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  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/ptlrpc/sec_bulk.c
  33  *
  34  * Author: Eric Mei <ericm@clusterfs.com>
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_SEC
  38
  39 #include <libcfs/linux/linux-mem.h>
  40
  41 #include <obd.h>
  42 #include <obd_cksum.h>
  43 #include <obd_class.h>
  44 #include <obd_support.h>
  45 #include <lustre_net.h>
  46 #include <lustre_import.h>
  47 #include <lustre_dlm.h>
  48 #include <lustre_sec.h>
  49
  50 #include "ptlrpc_internal.h"
  51
  52 static int mult = 20 - PAGE_SHIFT;
  53 static int enc_pool_max_memory_mb;
  54 module_param(enc_pool_max_memory_mb, int, 0644);
  55 MODULE_PARM_DESC(enc_pool_max_memory_mb,
  56                  "Encoding pool max memory (MB), 1/8 of total physical memory by default");
  57
  58 /*
  59  * bulk encryption page pools
  60  */
  61
  62 #define PTRS_PER_PAGE   (PAGE_SIZE / sizeof(void *))
  63 #define PAGES_PER_POOL  (PTRS_PER_PAGE)
  64
  65 #define IDLE_IDX_MAX            (100)
  66 #define IDLE_IDX_WEIGHT         (3)
  67
  68 #define CACHE_QUIESCENT_PERIOD  (20)
  69
  70 static struct ptlrpc_enc_page_pool {
  71         unsigned long epp_max_pages;   /* maximum pages can hold, const */
  72         unsigned int epp_max_pools;   /* number of pools, const */
  73
  74         /*
  75          * wait queue in case of not enough free pages.
  76          */
  77         wait_queue_head_t epp_waitq;   /* waiting threads */
  78         unsigned int epp_waitqlen;    /* wait queue length */
  79         unsigned long epp_pages_short; /* # of pages wanted of in-q users */
  80         unsigned int epp_growing:1;   /* during adding pages */
  81
  82         /*
  83          * indicating how idle the pools are, from 0 to MAX_IDLE_IDX
  84          * this is counted based on each time when getting pages from
  85          * the pools, not based on time. which means in case that system
  86          * is idled for a while but the idle_idx might still be low if no
  87          * activities happened in the pools.
  88          */
  89         unsigned long epp_idle_idx;
  90
  91         /* last shrink time due to mem tight */
  92         time64_t epp_last_shrink;
  93         time64_t epp_last_access;
  94
  95         /* in-pool pages bookkeeping */
  96         spinlock_t epp_lock; /* protect following fields */
  97         unsigned long epp_total_pages; /* total pages in pools */
  98         unsigned long epp_free_pages;  /* current pages available */
  99
 100         /* statistics */
 101         unsigned long epp_st_max_pages;      /* # of pages ever reached */
 102         unsigned int epp_st_grows;          /* # of grows */
 103         unsigned int epp_st_grow_fails;     /* # of add pages failures */
 104         unsigned int epp_st_shrinks;        /* # of shrinks */
 105         unsigned long epp_st_access;         /* # of access */
 106         unsigned long epp_st_missings;       /* # of cache missing */
 107         unsigned long epp_st_lowfree;        /* lowest free pages reached */
 108         unsigned int epp_st_max_wqlen;      /* highest waitqueue length */
 109         ktime_t epp_st_max_wait; /* in nanoseconds */
 110         unsigned long epp_st_outofmem; /* # of out of mem requests */
 111         /*
 112          * pointers to pools, may be vmalloc'd
 113          */
 114         struct page ***epp_pools;
 115 } page_pools;
 116
 117 /*
 118  * memory shrinker
 119  */
 120 static const int pools_shrinker_seeks = DEFAULT_SEEKS;
 121 static struct shrinker *pools_shrinker;
 122
 123
 124 /*
 125  * /proc/fs/lustre/sptlrpc/encrypt_page_pools
 126  */
 127 int sptlrpc_proc_enc_pool_seq_show(struct seq_file *m, void *v)
 128 {
 129         spin_lock(&page_pools.epp_lock);
 130
 131         seq_printf(m, "physical pages:          %lu\n"
 132                    "pages per pool:          %lu\n"
 133                    "max pages:               %lu\n"
 134                    "max pools:               %u\n"
 135                    "total pages:             %lu\n"
 136                    "total free:              %lu\n"
 137                    "idle index:              %lu/100\n"
 138                    "last shrink:             %llds\n"
 139                    "last access:             %llds\n"
 140                    "max pages reached:       %lu\n"
 141                    "grows:                   %u\n"
