lustre/ptlrpc/sec_bulk.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  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 #ifndef EXPORT_SYMTAB
  42 #define EXPORT_SYMTAB
  43 #endif
  44 #define DEBUG_SUBSYSTEM S_SEC
  45
  46 #include <libcfs/libcfs.h>
  47 #ifndef __KERNEL__
  48 #include <liblustre.h>
  49 #include <libcfs/list.h>
  50 #else
  51 #include <linux/crypto.h>
  52 #endif
  53
  54 #include <obd.h>
  55 #include <obd_cksum.h>
  56 #include <obd_class.h>
  57 #include <obd_support.h>
  58 #include <lustre_net.h>
  59 #include <lustre_import.h>
  60 #include <lustre_dlm.h>
  61 #include <lustre_sec.h>
  62
  63 #include "ptlrpc_internal.h"
  64
  65 /****************************************
  66  * bulk encryption page pools           *
  67  ****************************************/
  68
  69 #ifdef __KERNEL__
  70
  71 #define PTRS_PER_PAGE   (CFS_PAGE_SIZE / sizeof(void *))
  72 #define PAGES_PER_POOL  (PTRS_PER_PAGE)
  73
  74 #define IDLE_IDX_MAX            (100)
  75 #define IDLE_IDX_WEIGHT         (3)
  76
  77 #define CACHE_QUIESCENT_PERIOD  (20)
  78
  79 static struct ptlrpc_enc_page_pool {
  80         /*
  81          * constants
  82          */
  83         unsigned long    epp_max_pages;   /* maximum pages can hold, const */
  84         unsigned int     epp_max_pools;   /* number of pools, const */
  85
  86         /*
  87          * wait queue in case of not enough free pages.
  88          */
  89         cfs_waitq_t      epp_waitq;       /* waiting threads */
  90         unsigned int     epp_waitqlen;    /* wait queue length */
  91         unsigned long    epp_pages_short; /* # of pages wanted of in-q users */
  92         unsigned int     epp_growing:1;   /* during adding pages */
  93
  94         /*
  95          * indicating how idle the pools are, from 0 to MAX_IDLE_IDX
  96          * this is counted based on each time when getting pages from
  97          * the pools, not based on time. which means in case that system
  98          * is idled for a while but the idle_idx might still be low if no
  99          * activities happened in the pools.
 100          */
 101         unsigned long    epp_idle_idx;
 102
 103         /* last shrink time due to mem tight */
 104         long             epp_last_shrink;
 105         long             epp_last_access;
 106
 107         /*
 108          * in-pool pages bookkeeping
 109          */
 110         cfs_spinlock_t   epp_lock;        /* protect following fields */
 111         unsigned long    epp_total_pages; /* total pages in pools */
 112         unsigned long    epp_free_pages;  /* current pages available */
 113
 114         /*
 115          * statistics
 116          */
 117         unsigned long    epp_st_max_pages;      /* # of pages ever reached */
 118         unsigned int     epp_st_grows;          /* # of grows */
 119         unsigned int     epp_st_grow_fails;     /* # of add pages failures */
 120         unsigned int     epp_st_shrinks;        /* # of shrinks */
 121         unsigned long    epp_st_access;         /* # of access */
 122         unsigned long    epp_st_missings;       /* # of cache missing */
 123         unsigned long    epp_st_lowfree;        /* lowest free pages reached */
 124         unsigned int     epp_st_max_wqlen;      /* highest waitqueue length */
 125         cfs_time_t       epp_st_max_wait;       /* in jeffies */
 126         /*
 127          * pointers to pools
 128          */
 129         cfs_page_t    ***epp_pools;
 130 } page_pools;
 131
 132 /*
 133  * memory shrinker
 134  */
 135 const int pools_shrinker_seeks = CFS_DEFAULT_SEEKS;
 136 static struct cfs_shrinker *pools_shrinker = NULL;
 137
 138
 139 /*
 140  * /proc/fs/lustre/sptlrpc/encrypt_page_pools
 141  */
 142 int sptlrpc_proc_read_enc_pool(char *page, char **start, off_t off, int count,
 143                                int *eof, void *data)
 144 {
 145         int     rc;
 146
 147         cfs_spin_lock(&page_pools.epp_lock);
 148
 149         rc = snprintf(page, count,
 150                       "physical pages:          %lu\n"
 151                       "pages per pool:          %lu\n"
 152                       "max pages:               %lu\n"
 153                       "max pools:               %u\n"
 154                       "total pages:             %lu\n"
 155                       "total free:              %lu\n"
 156                       "idle index:              %lu/100\n"
 157                       "last shrink:             %lds\n"
 158                       "last access:             %lds\n"
 159                       "max pages reached:       %lu\n"
 160                       "grows:                   %u\n"
 161                       "grows failure:           %u\n"
 162                       "shrinks:                 %u\n"
 163                       "cache access:            %lu\n"
 164                       "cache missing:           %lu\n"
 165                       "low free mark:           %lu\n"
 166                       "max waitqueue depth:     %u\n"
 167                       "max wait time:           "CFS_TIME_T"/%u\n"
