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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  *
  31  * lustre/ptlrpc/sec_bulk.c
  32  *
  33  * Author: Eric Mei <ericm@clusterfs.com>
  34  */
  35
  36 #define DEBUG_SUBSYSTEM S_SEC
  37
  38 #include <libcfs/linux/linux-mem.h>
  39
  40 #include <obd.h>
  41 #include <obd_cksum.h>
  42 #include <obd_class.h>
  43 #include <obd_support.h>
  44 #include <lustre_net.h>
  45 #include <lustre_import.h>
  46 #include <lustre_dlm.h>
  47 #include <lustre_sec.h>
  48
  49 #include "ptlrpc_internal.h"
  50
  51 static int mult = 20 - PAGE_SHIFT;
  52 static int enc_pool_max_memory_mb;
  53 module_param(enc_pool_max_memory_mb, int, 0644);
  54 MODULE_PARM_DESC(enc_pool_max_memory_mb,
  55                  "Encoding pool max memory (MB), 1/8 of total physical memory by default");
  56
  57 /*
  58  * bulk encryption page pools
  59  */
  60
  61 #define PTRS_PER_PAGE   (PAGE_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         unsigned long epp_max_pages;   /* maximum pages can hold, const */
  71         unsigned int epp_max_pools;   /* number of pools, const */
  72
  73         /*
  74          * wait queue in case of not enough free pages.
  75          */
  76         wait_queue_head_t epp_waitq;   /* waiting threads */
  77         unsigned int epp_waitqlen;    /* wait queue length */
  78         unsigned long epp_pages_short; /* # of pages wanted of in-q users */
  79         unsigned int epp_growing:1;   /* during adding pages */
  80
  81         /*
  82          * indicating how idle the pools are, from 0 to MAX_IDLE_IDX
  83          * this is counted based on each time when getting pages from
  84          * the pools, not based on time. which means in case that system
  85          * is idled for a while but the idle_idx might still be low if no
  86          * activities happened in the pools.
  87          */
  88         unsigned long epp_idle_idx;
  89
  90         /* last shrink time due to mem tight */
  91         time64_t epp_last_shrink;
  92         time64_t epp_last_access;
  93
  94         /* in-pool pages bookkeeping */
  95         spinlock_t epp_lock; /* protect following fields */
  96         unsigned long epp_total_pages; /* total pages in pools */
  97         unsigned long epp_free_pages;  /* current pages available */
  98
  99         /* statistics */
 100         unsigned long epp_st_max_pages;      /* # of pages ever reached */
 101         unsigned int epp_st_grows;          /* # of grows */
 102         unsigned int epp_st_grow_fails;     /* # of add pages failures */
 103         unsigned int epp_st_shrinks;        /* # of shrinks */
 104         unsigned long epp_st_access;         /* # of access */
 105         unsigned long epp_st_missings;       /* # of cache missing */
 106         unsigned long epp_st_lowfree;        /* lowest free pages reached */
 107         unsigned int epp_st_max_wqlen;      /* highest waitqueue length */
 108         ktime_t epp_st_max_wait; /* in nanoseconds */
 109         unsigned long epp_st_outofmem; /* # of out of mem requests */
 110         /*
 111          * pointers to pools, may be vmalloc'd
 112          */
 113         struct page ***epp_pools;
 114 } page_pools;
 115
 116 /*
 117  * /proc/fs/lustre/sptlrpc/encrypt_page_pools
 118  */
 119 int sptlrpc_proc_enc_pool_seq_show(struct seq_file *m, void *v)
 120 {
 121         spin_lock(&page_pools.epp_lock);
 122
 123         seq_printf(m, "physical pages:          %lu\n"
 124                    "pages per pool:          %lu\n"
 125                    "max pages:               %lu\n"
 126                    "max pools:               %u\n"
 127                    "total pages:             %lu\n"
 128                    "total free:              %lu\n"
 129                    "idle index:              %lu/100\n"
 130                    "last shrink:             %llds\n"
 131                    "last access:             %llds\n"
 132                    "max pages reached:       %lu\n"
 133                    "grows:                   %u\n"
 134                    "grows failure:           %u\n"
 135                    "shrinks:                 %u\n"
 136                    "cache access:            %lu\n"
 137                    "cache missing:           %lu\n"
 138                    "low free mark:           %lu\n"
 139                    "max waitqueue depth:     %u\n"
 140                    "max wait time ms:        %lld\n"
 141                    "out of mem:              %lu\n",
 142                    cfs_totalram_pages(), PAGES_PER_POOL,
 143                    page_pools.epp_max_pages,
 144                    page_pools.epp_max_pools,
 145                    page_pools.epp_total_pages,
 146                    page_pools.epp_free_pages,