 142                    "grows failure:           %u\n"
 143                    "shrinks:                 %u\n"
 144                    "cache access:            %lu\n"
 145                    "cache missing:           %lu\n"
 146                    "low free mark:           %lu\n"
 147                    "max waitqueue depth:     %u\n"
 148                    "max wait time ms:        %lld\n"
 149                    "out of mem:              %lu\n",
 150                    cfs_totalram_pages(), PAGES_PER_POOL,
 151                    page_pools.epp_max_pages,
 152                    page_pools.epp_max_pools,
 153                    page_pools.epp_total_pages,
 154                    page_pools.epp_free_pages,
 155                    page_pools.epp_idle_idx,
 156                    ktime_get_seconds() - page_pools.epp_last_shrink,
 157                    ktime_get_seconds() - page_pools.epp_last_access,
 158                    page_pools.epp_st_max_pages,
 159                    page_pools.epp_st_grows,
 160                    page_pools.epp_st_grow_fails,
 161                    page_pools.epp_st_shrinks,
 162                    page_pools.epp_st_access,
 163                    page_pools.epp_st_missings,
 164                    page_pools.epp_st_lowfree,
 165                    page_pools.epp_st_max_wqlen,
 166                    ktime_to_ms(page_pools.epp_st_max_wait),
 167                    page_pools.epp_st_outofmem);
 168
 169         spin_unlock(&page_pools.epp_lock);
 170         return 0;
 171 }
 172
 173 static void enc_pools_release_free_pages(long npages)
 174 {
 175         int p_idx, g_idx;
 176         int p_idx_max1, p_idx_max2;
 177
 178         LASSERT(npages > 0);
 179         LASSERT(npages <= page_pools.epp_free_pages);
 180         LASSERT(page_pools.epp_free_pages <= page_pools.epp_total_pages);
 181
 182         /* max pool index before the release */
 183         p_idx_max2 = (page_pools.epp_total_pages - 1) / PAGES_PER_POOL;
 184
 185         page_pools.epp_free_pages -= npages;
 186         page_pools.epp_total_pages -= npages;
 187
 188         /* max pool index after the release */
 189         p_idx_max1 = page_pools.epp_total_pages == 0 ? -1 :
 190                 ((page_pools.epp_total_pages - 1) / PAGES_PER_POOL);
 191
 192         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 193         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 194         LASSERT(page_pools.epp_pools[p_idx]);
 195
 196         while (npages--) {
 197                 LASSERT(page_pools.epp_pools[p_idx]);
 198                 LASSERT(page_pools.epp_pools[p_idx][g_idx] != NULL);
 199
 200                 __free_page(page_pools.epp_pools[p_idx][g_idx]);
 201                 page_pools.epp_pools[p_idx][g_idx] = NULL;
 202
 203                 if (++g_idx == PAGES_PER_POOL) {
 204                         p_idx++;
 205                         g_idx = 0;
 206                 }
 207         }
 208
 209         /* free unused pools */
 210         while (p_idx_max1 < p_idx_max2) {
 211                 LASSERT(page_pools.epp_pools[p_idx_max2]);
 212                 OBD_FREE(page_pools.epp_pools[p_idx_max2], PAGE_SIZE);
 213                 page_pools.epp_pools[p_idx_max2] = NULL;
 214                 p_idx_max2--;
 215         }
 216 }
 217
 218 /*
 219  * we try to keep at least PTLRPC_MAX_BRW_PAGES pages in the pool.
 220  */
 221 static unsigned long enc_pools_shrink_count(struct shrinker *s,
 222                                             struct shrink_control *sc)
 223 {
 224         /*
 225          * if no pool access for a long time, we consider it's fully idle.
 226          * a little race here is fine.
 227          */
 228         if (unlikely(ktime_get_seconds() - page_pools.epp_last_access >
 229                      CACHE_QUIESCENT_PERIOD)) {
 230                 spin_lock(&page_pools.epp_lock);
 231                 page_pools.epp_idle_idx = IDLE_IDX_MAX;
 232                 spin_unlock(&page_pools.epp_lock);
 233         }
 234
 235         LASSERT(page_pools.epp_idle_idx <= IDLE_IDX_MAX);
 236         return (page_pools.epp_free_pages <= PTLRPC_MAX_BRW_PAGES) ? 0 :
 237                 (page_pools.epp_free_pages - PTLRPC_MAX_BRW_PAGES) *
 238                 (IDLE_IDX_MAX - page_pools.epp_idle_idx) / IDLE_IDX_MAX;
 239 }
 240
 241 /*
 242  * we try to keep at least PTLRPC_MAX_BRW_PAGES pages in the pool.