 168                       ,
 169                       cfs_num_physpages,
 170                       PAGES_PER_POOL,
 171                       page_pools.epp_max_pages,
 172                       page_pools.epp_max_pools,
 173                       page_pools.epp_total_pages,
 174                       page_pools.epp_free_pages,
 175                       page_pools.epp_idle_idx,
 176                       cfs_time_current_sec() - page_pools.epp_last_shrink,
 177                       cfs_time_current_sec() - page_pools.epp_last_access,
 178                       page_pools.epp_st_max_pages,
 179                       page_pools.epp_st_grows,
 180                       page_pools.epp_st_grow_fails,
 181                       page_pools.epp_st_shrinks,
 182                       page_pools.epp_st_access,
 183                       page_pools.epp_st_missings,
 184                       page_pools.epp_st_lowfree,
 185                       page_pools.epp_st_max_wqlen,
 186                       page_pools.epp_st_max_wait, CFS_HZ
 187                      );
 188
 189         cfs_spin_unlock(&page_pools.epp_lock);
 190         return rc;
 191 }
 192
 193 static void enc_pools_release_free_pages(long npages)
 194 {
 195         int     p_idx, g_idx;
 196         int     p_idx_max1, p_idx_max2;
 197
 198         LASSERT(npages > 0);
 199         LASSERT(npages <= page_pools.epp_free_pages);
 200         LASSERT(page_pools.epp_free_pages <= page_pools.epp_total_pages);
 201
 202         /* max pool index before the release */
 203         p_idx_max2 = (page_pools.epp_total_pages - 1) / PAGES_PER_POOL;
 204
 205         page_pools.epp_free_pages -= npages;
 206         page_pools.epp_total_pages -= npages;
 207
 208         /* max pool index after the release */
 209         p_idx_max1 = page_pools.epp_total_pages == 0 ? -1 :
 210                      ((page_pools.epp_total_pages - 1) / PAGES_PER_POOL);
 211
 212         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 213         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 214         LASSERT(page_pools.epp_pools[p_idx]);
 215
 216         while (npages--) {
 217                 LASSERT(page_pools.epp_pools[p_idx]);
 218                 LASSERT(page_pools.epp_pools[p_idx][g_idx] != NULL);
 219
 220                 cfs_free_page(page_pools.epp_pools[p_idx][g_idx]);
 221                 page_pools.epp_pools[p_idx][g_idx] = NULL;
 222
 223                 if (++g_idx == PAGES_PER_POOL) {
 224                         p_idx++;
 225                         g_idx = 0;
 226                 }
 227         };
 228
 229         /* free unused pools */
 230         while (p_idx_max1 < p_idx_max2) {
 231                 LASSERT(page_pools.epp_pools[p_idx_max2]);
 232                 OBD_FREE(page_pools.epp_pools[p_idx_max2], CFS_PAGE_SIZE);
 233                 page_pools.epp_pools[p_idx_max2] = NULL;
 234                 p_idx_max2--;
 235         }
 236 }
 237
 238 /*
 239  * could be called frequently for query (@nr_to_scan == 0).
 240  * we try to keep at least PTLRPC_MAX_BRW_PAGES pages in the pool.
 241  */
 242 static int enc_pools_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
 243 {
 244         if (unlikely(shrink_param(sc, nr_to_scan) != 0)) {
 245                 cfs_spin_lock(&page_pools.epp_lock);
 246                 shrink_param(sc, nr_to_scan) = min_t(unsigned long,
 247                                                    shrink_param(sc, nr_to_scan),
 248                                                    page_pools.epp_free_pages -
 249                                                    PTLRPC_MAX_BRW_PAGES);
 250                 if (shrink_param(sc, nr_to_scan) > 0) {
 251                         enc_pools_release_free_pages(shrink_param(sc,
 252                                                                   nr_to_scan));
 253                         CDEBUG(D_SEC, "released %ld pages, %ld left\n",
 254                                (long)shrink_param(sc, nr_to_scan),
 255                                page_pools.epp_free_pages);
 256
 257                         page_pools.epp_st_shrinks++;
 258                         page_pools.epp_last_shrink = cfs_time_current_sec();
 259                 }
 260                 cfs_spin_unlock(&page_pools.epp_lock);
 261         }
 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(cfs_time_current_sec() - page_pools.epp_last_access >
 268                      CACHE_QUIESCENT_PERIOD)) {
 269                 cfs_spin_lock(&page_pools.epp_lock);
 270                 page_pools.epp_idle_idx = IDLE_IDX_MAX;
 271                 cfs_spin_unlock(&page_pools.epp_lock);
 272         }
 273
 274         LASSERT(page_pools.epp_idle_idx <= IDLE_IDX_MAX);
 275         return max((int) page_pools.epp_free_pages - PTLRPC_MAX_BRW_PAGES, 0) *
 276                (IDLE_IDX_MAX - page_pools.epp_idle_idx) / IDLE_IDX_MAX;
 277 }
 278
 279 static inline
 280 int npages_to_npools(unsigned long npages)
 281 {
 282         return (int) ((npages + PAGES_PER_POOL - 1) / PAGES_PER_POOL);
 283 }
 284
 285 /*
 286  * return how many pages cleaned up.