 147                    page_pools.epp_idle_idx,
 148                    ktime_get_seconds() - page_pools.epp_last_shrink,
 149                    ktime_get_seconds() - page_pools.epp_last_access,
 150                    page_pools.epp_st_max_pages,
 151                    page_pools.epp_st_grows,
 152                    page_pools.epp_st_grow_fails,
 153                    page_pools.epp_st_shrinks,
 154                    page_pools.epp_st_access,
 155                    page_pools.epp_st_missings,
 156                    page_pools.epp_st_lowfree,
 157                    page_pools.epp_st_max_wqlen,
 158                    ktime_to_ms(page_pools.epp_st_max_wait),
 159                    page_pools.epp_st_outofmem);
 160
 161         spin_unlock(&page_pools.epp_lock);
 162         return 0;
 163 }
 164
 165 static void enc_pools_release_free_pages(long npages)
 166 {
 167         int p_idx, g_idx;
 168         int p_idx_max1, p_idx_max2;
 169
 170         LASSERT(npages > 0);
 171         LASSERT(npages <= page_pools.epp_free_pages);
 172         LASSERT(page_pools.epp_free_pages <= page_pools.epp_total_pages);
 173
 174         /* max pool index before the release */
 175         p_idx_max2 = (page_pools.epp_total_pages - 1) / PAGES_PER_POOL;
 176
 177         page_pools.epp_free_pages -= npages;
 178         page_pools.epp_total_pages -= npages;
 179
 180         /* max pool index after the release */
 181         p_idx_max1 = page_pools.epp_total_pages == 0 ? -1 :
 182                 ((page_pools.epp_total_pages - 1) / PAGES_PER_POOL);
 183
 184         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 185         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 186         LASSERT(page_pools.epp_pools[p_idx]);
 187
 188         while (npages--) {
 189                 LASSERT(page_pools.epp_pools[p_idx]);
 190                 LASSERT(page_pools.epp_pools[p_idx][g_idx] != NULL);
 191
 192                 __free_page(page_pools.epp_pools[p_idx][g_idx]);
 193                 page_pools.epp_pools[p_idx][g_idx] = NULL;
 194
 195                 if (++g_idx == PAGES_PER_POOL) {
 196                         p_idx++;
 197                         g_idx = 0;
 198                 }
 199         }
 200
 201         /* free unused pools */
 202         while (p_idx_max1 < p_idx_max2) {
 203                 LASSERT(page_pools.epp_pools[p_idx_max2]);
 204                 OBD_FREE(page_pools.epp_pools[p_idx_max2], PAGE_SIZE);
 205                 page_pools.epp_pools[p_idx_max2] = NULL;
 206                 p_idx_max2--;
 207         }
 208 }
 209
 210 /*
 211  * we try to keep at least PTLRPC_MAX_BRW_PAGES pages in the pool.
 212  */
 213 static unsigned long enc_pools_shrink_count(struct shrinker *s,
 214                                             struct shrink_control *sc)
 215 {
 216         /*
 217          * if no pool access for a long time, we consider it's fully idle.
 218          * a little race here is fine.
 219          */
 220         if (unlikely(ktime_get_seconds() - page_pools.epp_last_access >
 221                      CACHE_QUIESCENT_PERIOD)) {
 222                 spin_lock(&page_pools.epp_lock);
 223                 page_pools.epp_idle_idx = IDLE_IDX_MAX;
 224                 spin_unlock(&page_pools.epp_lock);
 225         }
 226
 227         LASSERT(page_pools.epp_idle_idx <= IDLE_IDX_MAX);
 228         return (page_pools.epp_free_pages <= PTLRPC_MAX_BRW_PAGES) ? 0 :
 229                 (page_pools.epp_free_pages - PTLRPC_MAX_BRW_PAGES) *
 230                 (IDLE_IDX_MAX - page_pools.epp_idle_idx) / IDLE_IDX_MAX;
 231 }
 232
 233 /*
 234  * we try to keep at least PTLRPC_MAX_BRW_PAGES pages in the pool.
 235  */
 236 static unsigned long enc_pools_shrink_scan(struct shrinker *s,
 237                                            struct shrink_control *sc)
 238 {
 239         spin_lock(&page_pools.epp_lock);
 240         if (page_pools.epp_free_pages <= PTLRPC_MAX_BRW_PAGES)
 241                 sc->nr_to_scan = 0;
 242         else
 243                 sc->nr_to_scan = min_t(unsigned long, sc->nr_to_scan,
 244                               page_pools.epp_free_pages - PTLRPC_MAX_BRW_PAGES);
 245         if (sc->nr_to_scan > 0) {
 246                 enc_pools_release_free_pages(sc->nr_to_scan);
 247                 CDEBUG(D_SEC, "released %ld pages, %ld left\n",
 248                        (long)sc->nr_to_scan, page_pools.epp_free_pages);
 249
 250                 page_pools.epp_st_shrinks++;
 251                 page_pools.epp_last_shrink = ktime_get_seconds();
 252         }
 253         spin_unlock(&page_pools.epp_lock);
 254
 255         /*
 256          * if no pool access for a long time, we consider it's fully idle.
 257          * a little race here is fine.
