kernel/drivers/staging/lustre/lustre/ptlrpc/sec_bulk.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/ptlrpc/sec_bulk.c
  37  *
  38  * Author: Eric Mei <ericm@clusterfs.com>
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_SEC
  42
  43 #include "../../include/linux/libcfs/libcfs.h"
  44 #include <linux/crypto.h>
  45
  46 #include "../include/obd.h"
  47 #include "../include/obd_cksum.h"
  48 #include "../include/obd_class.h"
  49 #include "../include/obd_support.h"
  50 #include "../include/lustre_net.h"
  51 #include "../include/lustre_import.h"
  52 #include "../include/lustre_dlm.h"
  53 #include "../include/lustre_sec.h"
  54
  55 #include "ptlrpc_internal.h"
  56
  57 /****************************************
  58  * bulk encryption page pools      *
  59  ****************************************/
  60
  61
  62 #define POINTERS_PER_PAGE       (PAGE_CACHE_SIZE / sizeof(void *))
  63 #define PAGES_PER_POOL          (POINTERS_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         /*
  72          * constants
  73          */
  74         unsigned long    epp_max_pages;   /* maximum pages can hold, const */
  75         unsigned int     epp_max_pools;   /* number of pools, const */
  76
  77         /*
  78          * wait queue in case of not enough free pages.
  79          */
  80         wait_queue_head_t      epp_waitq;       /* waiting threads */
  81         unsigned int     epp_waitqlen;    /* wait queue length */
  82         unsigned long    epp_pages_short; /* # of pages wanted of in-q users */
  83         unsigned int     epp_growing:1;   /* during adding pages */
  84
  85         /*
  86          * indicating how idle the pools are, from 0 to MAX_IDLE_IDX
  87          * this is counted based on each time when getting pages from
  88          * the pools, not based on time. which means in case that system
  89          * is idled for a while but the idle_idx might still be low if no
  90          * activities happened in the pools.
  91          */
  92         unsigned long    epp_idle_idx;
  93
  94         /* last shrink time due to mem tight */
  95         long         epp_last_shrink;
  96         long         epp_last_access;
  97
  98         /*
  99          * in-pool pages bookkeeping
 100          */
 101         spinlock_t       epp_lock;         /* protect following fields */
 102         unsigned long    epp_total_pages; /* total pages in pools */
 103         unsigned long    epp_free_pages;  /* current pages available */
 104
 105         /*
 106          * statistics
 107          */
 108         unsigned long    epp_st_max_pages;      /* # of pages ever reached */
 109         unsigned int     epp_st_grows;    /* # of grows */
 110         unsigned int     epp_st_grow_fails;     /* # of add pages failures */
 111         unsigned int     epp_st_shrinks;        /* # of shrinks */
 112         unsigned long    epp_st_access;  /* # of access */
 113         unsigned long    epp_st_missings;       /* # of cache missing */
 114         unsigned long    epp_st_lowfree;        /* lowest free pages reached */
 115         unsigned int     epp_st_max_wqlen;      /* highest waitqueue length */
 116         unsigned long       epp_st_max_wait;       /* in jiffies */
 117         /*
 118          * pointers to pools
 119          */
 120         struct page    ***epp_pools;
 121 } page_pools;
 122
 123 /*
 124  * /proc/fs/lustre/sptlrpc/encrypt_page_pools
 125  */
 126 int sptlrpc_proc_enc_pool_seq_show(struct seq_file *m, void *v)
 127 {
 128         spin_lock(&page_pools.epp_lock);
 129
 130         seq_printf(m,
 131                    "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:             %lds\n"
 139                    "last access:             %lds\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:         " CFS_TIME_T "/%u\n",
 149                    totalram_pages,
 150                    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                    get_seconds() - page_pools.epp_last_shrink,
 157                    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                    page_pools.epp_st_max_wait,
 167                    HZ);
