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, 2015, 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 static int mult = 20 - PAGE_CACHE_SHIFT;
  57 static int enc_pool_max_memory_mb;
  58 CFS_MODULE_PARM(enc_pool_max_memory_mb, "i", int, 0644,
  59                 "Encoding pool max memory (MB), 1/8 of total physical memory by default");
  60
  61
  62 /****************************************
  63  * bulk encryption page pools           *
  64  ****************************************/
  65
  66
  67 #define PTRS_PER_PAGE   (PAGE_CACHE_SIZE / sizeof(void *))
  68 #define PAGES_PER_POOL  (PTRS_PER_PAGE)
  69
  70 #define IDLE_IDX_MAX            (100)
  71 #define IDLE_IDX_WEIGHT         (3)
  72
  73 #define CACHE_QUIESCENT_PERIOD  (20)
  74
  75 static struct ptlrpc_enc_page_pool {
  76         /*
  77          * constants
  78          */
  79         unsigned long    epp_max_pages;   /* maximum pages can hold, const */
  80         unsigned int     epp_max_pools;   /* number of pools, const */
  81
  82         /*
  83          * wait queue in case of not enough free pages.
  84          */
  85         wait_queue_head_t    epp_waitq;       /* waiting threads */
  86         unsigned int     epp_waitqlen;    /* wait queue length */
  87         unsigned long    epp_pages_short; /* # of pages wanted of in-q users */
  88         unsigned int     epp_growing:1;   /* during adding pages */
  89
  90         /*
  91          * indicating how idle the pools are, from 0 to MAX_IDLE_IDX
  92          * this is counted based on each time when getting pages from
  93          * the pools, not based on time. which means in case that system
  94          * is idled for a while but the idle_idx might still be low if no
  95          * activities happened in the pools.
  96          */
  97         unsigned long    epp_idle_idx;
  98
  99         /* last shrink time due to mem tight */
 100         long             epp_last_shrink;
 101         long             epp_last_access;
 102
 103         /*
 104          * in-pool pages bookkeeping
 105          */
 106         spinlock_t       epp_lock;         /* protect following fields */
 107         unsigned long    epp_total_pages; /* total pages in pools */
 108         unsigned long    epp_free_pages;  /* current pages available */
 109
 110         /*
 111          * statistics
 112          */
 113         unsigned long    epp_st_max_pages;      /* # of pages ever reached */
 114         unsigned int     epp_st_grows;          /* # of grows */
 115         unsigned int     epp_st_grow_fails;     /* # of add pages failures */
 116         unsigned int     epp_st_shrinks;        /* # of shrinks */
 117         unsigned long    epp_st_access;         /* # of access */
 118         unsigned long    epp_st_missings;       /* # of cache missing */
 119         unsigned long    epp_st_lowfree;        /* lowest free pages reached */
 120         unsigned int     epp_st_max_wqlen;      /* highest waitqueue length */
 121         cfs_time_t       epp_st_max_wait;       /* in jeffies */
 122         unsigned long    epp_st_outofmem;       /* # of out of mem requests */
 123         /*
 124          * pointers to pools, may be vmalloc'd
 125          */
 126         struct page    ***epp_pools;
 127 } page_pools;
 128
 129 /*
 130  * memory shrinker
 131  */
 132 static const int pools_shrinker_seeks = DEFAULT_SEEKS;
 133 static struct shrinker *pools_shrinker;
 134
 135
 136 /*
 137  * /proc/fs/lustre/sptlrpc/encrypt_page_pools
 138  */
 139 int sptlrpc_proc_enc_pool_seq_show(struct seq_file *m, void *v)
 140 {
 141         int     rc;
 142
 143         spin_lock(&page_pools.epp_lock);
 144
 145         rc = seq_printf(m,
 146                       "physical pages:          %lu\n"
 147                       "pages per pool:          %lu\n"
 148                       "max pages:               %lu\n"
 149                       "max pools:               %u\n"
 150                       "total pages:             %lu\n"
 151                       "total free:              %lu\n"
 152                       "idle index:              %lu/100\n"
 153                       "last shrink:             %lds\n"
 154                       "last access:             %lds\n"
 155                       "max pages reached:       %lu\n"
 156                       "grows:                   %u\n"
 157                       "grows failure:           %u\n"
 158                       "shrinks:                 %u\n"
 159                       "cache access:            %lu\n"
 160                       "cache missing:           %lu\n"
 161                       "low free mark:           %lu\n"
 162                       "max waitqueue depth:     %u\n"