 258          */
 259         if (unlikely(ktime_get_seconds() - page_pools.epp_last_access >
 260                      CACHE_QUIESCENT_PERIOD)) {
 261                 spin_lock(&page_pools.epp_lock);
 262                 page_pools.epp_idle_idx = IDLE_IDX_MAX;
 263                 spin_unlock(&page_pools.epp_lock);
 264         }
 265
 266         LASSERT(page_pools.epp_idle_idx <= IDLE_IDX_MAX);
 267         return sc->nr_to_scan;
 268 }
 269
 270 #ifdef HAVE_SHRINKER_COUNT
 271 static struct shrinker pools_shrinker = {
 272         .count_objects  = enc_pools_shrink_count,
 273         .scan_objects   = enc_pools_shrink_scan,
 274         .seeks          = DEFAULT_SEEKS,
 275 };
 276 #else
 277 /*
 278  * could be called frequently for query (@nr_to_scan == 0).
 279  * we try to keep at least PTLRPC_MAX_BRW_PAGES pages in the pool.
 280  */
 281 static int enc_pools_shrink(struct shrinker *shrinker,
 282                             struct shrink_control *sc)
 283 {
 284         enc_pools_shrink_scan(shrinker, sc);
 285
 286         return enc_pools_shrink_count(shrinker, sc);
 287 }
 288
 289 static struct shrinker pools_shrinker = {
 290         .shrink  = enc_pools_shrink,
 291         .seeks   = DEFAULT_SEEKS,
 292 };
 293 #endif /* HAVE_SHRINKER_COUNT */
 294
 295 static inline
 296 int npages_to_npools(unsigned long npages)
 297 {
 298         return (int) ((npages + PAGES_PER_POOL - 1) / PAGES_PER_POOL);
 299 }
 300
 301 /*
 302  * return how many pages cleaned up.
 303  */
 304 static unsigned long enc_pools_cleanup(struct page ***pools, int npools)
 305 {
 306         unsigned long cleaned = 0;
 307         int i, j;
 308
 309         for (i = 0; i < npools; i++) {
 310                 if (pools[i]) {
 311                         for (j = 0; j < PAGES_PER_POOL; j++) {
 312                                 if (pools[i][j]) {
 313                                         __free_page(pools[i][j]);
 314                                         cleaned++;
 315                                 }
 316                         }
 317                         OBD_FREE(pools[i], PAGE_SIZE);
 318                         pools[i] = NULL;
 319                 }
 320         }
 321
 322         return cleaned;
 323 }
 324
 325 /*
 326  * merge @npools pointed by @pools which contains @npages new pages
 327  * into current pools.
 328  *
 329  * we have options to avoid most memory copy with some tricks. but we choose
 330  * the simplest way to avoid complexity. It's not frequently called.
 331  */
 332 static void enc_pools_insert(struct page ***pools, int npools, int npages)
 333 {
 334         int freeslot;
 335         int op_idx, np_idx, og_idx, ng_idx;
 336         int cur_npools, end_npools;
 337
 338         LASSERT(npages > 0);
 339         LASSERT(page_pools.epp_total_pages+npages <= page_pools.epp_max_pages);
 340         LASSERT(npages_to_npools(npages) == npools);
 341         LASSERT(page_pools.epp_growing);
 342
 343         spin_lock(&page_pools.epp_lock);
 344
 345         /*
 346          * (1) fill all the free slots of current pools.
 347          */
 348         /*
 349          * free slots are those left by rent pages, and the extra ones with
 350          * index >= total_pages, locate at the tail of last pool.
 351          */
 352         freeslot = page_pools.epp_total_pages % PAGES_PER_POOL;
 353         if (freeslot != 0)
 354                 freeslot = PAGES_PER_POOL - freeslot;
 355         freeslot += page_pools.epp_total_pages - page_pools.epp_free_pages;
 356
 357         op_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 358         og_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 359         np_idx = npools - 1;
 360         ng_idx = (npages - 1) % PAGES_PER_POOL;
 361
 362         while (freeslot) {
 363                 LASSERT(page_pools.epp_pools[op_idx][og_idx] == NULL);
 364                 LASSERT(pools[np_idx][ng_idx] != NULL);
 365
 366                 page_pools.epp_pools[op_idx][og_idx] = pools[np_idx][ng_idx];
 367                 pools[np_idx][ng_idx] = NULL;
 368
 369                 freeslot--;
 370
 371                 if (++og_idx == PAGES_PER_POOL) {
 372                         op_idx++;
 373                         og_idx = 0;
 374                 }
 375                 if (--ng_idx < 0) {
 376                         if (np_idx == 0)
 377                                 break;
 378                         np_idx--;
 379                         ng_idx = PAGES_PER_POOL - 1;
 380                 }
 381         }
 382
 383         /*
 384          * (2) add pools if needed.
