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  * Copyright (C) 2006-2007 Cluster File Systems, Inc.
   5  *   Author: Eric Mei <ericm@clusterfs.com>
   6  *
   7  *   This file is part of Lustre, http://www.lustre.org.
   8  *
   9  *   Lustre is free software; you can redistribute it and/or
  10  *   modify it under the terms of version 2 of the GNU General Public
  11  *   License as published by the Free Software Foundation.
  12  *
  13  *   Lustre is distributed in the hope that it will be useful,
  14  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  *   GNU General Public License for more details.
  17  *
  18  *   You should have received a copy of the GNU General Public License
  19  *   along with Lustre; if not, write to the Free Software
  20  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  21  */
  22
  23 #ifndef EXPORT_SYMTAB
  24 #define EXPORT_SYMTAB
  25 #endif
  26 #define DEBUG_SUBSYSTEM S_SEC
  27
  28 #include <libcfs/libcfs.h>
  29 #ifndef __KERNEL__
  30 #include <liblustre.h>
  31 #include <libcfs/list.h>
  32 #else
  33 #include <linux/crypto.h>
  34 #endif
  35
  36 #include <obd.h>
  37 #include <obd_class.h>
  38 #include <obd_support.h>
  39 #include <lustre_net.h>
  40 #include <lustre_import.h>
  41 #include <lustre_dlm.h>
  42 #include <lustre_sec.h>
  43
  44 #include "ptlrpc_internal.h"
  45
  46 /****************************************
  47  * bulk encryption page pools           *
  48  ****************************************/
  49
  50 #ifdef __KERNEL__
  51
  52 #define PTRS_PER_PAGE   (CFS_PAGE_SIZE / sizeof(void *))
  53 #define PAGES_PER_POOL  (PTRS_PER_PAGE)
  54
  55 #define IDLE_IDX_MAX            (100)
  56 #define IDLE_IDX_WEIGHT         (3)
  57
  58 #define CACHE_QUIESCENCE_PERIOD (20)
  59
  60 static struct ptlrpc_enc_page_pool {
  61         /*
  62          * constants
  63          */
  64         unsigned long    epp_max_pages;   /* maximum pages can hold, const */
  65         unsigned int     epp_max_pools;   /* number of pools, const */
  66
  67         /*
  68          * wait queue in case of not enough free pages.
  69          */
  70         cfs_waitq_t      epp_waitq;       /* waiting threads */
  71         unsigned int     epp_waitqlen;    /* wait queue length */
  72         unsigned long    epp_pages_short; /* # of pages wanted of in-q users */
  73         unsigned int     epp_growing:1;   /* during adding pages */
  74
  75         /*
  76          * indicating how idle the pools are, from 0 to MAX_IDLE_IDX
  77          * this is counted based on each time when getting pages from
  78          * the pools, not based on time. which means in case that system
  79          * is idled for a while but the idle_idx might still be low if no
  80          * activities happened in the pools.
  81          */
  82         unsigned long    epp_idle_idx;
  83
  84         /* last shrink time due to mem tight */
  85         long             epp_last_shrink;
  86         long             epp_last_access;
  87
  88         /*
  89          * in-pool pages bookkeeping
  90          */
  91         spinlock_t       epp_lock;        /* protect following fields */
  92         unsigned long    epp_total_pages; /* total pages in pools */
  93         unsigned long    epp_free_pages;  /* current pages available */
  94
  95         /*
  96          * statistics
  97          */
  98         unsigned int     epp_st_grows;          /* # of grows */
  99         unsigned int     epp_st_grow_fails;     /* # of add pages failures */
 100         unsigned int     epp_st_shrinks;        /* # of shrinks */
 101         unsigned long    epp_st_access;         /* # of access */
 102         unsigned long    epp_st_missings;       /* # of cache missing */
 103         unsigned long    epp_st_lowfree;        /* lowest free pages reached */
 104         unsigned long    epp_st_max_wqlen;      /* highest waitqueue length */
 105         cfs_time_t       epp_st_max_wait;       /* in jeffies */
 106         /*
 107          * pointers to pools
 108          */
 109         cfs_page_t    ***epp_pools;
 110 } page_pools;
 111
 112 /*
 113  * memory shrinker
 114  */
 115 const int pools_shrinker_seeks = DEFAULT_SEEKS;
 116 static struct shrinker *pools_shrinker = NULL;
 117
 118
 119 /*
 120  * /proc/fs/lustre/sptlrpc/encrypt_page_pools
 121  */
 122 int sptlrpc_proc_read_enc_pool(char *page, char **start, off_t off, int count,
 123                                int *eof, void *data)
 124 {
 125         int     rc;
 126
 127         spin_lock(&page_pools.epp_lock);
 128
 129         rc = snprintf(page, count,
 130                       "physical pages:          %lu\n"
 131                       "pages per pool:          %lu\n"
 132                       "max pages:               %lu\n"
 133                       "max pools:               %u\n"
 134                       "total pages:             %lu\n"
 135                       "total free:              %lu\n"
 136                       "idle index:              %lu/100\n"
 137                       "last shrink:             %lds\n"
 138                       "last access:             %lds\n"
 139                       "grows:                   %u\n"
 140                       "grows failure:           %u\n"
 141                       "shrinks:                 %u\n"
 142                       "cache access:            %lu\n"
 143                       "cache missing:           %lu\n"
 144                       "low free mark:           %lu\n"
 145                       "max waitqueue depth:     %lu\n"
 146                       "max wait time:           "CFS_TIME_T"/%u\n"
