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