 385          */
 386         cur_npools = (page_pools.epp_total_pages + PAGES_PER_POOL - 1) /
 387                       PAGES_PER_POOL;
 388         end_npools = (page_pools.epp_total_pages + npages +
 389                       PAGES_PER_POOL - 1) / PAGES_PER_POOL;
 390         LASSERT(end_npools <= page_pools.epp_max_pools);
 391
 392         np_idx = 0;
 393         while (cur_npools < end_npools) {
 394                 LASSERT(page_pools.epp_pools[cur_npools] == NULL);
 395                 LASSERT(np_idx < npools);
 396                 LASSERT(pools[np_idx] != NULL);
 397
 398                 page_pools.epp_pools[cur_npools++] = pools[np_idx];
 399                 pools[np_idx++] = NULL;
 400         }
 401
 402         page_pools.epp_total_pages += npages;
 403         page_pools.epp_free_pages += npages;
 404         page_pools.epp_st_lowfree = page_pools.epp_free_pages;
 405
 406         if (page_pools.epp_total_pages > page_pools.epp_st_max_pages)
 407                 page_pools.epp_st_max_pages = page_pools.epp_total_pages;
 408
 409         CDEBUG(D_SEC, "add %d pages to total %lu\n", npages,
 410                page_pools.epp_total_pages);
 411
 412         spin_unlock(&page_pools.epp_lock);
 413 }
 414
 415 static int enc_pools_add_pages(int npages)
 416 {
 417         static DEFINE_MUTEX(add_pages_mutex);
 418         struct page ***pools;
 419         int npools, alloced = 0;
 420         int i, j, rc = -ENOMEM;
 421
 422         if (npages < PTLRPC_MAX_BRW_PAGES)
 423                 npages = PTLRPC_MAX_BRW_PAGES;
 424
 425         mutex_lock(&add_pages_mutex);
 426
 427         if (npages + page_pools.epp_total_pages > page_pools.epp_max_pages)
 428                 npages = page_pools.epp_max_pages - page_pools.epp_total_pages;
 429         LASSERT(npages > 0);
 430
 431         page_pools.epp_st_grows++;
 432
 433         npools = npages_to_npools(npages);
 434         OBD_ALLOC_PTR_ARRAY(pools, npools);
 435         if (pools == NULL)
 436                 goto out;
 437
 438         for (i = 0; i < npools; i++) {
 439                 OBD_ALLOC(pools[i], PAGE_SIZE);
 440                 if (pools[i] == NULL)
 441                         goto out_pools;
 442
 443                 for (j = 0; j < PAGES_PER_POOL && alloced < npages; j++) {
 444                         pools[i][j] = alloc_page(GFP_NOFS |
 445                                                  __GFP_HIGHMEM);
 446                         if (pools[i][j] == NULL)
 447                                 goto out_pools;
 448
 449                         alloced++;
 450                 }
 451         }
 452         LASSERT(alloced == npages);
 453
 454         enc_pools_insert(pools, npools, npages);
 455         CDEBUG(D_SEC, "added %d pages into pools\n", npages);
 456         rc = 0;
 457
 458 out_pools:
 459         enc_pools_cleanup(pools, npools);
 460         OBD_FREE_PTR_ARRAY(pools, npools);
 461 out:
 462         if (rc) {
 463                 page_pools.epp_st_grow_fails++;
 464                 CERROR("Failed to allocate %d enc pages\n", npages);
 465         }
 466
 467         mutex_unlock(&add_pages_mutex);
 468         return rc;
 469 }
 470
 471 static inline void enc_pools_wakeup(void)
 472 {
 473         assert_spin_locked(&page_pools.epp_lock);
 474
 475         /* waitqueue_active */
 476         if (unlikely(waitqueue_active(&page_pools.epp_waitq)))
 477                 wake_up(&page_pools.epp_waitq);
 478 }
 479
 480 static int enc_pools_should_grow(int page_needed, time64_t now)
 481 {
 482         /*
 483          * don't grow if someone else is growing the pools right now,
 484          * or the pools has reached its full capacity
 485          */
 486         if (page_pools.epp_growing ||
 487             page_pools.epp_total_pages == page_pools.epp_max_pages)
 488                 return 0;
 489
 490         /* if total pages is not enough, we need to grow */
 491         if (page_pools.epp_total_pages < page_needed)
 492                 return 1;
 493
 494         /*
 495          * we wanted to return 0 here if there was a shrink just
 496          * happened a moment ago, but this may cause deadlock if both
 497          * client and ost live on single node.
 498          */
 499
 500         /*
 501          * here we perhaps need consider other factors like wait queue
 502          * length, idle index, etc. ?