 147                       ,
 148                       num_physpages,
 149                       PAGES_PER_POOL,
 150                       page_pools.epp_max_pages,
 151                       page_pools.epp_max_pools,
 152                       page_pools.epp_total_pages,
 153                       page_pools.epp_free_pages,
 154                       page_pools.epp_idle_idx,
 155                       cfs_time_current_sec() - page_pools.epp_last_shrink,
 156                       cfs_time_current_sec() - page_pools.epp_last_access,
 157                       page_pools.epp_st_grows,
 158                       page_pools.epp_st_grow_fails,
 159                       page_pools.epp_st_shrinks,
 160                       page_pools.epp_st_access,
 161                       page_pools.epp_st_missings,
 162                       page_pools.epp_st_lowfree,
 163                       page_pools.epp_st_max_wqlen,
 164                       page_pools.epp_st_max_wait, HZ
 165                      );
 166
 167         spin_unlock(&page_pools.epp_lock);
 168         return rc;
 169 }
 170
 171 static void enc_pools_release_free_pages(long npages)
 172 {
 173         int     p_idx, g_idx;
 174
 175         LASSERT(npages <= page_pools.epp_free_pages);
 176
 177         p_idx = (page_pools.epp_free_pages - 1) / PAGES_PER_POOL;
 178         g_idx = (page_pools.epp_free_pages - 1) % PAGES_PER_POOL;
 179         LASSERT(page_pools.epp_pools[p_idx]);
 180
 181         page_pools.epp_free_pages -= npages;
 182         page_pools.epp_total_pages -= npages;
 183
 184         while (npages-- > 0) {
 185                 LASSERT(page_pools.epp_pools[p_idx][g_idx] != NULL);
 186
 187                 cfs_free_page(page_pools.epp_pools[p_idx][g_idx]);
 188                 page_pools.epp_pools[p_idx][g_idx] = NULL;
 189
 190                 if (g_idx-- == 0) {
 191                         p_idx--;
 192                         g_idx = PAGES_PER_POOL - 1;
 193
 194                         LASSERT(page_pools.epp_pools[p_idx]);
 195                 }
 196         }
 197 }
 198
 199 /*
 200  * could be called frequently for query (@nr_to_scan == 0)
 201  */
 202 static int enc_pools_shrink(int nr_to_scan, unsigned int gfp_mask)
 203 {
 204         unsigned long   ret;
 205
 206         spin_lock(&page_pools.epp_lock);
 207
 208         if (nr_to_scan) {
 209                 if (nr_to_scan > page_pools.epp_free_pages)
 210                         nr_to_scan = page_pools.epp_free_pages;
 211
 212                 enc_pools_release_free_pages(nr_to_scan);
 213                 CDEBUG(D_SEC, "released %d pages, %ld left\n",
 214                        nr_to_scan, page_pools.epp_free_pages);
 215
 216                 page_pools.epp_st_shrinks++;
 217                 page_pools.epp_last_shrink = cfs_time_current_sec();
 218         }
 219
 220         /*
 221          * try to keep at least PTLRPC_MAX_BRW_PAGES pages in the pool
 222          */
 223         if (page_pools.epp_free_pages <= PTLRPC_MAX_BRW_PAGES) {
 224                 ret = 0;
 225                 goto out_unlock;
 226         }
 227
 228         /*
 229          * if no pool access for a long time, we consider it's fully idle
 230          */
 231         if (cfs_time_current_sec() - page_pools.epp_last_access >
 232             CACHE_QUIESCENCE_PERIOD)
 233                 page_pools.epp_idle_idx = IDLE_IDX_MAX;
 234
 235         LASSERT(page_pools.epp_idle_idx <= IDLE_IDX_MAX);
 236         ret = (page_pools.epp_free_pages * page_pools.epp_idle_idx /
 237                IDLE_IDX_MAX);
 238         if (page_pools.epp_free_pages - ret < PTLRPC_MAX_BRW_PAGES)
 239                 ret = page_pools.epp_free_pages - PTLRPC_MAX_BRW_PAGES;
 240
 241 out_unlock:
 242         spin_unlock(&page_pools.epp_lock);
 243         return ret;
 244 }
 245
 246 static inline
 247 int npages_to_npools(unsigned long npages)
 248 {
 249         return (int) ((npages + PAGES_PER_POOL - 1) / PAGES_PER_POOL);
 250 }
 251
 252 /*
 253  * return how many pages cleaned up.
 254  */
 255 static unsigned long enc_pools_cleanup(cfs_page_t ***pools, int npools)
 256 {
 257         unsigned long cleaned = 0;
 258         int           i, j;
 259
 260         for (i = 0; i < npools; i++) {
 261                 if (pools[i]) {
 262                         for (j = 0; j < PAGES_PER_POOL; j++) {
 263                                 if (pools[i][j]) {
 264                                         cfs_free_page(pools[i][j]);
 265                                         cleaned++;
 266                                 }
 267                         }
 268                         OBD_FREE(pools[i], CFS_PAGE_SIZE);
 269                         pools[i] = NULL;
 270                 }
 271         }
 272
 273         return cleaned;
 274 }
 275
 276 /*
 277  * merge @npools pointed by @pools which contains @npages new pages
 278  * into current pools.
 279  *
 280  * we have options to avoid most memory copy with some tricks. but we choose
 281  * the simplest way to avoid complexity. It's not frequently called.
 282  */
 283 static void enc_pools_insert(cfs_page_t ***pools, int npools, int npages)
 284 {
 285         int     freeslot;
 286         int     op_idx, np_idx, og_idx, ng_idx;
 287         int     cur_npools, end_npools;
 288
 289         LASSERT(npages > 0);
 290         LASSERT(page_pools.epp_total_pages+npages <= page_pools.epp_max_pages);
 291         LASSERT(npages_to_npools(npages) == npools);
 292
 293         spin_lock(&page_pools.epp_lock);
 294
 295         /*
 296          * (1) fill all the free slots of current pools.
