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