 163                       "max wait time:           "CFS_TIME_T"/%lu\n"
 164                       "out of mem:             %lu\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,
 184                       msecs_to_jiffies(MSEC_PER_SEC),
 185                       page_pools.epp_st_outofmem
 186                      );
 187
 188         spin_unlock(&page_pools.epp_lock);
 189         return rc;
 190 }
 191
 192 static void enc_pools_release_free_pages(long npages)
 193 {
 194         int     p_idx, g_idx;
 195         int     p_idx_max1, p_idx_max2;
 196
 197         LASSERT(npages > 0);
 198         LASSERT(npages <= page_pools.epp_free_pages);
 199         LASSERT(page_pools.epp_free_pages <= page_pools.epp_total_pages);
 200
 201         /* max pool index before the release */
 202         p_idx_max2 = (page_pools.epp_total_pages - 1) / PAGES_PER_POOL;
 203
 204         page_pools.epp_free_pages -= npages;
 205         page_pools.epp_total_pages -= npages;
 206
 207         /* max pool index after the release */
 208         p_idx_max1 = page_pools.epp_total_pages == 0 ? -1 :
 209                      ((page_pools.epp_total_pages - 1) / PAGES_PER_POOL);
 210
 211         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 212         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 213         LASSERT(page_pools.epp_pools[p_idx]);
 214
 215         while (npages--) {
 216                 LASSERT(page_pools.epp_pools[p_idx]);
 217                 LASSERT(page_pools.epp_pools[p_idx][g_idx] != NULL);
 218
 219                 __free_page(page_pools.epp_pools[p_idx][g_idx]);
 220                 page_pools.epp_pools[p_idx][g_idx] = NULL;
 221
 222                 if (++g_idx == PAGES_PER_POOL) {
 223                         p_idx++;
 224                         g_idx = 0;
 225                 }
 226         }
 227
 228         /* free unused pools */
 229         while (p_idx_max1 < p_idx_max2) {
 230                 LASSERT(page_pools.epp_pools[p_idx_max2]);
 231                 OBD_FREE(page_pools.epp_pools[p_idx_max2], PAGE_CACHE_SIZE);
 232                 page_pools.epp_pools[p_idx_max2] = NULL;
 233                 p_idx_max2--;
 234         }
 235 }
 236
 237 /*
 238  * we try to keep at least PTLRPC_MAX_BRW_PAGES pages in the pool.
 239  */
 240 static unsigned long enc_pools_shrink_count(struct shrinker *s,
 241                                             struct shrink_control *sc)
 242 {
 243         /*
 244          * if no pool access for a long time, we consider it's fully idle.
 245          * a little race here is fine.
 246          */
 247         if (unlikely(cfs_time_current_sec() - page_pools.epp_last_access >
 248                      CACHE_QUIESCENT_PERIOD)) {
 249                 spin_lock(&page_pools.epp_lock);
 250                 page_pools.epp_idle_idx = IDLE_IDX_MAX;
 251                 spin_unlock(&page_pools.epp_lock);
 252         }
 253
 254         LASSERT(page_pools.epp_idle_idx <= IDLE_IDX_MAX);
 255         return max((int)page_pools.epp_free_pages - PTLRPC_MAX_BRW_PAGES, 0) *
 256                 (IDLE_IDX_MAX - page_pools.epp_idle_idx) / IDLE_IDX_MAX;
 257 }
 258
 259 /*
 260  * we try to keep at least PTLRPC_MAX_BRW_PAGES pages in the pool.
 261  */
 262 static unsigned long enc_pools_shrink_scan(struct shrinker *s,
 263                                            struct shrink_control *sc)
 264 {
 265         spin_lock(&page_pools.epp_lock);
 266         sc->nr_to_scan = min_t(unsigned long, sc->nr_to_scan,
 267                               page_pools.epp_free_pages - PTLRPC_MAX_BRW_PAGES);
 268         if (sc->nr_to_scan > 0) {
 269                 enc_pools_release_free_pages(sc->nr_to_scan);
 270                 CDEBUG(D_SEC, "released %ld pages, %ld left\n",
 271                        (long)sc->nr_to_scan, page_pools.epp_free_pages);
 272
 273                 page_pools.epp_st_shrinks++;
 274                 page_pools.epp_last_shrink = cfs_time_current_sec();
 275         }
 276         spin_unlock(&page_pools.epp_lock);
 277
 278         /*
 279          * if no pool access for a long time, we consider it's fully idle.
 280          * a little race here is fine.
 281          */
 282         if (unlikely(cfs_time_current_sec() - page_pools.epp_last_access >
 283                      CACHE_QUIESCENT_PERIOD)) {
 284                 spin_lock(&page_pools.epp_lock);
 285                 page_pools.epp_idle_idx = IDLE_IDX_MAX;
 286                 spin_unlock(&page_pools.epp_lock);
 287         }
 288
 289         LASSERT(page_pools.epp_idle_idx <= IDLE_IDX_MAX);
 290         return sc->nr_to_scan;
 291 }
 292
 293 #ifndef HAVE_SHRINKER_COUNT
 294 /*
 295  * could be called frequently for query (@nr_to_scan == 0).