 297          */
 298         /* free slots are those left by rent pages, and the extra ones with
 299          * index >= eep_total_pages, locate at the tail of last pool. */
 300         freeslot = page_pools.epp_total_pages % PAGES_PER_POOL;
 301         if (freeslot != 0)
 302                 freeslot = PAGES_PER_POOL - freeslot;
 303         freeslot += page_pools.epp_total_pages - page_pools.epp_free_pages;
 304
 305         op_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 306         og_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 307         np_idx = npools - 1;
 308         ng_idx = (npages - 1) % PAGES_PER_POOL;
 309
 310         while (freeslot) {
 311                 LASSERT(page_pools.epp_pools[op_idx][og_idx] == NULL);
 312                 LASSERT(pools[np_idx][ng_idx] != NULL);
 313
 314                 page_pools.epp_pools[op_idx][og_idx] = pools[np_idx][ng_idx];
 315                 pools[np_idx][ng_idx] = NULL;
 316
 317                 freeslot--;
 318
 319                 if (++og_idx == PAGES_PER_POOL) {
 320                         op_idx++;
 321                         og_idx = 0;
 322                 }
 323                 if (--ng_idx < 0) {
 324                         if (np_idx == 0)
 325                                 break;
 326                         np_idx--;
 327                         ng_idx = PAGES_PER_POOL - 1;
 328                 }
 329         }
 330
 331         /*
 332          * (2) add pools if needed.
 333          */
 334         cur_npools = (page_pools.epp_total_pages + PAGES_PER_POOL - 1) /
 335                      PAGES_PER_POOL;
 336         end_npools = (page_pools.epp_total_pages + npages + PAGES_PER_POOL -1) /
 337                      PAGES_PER_POOL;
 338         LASSERT(end_npools <= page_pools.epp_max_pools);
 339
 340         np_idx = 0;
 341         while (cur_npools < end_npools) {
 342                 LASSERT(page_pools.epp_pools[cur_npools] == NULL);
 343                 LASSERT(np_idx < npools);
 344                 LASSERT(pools[np_idx] != NULL);
 345
 346                 page_pools.epp_pools[cur_npools++] = pools[np_idx];
 347                 pools[np_idx++] = NULL;
 348         }
 349
 350         page_pools.epp_total_pages += npages;
 351         page_pools.epp_free_pages += npages;
 352         page_pools.epp_st_lowfree = page_pools.epp_free_pages;
 353
 354         CDEBUG(D_SEC, "add %d pages to total %lu\n", npages,
 355                page_pools.epp_total_pages);
 356
 357         spin_unlock(&page_pools.epp_lock);
 358 }
 359
 360 static int enc_pools_add_pages(int npages)
 361 {
 362         static DECLARE_MUTEX(sem_add_pages);
 363         cfs_page_t   ***pools;
 364         int             npools, alloced = 0;
 365         int             i, j, rc = -ENOMEM;
 366
 367         if (npages < PTLRPC_MAX_BRW_PAGES)
 368                 npages = PTLRPC_MAX_BRW_PAGES;
 369
 370         down(&sem_add_pages);
 371
 372         if (npages + page_pools.epp_total_pages > page_pools.epp_max_pages)
 373                 npages = page_pools.epp_max_pages - page_pools.epp_total_pages;
 374         LASSERT(npages > 0);
 375
 376         page_pools.epp_st_grows++;
 377
 378         npools = npages_to_npools(npages);
 379         OBD_ALLOC(pools, npools * sizeof(*pools));
 380         if (pools == NULL)
 381                 goto out;
 382
 383         for (i = 0; i < npools; i++) {
 384                 OBD_ALLOC(pools[i], CFS_PAGE_SIZE);
 385                 if (pools[i] == NULL)
 386                         goto out_pools;
 387
 388                 for (j = 0; j < PAGES_PER_POOL && alloced < npages; j++) {
 389                         pools[i][j] = cfs_alloc_page(CFS_ALLOC_IO |
 390                                                      CFS_ALLOC_HIGH);
 391                         if (pools[i][j] == NULL)
 392                                 goto out_pools;
 393
 394                         alloced++;
 395                 }
 396         }
 397
 398         enc_pools_insert(pools, npools, npages);
 399         CDEBUG(D_SEC, "added %d pages into pools\n", npages);
 400         rc = 0;
 401
 402 out_pools:
 403         enc_pools_cleanup(pools, npools);
 404         OBD_FREE(pools, npools * sizeof(*pools));
 405 out:
 406         if (rc) {
 407                 page_pools.epp_st_grow_fails++;
 408                 CERROR("Failed to allocate %d enc pages\n", npages);
 409         }
 410
 411         up(&sem_add_pages);
 412         return rc;
 413 }
 414
 415 static inline void enc_pools_wakeup(void)
 416 {
 417         if (unlikely(page_pools.epp_waitqlen)) {
 418                 LASSERT(page_pools.epp_waitqlen > 0);
 419                 LASSERT(cfs_waitq_active(&page_pools.epp_waitq));
 420                 cfs_waitq_broadcast(&page_pools.epp_waitq);
 421         }
 422 }
 423
 424 static int enc_pools_should_grow(int page_needed, long now)
 425 {
 426         /* don't grow if someone else is growing the pools right now,
 427          * or the pools has reached its full capacity
 428          */
 429         if (page_pools.epp_growing ||
 430             page_pools.epp_total_pages == page_pools.epp_max_pages)
 431                 return 0;
 432
 433         /* if total pages is not enough, we need to grow */
 434         if (page_pools.epp_total_pages < page_needed)
 435                 return 1;
 436
 437         /* if we just did a shrink due to memory tight, we'd better
 438          * wait a while to grow again.
 439          */
 440         if (now - page_pools.epp_last_shrink < 2)
 441                 return 0;
 442
 443         /*
 444          * here we perhaps need consider other factors like wait queue
 445          * length, idle index, etc. ?
 446          */
 447
 448         /* grow the pools in any other cases */
 449         return 1;
 450 }
 451
 452 /*
 453  * we allocate the requested pages atomically.