 287  */
 288 static unsigned long enc_pools_cleanup(cfs_page_t ***pools, int npools)
 289 {
 290         unsigned long cleaned = 0;
 291         int           i, j;
 292
 293         for (i = 0; i < npools; i++) {
 294                 if (pools[i]) {
 295                         for (j = 0; j < PAGES_PER_POOL; j++) {
 296                                 if (pools[i][j]) {
 297                                         cfs_free_page(pools[i][j]);
 298                                         cleaned++;
 299                                 }
 300                         }
 301                         OBD_FREE(pools[i], CFS_PAGE_SIZE);
 302                         pools[i] = NULL;
 303                 }
 304         }
 305
 306         return cleaned;
 307 }
 308
 309 /*
 310  * merge @npools pointed by @pools which contains @npages new pages
 311  * into current pools.
 312  *
 313  * we have options to avoid most memory copy with some tricks. but we choose
 314  * the simplest way to avoid complexity. It's not frequently called.
 315  */
 316 static void enc_pools_insert(cfs_page_t ***pools, int npools, int npages)
 317 {
 318         int     freeslot;
 319         int     op_idx, np_idx, og_idx, ng_idx;
 320         int     cur_npools, end_npools;
 321
 322         LASSERT(npages > 0);
 323         LASSERT(page_pools.epp_total_pages+npages <= page_pools.epp_max_pages);
 324         LASSERT(npages_to_npools(npages) == npools);
 325         LASSERT(page_pools.epp_growing);
 326
 327         cfs_spin_lock(&page_pools.epp_lock);
 328
 329         /*
 330          * (1) fill all the free slots of current pools.
 331          */
 332         /* free slots are those left by rent pages, and the extra ones with
 333          * index >= total_pages, locate at the tail of last pool. */
 334         freeslot = page_pools.epp_total_pages % PAGES_PER_POOL;
 335         if (freeslot != 0)
 336                 freeslot = PAGES_PER_POOL - freeslot;
 337         freeslot += page_pools.epp_total_pages - page_pools.epp_free_pages;
 338
 339         op_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 340         og_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 341         np_idx = npools - 1;
 342         ng_idx = (npages - 1) % PAGES_PER_POOL;
 343
 344         while (freeslot) {
 345                 LASSERT(page_pools.epp_pools[op_idx][og_idx] == NULL);
 346                 LASSERT(pools[np_idx][ng_idx] != NULL);
 347
 348                 page_pools.epp_pools[op_idx][og_idx] = pools[np_idx][ng_idx];
 349                 pools[np_idx][ng_idx] = NULL;
 350
 351                 freeslot--;
 352
 353                 if (++og_idx == PAGES_PER_POOL) {
 354                         op_idx++;
 355                         og_idx = 0;
 356                 }
 357                 if (--ng_idx < 0) {
 358                         if (np_idx == 0)
 359                                 break;
 360                         np_idx--;
 361                         ng_idx = PAGES_PER_POOL - 1;
 362                 }
 363         }
 364
 365         /*
 366          * (2) add pools if needed.
 367          */
 368         cur_npools = (page_pools.epp_total_pages + PAGES_PER_POOL - 1) /
 369                      PAGES_PER_POOL;
 370         end_npools = (page_pools.epp_total_pages + npages + PAGES_PER_POOL -1) /
 371                      PAGES_PER_POOL;
 372         LASSERT(end_npools <= page_pools.epp_max_pools);
 373
 374         np_idx = 0;
 375         while (cur_npools < end_npools) {
 376                 LASSERT(page_pools.epp_pools[cur_npools] == NULL);
 377                 LASSERT(np_idx < npools);
 378                 LASSERT(pools[np_idx] != NULL);
 379
 380                 page_pools.epp_pools[cur_npools++] = pools[np_idx];
 381                 pools[np_idx++] = NULL;
 382         }
 383
 384         page_pools.epp_total_pages += npages;
 385         page_pools.epp_free_pages += npages;
 386         page_pools.epp_st_lowfree = page_pools.epp_free_pages;
 387
 388         if (page_pools.epp_total_pages > page_pools.epp_st_max_pages)
 389                 page_pools.epp_st_max_pages = page_pools.epp_total_pages;
 390
 391         CDEBUG(D_SEC, "add %d pages to total %lu\n", npages,
 392                page_pools.epp_total_pages);
 393
 394         cfs_spin_unlock(&page_pools.epp_lock);
 395 }
 396
 397 static int enc_pools_add_pages(int npages)
 398 {
 399         static CFS_DECLARE_MUTEX(sem_add_pages);
 400         cfs_page_t   ***pools;
 401         int             npools, alloced = 0;
 402         int             i, j, rc = -ENOMEM;
 403
 404         if (npages < PTLRPC_MAX_BRW_PAGES)
 405                 npages = PTLRPC_MAX_BRW_PAGES;
 406
 407         cfs_down(&sem_add_pages);
 408
 409         if (npages + page_pools.epp_total_pages > page_pools.epp_max_pages)