 243  */
 244 static unsigned long enc_pools_shrink_scan(struct shrinker *s,
 245                                            struct shrink_control *sc)
 246 {
 247         spin_lock(&page_pools.epp_lock);
 248         if (page_pools.epp_free_pages <= PTLRPC_MAX_BRW_PAGES)
 249                 sc->nr_to_scan = 0;
 250         else
 251                 sc->nr_to_scan = min_t(unsigned long, sc->nr_to_scan,
 252                               page_pools.epp_free_pages - PTLRPC_MAX_BRW_PAGES);
 253         if (sc->nr_to_scan > 0) {
 254                 enc_pools_release_free_pages(sc->nr_to_scan);
 255                 CDEBUG(D_SEC, "released %ld pages, %ld left\n",
 256                        (long)sc->nr_to_scan, page_pools.epp_free_pages);
 257
 258                 page_pools.epp_st_shrinks++;
 259                 page_pools.epp_last_shrink = ktime_get_seconds();
 260         }
 261         spin_unlock(&page_pools.epp_lock);
 262
 263         /*
 264          * if no pool access for a long time, we consider it's fully idle.
 265          * a little race here is fine.
 266          */
 267         if (unlikely(ktime_get_seconds() - page_pools.epp_last_access >
 268                      CACHE_QUIESCENT_PERIOD)) {
 269                 spin_lock(&page_pools.epp_lock);
 270                 page_pools.epp_idle_idx = IDLE_IDX_MAX;
 271                 spin_unlock(&page_pools.epp_lock);
 272         }
 273
 274         LASSERT(page_pools.epp_idle_idx <= IDLE_IDX_MAX);
 275         return sc->nr_to_scan;
 276 }
 277
 278 #ifndef HAVE_SHRINKER_COUNT
 279 /*
 280  * could be called frequently for query (@nr_to_scan == 0).
 281  * we try to keep at least PTLRPC_MAX_BRW_PAGES pages in the pool.
 282  */
 283 static int enc_pools_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
 284 {
 285         struct shrink_control scv = {
 286                 .nr_to_scan = shrink_param(sc, nr_to_scan),
 287                 .gfp_mask   = shrink_param(sc, gfp_mask)
 288         };
 289 #if !defined(HAVE_SHRINKER_WANT_SHRINK_PTR) && !defined(HAVE_SHRINK_CONTROL)
 290         struct shrinker *shrinker = NULL;
 291 #endif
 292
 293         enc_pools_shrink_scan(shrinker, &scv);
 294
 295         return enc_pools_shrink_count(shrinker, &scv);
 296 }
 297
 298 #endif /* HAVE_SHRINKER_COUNT */
 299
 300 static inline
 301 int npages_to_npools(unsigned long npages)
 302 {
 303         return (int) ((npages + PAGES_PER_POOL - 1) / PAGES_PER_POOL);
 304 }
 305
 306 /*
 307  * return how many pages cleaned up.
 308  */
 309 static unsigned long enc_pools_cleanup(struct page ***pools, int npools)
 310 {
 311         unsigned long cleaned = 0;
 312         int i, j;
 313
 314         for (i = 0; i < npools; i++) {
 315                 if (pools[i]) {
 316                         for (j = 0; j < PAGES_PER_POOL; j++) {
 317                                 if (pools[i][j]) {
 318                                         __free_page(pools[i][j]);
 319                                         cleaned++;
 320                                 }
 321                         }
 322                         OBD_FREE(pools[i], PAGE_SIZE);
 323                         pools[i] = NULL;
 324                 }
 325         }
 326
 327         return cleaned;
 328 }
 329
 330 /*
 331  * merge @npools pointed by @pools which contains @npages new pages
 332  * into current pools.
 333  *
 334  * we have options to avoid most memory copy with some tricks. but we choose
 335  * the simplest way to avoid complexity. It's not frequently called.
 336  */
 337 static void enc_pools_insert(struct page ***pools, int npools, int npages)
 338 {
 339         int freeslot;
 340         int op_idx, np_idx, og_idx, ng_idx;
 341         int cur_npools, end_npools;
 342
 343         LASSERT(npages > 0);
 344         LASSERT(page_pools.epp_total_pages+npages <= page_pools.epp_max_pages);
 345         LASSERT(npages_to_npools(npages) == npools);
 346         LASSERT(page_pools.epp_growing);
 347
 348         spin_lock(&page_pools.epp_lock);
 349
 350         /*
 351          * (1) fill all the free slots of current pools.
 352          */
 353         /*
 354          * free slots are those left by rent pages, and the extra ones with
 355          * index >= total_pages, locate at the tail of last pool.
 356          */
 357         freeslot = page_pools.epp_total_pages % PAGES_PER_POOL;
 358         if (freeslot != 0)
 359                 freeslot = PAGES_PER_POOL - freeslot;
 360         freeslot += page_pools.epp_total_pages - page_pools.epp_free_pages;
 361
 362         op_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 363         og_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 364         np_idx = npools - 1;
 365         ng_idx = (npages - 1) % PAGES_PER_POOL;
 366
 367         while (freeslot) {
 368                 LASSERT(page_pools.epp_pools[op_idx][og_idx] == NULL);
 369                 LASSERT(pools[np_idx][ng_idx] != NULL);
 370
 371                 page_pools.epp_pools[op_idx][og_idx] = pools[np_idx][ng_idx];
 372                 pools[np_idx][ng_idx] = NULL;
 373
 374                 freeslot--;
 375
 376                 if (++og_idx == PAGES_PER_POOL) {
 377                         op_idx++;
 378                         og_idx = 0;
 379                 }
 380                 if (--ng_idx < 0) {
 381                         if (np_idx == 0)
 382                                 break;
 383                         np_idx--;
 384                         ng_idx = PAGES_PER_POOL - 1;
 385                 }
 386         }
 387
 388         /*
 389          * (2) add pools if needed.