 454  */
 455 int sptlrpc_enc_pool_get_pages(struct ptlrpc_bulk_desc *desc)
 456 {
 457         cfs_waitlink_t  waitlink;
 458         unsigned long   this_idle = -1;
 459         cfs_time_t      tick = 0;
 460         long            now;
 461         int             p_idx, g_idx;
 462         int             i;
 463
 464         LASSERT(desc->bd_max_iov > 0);
 465         LASSERT(desc->bd_max_iov <= page_pools.epp_max_pages);
 466
 467         /* resent bulk, enc pages might have been allocated previously */
 468         if (desc->bd_enc_pages != NULL)
 469                 return 0;
 470
 471         OBD_ALLOC(desc->bd_enc_pages,
 472                   desc->bd_max_iov * sizeof(*desc->bd_enc_pages));
 473         if (desc->bd_enc_pages == NULL)
 474                 return -ENOMEM;
 475
 476         spin_lock(&page_pools.epp_lock);
 477
 478         page_pools.epp_st_access++;
 479 again:
 480         if (unlikely(page_pools.epp_free_pages < desc->bd_max_iov)) {
 481                 if (tick == 0)
 482                         tick = cfs_time_current();
 483
 484                 now = cfs_time_current_sec();
 485
 486                 page_pools.epp_st_missings++;
 487                 page_pools.epp_pages_short += desc->bd_max_iov;
 488
 489                 if (enc_pools_should_grow(desc->bd_max_iov, now)) {
 490                         page_pools.epp_growing = 1;
 491
 492                         spin_unlock(&page_pools.epp_lock);
 493                         enc_pools_add_pages(page_pools.epp_pages_short / 2);
 494                         spin_lock(&page_pools.epp_lock);
 495
 496                         page_pools.epp_growing = 0;
 497                 } else {
 498                         if (++page_pools.epp_waitqlen >
 499                             page_pools.epp_st_max_wqlen)
 500                                 page_pools.epp_st_max_wqlen =
 501                                                 page_pools.epp_waitqlen;
 502
 503                         set_current_state(TASK_UNINTERRUPTIBLE);
 504                         cfs_waitlink_init(&waitlink);
 505                         cfs_waitq_add(&page_pools.epp_waitq, &waitlink);
 506
 507                         spin_unlock(&page_pools.epp_lock);
 508                         cfs_schedule();
 509                         spin_lock(&page_pools.epp_lock);
 510
 511                         LASSERT(page_pools.epp_waitqlen > 0);
 512                         page_pools.epp_waitqlen--;
 513                 }
 514
 515                 LASSERT(page_pools.epp_pages_short >= desc->bd_max_iov);
 516                 page_pools.epp_pages_short -= desc->bd_max_iov;
 517
 518                 this_idle = 0;
 519                 goto again;
 520         }
 521
 522         /* record max wait time */
 523         if (unlikely(tick != 0)) {
 524                 tick = cfs_time_current() - tick;
 525                 if (tick > page_pools.epp_st_max_wait)
 526                         page_pools.epp_st_max_wait = tick;
 527         }
 528
 529         /* proceed with rest of allocation */
 530         page_pools.epp_free_pages -= desc->bd_max_iov;
 531
 532         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 533         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 534
 535         for (i = 0; i < desc->bd_max_iov; i++) {
 536                 LASSERT(page_pools.epp_pools[p_idx][g_idx] != NULL);
 537                 desc->bd_enc_pages[i] = page_pools.epp_pools[p_idx][g_idx];
 538                 page_pools.epp_pools[p_idx][g_idx] = NULL;
 539
 540                 if (++g_idx == PAGES_PER_POOL) {
 541                         p_idx++;
 542                         g_idx = 0;
 543                 }
 544         }
 545
 546         if (page_pools.epp_free_pages < page_pools.epp_st_lowfree)
 547                 page_pools.epp_st_lowfree = page_pools.epp_free_pages;
 548
 549         /*
 550          * new idle index = (old * weight + new) / (weight + 1)
 551          */
 552         if (this_idle == -1) {
 553                 this_idle = page_pools.epp_free_pages * IDLE_IDX_MAX /
 554                             page_pools.epp_total_pages;
 555         }
 556         page_pools.epp_idle_idx = (page_pools.epp_idle_idx * IDLE_IDX_WEIGHT +
 557                                    this_idle) /
 558                                   (IDLE_IDX_WEIGHT + 1);
 559
 560         page_pools.epp_last_access = cfs_time_current_sec();
 561
 562         spin_unlock(&page_pools.epp_lock);
 563         return 0;
 564 }
 565 EXPORT_SYMBOL(sptlrpc_enc_pool_get_pages);
 566
 567 void sptlrpc_enc_pool_put_pages(struct ptlrpc_bulk_desc *desc)
 568 {
 569         int     p_idx, g_idx;
 570         int     i;
 571
 572         if (desc->bd_enc_pages == NULL)
 573                 return;
 574         if (desc->bd_max_iov == 0)
 575                 return;
 576
 577         spin_lock(&page_pools.epp_lock);
 578
 579         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 580         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 581
 582         LASSERT(page_pools.epp_free_pages + desc->bd_max_iov <=
 583                 page_pools.epp_total_pages);
 584         LASSERT(page_pools.epp_pools[p_idx]);
 585
 586         for (i = 0; i < desc->bd_max_iov; i++) {
 587                 LASSERT(desc->bd_enc_pages[i] != NULL);
 588                 LASSERT(g_idx != 0 || page_pools.epp_pools[p_idx]);
 589                 LASSERT(page_pools.epp_pools[p_idx][g_idx] == NULL);
 590
 591                 page_pools.epp_pools[p_idx][g_idx] = desc->bd_enc_pages[i];
 592
 593                 if (++g_idx == PAGES_PER_POOL) {
 594                         p_idx++;
 595                         g_idx = 0;
 596                 }
 597         }
 598
 599         page_pools.epp_free_pages += desc->bd_max_iov;
 600
 601         enc_pools_wakeup();
 602
 603         spin_unlock(&page_pools.epp_lock);
 604
 605         OBD_FREE(desc->bd_enc_pages,
 606                  desc->bd_max_iov * sizeof(*desc->bd_enc_pages));
 607         desc->bd_enc_pages = NULL;
 608 }
 609 EXPORT_SYMBOL(sptlrpc_enc_pool_put_pages);
 610
 611 /*
 612  * we don't do much stuff for add_user/del_user anymore, except adding some
 613  * initial pages in add_user() if current pools are empty, rest would be
 614  * handled by the pools's self-adaption.