 410                 npages = page_pools.epp_max_pages - page_pools.epp_total_pages;
 411         LASSERT(npages > 0);
 412
 413         page_pools.epp_st_grows++;
 414
 415         npools = npages_to_npools(npages);
 416         OBD_ALLOC(pools, npools * sizeof(*pools));
 417         if (pools == NULL)
 418                 goto out;
 419
 420         for (i = 0; i < npools; i++) {
 421                 OBD_ALLOC(pools[i], CFS_PAGE_SIZE);
 422                 if (pools[i] == NULL)
 423                         goto out_pools;
 424
 425                 for (j = 0; j < PAGES_PER_POOL && alloced < npages; j++) {
 426                         pools[i][j] = cfs_alloc_page(CFS_ALLOC_IO |
 427                                                      CFS_ALLOC_HIGH);
 428                         if (pools[i][j] == NULL)
 429                                 goto out_pools;
 430
 431                         alloced++;
 432                 }
 433         }
 434         LASSERT(alloced == npages);
 435
 436         enc_pools_insert(pools, npools, npages);
 437         CDEBUG(D_SEC, "added %d pages into pools\n", npages);
 438         rc = 0;
 439
 440 out_pools:
 441         enc_pools_cleanup(pools, npools);
 442         OBD_FREE(pools, npools * sizeof(*pools));
 443 out:
 444         if (rc) {
 445                 page_pools.epp_st_grow_fails++;
 446                 CERROR("Failed to allocate %d enc pages\n", npages);
 447         }
 448
 449         cfs_up(&sem_add_pages);
 450         return rc;
 451 }
 452
 453 static inline void enc_pools_wakeup(void)
 454 {
 455         LASSERT_SPIN_LOCKED(&page_pools.epp_lock);
 456         LASSERT(page_pools.epp_waitqlen >= 0);
 457
 458         if (unlikely(page_pools.epp_waitqlen)) {
 459                 LASSERT(cfs_waitq_active(&page_pools.epp_waitq));
 460                 cfs_waitq_broadcast(&page_pools.epp_waitq);
 461         }
 462 }
 463
 464 static int enc_pools_should_grow(int page_needed, long now)
 465 {
 466         /* don't grow if someone else is growing the pools right now,
 467          * or the pools has reached its full capacity
 468          */
 469         if (page_pools.epp_growing ||
 470             page_pools.epp_total_pages == page_pools.epp_max_pages)
 471                 return 0;
 472
 473         /* if total pages is not enough, we need to grow */
 474         if (page_pools.epp_total_pages < page_needed)
 475                 return 1;
 476
 477         /*
 478          * we wanted to return 0 here if there was a shrink just happened
 479          * moment ago, but this may cause deadlock if both client and ost
 480          * live on single node.
 481          */
 482 #if 0
 483         if (now - page_pools.epp_last_shrink < 2)
 484                 return 0;
 485 #endif
 486
 487         /*
 488          * here we perhaps need consider other factors like wait queue
 489          * length, idle index, etc. ?
 490          */
 491
 492         /* grow the pools in any other cases */
 493         return 1;
 494 }
 495
 496 /*
 497  * we allocate the requested pages atomically.
 498  */
 499 int sptlrpc_enc_pool_get_pages(struct ptlrpc_bulk_desc *desc)
 500 {
 501         cfs_waitlink_t  waitlink;
 502         unsigned long   this_idle = -1;
 503         cfs_time_t      tick = 0;
 504         long            now;
 505         int             p_idx, g_idx;
 506         int             i;
 507
 508         LASSERT(desc->bd_iov_count > 0);
 509         LASSERT(desc->bd_iov_count <= page_pools.epp_max_pages);
 510
 511         /* resent bulk, enc iov might have been allocated previously */
 512         if (desc->bd_enc_iov != NULL)
 513                 return 0;
 514
 515         OBD_ALLOC(desc->bd_enc_iov,
 516                   desc->bd_iov_count * sizeof(*desc->bd_enc_iov));
 517         if (desc->bd_enc_iov == NULL)
 518                 return -ENOMEM;
 519
 520         cfs_spin_lock(&page_pools.epp_lock);
 521
 522         page_pools.epp_st_access++;
 523 again:
 524         if (unlikely(page_pools.epp_free_pages < desc->bd_iov_count)) {
 525                 if (tick == 0)
 526                         tick = cfs_time_current();
 527
 528                 now = cfs_time_current_sec();
 529
 530                 page_pools.epp_st_missings++;
 531                 page_pools.epp_pages_short += desc->bd_iov_count;
 532
 533                 if (enc_pools_should_grow(desc->bd_iov_count, now)) {
 534                         page_pools.epp_growing = 1;
 535
 536                         cfs_spin_unlock(&page_pools.epp_lock);
 537                         enc_pools_add_pages(page_pools.epp_pages_short / 2);
 538                         cfs_spin_lock(&page_pools.epp_lock);
 539
 540                         page_pools.epp_growing = 0;
 541
 542                         enc_pools_wakeup();
 543                 } else {