 168
 169         spin_unlock(&page_pools.epp_lock);
 170
 171         return 0;
 172 }
 173
 174 static void enc_pools_release_free_pages(long npages)
 175 {
 176         int     p_idx, g_idx;
 177         int     p_idx_max1, p_idx_max2;
 178
 179         LASSERT(npages > 0);
 180         LASSERT(npages <= page_pools.epp_free_pages);
 181         LASSERT(page_pools.epp_free_pages <= page_pools.epp_total_pages);
 182
 183         /* max pool index before the release */
 184         p_idx_max2 = (page_pools.epp_total_pages - 1) / PAGES_PER_POOL;
 185
 186         page_pools.epp_free_pages -= npages;
 187         page_pools.epp_total_pages -= npages;
 188
 189         /* max pool index after the release */
 190         p_idx_max1 = page_pools.epp_total_pages == 0 ? -1 :
 191                      ((page_pools.epp_total_pages - 1) / PAGES_PER_POOL);
 192
 193         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 194         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 195         LASSERT(page_pools.epp_pools[p_idx]);
 196
 197         while (npages--) {
 198                 LASSERT(page_pools.epp_pools[p_idx]);
 199                 LASSERT(page_pools.epp_pools[p_idx][g_idx] != NULL);
 200
 201                 __free_page(page_pools.epp_pools[p_idx][g_idx]);
 202                 page_pools.epp_pools[p_idx][g_idx] = NULL;
 203
 204                 if (++g_idx == PAGES_PER_POOL) {
 205                         p_idx++;
 206                         g_idx = 0;
 207                 }
 208         }
 209
 210         /* free unused pools */
 211         while (p_idx_max1 < p_idx_max2) {
 212                 LASSERT(page_pools.epp_pools[p_idx_max2]);
 213                 OBD_FREE(page_pools.epp_pools[p_idx_max2], PAGE_CACHE_SIZE);
 214                 page_pools.epp_pools[p_idx_max2] = NULL;
 215                 p_idx_max2--;
 216         }
 217 }
 218
 219 /*
 220  * we try to keep at least PTLRPC_MAX_BRW_PAGES pages in the pool.
 221  */
 222 static unsigned long enc_pools_shrink_count(struct shrinker *s,
 223                                             struct shrink_control *sc)
 224 {
 225         /*
 226          * if no pool access for a long time, we consider it's fully idle.
 227          * a little race here is fine.
 228          */
 229         if (unlikely(get_seconds() - page_pools.epp_last_access >
 230                      CACHE_QUIESCENT_PERIOD)) {
 231                 spin_lock(&page_pools.epp_lock);
 232                 page_pools.epp_idle_idx = IDLE_IDX_MAX;
 233                 spin_unlock(&page_pools.epp_lock);
 234         }
 235
 236         LASSERT(page_pools.epp_idle_idx <= IDLE_IDX_MAX);
 237         return max((int)page_pools.epp_free_pages - PTLRPC_MAX_BRW_PAGES, 0) *
 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         sc->nr_to_scan = min_t(unsigned long, sc->nr_to_scan,
 249                               page_pools.epp_free_pages - PTLRPC_MAX_BRW_PAGES);
 250         if (sc->nr_to_scan > 0) {
 251                 enc_pools_release_free_pages(sc->nr_to_scan);
 252                 CDEBUG(D_SEC, "released %ld pages, %ld left\n",
 253                        (long)sc->nr_to_scan, page_pools.epp_free_pages);
 254
 255                 page_pools.epp_st_shrinks++;
 256                 page_pools.epp_last_shrink = get_seconds();
 257         }
 258         spin_unlock(&page_pools.epp_lock);
 259
 260         /*
 261          * if no pool access for a long time, we consider it's fully idle.
 262          * a little race here is fine.
 263          */
 264         if (unlikely(get_seconds() - page_pools.epp_last_access >
 265                      CACHE_QUIESCENT_PERIOD)) {
 266                 spin_lock(&page_pools.epp_lock);
 267                 page_pools.epp_idle_idx = IDLE_IDX_MAX;
 268                 spin_unlock(&page_pools.epp_lock);
 269         }
 270
 271         LASSERT(page_pools.epp_idle_idx <= IDLE_IDX_MAX);
 272         return sc->nr_to_scan;
 273 }
 274
 275 static inline
 276 int npages_to_npools(unsigned long npages)
 277 {
 278         return (int) ((npages + PAGES_PER_POOL - 1) / PAGES_PER_POOL);
 279 }
 280
 281 /*
 282  * return how many pages cleaned up.