 296  * we try to keep at least PTLRPC_MAX_BRW_PAGES pages in the pool.
 297  */
 298 static int enc_pools_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
 299 {
 300         struct shrink_control scv = {
 301                 .nr_to_scan = shrink_param(sc, nr_to_scan),
 302                 .gfp_mask   = shrink_param(sc, gfp_mask)
 303         };
 304 #if !defined(HAVE_SHRINKER_WANT_SHRINK_PTR) && !defined(HAVE_SHRINK_CONTROL)
 305         struct shrinker* shrinker = NULL;
 306 #endif
 307
 308         enc_pools_shrink_scan(shrinker, &scv);
 309
 310         return enc_pools_shrink_count(shrinker, &scv);
 311 }
 312
 313 #endif /* HAVE_SHRINKER_COUNT */
 314
 315 static inline
 316 int npages_to_npools(unsigned long npages)
 317 {
 318         return (int) ((npages + PAGES_PER_POOL - 1) / PAGES_PER_POOL);
 319 }
 320
 321 /*
 322  * return how many pages cleaned up.
 323  */
 324 static unsigned long enc_pools_cleanup(struct page ***pools, int npools)
 325 {
 326         unsigned long cleaned = 0;
 327         int           i, j;
 328
 329         for (i = 0; i < npools; i++) {
 330                 if (pools[i]) {
 331                         for (j = 0; j < PAGES_PER_POOL; j++) {
 332                                 if (pools[i][j]) {
 333                                         __free_page(pools[i][j]);
 334                                         cleaned++;
 335                                 }
 336                         }
 337                         OBD_FREE(pools[i], PAGE_CACHE_SIZE);
 338                         pools[i] = NULL;
 339                 }
 340         }
 341
 342         return cleaned;
 343 }
 344
 345 /*
 346  * merge @npools pointed by @pools which contains @npages new pages
 347  * into current pools.
 348  *
 349  * we have options to avoid most memory copy with some tricks. but we choose
 350  * the simplest way to avoid complexity. It's not frequently called.
 351  */
 352 static void enc_pools_insert(struct page ***pools, int npools, int npages)
 353 {
 354         int     freeslot;
 355         int     op_idx, np_idx, og_idx, ng_idx;
 356         int     cur_npools, end_npools;
 357
 358         LASSERT(npages > 0);
 359         LASSERT(page_pools.epp_total_pages+npages <= page_pools.epp_max_pages);
 360         LASSERT(npages_to_npools(npages) == npools);
 361         LASSERT(page_pools.epp_growing);
 362
 363         spin_lock(&page_pools.epp_lock);
 364
 365         /*
 366          * (1) fill all the free slots of current pools.
 367          */
 368         /* free slots are those left by rent pages, and the extra ones with
 369          * index >= total_pages, locate at the tail of last pool. */
 370         freeslot = page_pools.epp_total_pages % PAGES_PER_POOL;
 371         if (freeslot != 0)
 372                 freeslot = PAGES_PER_POOL - freeslot;
 373         freeslot += page_pools.epp_total_pages - page_pools.epp_free_pages;
 374
 375         op_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 376         og_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 377         np_idx = npools - 1;
 378         ng_idx = (npages - 1) % PAGES_PER_POOL;
 379
 380         while (freeslot) {
 381                 LASSERT(page_pools.epp_pools[op_idx][og_idx] == NULL);
 382                 LASSERT(pools[np_idx][ng_idx] != NULL);
 383
 384                 page_pools.epp_pools[op_idx][og_idx] = pools[np_idx][ng_idx];
 385                 pools[np_idx][ng_idx] = NULL;
 386
 387                 freeslot--;
 388
 389                 if (++og_idx == PAGES_PER_POOL) {
 390                         op_idx++;
 391                         og_idx = 0;
 392                 }
 393                 if (--ng_idx < 0) {
 394                         if (np_idx == 0)
 395                                 break;
 396                         np_idx--;
 397                         ng_idx = PAGES_PER_POOL - 1;
 398                 }
 399         }
 400
 401         /*
 402          * (2) add pools if needed.