 544                         if (++page_pools.epp_waitqlen >
 545                             page_pools.epp_st_max_wqlen)
 546                                 page_pools.epp_st_max_wqlen =
 547                                                 page_pools.epp_waitqlen;
 548
 549                         cfs_set_current_state(CFS_TASK_UNINT);
 550                         cfs_waitlink_init(&waitlink);
 551                         cfs_waitq_add(&page_pools.epp_waitq, &waitlink);
 552
 553                         cfs_spin_unlock(&page_pools.epp_lock);
 554                         cfs_waitq_wait(&waitlink, CFS_TASK_UNINT);
 555                         cfs_waitq_del(&page_pools.epp_waitq, &waitlink);
 556                         LASSERT(page_pools.epp_waitqlen > 0);
 557                         cfs_spin_lock(&page_pools.epp_lock);
 558                         page_pools.epp_waitqlen--;
 559                 }
 560
 561                 LASSERT(page_pools.epp_pages_short >= desc->bd_iov_count);
 562                 page_pools.epp_pages_short -= desc->bd_iov_count;
 563
 564                 this_idle = 0;
 565                 goto again;
 566         }
 567
 568         /* record max wait time */
 569         if (unlikely(tick != 0)) {
 570                 tick = cfs_time_current() - tick;
 571                 if (tick > page_pools.epp_st_max_wait)
 572                         page_pools.epp_st_max_wait = tick;
 573         }
 574
 575         /* proceed with rest of allocation */
 576         page_pools.epp_free_pages -= desc->bd_iov_count;
 577
 578         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 579         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 580
 581         for (i = 0; i < desc->bd_iov_count; i++) {
 582                 LASSERT(page_pools.epp_pools[p_idx][g_idx] != NULL);
 583                 desc->bd_enc_iov[i].kiov_page =
 584                                         page_pools.epp_pools[p_idx][g_idx];
 585                 page_pools.epp_pools[p_idx][g_idx] = NULL;
 586
 587                 if (++g_idx == PAGES_PER_POOL) {
 588                         p_idx++;
 589                         g_idx = 0;
 590                 }
 591         }
 592
 593         if (page_pools.epp_free_pages < page_pools.epp_st_lowfree)
 594                 page_pools.epp_st_lowfree = page_pools.epp_free_pages;
 595
 596         /*
 597          * new idle index = (old * weight + new) / (weight + 1)
 598          */
 599         if (this_idle == -1) {
 600                 this_idle = page_pools.epp_free_pages * IDLE_IDX_MAX /
 601                             page_pools.epp_total_pages;
 602         }
 603         page_pools.epp_idle_idx = (page_pools.epp_idle_idx * IDLE_IDX_WEIGHT +
 604                                    this_idle) /
 605                                   (IDLE_IDX_WEIGHT + 1);
 606
 607         page_pools.epp_last_access = cfs_time_current_sec();
 608
 609         cfs_spin_unlock(&page_pools.epp_lock);
 610         return 0;
 611 }
 612 EXPORT_SYMBOL(sptlrpc_enc_pool_get_pages);
 613
 614 void sptlrpc_enc_pool_put_pages(struct ptlrpc_bulk_desc *desc)
 615 {
 616         int     p_idx, g_idx;
 617         int     i;
 618
 619         if (desc->bd_enc_iov == NULL)
 620                 return;
 621
 622         LASSERT(desc->bd_iov_count > 0);
 623
 624         cfs_spin_lock(&page_pools.epp_lock);
 625
 626         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 627         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 628
 629         LASSERT(page_pools.epp_free_pages + desc->bd_iov_count <=
 630                 page_pools.epp_total_pages);
 631         LASSERT(page_pools.epp_pools[p_idx]);
 632
 633         for (i = 0; i < desc->bd_iov_count; i++) {
 634                 LASSERT(desc->bd_enc_iov[i].kiov_page != NULL);
 635                 LASSERT(g_idx != 0 || page_pools.epp_pools[p_idx]);
 636                 LASSERT(page_pools.epp_pools[p_idx][g_idx] == NULL);
 637
 638                 page_pools.epp_pools[p_idx][g_idx] =
 639                                         desc->bd_enc_iov[i].kiov_page;
 640
 641                 if (++g_idx == PAGES_PER_POOL) {
 642                         p_idx++;
 643                         g_idx = 0;
 644                 }
 645         }
 646
 647         page_pools.epp_free_pages += desc->bd_iov_count;
 648
 649         enc_pools_wakeup();
 650
 651         cfs_spin_unlock(&page_pools.epp_lock);
 652
 653         OBD_FREE(desc->bd_enc_iov,
 654                  desc->bd_iov_count * sizeof(*desc->bd_enc_iov));
 655         desc->bd_enc_iov = NULL;
 656 }
 657 EXPORT_SYMBOL(sptlrpc_enc_pool_put_pages);
 658
 659 /*
 660  * we don't do much stuff for add_user/del_user anymore, except adding some
 661  * initial pages in add_user() if current pools are empty, rest would be
 662  * handled by the pools's self-adaption.
