1;
 472
 473         /*
 474          * we wanted to return 0 here if there was a shrink just happened
 475          * moment ago, but this may cause deadlock if both client and ost
 476          * live on single node.
 477          */
 478 #if 0
 479         if (now - page_pools.epp_last_shrink < 2)
 480                 return 0;
 481 #endif
 482
 483         /*
 484          * here we perhaps need consider other factors like wait queue
 485          * length, idle index, etc. ?
 486          */
 487
 488         /* grow the pools in any other cases */
 489         return 1;
 490 }
 491
 492 /*
 493  * we allocate the requested pages atomically.
 494  */
 495 int sptlrpc_enc_pool_get_pages(struct ptlrpc_bulk_desc *desc)
 496 {
 497         wait_queue_t  waitlink;
 498         unsigned long   this_idle = -1;
 499         cfs_time_t      tick = 0;
 500         long            now;
 501         int             p_idx, g_idx;
 502         int             i;
 503
 504         LASSERT(desc->bd_iov_count > 0);
 505         LASSERT(desc->bd_iov_count <= page_pools.epp_max_pages);
 506
 507         /* resent bulk, enc iov might have been allocated previously */
 508         if (desc->bd_enc_iov != NULL)
 509                 return 0;
 510
 511         OBD_ALLOC(desc->bd_enc_iov,
 512                   desc->bd_iov_count * sizeof(*desc->bd_enc_iov));
 513         if (desc->bd_enc_iov == NULL)
 514                 return -ENOMEM;
 515
 516         spin_lock(&page_pools.epp_lock);
 517
 518         page_pools.epp_st_access++;
 519 again:
 520         if (unlikely(page_pools.epp_free_pages < desc->bd_iov_count)) {
 521                 if (tick == 0)
 522                         tick = cfs_time_current();
 523
 524                 now = cfs_time_current_sec();
 525
 526                 page_pools.epp_st_missings++;
 527                 page_pools.epp_pages_short += desc->bd_iov_count;
 528
 529                 if (enc_pools_should_grow(desc->bd_iov_count, now)) {
 530                         page_pools.epp_growing = 1;
 531
 532                         spin_unlock(&page_pools.epp_lock);
 533                         enc_pools_add_pages(page_pools.epp_pages_short / 2);
 534                         spin_lock(&page_pools.epp_lock);
 535
 536                         page_pools.epp_growing = 0;
 537
 538                         enc_pools_wakeup();
 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                         init_waitqueue_entry_current(&waitlink);
 547                         add_wait_queue(&page_pools.epp_waitq, &waitlink);
 548
 549                         spin_unlock(&page_pools.epp_lock);
 550                         waitq_wait(&waitlink, TASK_UNINTERRUPTIBLE);
 551                         remove_wait_queue(&page_pools.epp_waitq, &waitlink);
 552                         LASSERT(page_pools.epp_waitqlen > 0);
 553                         spin_lock(&page_pools.epp_lock);
 554                         page_pools.epp_waitqlen--;
 555                 }
 556
 557                 LASSERT(page_pools.epp_pages_short >= desc->bd_iov_count);
 558                 page_pools.epp_pages_short -= desc->bd_iov_count;
 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_iov_count;
 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_iov_count; i++) {
 578                 LASSERT(page_pools.epp_pools[p_idx][g_idx] != NULL);
 579                 desc->bd_enc_iov[i].kiov_page =
 580                                         page_pools.epp_pools[p_idx][g_idx];
 581                 page_pools.epp_pools[p_idx][g_idx] = NULL;
 582