 615  */
 616 int sptlrpc_enc_pool_add_user(void)
 617 {
 618         int     need_grow = 0;
 619
 620         spin_lock(&page_pools.epp_lock);
 621         if (page_pools.epp_growing == 0 && page_pools.epp_total_pages == 0) {
 622                 page_pools.epp_growing = 1;
 623                 need_grow = 1;
 624         }
 625         spin_unlock(&page_pools.epp_lock);
 626
 627         if (need_grow) {
 628                 enc_pools_add_pages(PTLRPC_MAX_BRW_PAGES);
 629
 630                 spin_lock(&page_pools.epp_lock);
 631                 page_pools.epp_growing = 0;
 632                 enc_pools_wakeup();
 633                 spin_unlock(&page_pools.epp_lock);
 634         }
 635         return 0;
 636 }
 637 EXPORT_SYMBOL(sptlrpc_enc_pool_add_user);
 638
 639 int sptlrpc_enc_pool_del_user(void)
 640 {
 641         return 0;
 642 }
 643 EXPORT_SYMBOL(sptlrpc_enc_pool_del_user);
 644
 645 static inline void enc_pools_alloc(void)
 646 {
 647         LASSERT(page_pools.epp_max_pools);
 648         /*
 649          * on system with huge memory but small page size, this might lead to
 650          * high-order allocation. but it's not common, and we suppose memory
 651          * be not too much fragmented at module loading time.
 652          */
 653         OBD_ALLOC(page_pools.epp_pools,
 654                   page_pools.epp_max_pools * sizeof(*page_pools.epp_pools));
 655 }
 656
 657 static inline void enc_pools_free(void)
 658 {
 659         LASSERT(page_pools.epp_max_pools);
 660         LASSERT(page_pools.epp_pools);
 661
 662         OBD_FREE(page_pools.epp_pools,
 663                  page_pools.epp_max_pools * sizeof(*page_pools.epp_pools));
 664 }
 665
 666 int sptlrpc_enc_pool_init(void)
 667 {
 668         /*
 669          * maximum capacity is 1/8 of total physical memory.
 670          * is the 1/8 a good number?
 671          */
 672         page_pools.epp_max_pages = num_physpages / 8;
 673         page_pools.epp_max_pools = npages_to_npools(page_pools.epp_max_pages);
 674
 675         cfs_waitq_init(&page_pools.epp_waitq);
 676         page_pools.epp_waitqlen = 0;
 677         page_pools.epp_pages_short = 0;
 678
 679         page_pools.epp_growing = 0;
 680
 681         page_pools.epp_idle_idx = 0;
 682         page_pools.epp_last_shrink = cfs_time_current_sec();
 683         page_pools.epp_last_access = cfs_time_current_sec();
 684
 685         spin_lock_init(&page_pools.epp_lock);
 686         page_pools.epp_total_pages = 0;
 687         page_pools.epp_free_pages = 0;
 688
 689         page_pools.epp_st_grows = 0;
 690         page_pools.epp_st_grow_fails = 0;
 691         page_pools.epp_st_shrinks = 0;
 692         page_pools.epp_st_access = 0;
 693         page_pools.epp_st_missings = 0;
 694         page_pools.epp_st_lowfree = 0;
 695         page_pools.epp_st_max_wqlen = 0;
 696         page_pools.epp_st_max_wait = 0;
 697
 698         enc_pools_alloc();
 699         if (page_pools.epp_pools == NULL)
 700                 return -ENOMEM;
 701
 702         pools_shrinker = set_shrinker(pools_shrinker_seeks, enc_pools_shrink);
 703         if (pools_shrinker == NULL) {
 704                 enc_pools_free();
 705                 return -ENOMEM;
 706         }
 707
 708         return 0;
 709 }
 710
 711 void sptlrpc_enc_pool_fini(void)
 712 {
 713         unsigned long cleaned, npools;
 714
 715         LASSERT(pools_shrinker);
 716         LASSERT(page_pools.epp_pools);
 717         LASSERT(page_pools.epp_total_pages == page_pools.epp_free_pages);
 718
 719         remove_shrinker(pools_shrinker);
 720
 721         npools = npages_to_npools(page_pools.epp_total_pages);
 722         cleaned = enc_pools_cleanup(page_pools.epp_pools, npools);
 723         LASSERT(cleaned == page_pools.epp_total_pages);
 724
 725         enc_pools_free();
 726 }
 727
 728 #else /* !__KERNEL__ */
 729
 730 int sptlrpc_enc_pool_get_pages(struct ptlrpc_bulk_desc *desc)
 731 {
 732         return 0;
 733 }
 734
 735 void sptlrpc_enc_pool_put_pages(struct ptlrpc_bulk_desc *desc)
 736 {
 737 }
 738
 739 int sptlrpc_enc_pool_init(void)
 740 {
 741         return 0;
 742 }
 743
 744 void sptlrpc_enc_pool_fini(void)
 745 {
 746 }
 747 #endif
 748
 749 /****************************************
 750  * Helpers to assist policy modules to  *
 751  * implement checksum funcationality    *
 752  ****************************************/
 753
 754 static struct {
 755         char    *name;
 756         int      size;
 757 } csum_types[] = {
 758         [BULK_CSUM_ALG_NULL]    = { "null",     0 },
 759         [BULK_CSUM_ALG_CRC32]   = { "crc32",    4 },
 760         [BULK_CSUM_ALG_MD5]     = { "md5",     16 },
 761         [BULK_CSUM_ALG_SHA1]    = { "sha1",    20 },
 762         [BULK_CSUM_ALG_SHA256]  = { "sha256",  32 },
 763         [BULK_CSUM_ALG_SHA384]  = { "sha384",  48 },
 764         [BULK_CSUM_ALG_SHA512]  = { "sha512",  64 },
 765 };
 766
 767 const char * sptlrpc_bulk_csum_alg2name(__u8 csum_alg)
 768 {
 769         if (csum_alg < BULK_CSUM_ALG_MAX)