 283  */
 284 static unsigned long enc_pools_cleanup(struct page ***pools, int npools)
 285 {
 286         unsigned long cleaned = 0;
 287         int        i, j;
 288
 289         for (i = 0; i < npools; i++) {
 290                 if (pools[i]) {
 291                         for (j = 0; j < PAGES_PER_POOL; j++) {
 292                                 if (pools[i][j]) {
 293                                         __free_page(pools[i][j]);
 294                                         cleaned++;
 295                                 }
 296                         }
 297                         OBD_FREE(pools[i], PAGE_CACHE_SIZE);
 298                         pools[i] = NULL;
 299                 }
 300         }
 301
 302         return cleaned;
 303 }
 304
 305 /*
 306  * merge @npools pointed by @pools which contains @npages new pages
 307  * into current pools.
 308  *
 309  * we have options to avoid most memory copy with some tricks. but we choose
 310  * the simplest way to avoid complexity. It's not frequently called.
 311  */
 312 static void enc_pools_insert(struct page ***pools, int npools, int npages)
 313 {
 314         int     freeslot;
 315         int     op_idx, np_idx, og_idx, ng_idx;
 316         int     cur_npools, end_npools;
 317
 318         LASSERT(npages > 0);
 319         LASSERT(page_pools.epp_total_pages+npages <= page_pools.epp_max_pages);
 320         LASSERT(npages_to_npools(npages) == npools);
 321         LASSERT(page_pools.epp_growing);
 322
 323         spin_lock(&page_pools.epp_lock);
 324
 325         /*
 326          * (1) fill all the free slots of current pools.
 327          */
 328         /* free slots are those left by rent pages, and the extra ones with
 329          * index >= total_pages, locate at the tail of last pool. */
 330         freeslot = page_pools.epp_total_pages % PAGES_PER_POOL;
 331         if (freeslot != 0)
 332                 freeslot = PAGES_PER_POOL - freeslot;
 333         freeslot += page_pools.epp_total_pages - page_pools.epp_free_pages;
 334
 335         op_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 336         og_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 337         np_idx = npools - 1;
 338         ng_idx = (npages - 1) % PAGES_PER_POOL;
 339
 340         while (freeslot) {
 341                 LASSERT(page_pools.epp_pools[op_idx][og_idx] == NULL);
 342                 LASSERT(pools[np_idx][ng_idx] != NULL);
 343
 344                 page_pools.epp_pools[op_idx][og_idx] = pools[np_idx][ng_idx];
 345                 pools[np_idx][ng_idx] = NULL;
 346
 347                 freeslot--;
 348
 349                 if (++og_idx == PAGES_PER_POOL) {
 350                         op_idx++;
 351                         og_idx = 0;
 352                 }
 353                 if (--ng_idx < 0) {
 354                         if (np_idx == 0)
 355                                 break;
 356                         np_idx--;
 357                         ng_idx = PAGES_PER_POOL - 1;
 358                 }
 359         }
 360
 361         /*
 362          * (2) add pools if needed.