 663  */
 664 int sptlrpc_enc_pool_add_user(void)
 665 {
 666         int     need_grow = 0;
 667
 668         cfs_spin_lock(&page_pools.epp_lock);
 669         if (page_pools.epp_growing == 0 && page_pools.epp_total_pages == 0) {
 670                 page_pools.epp_growing = 1;
 671                 need_grow = 1;
 672         }
 673         cfs_spin_unlock(&page_pools.epp_lock);
 674
 675         if (need_grow) {
 676                 enc_pools_add_pages(PTLRPC_MAX_BRW_PAGES +
 677                                     PTLRPC_MAX_BRW_PAGES);
 678
 679                 cfs_spin_lock(&page_pools.epp_lock);
 680                 page_pools.epp_growing = 0;
 681                 enc_pools_wakeup();
 682                 cfs_spin_unlock(&page_pools.epp_lock);
 683         }
 684         return 0;
 685 }
 686 EXPORT_SYMBOL(sptlrpc_enc_pool_add_user);
 687
 688 int sptlrpc_enc_pool_del_user(void)
 689 {
 690         return 0;
 691 }
 692 EXPORT_SYMBOL(sptlrpc_enc_pool_del_user);
 693
 694 static inline void enc_pools_alloc(void)
 695 {
 696         LASSERT(page_pools.epp_max_pools);
 697         OBD_ALLOC_LARGE(page_pools.epp_pools,
 698                         page_pools.epp_max_pools *
 699                         sizeof(*page_pools.epp_pools));
 700 }
 701
 702 static inline void enc_pools_free(void)
 703 {
 704         LASSERT(page_pools.epp_max_pools);
 705         LASSERT(page_pools.epp_pools);
 706
 707         OBD_FREE_LARGE(page_pools.epp_pools,
 708                        page_pools.epp_max_pools *
 709                        sizeof(*page_pools.epp_pools));
 710 }
 711
 712 int sptlrpc_enc_pool_init(void)
 713 {
 714         /*
 715          * maximum capacity is 1/8 of total physical memory.
 716          * is the 1/8 a good number?
 717          */
 718         page_pools.epp_max_pages = cfs_num_physpages / 8;
 719         page_pools.epp_max_pools = npages_to_npools(page_pools.epp_max_pages);
 720
 721         cfs_waitq_init(&page_pools.epp_waitq);
 722         page_pools.epp_waitqlen = 0;
 723         page_pools.epp_pages_short = 0;
 724
 725         page_pools.epp_growing = 0;
 726
 727         page_pools.epp_idle_idx = 0;
 728         page_pools.epp_last_shrink = cfs_time_current_sec();
 729         page_pools.epp_last_access = cfs_time_current_sec();
 730
 731         cfs_spin_lock_init(&page_pools.epp_lock);
 732         page_pools.epp_total_pages = 0;
 733         page_pools.epp_free_pages = 0;
 734
 735         page_pools.epp_st_max_pages = 0;
 736         page_pools.epp_st_grows = 0;
 737         page_pools.epp_st_grow_fails = 0;
 738         page_pools.epp_st_shrinks = 0;
 739         page_pools.epp_st_access = 0;
 740         page_pools.epp_st_missings = 0;
 741         page_pools.epp_st_lowfree = 0;
 742         page_pools.epp_st_max_wqlen = 0;
 743         page_pools.epp_st_max_wait = 0;
 744
 745         enc_pools_alloc();
 746         if (page_pools.epp_pools == NULL)
 747                 return -ENOMEM;
 748
 749         pools_shrinker = cfs_set_shrinker(pools_shrinker_seeks,
 750                                           enc_pools_shrink);
 751         if (pools_shrinker == NULL) {
 752                 enc_pools_free();
 753                 return -ENOMEM;
 754         }
 755
 756         return 0;
 757 }
 758
 759 void sptlrpc_enc_pool_fini(void)
 760 {
 761         unsigned long cleaned, npools;
 762
 763         LASSERT(pools_shrinker);
 764         LASSERT(page_pools.epp_pools);
 765         LASSERT(page_pools.epp_total_pages == page_pools.epp_free_pages);
 766
 767         cfs_remove_shrinker(pools_shrinker);
 768
 769         npools = npages_to_npools(page_pools.epp_total_pages);
 770         cleaned = enc_pools_cleanup(page_pools.epp_pools, npools);
 771         LASSERT(cleaned == page_pools.epp_total_pages);
 772
 773         enc_pools_free();
 774
 775         if (page_pools.epp_st_access > 0) {
 776                 CWARN("max pages %lu, grows %u, grow fails %u, shrinks %u, "
 777                       "access %lu, missing %lu, max qlen %u, max wait "
 778                       CFS_TIME_T"/%d\n",
 779                       page_pools.epp_st_max_pages, page_pools.epp_st_grows,
 780                       page_pools.epp_st_grow_fails,
 781                       page_pools.epp_st_shrinks, page_pools.epp_st_access,
 782                       page_pools.epp_st_missings, page_pools.epp_st_max_wqlen,