 390          */
 391         cur_npools = (page_pools.epp_total_pages + PAGES_PER_POOL - 1) /
 392                       PAGES_PER_POOL;
 393         end_npools = (page_pools.epp_total_pages + npages +
 394                       PAGES_PER_POOL - 1) / PAGES_PER_POOL;
 395         LASSERT(end_npools <= page_pools.epp_max_pools);
 396
 397         np_idx = 0;
 398         while (cur_npools < end_npools) {
 399                 LASSERT(page_pools.epp_pools[cur_npools] == NULL);
 400                 LASSERT(np_idx < npools);
 401                 LASSERT(pools[np_idx] != NULL);
 402
 403                 page_pools.epp_pools[cur_npools++] = pools[np_idx];
 404                 pools[np_idx++] = NULL;
 405         }
 406
 407         page_pools.epp_total_pages += npages;
 408         page_pools.epp_free_pages += npages;
 409         page_pools.epp_st_lowfree = page_pools.epp_free_pages;
 410
 411         if (page_pools.epp_total_pages > page_pools.epp_st_max_pages)
 412                 page_pools.epp_st_max_pages = page_pools.epp_total_pages;
 413
 414         CDEBUG(D_SEC, "add %d pages to total %lu\n", npages,
 415                page_pools.epp_total_pages);
 416
 417         spin_unlock(&page_pools.epp_lock);
 418 }
 419
 420 static int enc_pools_add_pages(int npages)
 421 {
 422         static DEFINE_MUTEX(add_pages_mutex);
 423         struct page ***pools;
 424         int npools, alloced = 0;
 425         int i, j, rc = -ENOMEM;
 426
 427         if (npages < PTLRPC_MAX_BRW_PAGES)
 428                 npages = PTLRPC_MAX_BRW_PAGES;
 429
 430         mutex_lock(&add_pages_mutex);
 431
 432         if (npages + page_pools.epp_total_pages > page_pools.epp_max_pages)
 433                 npages = page_pools.epp_max_pages - page_pools.epp_total_pages;
 434         LASSERT(npages > 0);
 435
 436         page_pools.epp_st_grows++;
 437
 438         npools = npages_to_npools(npages);
 439         OBD_ALLOC(pools, npools * sizeof(*pools));
 440         if (pools == NULL)
 441                 goto out;
 442
 443         for (i = 0; i < npools; i++) {
 444                 OBD_ALLOC(pools[i], PAGE_SIZE);
 445                 if (pools[i] == NULL)
 446                         goto out_pools;
 447
 448                 for (j = 0; j < PAGES_PER_POOL && alloced < npages; j++) {
 449                         pools[i][j] = alloc_page(GFP_NOFS |
 450                                                  __GFP_HIGHMEM);
 451                         if (pools[i][j] == NULL)
 452                                 goto out_pools;
 453
 454                         alloced++;
 455                 }
 456         }
 457         LASSERT(alloced == npages);
 458
 459         enc_pools_insert(pools, npools, npages);
 460         CDEBUG(D_SEC, "added %d pages into pools\n", npages);
 461         rc = 0;
 462
 463 out_pools:
 464         enc_pools_cleanup(pools, npools);
 465         OBD_FREE(pools, npools * sizeof(*pools));
 466 out:
 467         if (rc) {
 468                 page_pools.epp_st_grow_fails++;
 469                 CERROR("Failed to allocate %d enc pages\n", npages);
 470         }
 471
 472         mutex_unlock(&add_pages_mutex);
 473         return rc;
 474 }
 475
 476 static inline void enc_pools_wakeup(void)
 477 {
 478         assert_spin_locked(&page_pools.epp_lock);
 479
 480         if (unlikely(page_pools.epp_waitqlen)) {
 481                 LASSERT(waitqueue_active(&page_pools.epp_waitq));
 482                 wake_up_all(&page_pools.epp_waitq);
 483         }
 484 }
 485
 486 static int enc_pools_should_grow(int page_needed, time64_t now)
 487 {
 488         /*
 489          * don't grow if someone else is growing the pools right now,
 490          * or the pools has reached its full capacity
 491          */
 492         if (page_pools.epp_growing ||
 493             page_pools.epp_total_pages == page_pools.epp_max_pages)
 494                 return 0;
 495
 496         /* if total pages is not enough, we need to grow */
 497         if (page_pools.epp_total_pages < page_needed)
 498                 return 1;
 499
 500         /*
 501          * we wanted to return 0 here if there was a shrink just
 502          * happened a moment ago, but this may cause deadlock if both
 503          * client and ost live on single node.