 583                 if (++g_idx == PAGES_PER_POOL) {
 584                         p_idx++;
 585                         g_idx = 0;
 586                 }
 587         }
 588
 589         if (page_pools.epp_free_pages < page_pools.epp_st_lowfree)
 590                 page_pools.epp_st_lowfree = page_pools.epp_free_pages;
 591
 592         /*
 593          * new idle index = (old * weight + new) / (weight + 1)
 594          */
 595         if (this_idle == -1) {
 596                 this_idle = page_pools.epp_free_pages * IDLE_IDX_MAX /
 597                             page_pools.epp_total_pages;
 598         }
 599         page_pools.epp_idle_idx = (page_pools.epp_idle_idx * IDLE_IDX_WEIGHT +
 600                                    this_idle) /
 601                                   (IDLE_IDX_WEIGHT + 1);
 602
 603         page_pools.epp_last_access = cfs_time_current_sec();
 604
 605         spin_unlock(&page_pools.epp_lock);
 606         return 0;
 607 }
 608 EXPORT_SYMBOL(sptlrpc_enc_pool_get_pages);
 609
 610 void sptlrpc_enc_pool_put_pages(struct ptlrpc_bulk_desc *desc)
 611 {
 612         int     p_idx, g_idx;
 613         int     i;
 614
 615         if (desc->bd_enc_iov == NULL)
 616                 return;
 617
 618         LASSERT(desc->bd_iov_count > 0);
 619
 620         spin_lock(&page_pools.epp_lock);
 621
 622         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 623         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 624
 625         LASSERT(page_pools.epp_free_pages + desc->bd_iov_count <=
 626                 page_pools.epp_total_pages);
 627         LASSERT(page_pools.epp_pools[p_idx]);
 628
 629         for (i = 0; i < desc->bd_iov_count; i++) {
 630                 LASSERT(desc->bd_enc_iov[i].kiov_page != NULL);
 631                 LASSERT(g_idx != 0 || page_pools.epp_pools[p_idx]);
 632                 LASSERT(page_pools.epp_pools[p_idx][g_idx] == NULL);
 633
 634                 page_pools.epp_pools[p_idx][g_idx] =
 635                                         desc->bd_enc_iov[i].kiov_page;
 636
 637                 if (++g_idx == PAGES_PER_POOL) {
 638                         p_idx++;
 639                         g_idx = 0;
 640                 }
 641         }
 642
 643         page_pools.epp_free_pages += desc->bd_iov_count;
 644
 645         enc_pools_wakeup();
 646
 647         spin_unlock(&page_pools.epp_lock);
 648
 649         OBD_FREE(desc->bd_enc_iov,
 650                  desc->bd_iov_count * sizeof(*desc->bd_enc_iov));
 651         desc->bd_enc_iov = NULL;
 652 }
 653 EXPORT_SYMBOL(sptlrpc_enc_pool_put_pages);
 654
 655 /*
 656  * we don't do much stuff for add_user/del_user anymore, except adding some
 657  * initial pages in add_user() if current pools are empty, rest would be
 658  * handled by the pools's self-adaption.
 659  */
 660 int sptlrpc_enc_pool_add_user(void)
 661 {
 662         int     need_grow = 0;
 663
 664         spin_lock(&page_pools.epp_lock);
 665         if (page_pools.epp_growing == 0 && page_pools.epp_total_pages == 0) {
 666                 page_pools.epp_growing = 1;
 667                 need_grow = 1;
 668         }
 669         spin_unlock(&page_pools.epp_lock);
 670
 671         if (need_grow) {
 672                 enc_pools_add_pages(PTLRPC_MAX_BRW_PAGES +
 673                                     PTLRPC_MAX_BRW_PAGES);
 674
 675                 spin_lock(&page_pools.epp_lock);
 676                 page_pools.epp_growing = 0;
 677                 enc_pools_wakeup();
 678                 spin_unlock(&page_pools.epp_lock);
 679         }
 680         return 0;