 783                       page_pools.epp_st_max_wait, CFS_HZ);
 784         }
 785 }
 786
 787 #else /* !__KERNEL__ */
 788
 789 int sptlrpc_enc_pool_get_pages(struct ptlrpc_bulk_desc *desc)
 790 {
 791         return 0;
 792 }
 793
 794 void sptlrpc_enc_pool_put_pages(struct ptlrpc_bulk_desc *desc)
 795 {
 796 }
 797
 798 int sptlrpc_enc_pool_init(void)
 799 {
 800         return 0;
 801 }
 802
 803 void sptlrpc_enc_pool_fini(void)
 804 {
 805 }
 806 #endif
 807
 808 /****************************************
 809  * Helpers to assist policy modules to  *
 810  * implement checksum funcationality    *
 811  ****************************************/
 812
 813 static struct sptlrpc_hash_type hash_types[] = {
 814         [BULK_HASH_ALG_NULL]    = { "null",     "null",         0 },
 815         [BULK_HASH_ALG_ADLER32] = { "adler32",  "adler32",      4 },
 816         [BULK_HASH_ALG_CRC32]   = { "crc32",    "crc32",        4 },
 817         [BULK_HASH_ALG_MD5]     = { "md5",      "md5",          16 },
 818         [BULK_HASH_ALG_SHA1]    = { "sha1",     "sha1",         20 },
 819         [BULK_HASH_ALG_SHA256]  = { "sha256",   "sha256",       32 },
 820         [BULK_HASH_ALG_SHA384]  = { "sha384",   "sha384",       48 },
 821         [BULK_HASH_ALG_SHA512]  = { "sha512",   "sha512",       64 },
 822 };
 823
 824 const struct sptlrpc_hash_type *sptlrpc_get_hash_type(__u8 hash_alg)
 825 {
 826         struct sptlrpc_hash_type *ht;
 827
 828         if (hash_alg < BULK_HASH_ALG_MAX) {
 829                 ht = &hash_types[hash_alg];
 830                 if (ht->sht_tfm_name)
 831                         return ht;
 832         }
 833         return NULL;
 834 }
 835 EXPORT_SYMBOL(sptlrpc_get_hash_type);
 836
 837 const char * sptlrpc_get_hash_name(__u8 hash_alg)
 838 {
 839         const struct sptlrpc_hash_type *ht;
 840
 841         ht = sptlrpc_get_hash_type(hash_alg);
 842         if (ht)
 843                 return ht->sht_name;
 844         else
 845                 return "unknown";
 846 }
 847 EXPORT_SYMBOL(sptlrpc_get_hash_name);
 848
 849 __u8 sptlrpc_get_hash_alg(const char *algname)
 850 {
 851         int     i;
 852
 853         for (i = 0; i < BULK_HASH_ALG_MAX; i++)
 854                 if (!strcmp(hash_types[i].sht_name, algname))
 855                         break;
 856         return i;
 857 }
 858 EXPORT_SYMBOL(sptlrpc_get_hash_alg);
 859
 860 int bulk_sec_desc_unpack(struct lustre_msg *msg, int offset, int swabbed)
 861 {
 862         struct ptlrpc_bulk_sec_desc *bsd;
 863         int                          size = msg->lm_buflens[offset];
 864
 865         bsd = lustre_msg_buf(msg, offset, sizeof(*bsd));
 866         if (bsd == NULL) {
 867                 CERROR("Invalid bulk sec desc: size %d\n", size);
 868                 return -EINVAL;
 869         }
 870
 871         if (swabbed) {
 872                 __swab32s(&bsd->bsd_nob);
 873         }
 874
 875         if (unlikely(bsd->bsd_version != 0)) {
 876                 CERROR("Unexpected version %u\n", bsd->bsd_version);
 877                 return -EPROTO;
 878         }
 879
 880         if (unlikely(bsd->bsd_type >= SPTLRPC_BULK_MAX)) {
 881                 CERROR("Invalid type %u\n", bsd->bsd_type);
 882                 return -EPROTO;
 883         }
 884
 885         /* FIXME more sanity check here */
 886
 887         if (unlikely(bsd->bsd_svc != SPTLRPC_BULK_SVC_NULL &&
 888                      bsd->bsd_svc != SPTLRPC_BULK_SVC_INTG &&
 889                      bsd->bsd_svc != SPTLRPC_BULK_SVC_PRIV)) {
 890                 CERROR("Invalid svc %u\n", bsd->bsd_svc);
 891                 return -EPROTO;
 892         }
 893
 894         return 0;
 895 }
 896 EXPORT_SYMBOL(bulk_sec_desc_unpack);
 897
 898 #ifdef __KERNEL__
 899
 900 #ifdef HAVE_ADLER
 901 static int do_bulk_checksum_adler32(struct ptlrpc_bulk_desc *desc, void *buf)
 902 {
 903         struct page    *page;
 904         int             off;
 905         char           *ptr;
 906         __u32           adler32 = 1;
 907         int             len, i;
 908
 909         for (i = 0; i < desc->bd_iov_count; i++) {
 910                 page = desc->bd_iov[i].kiov_page;
 911                 off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK;
 912                 ptr = cfs_kmap(page) + off;
 913                 len = desc->bd_iov[i].kiov_len;
 914