 504          */
 505
 506         /*
 507          * here we perhaps need consider other factors like wait queue
 508          * length, idle index, etc. ?
 509          */
 510
 511         /* grow the pools in any other cases */
 512         return 1;
 513 }
 514
 515 /*
 516  * Export the number of free pages in the pool
 517  */
 518 int get_free_pages_in_pool(void)
 519 {
 520         return page_pools.epp_free_pages;
 521 }
 522 EXPORT_SYMBOL(get_free_pages_in_pool);
 523
 524 /*
 525  * Let outside world know if enc_pool full capacity is reached
 526  */
 527 int pool_is_at_full_capacity(void)
 528 {
 529         return (page_pools.epp_total_pages == page_pools.epp_max_pages);
 530 }
 531 EXPORT_SYMBOL(pool_is_at_full_capacity);
 532
 533 /*
 534  * we allocate the requested pages atomically.
 535  */
 536 int sptlrpc_enc_pool_get_pages(struct ptlrpc_bulk_desc *desc)
 537 {
 538         wait_queue_entry_t waitlink;
 539         unsigned long this_idle = -1;
 540         u64 tick_ns = 0;
 541         time64_t now;
 542         int p_idx, g_idx;
 543         int i;
 544
 545         LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
 546         LASSERT(desc->bd_iov_count > 0);
 547         LASSERT(desc->bd_iov_count <= page_pools.epp_max_pages);
 548
 549         /* resent bulk, enc iov might have been allocated previously */
 550         if (GET_ENC_KIOV(desc) != NULL)
 551                 return 0;
 552
 553         OBD_ALLOC_LARGE(GET_ENC_KIOV(desc),
 554                   desc->bd_iov_count * sizeof(*GET_ENC_KIOV(desc)));
 555         if (GET_ENC_KIOV(desc) == NULL)
 556                 return -ENOMEM;
 557
 558         spin_lock(&page_pools.epp_lock);
 559
 560         page_pools.epp_st_access++;
 561 again:
 562         if (unlikely(page_pools.epp_free_pages < desc->bd_iov_count)) {
 563                 if (tick_ns == 0)
 564                         tick_ns = ktime_get_ns();
 565
 566                 now = ktime_get_real_seconds();
 567
 568                 page_pools.epp_st_missings++;
 569                 page_pools.epp_pages_short += desc->bd_iov_count;
 570
 571                 if (enc_pools_should_grow(desc->bd_iov_count, now)) {
 572                         page_pools.epp_growing = 1;
 573
 574                         spin_unlock(&page_pools.epp_lock);
 575                         enc_pools_add_pages(page_pools.epp_pages_short / 2);
 576                         spin_lock(&page_pools.epp_lock);
 577
 578                         page_pools.epp_growing = 0;
 579
 580                         enc_pools_wakeup();
 581                 } else {
 582                         if (page_pools.epp_growing) {
 583                                 if (++page_pools.epp_waitqlen >
 584                                     page_pools.epp_st_max_wqlen)
 585                                         page_pools.epp_st_max_wqlen =
 586                                                         page_pools.epp_waitqlen;
 587
 588                                 set_current_state(TASK_UNINTERRUPTIBLE);
 589                                 init_waitqueue_entry(&waitlink, current);
 590                                 add_wait_queue(&page_pools.epp_waitq,
 591                                                &waitlink);
 592
 593                                 spin_unlock(&page_pools.epp_lock);
 594                                 schedule();
 595                                 remove_wait_queue(&page_pools.epp_waitq,
 596                                                   &waitlink);
 597                                 LASSERT(page_pools.epp_waitqlen > 0);
 598                                 spin_lock(&page_pools.epp_lock);
 599                                 page_pools.epp_waitqlen--;
 600                         } else {
 601                                 /*
 602                                  * ptlrpcd thread should not sleep in that case,
 603                                  * or deadlock may occur!
 604                                  * Instead, return -ENOMEM so that upper layers
 605                                  * will put request back in queue.