 770                 return csum_types[csum_alg].name;
 771         return "unknown";
 772 }
 773 EXPORT_SYMBOL(sptlrpc_bulk_csum_alg2name);
 774
 775 int bulk_sec_desc_size(__u8 csum_alg, int request, int read)
 776 {
 777         int size = sizeof(struct ptlrpc_bulk_sec_desc);
 778
 779         LASSERT(csum_alg < BULK_CSUM_ALG_MAX);
 780
 781         /* read request don't need extra data */
 782         if (!(read && request))
 783                 size += csum_types[csum_alg].size;
 784
 785         return size;
 786 }
 787 EXPORT_SYMBOL(bulk_sec_desc_size);
 788
 789 int bulk_sec_desc_unpack(struct lustre_msg *msg, int offset)
 790 {
 791         struct ptlrpc_bulk_sec_desc *bsd;
 792         int    size = msg->lm_buflens[offset];
 793
 794         bsd = lustre_msg_buf(msg, offset, sizeof(*bsd));
 795         if (bsd == NULL) {
 796                 CERROR("Invalid bulk sec desc: size %d\n", size);
 797                 return -EINVAL;
 798         }
 799
 800         if (lustre_msg_swabbed(msg)) {
 801                 __swab32s(&bsd->bsd_version);
 802                 __swab16s(&bsd->bsd_pad);
 803         }
 804
 805         if (bsd->bsd_version != 0) {
 806                 CERROR("Unexpected version %u\n", bsd->bsd_version);
 807                 return -EPROTO;
 808         }
 809
 810         if (bsd->bsd_csum_alg >= BULK_CSUM_ALG_MAX) {
 811                 CERROR("Unsupported checksum algorithm %u\n",
 812                        bsd->bsd_csum_alg);
 813                 return -EINVAL;
 814         }
 815         if (bsd->bsd_priv_alg >= BULK_PRIV_ALG_MAX) {
 816                 CERROR("Unsupported cipher algorithm %u\n",
 817                        bsd->bsd_priv_alg);
 818                 return -EINVAL;
 819         }
 820
 821         if (size > sizeof(*bsd) &&
 822             size < sizeof(*bsd) + csum_types[bsd->bsd_csum_alg].size) {
 823                 CERROR("Mal-formed checksum data: csum alg %u, size %d\n",
 824                        bsd->bsd_csum_alg, size);
 825                 return -EINVAL;
 826         }
 827
 828         return 0;
 829 }
 830 EXPORT_SYMBOL(bulk_sec_desc_unpack);
 831
 832 #ifdef __KERNEL__
 833 static
 834 int do_bulk_checksum_crc32(struct ptlrpc_bulk_desc *desc, void *buf)
 835 {
 836         struct page *page;
 837         int off;
 838         char *ptr;
 839         __u32 crc32 = ~0;
 840         int len, i;
 841
 842         for (i = 0; i < desc->bd_iov_count; i++) {
 843                 page = desc->bd_iov[i].kiov_page;
 844                 off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK;
 845                 ptr = cfs_kmap(page) + off;
 846                 len = desc->bd_iov[i].kiov_len;
 847
 848                 crc32 = crc32_le(crc32, ptr, len);
 849
 850                 cfs_kunmap(page);
 851         }
 852
 853         *((__u32 *) buf) = crc32;
 854         return 0;
 855 }
 856
 857 static
 858 int do_bulk_checksum(struct ptlrpc_bulk_desc *desc, __u32 alg, void *buf)
 859 {
 860         struct crypto_tfm *tfm;
 861         struct scatterlist *sl;
 862         int i, rc = 0;
 863
 864         LASSERT(alg > BULK_CSUM_ALG_NULL &&
 865                 alg < BULK_CSUM_ALG_MAX);
 866
 867         if (alg == BULK_CSUM_ALG_CRC32)
 868                 return do_bulk_checksum_crc32(desc, buf);
 869
 870         tfm = crypto_alloc_tfm(csum_types[alg].name, 0);
 871         if (tfm == NULL) {
 872                 CERROR("Unable to allocate tfm %s\n", csum_types[alg].name);
 873                 return -ENOMEM;
 874         }
 875
 876         OBD_ALLOC(sl, sizeof(*sl) * desc->bd_iov_count);
 877         if (sl == NULL) {
 878                 rc = -ENOMEM;
 879                 goto out_tfm;
 880         }
 881
 882         for (i = 0; i < desc->bd_iov_count; i++) {
 883                 sl[i].page = desc->bd_iov[i].kiov_page;
 884                 sl[i].offset = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK;
 885                 sl[i].length = desc->bd_iov[i].kiov_len;
 886         }
 887
 888         crypto_digest_init(tfm);
 889         crypto_digest_update(tfm, sl, desc->bd_iov_count);
 890         crypto_digest_final(tfm, buf);
 891
 892         OBD_FREE(sl, sizeof(*sl) * desc->bd_iov_count);
 893
 894 out_tfm:
 895         crypto_free_tfm(tfm);
 896         return rc;
 897 }
 898
 899 #else /* !__KERNEL__ */
 900 static
 901 int do_bulk_checksum(struct ptlrpc_bulk_desc *desc, __u32 alg, void *buf)
 902 {
 903         __u32 crc32 = ~0;
 904         int i;
 905
 906         LASSERT(alg == BULK_CSUM_ALG_CRC32);
 907
 908         for (i = 0; i < desc->bd_iov_count; i++) {
 909                 char *ptr = desc->bd_iov[i].iov_base;
 910                 int len = desc->bd_iov[i].iov_len;
 911
 912                 crc32 = crc32_le(crc32, ptr, len);
 913         }
 914
 915         *((__u32 *) buf) = crc32;
 916         return 0;
 917 }
 918 #endif
 919
 920 /*
 921  * perform algorithm @alg checksum on @desc, store result in @buf.
 922  * if anything goes wrong, leave 'alg' be BULK_CSUM_ALG_NULL.