 403          */
 404         cur_npools = (page_pools.epp_total_pages + PAGES_PER_POOL - 1) /
 405                      PAGES_PER_POOL;
 406         end_npools = (page_pools.epp_total_pages + npages + PAGES_PER_POOL -1) /
 407                      PAGES_PER_POOL;
 408         LASSERT(end_npools <= page_pools.epp_max_pools);
 409
 410         np_idx = 0;
 411         while (cur_npools < end_npools) {
 412                 LASSERT(page_pools.epp_pools[cur_npools] == NULL);
 413                 LASSERT(np_idx < npools);
 414                 LASSERT(pools[np_idx] != NULL);
 415
 416                 page_pools.epp_pools[cur_npools++] = pools[np_idx];
 417                 pools[np_idx++] = NULL;
 418         }
 419
 420         page_pools.epp_total_pages += npages;
 421         page_pools.epp_free_pages += npages;
 422         page_pools.epp_st_lowfree = page_pools.epp_free_pages;
 423
 424         if (page_pools.epp_total_pages > page_pools.epp_st_max_pages)
 425                 page_pools.epp_st_max_pages = page_pools.epp_total_pages;
 426
 427         CDEBUG(D_SEC, "add %d pages to total %lu\n", npages,
 428                page_pools.epp_total_pages);
 429
 430         spin_unlock(&page_pools.epp_lock);
 431 }
 432
 433 static int enc_pools_add_pages(int npages)
 434 {
 435         static DEFINE_MUTEX(add_pages_mutex);
 436         struct page   ***pools;
 437         int             npools, alloced = 0;
 438         int             i, j, rc = -ENOMEM;
 439
 440         if (npages < PTLRPC_MAX_BRW_PAGES)
 441                 npages = PTLRPC_MAX_BRW_PAGES;
 442
 443         mutex_lock(&add_pages_mutex);
 444
 445         if (npages + page_pools.epp_total_pages > page_pools.epp_max_pages)
 446                 npages = page_pools.epp_max_pages - page_pools.epp_total_pages;
 447         LASSERT(npages > 0);
 448
 449         page_pools.epp_st_grows++;
 450
 451         npools = npages_to_npools(npages);
 452         OBD_ALLOC(pools, npools * sizeof(*pools));
 453         if (pools == NULL)
 454                 goto out;
 455
 456         for (i = 0; i < npools; i++) {
 457                 OBD_ALLOC(pools[i], PAGE_CACHE_SIZE);
 458                 if (pools[i] == NULL)
 459                         goto out_pools;
 460
 461                 for (j = 0; j < PAGES_PER_POOL && alloced < npages; j++) {
 462                         pools[i][j] = alloc_page(GFP_NOFS |
 463                                                  __GFP_HIGHMEM);
 464                         if (pools[i][j] == NULL)
 465                                 goto out_pools;
 466
 467                         alloced++;
 468                 }
 469         }
 470         LASSERT(alloced == npages);
 471
 472         enc_pools_insert(pools, npools, npages);
 473         CDEBUG(D_SEC, "added %d pages into pools\n", npages);
 474         rc = 0;
 475
 476 out_pools:
 477         enc_pools_cleanup(pools, npools);
 478         OBD_FREE(pools, npools * sizeof(*pools));
 479 out:
 480         if (rc) {
 481                 page_pools.epp_st_grow_fails++;
 482                 CERROR("Failed to allocate %d enc pages\n", npages);
 483         }
 484
 485         mutex_unlock(&add_pages_mutex);
 486         return rc;
 487 }
 488
 489 static inline void enc_pools_wakeup(void)
 490 {
 491         assert_spin_locked(&page_pools.epp_lock);
 492
 493         if (unlikely(page_pools.epp_waitqlen)) {
 494                 LASSERT(waitqueue_active(&page_pools.epp_waitq));
 495                 wake_up_all(&page_pools.epp_waitq);
 496         }
 497 }
 498
 499 static int enc_pools_should_grow(int page_needed, long now)
 500 {
 501         /* don't grow if someone else is growing the pools right now,
 502          * or the pools has reached its full capacity
 503          */
 504         if (page_pools.epp_growing ||
 505             page_pools.epp_total_pages == page_pools.epp_max_pages)
 506                 return 0;
 507
 508         /* if total pages is not enough, we need to grow */
 509         if (page_pools.epp_total_pages < page_needed)
 510                 return 1;
 511
 512         /*
 513          * we wanted to return 0 here if there was a shrink just happened
 514          * moment ago, but this may cause deadlock if both client and ost
 515          * live on single node.
 516          */
 517 #if 0
 518         if (now - page_pools.epp_last_shrink < 2)
 519                 return 0;
 520 #endif
 521
 522         /*
 523          * here we perhaps need consider other factors like wait queue
 524          * length, idle index, etc. ?