 606                                  */
 607                                 page_pools.epp_st_outofmem++;
 608                                 spin_unlock(&page_pools.epp_lock);
 609                                 OBD_FREE_LARGE(GET_ENC_KIOV(desc),
 610                                                desc->bd_iov_count *
 611                                                 sizeof(*GET_ENC_KIOV(desc)));
 612                                 GET_ENC_KIOV(desc) = NULL;
 613                                 return -ENOMEM;
 614                         }
 615                 }
 616
 617                 LASSERT(page_pools.epp_pages_short >= desc->bd_iov_count);
 618                 page_pools.epp_pages_short -= desc->bd_iov_count;
 619
 620                 this_idle = 0;
 621                 goto again;
 622         }
 623
 624         /* record max wait time */
 625         if (unlikely(tick_ns)) {
 626                 ktime_t tick = ktime_sub_ns(ktime_get(), tick_ns);
 627
 628                 if (ktime_after(tick, page_pools.epp_st_max_wait))
 629                         page_pools.epp_st_max_wait = tick;
 630         }
 631
 632         /* proceed with rest of allocation */
 633         page_pools.epp_free_pages -= desc->bd_iov_count;
 634
 635         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 636         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 637
 638         for (i = 0; i < desc->bd_iov_count; i++) {
 639                 LASSERT(page_pools.epp_pools[p_idx][g_idx] != NULL);
 640                 BD_GET_ENC_KIOV(desc, i).kiov_page =
 641                        page_pools.epp_pools[p_idx][g_idx];
 642                 page_pools.epp_pools[p_idx][g_idx] = NULL;
 643
 644                 if (++g_idx == PAGES_PER_POOL) {
 645                         p_idx++;
 646                         g_idx = 0;
 647                 }
 648         }
 649
 650         if (page_pools.epp_free_pages < page_pools.epp_st_lowfree)
 651                 page_pools.epp_st_lowfree = page_pools.epp_free_pages;
 652
 653         /*
 654          * new idle index = (old * weight + new) / (weight + 1)
 655          */
 656         if (this_idle == -1) {
 657                 this_idle = page_pools.epp_free_pages * IDLE_IDX_MAX /
 658                         page_pools.epp_total_pages;
 659         }
 660         page_pools.epp_idle_idx = (page_pools.epp_idle_idx * IDLE_IDX_WEIGHT +
 661                                    this_idle) /
 662                                    (IDLE_IDX_WEIGHT + 1);
 663
 664         page_pools.epp_last_access = ktime_get_seconds();
 665
 666         spin_unlock(&page_pools.epp_lock);
 667         return 0;
 668 }
 669 EXPORT_SYMBOL(sptlrpc_enc_pool_get_pages);
 670
 671 void sptlrpc_enc_pool_put_pages(struct ptlrpc_bulk_desc *desc)
 672 {
 673         int p_idx, g_idx;
 674         int i;
 675
 676         LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
 677
 678         if (GET_ENC_KIOV(desc) == NULL)
 679                 return;
 680
 681         LASSERT(desc->bd_iov_count > 0);
 682
 683         spin_lock(&page_pools.epp_lock);
 684
 685         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 686         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 687
 688         LASSERT(page_pools.epp_free_pages + desc->bd_iov_count <=
 689                 page_pools.epp_total_pages);
 690         LASSERT(page_pools.epp_pools[p_idx]);
 691
 692         for (i = 0; i < desc->bd_iov_count; i++) {
 693                 LASSERT(BD_GET_ENC_KIOV(desc, i).kiov_page != NULL);
 694                 LASSERT(g_idx != 0 || page_pools.epp_pools[p_idx]);
 695                 LASSERT(page_pools.epp_pools[p_idx][g_idx] == NULL);
 696
 697                 page_pools.epp_pools[p_idx][g_idx] =
 698                         BD_GET_ENC_KIOV(desc, i).kiov_page;
 699
 700                 if (++g_idx == PAGES_PER_POOL) {
 701                         p_idx++;
 702                         g_idx = 0;
 703                 }
 704         }
 705
 706         page_pools.epp_free_pages += desc->bd_iov_count;
 707
 708         enc_pools_wakeup();
 709
 710         spin_unlock(&page_pools.epp_lock);
 711
 712         OBD_FREE_LARGE(GET_ENC_KIOV(desc),
 713                  desc->bd_iov_count * sizeof(*GET_ENC_KIOV(desc)));
 714         GET_ENC_KIOV(desc) = NULL;
 715 }
 716
 717 /*
 718  * we don't do much stuff for add_user/del_user anymore, except adding some
 719  * initial pages in add_user() if current pools are empty, rest would be
 720  * handled by the pools's self-adaption.