 681 }
 682 EXPORT_SYMBOL(sptlrpc_enc_pool_add_user);
 683
 684 int sptlrpc_enc_pool_del_user(void)
 685 {
 686         return 0;
 687 }
 688 EXPORT_SYMBOL(sptlrpc_enc_pool_del_user);
 689
 690 static inline void enc_pools_alloc(void)
 691 {
 692         LASSERT(page_pools.epp_max_pools);
 693         OBD_ALLOC_LARGE(page_pools.epp_pools,
 694                         page_pools.epp_max_pools *
 695                         sizeof(*page_pools.epp_pools));
 696 }
 697
 698 static inline void enc_pools_free(void)
 699 {
 700         LASSERT(page_pools.epp_max_pools);
 701         LASSERT(page_pools.epp_pools);
 702
 703         OBD_FREE_LARGE(page_pools.epp_pools,
 704                        page_pools.epp_max_pools *
 705                        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 = totalram_pages / 8;
 715         page_pools.epp_max_pools = npages_to_npools(page_pools.epp_max_pages);
 716
 717         init_waitqueue_head(&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,
 746                                           enc_pools_shrink);
 747         if (pools_shrinker == NULL) {
 748                 enc_pools_free();
 749                 return -ENOMEM;
 750         }
 751
 752         return 0;
 753 }
 754
 755 void sptlrpc_enc_pool_fini(void)
 756 {
 757         unsigned long cleaned, npools;
 758
 759         LASSERT(pools_shrinker);
 760         LASSERT(page_pools.epp_pools);
 761         LASSERT(page_pools.epp_total_pages == page_pools.epp_free_pages);
 762
 763         remove_shrinker(pools_shrinker);
 764
 765         npools = npages_to_npools(page_pools.epp_total_pages);
 766         cleaned = enc_pools_cleanup(page_pools.epp_pools, npools);
 767         LASSERT(cleaned == page_pools.epp_total_pages);
 768
 769         enc_pools_free();
 770
 771         if (page_pools.epp_st_access > 0) {
 772                 CDEBUG(D_SEC,
 773                        "max pages %lu, grows %u, grow fails %u, shrinks %u, "
 774                        "access %lu, missing %lu, max qlen %u, max wait "
 775                        CFS_TIME_T"/%d\n",
 776                        page_pools.epp_st_max_pages, page_pools.epp_st_grows,
 777                        page_pools.epp_st_grow_fails,
 778                        page_pools.epp_st_shrinks, page_pools.epp_st_access,
 779                        page_pools.epp_st_missings, page_pools.epp_st_max_wqlen,
 780                        page_pools.epp_st_max_wait, HZ);
 781         }
 782 }
 783
 784 #else /* !__KERNEL__ */
 785
 786 int sptlrpc_enc_pool_get_pages(struct ptlrpc_bulk_desc *desc)
 787 {
 788         return 0;
 789 }
 790
 791 void sptlrpc_enc_pool_put_pages(struct ptlrpc_bulk_desc *desc)
 792 {
 793 }
 794
 795 int sptlrpc_enc_pool_init(void)
 796 {
 797         return 0;
 798 }
 799
 800 void sptlrpc_enc_pool_fini(void)
 801 {
 802 }
 803 #endif
 804
 805 static int cfs_hash_alg_id[] = {
 806         [BULK_HASH_ALG_NULL]    = CFS_HASH_ALG_NULL,
 807         [BULK_HASH_ALG_ADLER32] = CFS_HASH_ALG_ADLER32,
 808         [BULK_HASH_ALG_CRC32]   = CFS_HASH_ALG_CRC32,
 809         [BULK_HASH_ALG_MD5]     = CFS_HASH_ALG_MD5,
 810         [BULK_HASH_ALG_SHA1]    = CFS_HASH_ALG_SHA1,
 811         [BULK_HASH_ALG_SHA256]  = CFS_HASH_ALG_SHA256,
 812         [BULK_HASH_ALG_SHA384]  = CFS_HASH_ALG_SHA384,
 813         [BULK_HASH_ALG_SHA512]  = CFS_HASH_ALG_SHA512,
 814 };
 815 const char * sptlrpc_get_hash_name(__u8 hash_alg)
 816 {
 817         return cfs_crypto_hash_name(cfs_hash_alg_id[hash_alg]);
 818 }
 819 EXPORT_SYMBOL(sptlrpc_get_hash_name);
 820
 821 __u8 sptlrpc_get_hash_alg(const char *algname)