 525          */
 526
 527         /* grow the pools in any other cases */
 528         return 1;
 529 }
 530
 531 /*
 532  * Export the number of free pages in the pool
 533  */
 534 int get_free_pages_in_pool(void)
 535 {
 536         return page_pools.epp_free_pages;
 537 }
 538 EXPORT_SYMBOL(get_free_pages_in_pool);
 539
 540 /*
 541  * Let outside world know if enc_pool full capacity is reached
 542  */
 543 int pool_is_at_full_capacity(void)
 544 {
 545         return (page_pools.epp_total_pages == page_pools.epp_max_pages);
 546 }
 547 EXPORT_SYMBOL(pool_is_at_full_capacity);
 548
 549 /*
 550  * we allocate the requested pages atomically.
 551  */
 552 int sptlrpc_enc_pool_get_pages(struct ptlrpc_bulk_desc *desc)
 553 {
 554         wait_queue_t  waitlink;
 555         unsigned long   this_idle = -1;
 556         cfs_time_t      tick = 0;
 557         long            now;
 558         int             p_idx, g_idx;
 559         int             i;
 560
 561         LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
 562         LASSERT(desc->bd_iov_count > 0);
 563         LASSERT(desc->bd_iov_count <= page_pools.epp_max_pages);
 564
 565         /* resent bulk, enc iov might have been allocated previously */
 566         if (GET_ENC_KIOV(desc) != NULL)
 567                 return 0;
 568
 569         OBD_ALLOC(GET_ENC_KIOV(desc),
 570                   desc->bd_iov_count * sizeof(*GET_ENC_KIOV(desc)));
 571         if (GET_ENC_KIOV(desc) == NULL)
 572                 return -ENOMEM;
 573
 574         spin_lock(&page_pools.epp_lock);
 575
 576         page_pools.epp_st_access++;
 577 again:
 578         if (unlikely(page_pools.epp_free_pages < desc->bd_iov_count)) {
 579                 if (tick == 0)
 580                         tick = cfs_time_current();
 581
 582                 now = cfs_time_current_sec();
 583
 584                 page_pools.epp_st_missings++;
 585                 page_pools.epp_pages_short += desc->bd_iov_count;
 586
 587                 if (enc_pools_should_grow(desc->bd_iov_count, now)) {
 588                         page_pools.epp_growing = 1;
 589
 590                         spin_unlock(&page_pools.epp_lock);
 591                         enc_pools_add_pages(page_pools.epp_pages_short / 2);
 592                         spin_lock(&page_pools.epp_lock);
 593
 594                         page_pools.epp_growing = 0;
 595
 596                         enc_pools_wakeup();
 597                 } else {
 598                         if (page_pools.epp_growing) {
 599                                 if (++page_pools.epp_waitqlen >
 600                                     page_pools.epp_st_max_wqlen)
 601                                         page_pools.epp_st_max_wqlen =
 602                                                         page_pools.epp_waitqlen;
 603
 604                                 set_current_state(TASK_UNINTERRUPTIBLE);
 605                                 init_waitqueue_entry(&waitlink, current);
 606                                 add_wait_queue(&page_pools.epp_waitq,
 607                                                &waitlink);
 608
 609                                 spin_unlock(&page_pools.epp_lock);
 610                                 schedule();
 611                                 remove_wait_queue(&page_pools.epp_waitq,
 612                                                   &waitlink);
 613                                 LASSERT(page_pools.epp_waitqlen > 0);
 614                                 spin_lock(&page_pools.epp_lock);
 615                                 page_pools.epp_waitqlen--;
 616                         } else {
 617                                 /* ptlrpcd thread should not sleep in that case,
 618                                  * or deadlock may occur!