 503          */
 504
 505         /* grow the pools in any other cases */
 506         return 1;
 507 }
 508
 509 /*
 510  * Export the number of free pages in the pool
 511  */
 512 int get_free_pages_in_pool(void)
 513 {
 514         return page_pools.epp_free_pages;
 515 }
 516 EXPORT_SYMBOL(get_free_pages_in_pool);
 517
 518 /*
 519  * Let outside world know if enc_pool full capacity is reached
 520  */
 521 int pool_is_at_full_capacity(void)
 522 {
 523         return (page_pools.epp_total_pages == page_pools.epp_max_pages);
 524 }
 525 EXPORT_SYMBOL(pool_is_at_full_capacity);
 526
 527 static inline struct page **page_from_bulkdesc(void *array, int index)
 528 {
 529         struct ptlrpc_bulk_desc *desc = (struct ptlrpc_bulk_desc *)array;
 530
 531         return &desc->bd_enc_vec[index].bv_page;
 532 }
 533
 534 static inline struct page **page_from_pagearray(void *array, int index)
 535 {
 536         struct page **pa = (struct page **)array;
 537
 538         return &pa[index];
 539 }
 540
 541 /*
 542  * we allocate the requested pages atomically.
 543  */
 544 static inline int __sptlrpc_enc_pool_get_pages(void *array, unsigned int count,
 545                                         struct page **(*page_from)(void *, int))
 546 {
 547         wait_queue_entry_t waitlink;
 548         unsigned long this_idle = -1;
 549         u64 tick_ns = 0;
 550         time64_t now;
 551         int p_idx, g_idx;
 552         int i, rc = 0;
 553
 554         if (!array || count <= 0 || count > page_pools.epp_max_pages)
 555                 return -EINVAL;
 556
 557         spin_lock(&page_pools.epp_lock);
 558
 559         page_pools.epp_st_access++;
 560 again:
 561         if (unlikely(page_pools.epp_free_pages < count)) {
 562                 if (tick_ns == 0)
 563                         tick_ns = ktime_get_ns();
 564
 565                 now = ktime_get_real_seconds();
 566
 567                 page_pools.epp_st_missings++;
 568                 page_pools.epp_pages_short += count;
 569
 570                 if (enc_pools_should_grow(count, now)) {
 571                         page_pools.epp_growing = 1;
 572
 573                         spin_unlock(&page_pools.epp_lock);
 574                         enc_pools_add_pages(page_pools.epp_pages_short / 2);
 575                         spin_lock(&page_pools.epp_lock);
 576
 577                         page_pools.epp_growing = 0;
 578
 579                         enc_pools_wakeup();
 580                 } else {
 581                         if (page_pools.epp_growing) {
 582                                 if (++page_pools.epp_waitqlen >
 583                                     page_pools.epp_st_max_wqlen)
 584                                         page_pools.epp_st_max_wqlen =
 585                                                 page_pools.epp_waitqlen;
 586
 587                                 set_current_state(TASK_UNINTERRUPTIBLE);
 588                                 init_wait(&waitlink);
 589                                 add_wait_queue(&page_pools.epp_waitq,
 590                                                &waitlink);
 591
 592                                 spin_unlock(&page_pools.epp_lock);
 593                                 schedule();
 594                                 remove_wait_queue(&page_pools.epp_waitq,
 595                                                   &waitlink);
 596                                 spin_lock(&page_pools.epp_lock);
 597                                 page_pools.epp_waitqlen--;
 598                         } else {
 599                                 /*
 600                                  * ptlrpcd thread should not sleep in that case,
 601                                  * or deadlock may occur!
 602                                  * Instead, return -ENOMEM so that upper layers
 603                                  * will put request back in queue.
 604                                  */
 605                                 page_pools.epp_st_outofmem++;
 606                                 GOTO(out_unlock, rc = -ENOMEM);
 607                         }
 608                 }
 609
 610                 if (page_pools.epp_pages_short < count)
 611                         GOTO(out_unlock, rc = -EPROTO);
 612                 page_pools.epp_pages_short -= count;
 613
 614                 this_idle = 0;
 615                 goto again;
 616         }
 617
 618         /* record max wait time */
 619         if (unlikely(tick_ns)) {
 620                 ktime_t tick = ktime_sub_ns(ktime_get(), tick_ns);
 621
 622                 if (ktime_after(tick, page_pools.epp_st_max_wait))
 623                         page_pools.epp_st_max_wait = tick;
 624         }
 625
 626         /* proceed with rest of allocation */
 627         page_pools.epp_free_pages -= count;
 628
 629         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 630         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 631
 632         for (i = 0; i < count; i++) {
 633                 struct page **pagep = page_from(array, i);
 634
 635                 if (page_pools.epp_pools[p_idx][g_idx] == NULL)
 636                         GOTO(out_unlock, rc = -EPROTO);
