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