 822 {
 823         return cfs_crypto_hash_alg(algname);
 824 }
 825 EXPORT_SYMBOL(sptlrpc_get_hash_alg);
 826
 827 int bulk_sec_desc_unpack(struct lustre_msg *msg, int offset, int swabbed)
 828 {
 829         struct ptlrpc_bulk_sec_desc *bsd;
 830         int                          size = msg->lm_buflens[offset];
 831
 832         bsd = lustre_msg_buf(msg, offset, sizeof(*bsd));
 833         if (bsd == NULL) {
 834                 CERROR("Invalid bulk sec desc: size %d\n", size);
 835                 return -EINVAL;
 836         }
 837
 838         if (swabbed) {
 839                 __swab32s(&bsd->bsd_nob);
 840         }
 841
 842         if (unlikely(bsd->bsd_version != 0)) {
 843                 CERROR("Unexpected version %u\n", bsd->bsd_version);
 844                 return -EPROTO;
 845         }
 846
 847         if (unlikely(bsd->bsd_type >= SPTLRPC_BULK_MAX)) {
 848                 CERROR("Invalid type %u\n", bsd->bsd_type);
 849                 return -EPROTO;
 850         }
 851
 852         /* FIXME more sanity check here */
 853
 854         if (unlikely(bsd->bsd_svc != SPTLRPC_BULK_SVC_NULL &&
 855                      bsd->bsd_svc != SPTLRPC_BULK_SVC_INTG &&
 856                      bsd->bsd_svc != SPTLRPC_BULK_SVC_PRIV)) {
 857                 CERROR("Invalid svc %u\n", bsd->bsd_svc);
 858                 return -EPROTO;
 859         }
 860
 861         return 0;
 862 }
 863 EXPORT_SYMBOL(bulk_sec_desc_unpack);
 864
 865 int sptlrpc_get_bulk_checksum(struct ptlrpc_bulk_desc *desc, __u8 alg,
 866                               void *buf, int buflen)
 867 {
 868         struct cfs_crypto_hash_desc     *hdesc;
 869         int                             hashsize;
 870         char                            hashbuf[64];
 871         unsigned int                    bufsize;
 872         int                             i, err;
 873
 874         LASSERT(alg > BULK_HASH_ALG_NULL && alg < BULK_HASH_ALG_MAX);
 875         LASSERT(buflen >= 4);
 876
 877         hdesc = cfs_crypto_hash_init(cfs_hash_alg_id[alg], NULL, 0);
 878         if (IS_ERR(hdesc)) {
 879                 CERROR("Unable to initialize checksum hash %s\n",
 880                        cfs_crypto_hash_name(cfs_hash_alg_id[alg]));
 881                 return PTR_ERR(hdesc);
 882         }
 883
 884         hashsize = cfs_crypto_hash_digestsize(cfs_hash_alg_id[alg]);
 885
 886         for (i = 0; i < desc->bd_iov_count; i++) {
 887 #ifdef __KERNEL__
 888                 cfs_crypto_hash_update_page(hdesc, desc->bd_iov[i].kiov_page,
 889                                   desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK,
 890                                   desc->bd_iov[i].kiov_len);
 891 #else
 892                 cfs_crypto_hash_update(hdesc, desc->bd_iov[i].iov_base,
 893                                   desc->bd_iov[i].iov_len);
 894 #endif
 895         }
 896         if (hashsize > buflen) {
 897                 bufsize = sizeof(hashbuf);
 898                 err = cfs_crypto_hash_final(hdesc, (unsigned char *)hashbuf,
 899                                             &bufsize);
 900                 memcpy(buf, hashbuf, buflen);
 901         } else {
 902                 bufsize = buflen;
 903                 err = cfs_crypto_hash_final(hdesc, (unsigned char *)buf,
 904                                             &bufsize);
 905         }
 906
 907         if (err)
 908                 cfs_crypto_hash_final(hdesc, NULL, NULL);
 909         return err;
 910 }
 911 EXPORT_SYMBOL(sptlrpc_get_bulk_checksum);
 912
 913