 637                 *pagep = page_pools.epp_pools[p_idx][g_idx];
 638                 page_pools.epp_pools[p_idx][g_idx] = NULL;
 639
 640                 if (++g_idx == PAGES_PER_POOL) {
 641                         p_idx++;
 642                         g_idx = 0;
 643                 }
 644         }
 645
 646         if (page_pools.epp_free_pages < page_pools.epp_st_lowfree)
 647                 page_pools.epp_st_lowfree = page_pools.epp_free_pages;
 648
 649         /*
 650          * new idle index = (old * weight + new) / (weight + 1)
 651          */
 652         if (this_idle == -1) {
 653                 this_idle = page_pools.epp_free_pages * IDLE_IDX_MAX /
 654                         page_pools.epp_total_pages;
 655         }
 656         page_pools.epp_idle_idx = (page_pools.epp_idle_idx * IDLE_IDX_WEIGHT +
 657                                    this_idle) /
 658                 (IDLE_IDX_WEIGHT + 1);
 659
 660         page_pools.epp_last_access = ktime_get_seconds();
 661
 662 out_unlock:
 663         spin_unlock(&page_pools.epp_lock);
 664         return rc;
 665 }
 666
 667 int sptlrpc_enc_pool_get_pages(struct ptlrpc_bulk_desc *desc)
 668 {
 669         int rc;
 670
 671         LASSERT(desc->bd_iov_count > 0);
 672         LASSERT(desc->bd_iov_count <= page_pools.epp_max_pages);
 673
 674         /* resent bulk, enc iov might have been allocated previously */
 675         if (desc->bd_enc_vec != NULL)
 676                 return 0;
 677
 678         OBD_ALLOC_LARGE(desc->bd_enc_vec,
 679                         desc->bd_iov_count * sizeof(*desc->bd_enc_vec));
 680         if (desc->bd_enc_vec == NULL)
 681                 return -ENOMEM;
 682
 683         rc = __sptlrpc_enc_pool_get_pages((void *)desc, desc->bd_iov_count,
 684                                           page_from_bulkdesc);
 685         if (rc) {
 686                 OBD_FREE_LARGE(desc->bd_enc_vec,
 687                                desc->bd_iov_count *
 688                                sizeof(*desc->bd_enc_vec));
 689                 desc->bd_enc_vec = NULL;
 690         }
 691         return rc;
 692 }
 693 EXPORT_SYMBOL(sptlrpc_enc_pool_get_pages);
 694
 695 int sptlrpc_enc_pool_get_pages_array(struct page **pa, unsigned int count)
 696 {
 697         return __sptlrpc_enc_pool_get_pages((void *)pa, count,
 698                                             page_from_pagearray);
 699 }
 700 EXPORT_SYMBOL(sptlrpc_enc_pool_get_pages_array);
 701
 702 static int __sptlrpc_enc_pool_put_pages(void *array, unsigned int count,
 703                                         struct page **(*page_from)(void *, int))
 704 {
 705         int p_idx, g_idx;
 706         int i, rc = 0;
 707
 708         if (!array || count <= 0)
 709                 return -EINVAL;
 710
 711         spin_lock(&page_pools.epp_lock);
 712
 713         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 714         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 715
 716         if (page_pools.epp_free_pages + count > page_pools.epp_total_pages)
 717                 GOTO(out_unlock, rc = -EPROTO);
 718         if (!page_pools.epp_pools[p_idx])
 719                 GOTO(out_unlock, rc = -EPROTO);
 720
 721         for (i = 0; i < count; i++) {
 722                 struct page **pagep = page_from(array, i);
 723
 724                 if (!*pagep ||
 725                     page_pools.epp_pools[p_idx][g_idx] != NULL)
 726                         GOTO(out_unlock, rc = -EPROTO);
 727
 728                 page_pools.epp_pools[p_idx][g_idx] = *pagep;
 729                 if (++g_idx == PAGES_PER_POOL) {
 730                         p_idx++;
 731                         g_idx = 0;
 732                 }
 733         }
 734
 735         page_pools.epp_free_pages += count;
 736         enc_pools_wakeup();
 737
 738 out_unlock:
 739         spin_unlock(&page_pools.epp_lock);
 740         return rc;
 741 }
 742
 743 void sptlrpc_enc_pool_put_pages(struct ptlrpc_bulk_desc *desc)
 744 {
 745         int rc;
 746
 747         if (desc->bd_enc_vec == NULL)
 748                 return;
 749
 750         rc = __sptlrpc_enc_pool_put_pages((void *)desc, desc->bd_iov_count,
 751                                           page_from_bulkdesc);
 752         if (rc)
 753                 CDEBUG(D_SEC, "error putting pages in enc pool: %d\n", rc);
 754
 755         OBD_FREE_LARGE(desc->bd_enc_vec,
 756                        desc->bd_iov_count * sizeof(*desc->bd_enc_vec));
 757         desc->bd_enc_vec = NULL;
 758 }
 759
 760 void sptlrpc_enc_pool_put_pages_array(struct page **pa, unsigned int count)
 761 {
 762         int rc;
 763
 764         rc = __sptlrpc_enc_pool_put_pages((void *)pa, count,
 765                                           page_from_pagearray);
 766
 767         if (rc)
 768                 CDEBUG(D_SEC, "error putting pages in enc pool: %d\n", rc);
 769 }
 770 EXPORT_SYMBOL(sptlrpc_enc_pool_put_pages_array);
 771
 772 /*
 773  * we don't do much stuff for add_user/del_user anymore, except adding some
 774  * initial pages in add_user() if current pools are empty, rest would be
 775  * handled by the pools's self-adaption.