 923  */
 924 static
 925 int generate_bulk_csum(struct ptlrpc_bulk_desc *desc, __u32 alg,
 926                        struct ptlrpc_bulk_sec_desc *bsd, int bsdsize)
 927 {
 928         int rc;
 929
 930         LASSERT(bsd);
 931         LASSERT(alg < BULK_CSUM_ALG_MAX);
 932
 933         bsd->bsd_csum_alg = BULK_CSUM_ALG_NULL;
 934
 935         if (alg == BULK_CSUM_ALG_NULL)
 936                 return 0;
 937
 938         LASSERT(bsdsize >= sizeof(*bsd) + csum_types[alg].size);
 939
 940         rc = do_bulk_checksum(desc, alg, bsd->bsd_csum);
 941         if (rc == 0)
 942                 bsd->bsd_csum_alg = alg;
 943
 944         return rc;
 945 }
 946
 947 static
 948 int verify_bulk_csum(struct ptlrpc_bulk_desc *desc, int read,
 949                      struct ptlrpc_bulk_sec_desc *bsdv, int bsdvsize,
 950                      struct ptlrpc_bulk_sec_desc *bsdr, int bsdrsize)
 951 {
 952         char *csum_p;
 953         char *buf = NULL;
 954         int   csum_size, rc = 0;
 955
 956         LASSERT(bsdv);
 957         LASSERT(bsdv->bsd_csum_alg < BULK_CSUM_ALG_MAX);
 958
 959         if (bsdr)
 960                 bsdr->bsd_csum_alg = BULK_CSUM_ALG_NULL;
 961
 962         if (bsdv->bsd_csum_alg == BULK_CSUM_ALG_NULL)
 963                 return 0;
 964
 965         /* for all supported algorithms */
 966         csum_size = csum_types[bsdv->bsd_csum_alg].size;
 967
 968         if (bsdvsize < sizeof(*bsdv) + csum_size) {
 969                 CERROR("verifier size %d too small, require %d\n",
 970                        bsdvsize, (int) sizeof(*bsdv) + csum_size);
 971                 return -EINVAL;
 972         }
 973
 974         if (bsdr) {
 975                 LASSERT(bsdrsize >= sizeof(*bsdr) + csum_size);
 976                 csum_p = (char *) bsdr->bsd_csum;
 977         } else {
 978                 OBD_ALLOC(buf, csum_size);
 979                 if (buf == NULL)
 980                         return -EINVAL;
 981                 csum_p = buf;
 982         }
 983
 984         rc = do_bulk_checksum(desc, bsdv->bsd_csum_alg, csum_p);
 985
 986         if (memcmp(bsdv->bsd_csum, csum_p, csum_size)) {
 987                 CERROR("BAD %s CHECKSUM (%s), data mutated during "
 988                        "transfer!\n", read ? "READ" : "WRITE",
 989                        csum_types[bsdv->bsd_csum_alg].name);
 990                 rc = -EINVAL;
 991         } else {
 992                 CDEBUG(D_SEC, "bulk %s checksum (%s) verified\n",
 993                       read ? "read" : "write",
 994                       csum_types[bsdv->bsd_csum_alg].name);
 995         }
 996
 997         if (bsdr) {
 998                 bsdr->bsd_csum_alg = bsdv->bsd_csum_alg;
 999                 memcpy(bsdr->bsd_csum, csum_p, csum_size);
1000         } else {
1001                 LASSERT(buf);
1002                 OBD_FREE(buf, csum_size);
1003         }
1004
1005         return rc;
1006 }
1007
1008 int bulk_csum_cli_request(struct ptlrpc_bulk_desc *desc, int read,
1009                           __u32 alg, struct lustre_msg *rmsg, int roff)
1010 {
1011         struct ptlrpc_bulk_sec_desc *bsdr;
1012         int    rsize, rc = 0;
1013
1014         rsize = rmsg->lm_buflens[roff];
1015         bsdr = lustre_msg_buf(rmsg, roff, sizeof(*bsdr));
1016
1017         LASSERT(bsdr);
1018         LASSERT(rsize >= sizeof(*bsdr));
1019         LASSERT(alg < BULK_CSUM_ALG_MAX);
1020
1021         if (read) {
1022                 bsdr->bsd_csum_alg = alg;
1023         } else {
1024                 rc = generate_bulk_csum(desc, alg, bsdr, rsize);
1025                 if (rc)
1026                         CERROR("bulk write: client failed to compute "
1027                                "checksum: %d\n", rc);
1028
1029                 /* For sending we only compute the wrong checksum instead
1030                  * of corrupting the data so it is still correct on a redo */
1031                 if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND) &&
1032                     bsdr->bsd_csum_alg != BULK_CSUM_ALG_NULL)
1033                         bsdr->bsd_csum[0] ^= 0x1;
1034         }
1035
1036         return rc;
1037 }
1038 EXPORT_SYMBOL(bulk_csum_cli_request);
1039
1040 int bulk_csum_cli_reply(struct ptlrpc_bulk_desc *desc, int read,
1041                         struct lustre_msg *rmsg, int roff,
1042                         struct lustre_msg *vmsg, int voff)
1043 {
1044         struct ptlrpc_bulk_sec_desc *bsdv, *bsdr;
1045         int    rsize, vsize;
1046
1047         rsize = rmsg->lm_buflens[roff];
1048         vsize = vmsg->lm_buflens[voff];
1049         bsdr = lustre_msg_buf(rmsg, roff, 0);
1050         bsdv = lustre_msg_buf(vmsg, voff, 0);
1051
1052         if (bsdv == NULL || vsize < sizeof(*bsdv)) {
1053                 CERROR("Invalid checksum verifier from server: size %d\n",
1054                        vsize);
1055                 return -EINVAL;
1056         }
1057
1058         LASSERT(bsdr);
1059         LASSERT(rsize >= sizeof(*bsdr));
1060         LASSERT(vsize >= sizeof(*bsdv));
1061
1062         if (bsdr->bsd_csum_alg != bsdv->bsd_csum_alg) {
1063                 CERROR("bulk %s: checksum algorithm mismatch: client request "
1064                        "%s but server reply with %s. try to use the new one "
1065                        "for checksum verification\n",
1066                        read ? "read" : "write",
1067                        csum_types[bsdr->bsd_csum_alg].name,
1068                        csum_types[bsdv->bsd_csum_alg].name);
1069         }
1070
1071         if (read)