 619                                  * Instead, return -ENOMEM so that upper layers
 620                                  * will put request back in queue. */
 621                                 page_pools.epp_st_outofmem++;
 622                                 spin_unlock(&page_pools.epp_lock);
 623                                 OBD_FREE(GET_ENC_KIOV(desc),
 624                                          desc->bd_iov_count *
 625                                                 sizeof(*GET_ENC_KIOV(desc)));
 626                                 GET_ENC_KIOV(desc) = NULL;
 627                                 return -ENOMEM;
 628                         }
 629                 }
 630
 631                 LASSERT(page_pools.epp_pages_short >= desc->bd_iov_count);
 632                 page_pools.epp_pages_short -= desc->bd_iov_count;
 633
 634                 this_idle = 0;
 635                 goto again;
 636         }
 637
 638         /* record max wait time */
 639         if (unlikely(tick != 0)) {
 640                 tick = cfs_time_current() - tick;
 641                 if (tick > page_pools.epp_st_max_wait)
 642                         page_pools.epp_st_max_wait = tick;
 643         }
 644
 645         /* proceed with rest of allocation */
 646         page_pools.epp_free_pages -= desc->bd_iov_count;
 647
 648         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 649         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 650
 651         for (i = 0; i < desc->bd_iov_count; i++) {
 652                 LASSERT(page_pools.epp_pools[p_idx][g_idx] != NULL);
 653                 BD_GET_ENC_KIOV(desc, i).kiov_page =
 654                        page_pools.epp_pools[p_idx][g_idx];
 655                 page_pools.epp_pools[p_idx][g_idx] = NULL;
 656
 657                 if (++g_idx == PAGES_PER_POOL) {
 658                         p_idx++;
 659                         g_idx = 0;
 660                 }
 661         }
 662
 663         if (page_pools.epp_free_pages < page_pools.epp_st_lowfree)
 664                 page_pools.epp_st_lowfree = page_pools.epp_free_pages;
 665
 666         /*
 667          * new idle index = (old * weight + new) / (weight + 1)
 668          */
 669         if (this_idle == -1) {
 670                 this_idle = page_pools.epp_free_pages * IDLE_IDX_MAX /
 671                             page_pools.epp_total_pages;
 672         }
 673         page_pools.epp_idle_idx = (page_pools.epp_idle_idx * IDLE_IDX_WEIGHT +
 674                                    this_idle) /
 675                                   (IDLE_IDX_WEIGHT + 1);
 676
 677         page_pools.epp_last_access = cfs_time_current_sec();
 678
 679         spin_unlock(&page_pools.epp_lock);
 680         return 0;
 681 }
 682 EXPORT_SYMBOL(sptlrpc_enc_pool_get_pages);
 683
 684 void sptlrpc_enc_pool_put_pages(struct ptlrpc_bulk_desc *desc)
 685 {
 686         int     p_idx, g_idx;
 687         int     i;
 688
 689         LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
 690
 691         if (GET_ENC_KIOV(desc) == NULL)
 692                 return;
 693
 694         LASSERT(desc->bd_iov_count > 0);
 695
 696         spin_lock(&page_pools.epp_lock);
 697
 698         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 699         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 700
 701         LASSERT(page_pools.epp_free_pages + desc->bd_iov_count <=
 702                 page_pools.epp_total_pages);
 703         LASSERT(page_pools.epp_pools[p_idx]);
 704
 705         for (i = 0; i < desc->bd_iov_count; i++) {
 706                 LASSERT(BD_GET_ENC_KIOV(desc, i).kiov_page != NULL);
 707                 LASSERT(g_idx != 0 || page_pools.epp_pools[p_idx]);
 708                 LASSERT(page_pools.epp_pools[p_idx][g_idx] == NULL);
 709
 710                 page_pools.epp_pools[p_idx][g_idx] =
 711                         BD_GET_ENC_KIOV(desc, i).kiov_page;
 712
 713                 if (++g_idx == PAGES_PER_POOL) {
 714                         p_idx++;
 715                         g_idx = 0;
 716                 }
 717         }
 718
 719         page_pools.epp_free_pages += desc->bd_iov_count;
 720
 721         enc_pools_wakeup();
 722
 723         spin_unlock(&page_pools.epp_lock);
 724
 725         OBD_FREE(GET_ENC_KIOV(desc),
 726                  desc->bd_iov_count * sizeof(*GET_ENC_KIOV(desc)));
 727         GET_ENC_KIOV(desc) = NULL;
 728 }
 729
 730 /*
 731  * we don't do much stuff for add_user/del_user anymore, except adding some
 732  * initial pages in add_user() if current pools are empty, rest would be
 733  * handled by the pools's self-adaption.