 776  */
 777 int sptlrpc_enc_pool_add_user(void)
 778 {
 779         int need_grow = 0;
 780
 781         spin_lock(&page_pools.epp_lock);
 782         if (page_pools.epp_growing == 0 && page_pools.epp_total_pages == 0) {
 783                 page_pools.epp_growing = 1;
 784                 need_grow = 1;
 785         }
 786         spin_unlock(&page_pools.epp_lock);
 787
 788         if (need_grow) {
 789                 enc_pools_add_pages(PTLRPC_MAX_BRW_PAGES +
 790                                     PTLRPC_MAX_BRW_PAGES);
 791
 792                 spin_lock(&page_pools.epp_lock);
 793                 page_pools.epp_growing = 0;
 794                 enc_pools_wakeup();
 795                 spin_unlock(&page_pools.epp_lock);
 796         }
 797         return 0;
 798 }
 799 EXPORT_SYMBOL(sptlrpc_enc_pool_add_user);
 800
 801 static inline void enc_pools_alloc(void)
 802 {
 803         LASSERT(page_pools.epp_max_pools);
 804         OBD_ALLOC_LARGE(page_pools.epp_pools,
 805                         page_pools.epp_max_pools *
 806                         sizeof(*page_pools.epp_pools));
 807 }
 808
 809 static inline void enc_pools_free(void)
 810 {
 811         LASSERT(page_pools.epp_max_pools);
 812         LASSERT(page_pools.epp_pools);
 813
 814         OBD_FREE_LARGE(page_pools.epp_pools,
 815                        page_pools.epp_max_pools *
 816                        sizeof(*page_pools.epp_pools));
 817 }
 818
 819 int sptlrpc_enc_pool_init(void)
 820 {
 821         int rc;
 822
 823         page_pools.epp_max_pages = cfs_totalram_pages() / 8;
 824         if (enc_pool_max_memory_mb > 0 &&
 825             enc_pool_max_memory_mb <= (cfs_totalram_pages() >> mult))
 826                 page_pools.epp_max_pages = enc_pool_max_memory_mb << mult;
 827
 828         page_pools.epp_max_pools = npages_to_npools(page_pools.epp_max_pages);
 829
 830         init_waitqueue_head(&page_pools.epp_waitq);
 831         page_pools.epp_waitqlen = 0;
 832         page_pools.epp_pages_short = 0;
 833
 834         page_pools.epp_growing = 0;
 835
 836         page_pools.epp_idle_idx = 0;
 837         page_pools.epp_last_shrink = ktime_get_seconds();
 838         page_pools.epp_last_access = ktime_get_seconds();
 839
 840         spin_lock_init(&page_pools.epp_lock);
 841         page_pools.epp_total_pages = 0;
 842         page_pools.epp_free_pages = 0;
 843
 844         page_pools.epp_st_max_pages = 0;
 845         page_pools.epp_st_grows = 0;
 846         page_pools.epp_st_grow_fails = 0;
 847         page_pools.epp_st_shrinks = 0;
 848         page_pools.epp_st_access = 0;
 849         page_pools.epp_st_missings = 0;
 850         page_pools.epp_st_lowfree = 0;
 851         page_pools.epp_st_max_wqlen = 0;
 852         page_pools.epp_st_max_wait = ktime_set(0, 0);
 853         page_pools.epp_st_outofmem = 0;
 854
 855         enc_pools_alloc();
 856         if (page_pools.epp_pools == NULL)
 857                 return -ENOMEM;
 858
 859         rc = register_shrinker(&pools_shrinker);
 860         if (rc)
 861                 enc_pools_free();
 862
 863         return rc;
 864 }
 865
 866 void sptlrpc_enc_pool_fini(void)
 867 {
 868         unsigned long cleaned, npools;
 869
 870         LASSERT(page_pools.epp_pools);
 871         LASSERT(page_pools.epp_total_pages == page_pools.epp_free_pages);
 872
 873         unregister_shrinker(&pools_shrinker);
 874
 875         npools = npages_to_npools(page_pools.epp_total_pages);
 876         cleaned = enc_pools_cleanup(page_pools.epp_pools, npools);
 877         LASSERT(cleaned == page_pools.epp_total_pages);
 878
 879         enc_pools_free();
 880
 881         if (page_pools.epp_st_access > 0) {
 882                 CDEBUG(D_SEC,
 883                        "max pages %lu, grows %u, grow fails %u, shrinks %u, access %lu, missing %lu, max qlen %u, max wait ms %lld, out of mem %lu\n",
 884                        page_pools.epp_st_max_pages, page_pools.epp_st_grows,
 885                        page_pools.epp_st_grow_fails,
 886                        page_pools.epp_st_shrinks, page_pools.epp_st_access,
 887                        page_pools.epp_st_missings, page_pools.epp_st_max_wqlen,
 888                        ktime_to_ms(page_pools.epp_st_max_wait),