 363          */
 364         cur_npools = (page_pools.epp_total_pages + PAGES_PER_POOL - 1) /
 365                      PAGES_PER_POOL;
 366         end_npools = (page_pools.epp_total_pages + npages + PAGES_PER_POOL - 1)
 367                      / PAGES_PER_POOL;
 368         LASSERT(end_npools <= page_pools.epp_max_pools);
 369
 370         np_idx = 0;
 371         while (cur_npools < end_npools) {
 372                 LASSERT(page_pools.epp_pools[cur_npools] == NULL);
 373                 LASSERT(np_idx < npools);
 374                 LASSERT(pools[np_idx] != NULL);
 375
 376                 page_pools.epp_pools[cur_npools++] = pools[np_idx];
 377                 pools[np_idx++] = NULL;
 378         }
 379
 380         page_pools.epp_total_pages += npages;
 381         page_pools.epp_free_pages += npages;
 382         page_pools.epp_st_lowfree = page_pools.epp_free_pages;
 383
 384         if (page_pools.epp_total_pages > page_pools.epp_st_max_pages)
 385                 page_pools.epp_st_max_pages = page_pools.epp_total_pages;
 386
 387         CDEBUG(D_SEC, "add %d pages to total %lu\n", npages,
 388                page_pools.epp_total_pages);
 389
 390         spin_unlock(&page_pools.epp_lock);
 391 }
 392
 393 static int enc_pools_add_pages(int npages)
 394 {
 395         static DEFINE_MUTEX(add_pages_mutex);
 396         struct page   ***pools;
 397         int          npools, alloced = 0;
 398         int          i, j, rc = -ENOMEM;
 399
 400         if (npages < PTLRPC_MAX_BRW_PAGES)
 401                 npages = PTLRPC_MAX_BRW_PAGES;
 402
 403         mutex_lock(&add_pages_mutex);
 404
 405         if (npages + page_pools.epp_total_pages > page_pools.epp_max_pages)
 406                 npages = page_pools.epp_max_pages - page_pools.epp_total_pages;
 407         LASSERT(npages > 0);
 408
 409         page_pools.epp_st_grows++;
 410
 411         npools = npages_to_npools(npages);
 412         OBD_ALLOC(pools, npools * sizeof(*pools));
 413         if (pools == NULL)
 414                 goto out;
 415
 416         for (i = 0; i < npools; i++) {
 417                 OBD_ALLOC(pools[i], PAGE_CACHE_SIZE);
 418                 if (pools[i] == NULL)
 419                         goto out_pools;
 420
 421                 for (j = 0; j < PAGES_PER_POOL && alloced < npages; j++) {
 422                         pools[i][j] = alloc_page(GFP_NOFS |
 423                                                      __GFP_HIGHMEM);
 424                         if (pools[i][j] == NULL)
 425                                 goto out_pools;
 426
 427                         alloced++;
 428                 }
 429         }
 430         LASSERT(alloced == npages);
 431
 432         enc_pools_insert(pools, npools, npages);
 433         CDEBUG(D_SEC, "added %d pages into pools\n", npages);
 434         rc = 0;
 435
 436 out_pools:
 437         enc_pools_cleanup(pools, npools);
 438         OBD_FREE(pools, npools * sizeof(*pools));
 439 out:
 440         if (rc) {
 441                 page_pools.epp_st_grow_fails++;
 442                 CERROR("Failed to allocate %d enc pages\n", npages);
 443         }
 444
 445         mutex_unlock(&add_pages_mutex);
 446         return rc;
 447 }
 448
 449 static inline void enc_pools_wakeup(void)
 450 {
 451         assert_spin_locked(&page_pools.epp_lock);
 452         LASSERT(page_pools.epp_waitqlen >= 0);
 453
 454         if (unlikely(page_pools.epp_waitqlen)) {
 455                 LASSERT(waitqueue_active(&page_pools.epp_waitq));
 456                 wake_up_all(&page_pools.epp_waitq);
 457         }
 458 }
 459
 460 static int enc_pools_should_grow(int page_needed, long now)
 461 {
 462         /* don't grow if someone else is growing the pools right now,
 463          * or the pools has reached its full capacity
 464          */
 465         if (page_pools.epp_growing ||
 466             page_pools.epp_total_pages == page_pools.epp_max_pages)
 467                 return 0;
 468
 469         /* if total pages is not enough, we need to grow */
 470         if (page_pools.epp_total_pages < page_needed)
 471                 return 1;
 472
 473         /*
 474          * we wanted to return 0 here if there was a shrink just happened
 475          * moment ago, but this may cause deadlock if both client and ost
 476          * live on single node.
 477          */
 478 #if 0
 479         if (now - page_pools.epp_last_shrink < 2)
 480                 return 0;
 481 #endif
 482
 483         /*
 484          * here we perhaps need consider other factors like wait queue
 485          * length, idle index, etc. ?
 486          */
 487
 488         /* grow the pools in any other cases */
 489         return 1;
 490 }
 491
 492 /*
 493  * we allocate the requested pages atomically.