 734  */
 735 int sptlrpc_enc_pool_add_user(void)
 736 {
 737         int     need_grow = 0;
 738
 739         spin_lock(&page_pools.epp_lock);
 740         if (page_pools.epp_growing == 0 && page_pools.epp_total_pages == 0) {
 741                 page_pools.epp_growing = 1;
 742                 need_grow = 1;
 743         }
 744         spin_unlock(&page_pools.epp_lock);
 745
 746         if (need_grow) {
 747                 enc_pools_add_pages(PTLRPC_MAX_BRW_PAGES +
 748                                     PTLRPC_MAX_BRW_PAGES);
 749
 750                 spin_lock(&page_pools.epp_lock);
 751                 page_pools.epp_growing = 0;
 752                 enc_pools_wakeup();
 753                 spin_unlock(&page_pools.epp_lock);
 754         }
 755         return 0;
 756 }
 757 EXPORT_SYMBOL(sptlrpc_enc_pool_add_user);
 758
 759 int sptlrpc_enc_pool_del_user(void)
 760 {
 761         return 0;
 762 }
 763 EXPORT_SYMBOL(sptlrpc_enc_pool_del_user);
 764
 765 static inline void enc_pools_alloc(void)
 766 {
 767         LASSERT(page_pools.epp_max_pools);
 768         OBD_ALLOC_LARGE(page_pools.epp_pools,
 769                         page_pools.epp_max_pools *
 770                         sizeof(*page_pools.epp_pools));
 771 }
 772
 773 static inline void enc_pools_free(void)
 774 {
 775         LASSERT(page_pools.epp_max_pools);
 776         LASSERT(page_pools.epp_pools);
 777
 778         OBD_FREE_LARGE(page_pools.epp_pools,
 779                        page_pools.epp_max_pools *
 780                        sizeof(*page_pools.epp_pools));
 781 }
 782
 783 int sptlrpc_enc_pool_init(void)
 784 {
 785         DEF_SHRINKER_VAR(shvar, enc_pools_shrink,
 786                          enc_pools_shrink_count, enc_pools_shrink_scan);
 787
 788         page_pools.epp_max_pages = totalram_pages / 8;
 789         if (enc_pool_max_memory_mb > 0 &&
 790             enc_pool_max_memory_mb <= (totalram_pages >> mult))
 791                 page_pools.epp_max_pages = enc_pool_max_memory_mb << mult;
 792
 793         page_pools.epp_max_pools = npages_to_npools(page_pools.epp_max_pages);
 794
 795         init_waitqueue_head(&page_pools.epp_waitq);
 796         page_pools.epp_waitqlen = 0;
 797         page_pools.epp_pages_short = 0;
 798
 799         page_pools.epp_growing = 0;
 800
 801         page_pools.epp_idle_idx = 0;
 802         page_pools.epp_last_shrink = cfs_time_current_sec();
 803         page_pools.epp_last_access = cfs_time_current_sec();
 804
 805         spin_lock_init(&page_pools.epp_lock);
 806         page_pools.epp_total_pages = 0;
 807         page_pools.epp_free_pages = 0;
 808
 809         page_pools.epp_st_max_pages = 0;
 810         page_pools.epp_st_grows = 0;
 811         page_pools.epp_st_grow_fails = 0;
 812         page_pools.epp_st_shrinks = 0;
 813         page_pools.epp_st_access = 0;
 814         page_pools.epp_st_missings = 0;
 815         page_pools.epp_st_lowfree = 0;
 816         page_pools.epp_st_max_wqlen = 0;
 817         page_pools.epp_st_max_wait = 0;
 818         page_pools.epp_st_outofmem = 0;
 819
 820         enc_pools_alloc();
 821         if (page_pools.epp_pools == NULL)
 822                 return -ENOMEM;
 823
 824         pools_shrinker = set_shrinker(pools_shrinker_seeks, &shvar);
 825         if (pools_shrinker == NULL) {
 826                 enc_pools_free();
 827                 return -ENOMEM;
 828         }
 829
 830         return 0;
 831 }
 832
 833 void sptlrpc_enc_pool_fini(void)
 834 {
 835         unsigned long cleaned, npools;
 836
 837         LASSERT(pools_shrinker);
 838         LASSERT(page_pools.epp_pools);
 839         LASSERT(page_pools.epp_total_pages == page_pools.epp_free_pages);
 840
 841         remove_shrinker(pools_shrinker);
 842
 843         npools = npages_to_npools(page_pools.epp_total_pages);
 844         cleaned = enc_pools_cleanup(page_pools.epp_pools, npools);
 845         LASSERT(cleaned == page_pools.epp_total_pages);
 846
 847         enc_pools_free();
 848
 849         if (page_pools.epp_st_access > 0) {
 850                 CDEBUG(D_SEC,
 851                        "max pages %lu, grows %u, grow fails %u, shrinks %u, "
 852                        "access %lu, missing %lu, max qlen %u, max wait "
 853                        CFS_TIME_T"/%lu, out of mem %lu\n",
 854                        page_pools.epp_st_max_pages, page_pools.epp_st_grows,
 855                        page_pools.epp_st_grow_fails,
 856                        page_pools.epp_st_shrinks, page_pools.epp_st_access,
 857                        page_pools.epp_st_missings, page_pools.epp_st_max_wqlen,