 889                        page_pools.epp_st_outofmem);
 890         }
 891 }
 892
 893
 894 static int cfs_hash_alg_id[] = {
 895         [BULK_HASH_ALG_NULL]    = CFS_HASH_ALG_NULL,
 896         [BULK_HASH_ALG_ADLER32] = CFS_HASH_ALG_ADLER32,
 897         [BULK_HASH_ALG_CRC32]   = CFS_HASH_ALG_CRC32,
 898         [BULK_HASH_ALG_MD5]     = CFS_HASH_ALG_MD5,
 899         [BULK_HASH_ALG_SHA1]    = CFS_HASH_ALG_SHA1,
 900         [BULK_HASH_ALG_SHA256]  = CFS_HASH_ALG_SHA256,
 901         [BULK_HASH_ALG_SHA384]  = CFS_HASH_ALG_SHA384,
 902         [BULK_HASH_ALG_SHA512]  = CFS_HASH_ALG_SHA512,
 903 };
 904 const char *sptlrpc_get_hash_name(__u8 hash_alg)
 905 {
 906         return cfs_crypto_hash_name(cfs_hash_alg_id[hash_alg]);
 907 }
 908
 909 __u8 sptlrpc_get_hash_alg(const char *algname)
 910 {
 911         return cfs_crypto_hash_alg(algname);
 912 }
 913
 914 int bulk_sec_desc_unpack(struct lustre_msg *msg, int offset, int swabbed)
 915 {
 916         struct ptlrpc_bulk_sec_desc *bsd;
 917         int size = msg->lm_buflens[offset];
 918
 919         bsd = lustre_msg_buf(msg, offset, sizeof(*bsd));
 920         if (bsd == NULL) {
 921                 CERROR("Invalid bulk sec desc: size %d\n", size);
 922                 return -EINVAL;
 923         }
 924
 925         if (swabbed)
 926                 __swab32s(&bsd->bsd_nob);
 927
 928         if (unlikely(bsd->bsd_version != 0)) {
 929                 CERROR("Unexpected version %u\n", bsd->bsd_version);
 930                 return -EPROTO;
 931         }
 932
 933         if (unlikely(bsd->bsd_type >= SPTLRPC_BULK_MAX)) {
 934                 CERROR("Invalid type %u\n", bsd->bsd_type);
 935                 return -EPROTO;
 936         }
 937
 938         /* FIXME more sanity check here */
 939
 940         if (unlikely(bsd->bsd_svc != SPTLRPC_BULK_SVC_NULL &&
 941                      bsd->bsd_svc != SPTLRPC_BULK_SVC_INTG &&
 942                      bsd->bsd_svc != SPTLRPC_BULK_SVC_PRIV)) {
 943                 CERROR("Invalid svc %u\n", bsd->bsd_svc);
 944                 return -EPROTO;
 945         }
 946
 947         return 0;
 948 }
 949 EXPORT_SYMBOL(bulk_sec_desc_unpack);
 950
 951 /*
 952  * Compute the checksum of an RPC buffer payload.  If the return \a buflen
 953  * is not large enough, truncate the result to fit so that it is possible
 954  * to use a hash function with a large hash space, but only use a part of
 955  * the resulting hash.
 956  */
 957 int sptlrpc_get_bulk_checksum(struct ptlrpc_bulk_desc *desc, __u8 alg,
 958                               void *buf, int buflen)
 959 {
 960         struct ahash_request *req;
 961         int hashsize;
 962         unsigned int bufsize;
 963         int i, err;
 964
 965         LASSERT(alg > BULK_HASH_ALG_NULL && alg < BULK_HASH_ALG_MAX);
 966         LASSERT(buflen >= 4);
 967
 968         req = cfs_crypto_hash_init(cfs_hash_alg_id[alg], NULL, 0);
 969         if (IS_ERR(req)) {
 970                 CERROR("Unable to initialize checksum hash %s\n",
 971                        cfs_crypto_hash_name(cfs_hash_alg_id[alg]));
 972                 return PTR_ERR(req);
 973         }
 974
 975         hashsize = cfs_crypto_hash_digestsize(cfs_hash_alg_id[alg]);
 976
 977         for (i = 0; i < desc->bd_iov_count; i++) {
 978                 cfs_crypto_hash_update_page(req,
 979                                   desc->bd_vec[i].bv_page,
 980                                   desc->bd_vec[i].bv_offset &
 981                                               ~PAGE_MASK,
 982                                   desc->bd_vec[i].bv_len);
 983         }
 984
 985         if (hashsize > buflen) {
 986                 unsigned char hashbuf[CFS_CRYPTO_HASH_DIGESTSIZE_MAX];
 987
 988                 bufsize = sizeof(hashbuf);
 989                 LASSERTF(bufsize >= hashsize, "bufsize = %u < hashsize %u\n",
 990                          bufsize, hashsize);
 991                 err = cfs_crypto_hash_final(req, hashbuf, &bufsize);
 992                 memcpy(buf, hashbuf, buflen);
 993         } else {
 994                 bufsize = buflen;
 995                 err = cfs_crypto_hash_final(req, buf, &bufsize);
 996         }
 997
 998         return err;
 999 }