 494  */
 495 int sptlrpc_enc_pool_get_pages(struct ptlrpc_bulk_desc *desc)
 496 {
 497         wait_queue_t  waitlink;
 498         unsigned long   this_idle = -1;
 499         unsigned long      tick = 0;
 500         long        now;
 501         int          p_idx, g_idx;
 502         int          i;
 503
 504         LASSERT(desc->bd_iov_count > 0);
 505         LASSERT(desc->bd_iov_count <= page_pools.epp_max_pages);
 506
 507         /* resent bulk, enc iov might have been allocated previously */
 508         if (desc->bd_enc_iov != NULL)
 509                 return 0;
 510
 511         OBD_ALLOC(desc->bd_enc_iov,
 512                   desc->bd_iov_count * sizeof(*desc->bd_enc_iov));
 513         if (desc->bd_enc_iov == NULL)
 514                 return -ENOMEM;
 515
 516         spin_lock(&page_pools.epp_lock);
 517
 518         page_pools.epp_st_access++;
 519 again:
 520         if (unlikely(page_pools.epp_free_pages < desc->bd_iov_count)) {
 521                 if (tick == 0)
 522                         tick = cfs_time_current();
 523
 524                 now = get_seconds();
 525
 526                 page_pools.epp_st_missings++;
 527                 page_pools.epp_pages_short += desc->bd_iov_count;
 528
 529                 if (enc_pools_should_grow(desc->bd_iov_count, now)) {
 530                         page_pools.epp_growing = 1;
 531
 532                         spin_unlock(&page_pools.epp_lock);
 533                         enc_pools_add_pages(page_pools.epp_pages_short / 2);
 534                         spin_lock(&page_pools.epp_lock);
 535
 536                         page_pools.epp_growing = 0;
 537
 538                         enc_pools_wakeup();
 539                 } else {
 540                         if (++page_pools.epp_waitqlen >
 541                             page_pools.epp_st_max_wqlen)
 542                                 page_pools.epp_st_max_wqlen =
 543                                                 page_pools.epp_waitqlen;
 544
 545                         set_current_state(TASK_UNINTERRUPTIBLE);
 546                         init_waitqueue_entry(&waitlink, current);
 547                         add_wait_queue(&page_pools.epp_waitq, &waitlink);
 548
 549                         spin_unlock(&page_pools.epp_lock);
 550                         schedule();
 551                         remove_wait_queue(&page_pools.epp_waitq, &waitlink);
 552                         LASSERT(page_pools.epp_waitqlen > 0);
 553                         spin_lock(&page_pools.epp_lock);
 554                         page_pools.epp_waitqlen--;
 555                 }
 556
 557                 LASSERT(page_pools.epp_pages_short >= desc->bd_iov_count);
 558                 page_pools.epp_pages_short -= desc->bd_iov_count;
 559
 560                 this_idle = 0;
 561                 goto again;
 562         }
 563
 564         /* record max wait time */
 565         if (unlikely(tick != 0)) {
 566                 tick = cfs_time_current() - tick;
 567                 if (tick > page_pools.epp_st_max_wait)
 568                         page_pools.epp_st_max_wait = tick;
 569         }
 570
 571         /* proceed with rest of allocation */
 572         page_pools.epp_free_pages -= desc->bd_iov_count;
 573
 574         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 575         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 576
 577         for (i = 0; i < desc->bd_iov_count; i++) {
 578                 LASSERT(page_pools.epp_pools[p_idx][g_idx] != NULL);
 579                 desc->bd_enc_iov[i].kiov_page =
 580                                         page_pools.epp_pools[p_idx][g_idx];
 581                 page_pools.epp_pools[p_idx][g_idx] = NULL;
 582
 583                 if (++g_idx == PAGES_PER_POOL) {
 584                         p_idx++;
 585                         g_idx = 0;
 586                 }
 587         }
 588
 589         if (page_pools.epp_free_pages < page_pools.epp_st_lowfree)
 590                 page_pools.epp_st_lowfree = page_pools.epp_free_pages;
 591
 592         /*
 593          * new idle index = (old * weight + new) / (weight + 1)
 594          */
 595         if (this_idle == -1) {