 915                 adler32 = adler32(adler32, ptr, len);
 916
 917                 cfs_kunmap(page);
 918         }
 919
 920         adler32 = cpu_to_le32(adler32);
 921         memcpy(buf, &adler32, sizeof(adler32));
 922         return 0;
 923 }
 924 #endif
 925
 926 static int do_bulk_checksum_crc32(struct ptlrpc_bulk_desc *desc, void *buf)
 927 {
 928         struct page    *page;
 929         int             off;
 930         char           *ptr;
 931         __u32           crc32 = ~0;
 932         int             len, i;
 933
 934         for (i = 0; i < desc->bd_iov_count; i++) {
 935                 page = desc->bd_iov[i].kiov_page;
 936                 off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK;
 937                 ptr = cfs_kmap(page) + off;
 938                 len = desc->bd_iov[i].kiov_len;
 939
 940                 crc32 = crc32_le(crc32, ptr, len);
 941
 942                 cfs_kunmap(page);
 943         }
 944
 945         crc32 = cpu_to_le32(crc32);
 946         memcpy(buf, &crc32, sizeof(crc32));
 947         return 0;
 948 }
 949
 950 int sptlrpc_get_bulk_checksum(struct ptlrpc_bulk_desc *desc, __u8 alg,
 951                               void *buf, int buflen)
 952 {
 953         struct hash_desc    hdesc;
 954         int                 hashsize;
 955         char                hashbuf[64];
 956         struct scatterlist  sl;
 957         int                 i;
 958
 959         LASSERT(alg > BULK_HASH_ALG_NULL && alg < BULK_HASH_ALG_MAX);
 960         LASSERT(buflen >= 4);
 961
 962         switch (alg) {
 963         case BULK_HASH_ALG_ADLER32:
 964 #ifdef HAVE_ADLER
 965                 return do_bulk_checksum_adler32(desc, buf);
 966 #else
 967                 CERROR("Adler32 not supported\n");
 968                 return -EINVAL;
 969 #endif
 970         case BULK_HASH_ALG_CRC32:
 971                 return do_bulk_checksum_crc32(desc, buf);
 972         }
 973
 974         hdesc.tfm = ll_crypto_alloc_hash(hash_types[alg].sht_tfm_name, 0, 0);
 975         if (hdesc.tfm == NULL) {
 976                 CERROR("Unable to allocate TFM %s\n", hash_types[alg].sht_name);
 977                 return -ENOMEM;
 978         }
 979
 980         hdesc.flags = 0;
 981         ll_crypto_hash_init(&hdesc);
 982
 983         hashsize = ll_crypto_hash_digestsize(hdesc.tfm);
 984
 985         for (i = 0; i < desc->bd_iov_count; i++) {
 986                 sg_set_page(&sl, desc->bd_iov[i].kiov_page,
 987                              desc->bd_iov[i].kiov_len,
 988                              desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK);
 989                 ll_crypto_hash_update(&hdesc, &sl, sl.length);
 990         }
 991
 992         if (hashsize > buflen) {
 993                 ll_crypto_hash_final(&hdesc, hashbuf);
 994                 memcpy(buf, hashbuf, buflen);
 995         } else {
 996                 ll_crypto_hash_final(&hdesc, buf);
 997         }
 998
 999         ll_crypto_free_hash(hdesc.tfm);
1000         return 0;
1001 }
1002 EXPORT_SYMBOL(sptlrpc_get_bulk_checksum);
1003
1004 #else /* !__KERNEL__ */
1005
1006 int sptlrpc_get_bulk_checksum(struct ptlrpc_bulk_desc *desc, __u8 alg,
1007                               void *buf, int buflen)
1008 {
1009         __u32   csum32;
1010         int     i;
1011
1012         LASSERT(alg == BULK_HASH_ALG_ADLER32 || alg == BULK_HASH_ALG_CRC32);
1013
1014         if (alg == BULK_HASH_ALG_ADLER32)
1015                 csum32 = 1;
1016         else
1017                 csum32 = ~0;
1018
1019         for (i = 0; i < desc->bd_iov_count; i++) {
1020                 unsigned char *ptr = desc->bd_iov[i].iov_base;
1021                 int len = desc->bd_iov[i].iov_len;
1022
1023                 switch (alg) {
1024                 case BULK_HASH_ALG_ADLER32:
1025 #ifdef HAVE_ADLER
1026                         csum32 = adler32(csum32, ptr, len);
1027 #else
1028                         CERROR("Adler32 not supported\n");
1029                         return -EINVAL;
1030 #endif
1031                         break;
1032                 case BULK_HASH_ALG_CRC32:
1033                         csum32 = crc32_le(csum32, ptr, len);
1034                         break;
1035                 }
1036         }
1037
1038         csum32 = cpu_to_le32(csum32);
1039         memcpy(buf, &csum32, sizeof(csum32));
1040         return 0;
1041 }
1042
1043 #endif /* __KERNEL__ */