 721  */
 722 int sptlrpc_enc_pool_add_user(void)
 723 {
 724         int need_grow = 0;
 725
 726         spin_lock(&page_pools.epp_lock);
 727         if (page_pools.epp_growing == 0 && page_pools.epp_total_pages == 0) {
 728                 page_pools.epp_growing = 1;
 729                 need_grow = 1;
 730         }
 731         spin_unlock(&page_pools.epp_lock);
 732
 733         if (need_grow) {
 734                 enc_pools_add_pages(PTLRPC_MAX_BRW_PAGES +
 735                                     PTLRPC_MAX_BRW_PAGES);
 736
 737                 spin_lock(&page_pools.epp_lock);
 738                 page_pools.epp_growing = 0;
 739                 enc_pools_wakeup();
 740                 spin_unlock(&page_pools.epp_lock);
 741         }
 742         return 0;
 743 }
 744 EXPORT_SYMBOL(sptlrpc_enc_pool_add_user);
 745
 746 int sptlrpc_enc_pool_del_user(void)
 747 {
 748         return 0;
 749 }
 750 EXPORT_SYMBOL(sptlrpc_enc_pool_del_user);
 751
 752 static inline void enc_pools_alloc(void)
 753 {
 754         LASSERT(page_pools.epp_max_pools);
 755         OBD_ALLOC_LARGE(page_pools.epp_pools,
 756                         page_pools.epp_max_pools *
 757                         sizeof(*page_pools.epp_pools));
 758 }
 759
 760 static inline void enc_pools_free(void)
 761 {
 762         LASSERT(page_pools.epp_max_pools);
 763         LASSERT(page_pools.epp_pools);
 764
 765         OBD_FREE_LARGE(page_pools.epp_pools,
 766                        page_pools.epp_max_pools *
 767                        sizeof(*page_pools.epp_pools));
 768 }
 769
 770 int sptlrpc_enc_pool_init(void)
 771 {
 772         DEF_SHRINKER_VAR(shvar, enc_pools_shrink,
 773                          enc_pools_shrink_count, enc_pools_shrink_scan);
 774
 775         page_pools.epp_max_pages = cfs_totalram_pages() / 8;
 776         if (enc_pool_max_memory_mb > 0 &&
 777             enc_pool_max_memory_mb <= (cfs_totalram_pages() >> mult))
 778                 page_pools.epp_max_pages = enc_pool_max_memory_mb << mult;
 779
 780         page_pools.epp_max_pools = npages_to_npools(page_pools.epp_max_pages);
 781
 782         init_waitqueue_head(&page_pools.epp_waitq);
 783         page_pools.epp_waitqlen = 0;
 784         page_pools.epp_pages_short = 0;
 785
 786         page_pools.epp_growing = 0;
 787
 788         page_pools.epp_idle_idx = 0;
 789         page_pools.epp_last_shrink = ktime_get_seconds();
 790         page_pools.epp_last_access = ktime_get_seconds();
 791
 792         spin_lock_init(&page_pools.epp_lock);
 793         page_pools.epp_total_pages = 0;
 794         page_pools.epp_free_pages = 0;
 795
 796         page_pools.epp_st_max_pages = 0;
 797         page_pools.epp_st_grows = 0;
 798         page_pools.epp_st_grow_fails = 0;
 799         page_pools.epp_st_shrinks = 0;
 800         page_pools.epp_st_access = 0;
 801         page_pools.epp_st_missings = 0;
 802         page_pools.epp_st_lowfree = 0;
 803         page_pools.epp_st_max_wqlen = 0;
 804         page_pools.epp_st_max_wait = ktime_set(0, 0);
 805         page_pools.epp_st_outofmem = 0;
 806
 807         enc_pools_alloc();
 808         if (page_pools.epp_pools == NULL)
 809                 return -ENOMEM;
 810
 811         pools_shrinker = set_shrinker(pools_shrinker_seeks, &shvar);
 812         if (pools_shrinker == NULL) {
 813                 enc_pools_free();
 814                 return -ENOMEM;
 815         }
 816
 817         return 0;
 818 }
 819
 820 void sptlrpc_enc_pool_fini(void)
 821 {
 822         unsigned long cleaned, npools;
 823
 824         LASSERT(pools_shrinker);
 825         LASSERT(page_pools.epp_pools);
 826         LASSERT(page_pools.epp_total_pages == page_pools.epp_free_pages);
 827
 828         remove_shrinker(pools_shrinker);
 829
 830         npools = npages_to_npools(page_pools.epp_total_pages);
 831         cleaned = enc_pools_cleanup(page_pools.epp_pools, npools);
 832         LASSERT(cleaned == page_pools.epp_total_pages);
 833
 834         enc_pools_free();
 835
 836         if (page_pools.epp_st_access > 0) {
 837                 CDEBUG(D_SEC,
 838                        "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",
 839                        page_pools.epp_st_max_pages, page_pools.epp_st_grows,
 840                        page_pools.epp_st_grow_fails,
 841                        page_pools.epp_st_shrinks, page_pools.epp_st_access,