 596                 this_idle = page_pools.epp_free_pages * IDLE_IDX_MAX /
 597                             page_pools.epp_total_pages;
 598         }
 599         page_pools.epp_idle_idx = (page_pools.epp_idle_idx * IDLE_IDX_WEIGHT +
 600                                    this_idle) /
 601                                   (IDLE_IDX_WEIGHT + 1);
 602
 603         page_pools.epp_last_access = get_seconds();
 604
 605         spin_unlock(&page_pools.epp_lock);
 606         return 0;
 607 }
 608 EXPORT_SYMBOL(sptlrpc_enc_pool_get_pages);
 609
 610 void sptlrpc_enc_pool_put_pages(struct ptlrpc_bulk_desc *desc)
 611 {
 612         int     p_idx, g_idx;
 613         int     i;
 614
 615         if (desc->bd_enc_iov == NULL)
 616                 return;
 617
 618         LASSERT(desc->bd_iov_count > 0);
 619
 620         spin_lock(&page_pools.epp_lock);
 621
 622         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 623         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 624
 625         LASSERT(page_pools.epp_free_pages + desc->bd_iov_count <=
 626                 page_pools.epp_total_pages);
 627         LASSERT(page_pools.epp_pools[p_idx]);
 628
 629         for (i = 0; i < desc->bd_iov_count; i++) {
 630                 LASSERT(desc->bd_enc_iov[i].kiov_page != NULL);
 631                 LASSERT(g_idx != 0 || page_pools.epp_pools[p_idx]);
 632                 LASSERT(page_pools.epp_pools[p_idx][g_idx] == NULL);
 633
 634                 page_pools.epp_pools[p_idx][g_idx] =
 635                                         desc->bd_enc_iov[i].kiov_page;
 636
 637                 if (++g_idx == PAGES_PER_POOL) {
 638                         p_idx++;
 639                         g_idx = 0;
 640                 }
 641         }
 642
 643         page_pools.epp_free_pages += desc->bd_iov_count;
 644
 645         enc_pools_wakeup();
 646
 647         spin_unlock(&page_pools.epp_lock);
 648
 649         OBD_FREE(desc->bd_enc_iov,
 650                  desc->bd_iov_count * sizeof(*desc->bd_enc_iov));
 651         desc->bd_enc_iov = NULL;
 652 }
 653 EXPORT_SYMBOL(sptlrpc_enc_pool_put_pages);
 654
 655 /*
 656  * we don't do much stuff for add_user/del_user anymore, except adding some
 657  * initial pages in add_user() if current pools are empty, rest would be
 658  * handled by the pools's self-adaption.
 659  */
 660 int sptlrpc_enc_pool_add_user(void)
 661 {
 662         int     need_grow = 0;
 663
 664         spin_lock(&page_pools.epp_lock);
 665         if (page_pools.epp_growing == 0 && page_pools.epp_total_pages == 0) {
 666                 page_pools.epp_growing = 1;
 667                 need_grow = 1;
 668         }
 669         spin_unlock(&page_pools.epp_lock);
 670
 671         if (need_grow) {
 672                 enc_pools_add_pages(PTLRPC_MAX_BRW_PAGES +
 673                                     PTLRPC_MAX_BRW_PAGES);
 674
 675                 spin_lock(&page_pools.epp_lock);
 676                 page_pools.epp_growing = 0;
 677                 enc_pools_wakeup();
 678                 spin_unlock(&page_pools.epp_lock);
 679         }
 680         return 0;
 681 }
 682 EXPORT_SYMBOL(sptlrpc_enc_pool_add_user);
 683
 684 int sptlrpc_enc_pool_del_user(void)
 685 {
 686         return 0;
 687 }
 688 EXPORT_SYMBOL(sptlrpc_enc_pool_del_user);
 689
 690 static inline void enc_pools_alloc(void)
 691 {
 692         LASSERT(page_pools.epp_max_pools);
 693         OBD_ALLOC_LARGE(page_pools.epp_pools,
 694                         page_pools.epp_max_pools *
 695                         sizeof(*page_pools.epp_pools));
 696 }
 697
 698 static inline void enc_pools_free(void)
 699 {
 700         LASSERT(page_pools.epp_max_pools);
 701         LASSERT(page_pools.epp_pools);
 702
 703         OBD_FREE_LARGE(page_pools.epp_pools,
 704                        page_pools.epp_max_pools *
 705                        sizeof(*page_pools.epp_pools));
 706 }
 707
 708 static struct shrinker pools_shrinker = {
 709         .count_objects  = enc_pools_shrink_count,
 710         .scan_objects   = enc_pools_shrink_scan,
 711         .seeks          = DEFAULT_SEEKS,
 712 };
 713
 714 int sptlrpc_enc_pool_init(void)
 715 {
 716         /*
 717          * maximum capacity is 1/8 of total physical memory.