 858                        page_pools.epp_st_max_wait,
 859                        msecs_to_jiffies(MSEC_PER_SEC),
 860                        page_pools.epp_st_outofmem);
 861         }
 862 }
 863
 864
 865 static int cfs_hash_alg_id[] = {
 866         [BULK_HASH_ALG_NULL]    = CFS_HASH_ALG_NULL,
 867         [BULK_HASH_ALG_ADLER32] = CFS_HASH_ALG_ADLER32,
 868         [BULK_HASH_ALG_CRC32]   = CFS_HASH_ALG_CRC32,
 869         [BULK_HASH_ALG_MD5]     = CFS_HASH_ALG_MD5,
 870         [BULK_HASH_ALG_SHA1]    = CFS_HASH_ALG_SHA1,
 871         [BULK_HASH_ALG_SHA256]  = CFS_HASH_ALG_SHA256,
 872         [BULK_HASH_ALG_SHA384]  = CFS_HASH_ALG_SHA384,
 873         [BULK_HASH_ALG_SHA512]  = CFS_HASH_ALG_SHA512,
 874 };
 875 const char * sptlrpc_get_hash_name(__u8 hash_alg)
 876 {
 877         return cfs_crypto_hash_name(cfs_hash_alg_id[hash_alg]);
 878 }
 879
 880 __u8 sptlrpc_get_hash_alg(const char *algname)
 881 {
 882         return cfs_crypto_hash_alg(algname);
 883 }
 884
 885 int bulk_sec_desc_unpack(struct lustre_msg *msg, int offset, int swabbed)
 886 {
 887         struct ptlrpc_bulk_sec_desc *bsd;
 888         int                          size = msg->lm_buflens[offset];
 889
 890         bsd = lustre_msg_buf(msg, offset, sizeof(*bsd));
 891         if (bsd == NULL) {
 892                 CERROR("Invalid bulk sec desc: size %d\n", size);
 893                 return -EINVAL;
 894         }
 895
 896         if (swabbed) {
 897                 __swab32s(&bsd->bsd_nob);
 898         }
 899
 900         if (unlikely(bsd->bsd_version != 0)) {
 901                 CERROR("Unexpected version %u\n", bsd->bsd_version);
 902                 return -EPROTO;
 903         }
 904
 905         if (unlikely(bsd->bsd_type >= SPTLRPC_BULK_MAX)) {
 906                 CERROR("Invalid type %u\n", bsd->bsd_type);
 907                 return -EPROTO;
 908         }
 909
 910         /* FIXME more sanity check here */
 911
 912         if (unlikely(bsd->bsd_svc != SPTLRPC_BULK_SVC_NULL &&
 913                      bsd->bsd_svc != SPTLRPC_BULK_SVC_INTG &&
 914                      bsd->bsd_svc != SPTLRPC_BULK_SVC_PRIV)) {
 915                 CERROR("Invalid svc %u\n", bsd->bsd_svc);
 916                 return -EPROTO;
 917         }
 918
 919         return 0;
 920 }
 921 EXPORT_SYMBOL(bulk_sec_desc_unpack);
 922
 923 /*
 924  * Compute the checksum of an RPC buffer payload.  If the return \a buflen
 925  * is not large enough, truncate the result to fit so that it is possible
 926  * to use a hash function with a large hash space, but only use a part of
 927  * the resulting hash.
 928  */
 929 int sptlrpc_get_bulk_checksum(struct ptlrpc_bulk_desc *desc, __u8 alg,
 930                               void *buf, int buflen)
 931 {
 932         struct cfs_crypto_hash_desc     *hdesc;
 933         int                             hashsize;
 934         unsigned int                    bufsize;
 935         int                             i, err;
 936
 937         LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
 938         LASSERT(alg > BULK_HASH_ALG_NULL && alg < BULK_HASH_ALG_MAX);
 939         LASSERT(buflen >= 4);
 940
 941         hdesc = cfs_crypto_hash_init(cfs_hash_alg_id[alg], NULL, 0);
 942         if (IS_ERR(hdesc)) {
 943                 CERROR("Unable to initialize checksum hash %s\n",
 944                        cfs_crypto_hash_name(cfs_hash_alg_id[alg]));
 945                 return PTR_ERR(hdesc);
 946         }
 947
 948         hashsize = cfs_crypto_hash_digestsize(cfs_hash_alg_id[alg]);
 949
 950         for (i = 0; i < desc->bd_iov_count; i++) {
 951                 cfs_crypto_hash_update_page(hdesc,
 952                                   BD_GET_KIOV(desc, i).kiov_page,
 953                                   BD_GET_KIOV(desc, i).kiov_offset &
 954                                               ~PAGE_MASK,
 955                                   BD_GET_KIOV(desc, i).kiov_len);
 956         }
 957
 958         if (hashsize > buflen) {
 959                 unsigned char hashbuf[CFS_CRYPTO_HASH_DIGESTSIZE_MAX];
 960
 961                 bufsize = sizeof(hashbuf);
 962                 LASSERTF(bufsize >= hashsize, "bufsize = %u < hashsize %u\n",
 963                          bufsize, hashsize);
 964                 err = cfs_crypto_hash_final(hdesc, hashbuf, &bufsize);
 965                 memcpy(buf, hashbuf, buflen);
 966         } else {
 967                 bufsize = buflen;
 968                 err = cfs_crypto_hash_final(hdesc, buf, &bufsize);
 969         }
 970
 971         return err;
 972 }