 842                        page_pools.epp_st_missings, page_pools.epp_st_max_wqlen,
 843                        ktime_to_ms(page_pools.epp_st_max_wait),
 844                        page_pools.epp_st_outofmem);
 845         }
 846 }
 847
 848
 849 static int cfs_hash_alg_id[] = {
 850         [BULK_HASH_ALG_NULL]    = CFS_HASH_ALG_NULL,
 851         [BULK_HASH_ALG_ADLER32] = CFS_HASH_ALG_ADLER32,
 852         [BULK_HASH_ALG_CRC32]   = CFS_HASH_ALG_CRC32,
 853         [BULK_HASH_ALG_MD5]     = CFS_HASH_ALG_MD5,
 854         [BULK_HASH_ALG_SHA1]    = CFS_HASH_ALG_SHA1,
 855         [BULK_HASH_ALG_SHA256]  = CFS_HASH_ALG_SHA256,
 856         [BULK_HASH_ALG_SHA384]  = CFS_HASH_ALG_SHA384,
 857         [BULK_HASH_ALG_SHA512]  = CFS_HASH_ALG_SHA512,
 858 };
 859 const char *sptlrpc_get_hash_name(__u8 hash_alg)
 860 {
 861         return cfs_crypto_hash_name(cfs_hash_alg_id[hash_alg]);
 862 }
 863
 864 __u8 sptlrpc_get_hash_alg(const char *algname)
 865 {
 866         return cfs_crypto_hash_alg(algname);
 867 }
 868
 869 int bulk_sec_desc_unpack(struct lustre_msg *msg, int offset, int swabbed)
 870 {
 871         struct ptlrpc_bulk_sec_desc *bsd;
 872         int size = msg->lm_buflens[offset];
 873
 874         bsd = lustre_msg_buf(msg, offset, sizeof(*bsd));
 875         if (bsd == NULL) {
 876                 CERROR("Invalid bulk sec desc: size %d\n", size);
 877                 return -EINVAL;
 878         }
 879
 880         if (swabbed)
 881                 __swab32s(&bsd->bsd_nob);
 882
 883         if (unlikely(bsd->bsd_version != 0)) {
 884                 CERROR("Unexpected version %u\n", bsd->bsd_version);
 885                 return -EPROTO;
 886         }
 887
 888         if (unlikely(bsd->bsd_type >= SPTLRPC_BULK_MAX)) {
 889                 CERROR("Invalid type %u\n", bsd->bsd_type);
 890                 return -EPROTO;
 891         }
 892
 893         /* FIXME more sanity check here */
 894
 895         if (unlikely(bsd->bsd_svc != SPTLRPC_BULK_SVC_NULL &&
 896                      bsd->bsd_svc != SPTLRPC_BULK_SVC_INTG &&
 897                      bsd->bsd_svc != SPTLRPC_BULK_SVC_PRIV)) {
 898                 CERROR("Invalid svc %u\n", bsd->bsd_svc);
 899                 return -EPROTO;
 900         }
 901
 902         return 0;
 903 }
 904 EXPORT_SYMBOL(bulk_sec_desc_unpack);
 905
 906 /*
 907  * Compute the checksum of an RPC buffer payload.  If the return \a buflen
 908  * is not large enough, truncate the result to fit so that it is possible
 909  * to use a hash function with a large hash space, but only use a part of
 910  * the resulting hash.
 911  */
 912 int sptlrpc_get_bulk_checksum(struct ptlrpc_bulk_desc *desc, __u8 alg,
 913                               void *buf, int buflen)
 914 {
 915         struct ahash_request *req;
 916         int hashsize;
 917         unsigned int bufsize;
 918         int i, err;
 919
 920         LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
 921         LASSERT(alg > BULK_HASH_ALG_NULL && alg < BULK_HASH_ALG_MAX);
 922         LASSERT(buflen >= 4);
 923
 924         req = cfs_crypto_hash_init(cfs_hash_alg_id[alg], NULL, 0);
 925         if (IS_ERR(req)) {
 926                 CERROR("Unable to initialize checksum hash %s\n",
 927                        cfs_crypto_hash_name(cfs_hash_alg_id[alg]));
 928                 return PTR_ERR(req);
 929         }
 930
 931         hashsize = cfs_crypto_hash_digestsize(cfs_hash_alg_id[alg]);
 932
 933         for (i = 0; i < desc->bd_iov_count; i++) {
 934                 cfs_crypto_hash_update_page(req,
 935                                   BD_GET_KIOV(desc, i).kiov_page,
 936                                   BD_GET_KIOV(desc, i).kiov_offset &
 937                                               ~PAGE_MASK,
 938                                   BD_GET_KIOV(desc, i).kiov_len);
 939         }
 940
 941         if (hashsize > buflen) {
 942                 unsigned char hashbuf[CFS_CRYPTO_HASH_DIGESTSIZE_MAX];
 943
 944                 bufsize = sizeof(hashbuf);
 945                 LASSERTF(bufsize >= hashsize, "bufsize = %u < hashsize %u\n",
 946                          bufsize, hashsize);
 947                 err = cfs_crypto_hash_final(req, hashbuf, &bufsize);
 948                 memcpy(buf, hashbuf, buflen);
 949         } else {
 950                 bufsize = buflen;
 951                 err = cfs_crypto_hash_final(req, buf, &bufsize);
 952         }
 953
 954         return err;
 955 }