 718          * is the 1/8 a good number?
 719          */
 720         page_pools.epp_max_pages = totalram_pages / 8;
 721         page_pools.epp_max_pools = npages_to_npools(page_pools.epp_max_pages);
 722
 723         init_waitqueue_head(&page_pools.epp_waitq);
 724         page_pools.epp_waitqlen = 0;
 725         page_pools.epp_pages_short = 0;
 726
 727         page_pools.epp_growing = 0;
 728
 729         page_pools.epp_idle_idx = 0;
 730         page_pools.epp_last_shrink = get_seconds();
 731         page_pools.epp_last_access = get_seconds();
 732
 733         spin_lock_init(&page_pools.epp_lock);
 734         page_pools.epp_total_pages = 0;
 735         page_pools.epp_free_pages = 0;
 736
 737         page_pools.epp_st_max_pages = 0;
 738         page_pools.epp_st_grows = 0;
 739         page_pools.epp_st_grow_fails = 0;
 740         page_pools.epp_st_shrinks = 0;
 741         page_pools.epp_st_access = 0;
 742         page_pools.epp_st_missings = 0;
 743         page_pools.epp_st_lowfree = 0;
 744         page_pools.epp_st_max_wqlen = 0;
 745         page_pools.epp_st_max_wait = 0;
 746
 747         enc_pools_alloc();
 748         if (page_pools.epp_pools == NULL)
 749                 return -ENOMEM;
 750
 751         register_shrinker(&pools_shrinker);
 752
 753         return 0;
 754 }
 755
 756 void sptlrpc_enc_pool_fini(void)
 757 {
 758         unsigned long cleaned, npools;
 759
 760         LASSERT(page_pools.epp_pools);
 761         LASSERT(page_pools.epp_total_pages == page_pools.epp_free_pages);
 762
 763         unregister_shrinker(&pools_shrinker);
 764
 765         npools = npages_to_npools(page_pools.epp_total_pages);
 766         cleaned = enc_pools_cleanup(page_pools.epp_pools, npools);
 767         LASSERT(cleaned == page_pools.epp_total_pages);
 768
 769         enc_pools_free();
 770
 771         if (page_pools.epp_st_access > 0) {
 772                 CDEBUG(D_SEC,
 773                        "max pages %lu, grows %u, grow fails %u, shrinks %u, access %lu, missing %lu, max qlen %u, max wait "
 774                        CFS_TIME_T"/%d\n",
 775                        page_pools.epp_st_max_pages, page_pools.epp_st_grows,
 776                        page_pools.epp_st_grow_fails,
 777                        page_pools.epp_st_shrinks, page_pools.epp_st_access,
 778                        page_pools.epp_st_missings, page_pools.epp_st_max_wqlen,
 779                        page_pools.epp_st_max_wait, HZ);
 780         }
 781 }
 782
 783
 784 static int cfs_hash_alg_id[] = {
 785         [BULK_HASH_ALG_NULL]    = CFS_HASH_ALG_NULL,
 786         [BULK_HASH_ALG_ADLER32] = CFS_HASH_ALG_ADLER32,
 787         [BULK_HASH_ALG_CRC32]   = CFS_HASH_ALG_CRC32,
 788         [BULK_HASH_ALG_MD5]     = CFS_HASH_ALG_MD5,
 789         [BULK_HASH_ALG_SHA1]    = CFS_HASH_ALG_SHA1,
 790         [BULK_HASH_ALG_SHA256]  = CFS_HASH_ALG_SHA256,
 791         [BULK_HASH_ALG_SHA384]  = CFS_HASH_ALG_SHA384,
 792         [BULK_HASH_ALG_SHA512]  = CFS_HASH_ALG_SHA512,
 793 };
 794 const char *sptlrpc_get_hash_name(__u8 hash_alg)
 795 {
 796         return cfs_crypto_hash_name(cfs_hash_alg_id[hash_alg]);
 797 }
 798 EXPORT_SYMBOL(sptlrpc_get_hash_name);
 799
 800 __u8 sptlrpc_get_hash_alg(const char *algname)
 801 {
 802         return cfs_crypto_hash_alg(algname);