1072                 return verify_bulk_csum(desc, 1, bsdv, vsize, NULL, 0);
1073         else {
1074                 char *cli, *srv, *new = NULL;
1075                 int csum_size = csum_types[bsdr->bsd_csum_alg].size;
1076
1077                 LASSERT(bsdr->bsd_csum_alg < BULK_CSUM_ALG_MAX);
1078                 if (bsdr->bsd_csum_alg == BULK_CSUM_ALG_NULL)
1079                         return 0;
1080
1081                 if (vsize < sizeof(*bsdv) + csum_size) {
1082                         CERROR("verifier size %d too small, require %d\n",
1083                                vsize, (int) sizeof(*bsdv) + csum_size);
1084                         return -EINVAL;
1085                 }
1086
1087                 cli = (char *) (bsdr + 1);
1088                 srv = (char *) (bsdv + 1);
1089
1090                 if (!memcmp(cli, srv, csum_size)) {
1091                         /* checksum confirmed */
1092                         CDEBUG(D_SEC, "bulk write checksum (%s) confirmed\n",
1093                                csum_types[bsdr->bsd_csum_alg].name);
1094                         return 0;
1095                 }
1096
1097                 /* checksum mismatch, re-compute a new one and compare with
1098                  * others, give out proper warnings. */
1099                 OBD_ALLOC(new, csum_size);
1100                 if (new == NULL)
1101                         return -ENOMEM;
1102
1103                 do_bulk_checksum(desc, bsdr->bsd_csum_alg, new);
1104
1105                 if (!memcmp(new, srv, csum_size)) {
1106                         CERROR("BAD WRITE CHECKSUM (%s): pages were mutated "
1107                                "on the client after we checksummed them\n",
1108                                csum_types[bsdr->bsd_csum_alg].name);
1109                 } else if (!memcmp(new, cli, csum_size)) {
1110                         CERROR("BAD WRITE CHECKSUM (%s): pages were mutated "
1111                                "in transit\n",
1112                                csum_types[bsdr->bsd_csum_alg].name);
1113                 } else {
1114                         CERROR("BAD WRITE CHECKSUM (%s): pages were mutated "
1115                                "in transit, and the current page contents "
1116                                "don't match the originals and what the server "
1117                                "received\n",
1118                                csum_types[bsdr->bsd_csum_alg].name);
1119                 }
1120                 OBD_FREE(new, csum_size);
1121
1122                 return -EINVAL;
1123         }
1124 }
1125 EXPORT_SYMBOL(bulk_csum_cli_reply);
1126
1127 #ifdef __KERNEL__
1128 static void corrupt_bulk_data(struct ptlrpc_bulk_desc *desc)
1129 {
1130         char           *ptr;
1131         unsigned int    off, i;
1132
1133         for (i = 0; i < desc->bd_iov_count; i++) {
1134                 if (desc->bd_iov[i].kiov_len == 0)
1135                         continue;
1136
1137                 ptr = cfs_kmap(desc->bd_iov[i].kiov_page);
1138                 off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK;
1139                 ptr[off] ^= 0x1;
1140                 cfs_kunmap(desc->bd_iov[i].kiov_page);
1141                 return;
1142         }
1143 }
1144 #else
1145 static void corrupt_bulk_data(struct ptlrpc_bulk_desc *desc)
1146 {
1147 }
1148 #endif /* __KERNEL__ */
1149
1150 int bulk_csum_svc(struct ptlrpc_bulk_desc *desc, int read,
1151                   struct ptlrpc_bulk_sec_desc *bsdv, int vsize,
1152                   struct ptlrpc_bulk_sec_desc *bsdr, int rsize)
1153 {
1154         int    rc;
1155
1156         LASSERT(vsize >= sizeof(*bsdv));
1157         LASSERT(rsize >= sizeof(*bsdr));
1158         LASSERT(bsdv && bsdr);
1159
1160         if (read) {
1161                 rc = generate_bulk_csum(desc, bsdv->bsd_csum_alg, bsdr, rsize);
1162                 if (rc)
1163                         CERROR("bulk read: server failed to generate %s "
1164                                "checksum: %d\n",
1165                                csum_types[bsdv->bsd_csum_alg].name, rc);
1166
1167                 /* corrupt the data after we compute the checksum, to
1168                  * simulate an OST->client data error */
1169                 if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE))
1170                         corrupt_bulk_data(desc);
1171         } else {
1172                 rc = verify_bulk_csum(desc, 0, bsdv, vsize, bsdr, rsize);
1173         }
1174
1175         return rc;
1176 }
1177 EXPORT_SYMBOL(bulk_csum_svc);
1178
1179 /****************************************
1180  * Helpers to assist policy modules to  *
1181  * implement encryption funcationality  *
1182  ****************************************/
1183
1184 /*
1185  * NOTE: These algorithms must be stream cipher!
1186  */
1187 static struct {
1188         char    *name;
1189         __u32    flags;
1190 } priv_types[] = {
1191         [BULK_PRIV_ALG_NULL]   = { "null", 0   },
1192         [BULK_PRIV_ALG_ARC4]   = { "arc4", 0   },
1193 };
1194
1195 const char * sptlrpc_bulk_priv_alg2name(__u8 priv_alg)
1196 {
1197         if (priv_alg < BULK_PRIV_ALG_MAX)
1198                 return priv_types[priv_alg].name;
1199         return "unknown";
1200 }
1201 EXPORT_SYMBOL(sptlrpc_bulk_priv_alg2name);
1202
1203 __u32 sptlrpc_bulk_priv_alg2flags(__u8 priv_alg)
1204 {
1205         if (priv_alg < BULK_PRIV_ALG_MAX)
1206                 return priv_types[priv_alg].flags;
1207         return 0;
1208 }
1209 EXPORT_SYMBOL(sptlrpc_bulk_priv_alg2flags);