 803 }
 804 EXPORT_SYMBOL(sptlrpc_get_hash_alg);
 805
 806 int bulk_sec_desc_unpack(struct lustre_msg *msg, int offset, int swabbed)
 807 {
 808         struct ptlrpc_bulk_sec_desc *bsd;
 809         int                       size = msg->lm_buflens[offset];
 810
 811         bsd = lustre_msg_buf(msg, offset, sizeof(*bsd));
 812         if (bsd == NULL) {
 813                 CERROR("Invalid bulk sec desc: size %d\n", size);
 814                 return -EINVAL;
 815         }
 816
 817         if (swabbed)
 818                 __swab32s(&bsd->bsd_nob);
 819
 820         if (unlikely(bsd->bsd_version != 0)) {
 821                 CERROR("Unexpected version %u\n", bsd->bsd_version);
 822                 return -EPROTO;
 823         }
 824
 825         if (unlikely(bsd->bsd_type >= SPTLRPC_BULK_MAX)) {
 826                 CERROR("Invalid type %u\n", bsd->bsd_type);
 827                 return -EPROTO;
 828         }
 829
 830         /* FIXME more sanity check here */
 831
 832         if (unlikely(bsd->bsd_svc != SPTLRPC_BULK_SVC_NULL &&
 833                      bsd->bsd_svc != SPTLRPC_BULK_SVC_INTG &&
 834                      bsd->bsd_svc != SPTLRPC_BULK_SVC_PRIV)) {
 835                 CERROR("Invalid svc %u\n", bsd->bsd_svc);
 836                 return -EPROTO;
 837         }
 838
 839         return 0;
 840 }
 841 EXPORT_SYMBOL(bulk_sec_desc_unpack);
 842
 843 int sptlrpc_get_bulk_checksum(struct ptlrpc_bulk_desc *desc, __u8 alg,
 844                               void *buf, int buflen)
 845 {
 846         struct cfs_crypto_hash_desc     *hdesc;
 847         int                             hashsize;
 848         char                            hashbuf[64];
 849         unsigned int                    bufsize;
 850         int                             i, err;
 851
 852         LASSERT(alg > BULK_HASH_ALG_NULL && alg < BULK_HASH_ALG_MAX);
 853         LASSERT(buflen >= 4);
 854
 855         hdesc = cfs_crypto_hash_init(cfs_hash_alg_id[alg], NULL, 0);
 856         if (IS_ERR(hdesc)) {
 857                 CERROR("Unable to initialize checksum hash %s\n",
 858                        cfs_crypto_hash_name(cfs_hash_alg_id[alg]));
 859                 return PTR_ERR(hdesc);
 860         }
 861
 862         hashsize = cfs_crypto_hash_digestsize(cfs_hash_alg_id[alg]);
 863
 864         for (i = 0; i < desc->bd_iov_count; i++) {
 865                 cfs_crypto_hash_update_page(hdesc, desc->bd_iov[i].kiov_page,
 866                                   desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK,
 867                                   desc->bd_iov[i].kiov_len);
 868         }
 869         if (hashsize > buflen) {
 870                 bufsize = sizeof(hashbuf);
 871                 err = cfs_crypto_hash_final(hdesc, (unsigned char *)hashbuf,
 872                                             &bufsize);
 873                 memcpy(buf, hashbuf, buflen);
 874         } else {
 875                 bufsize = buflen;
 876                 err = cfs_crypto_hash_final(hdesc, (unsigned char *)buf,
 877                                             &bufsize);
 878         }
 879
 880         if (err)
 881                 cfs_crypto_hash_final(hdesc, NULL, NULL);
 882         return err;
 883 }
 884 EXPORT_SYMBOL(sptlrpc_get_bulk_checksum);