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  2008 Sun Microsystems, Inc. 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         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 = DEFAULT_SEEKS;
 136 static struct 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         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                       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, HZ
 187                      );
 188
 189         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(int nr_to_scan, unsigned int gfp_mask)
 243 {
 244         if (unlikely(nr_to_scan != 0)) {
 245                 spin_lock(&page_pools.epp_lock);
 246                 nr_to_scan = min(nr_to_scan, (int) page_pools.epp_free_pages -
 247                                              PTLRPC_MAX_BRW_PAGES);
 248                 if (nr_to_scan > 0) {
 249                         enc_pools_release_free_pages(nr_to_scan);
 250                         CDEBUG(D_SEC, "released %d pages, %ld left\n",
 251                                nr_to_scan, page_pools.epp_free_pages);
 252
 253                         page_pools.epp_st_shrinks++;
 254                         page_pools.epp_last_shrink = cfs_time_current_sec();
 255                 }
 256                 spin_unlock(&page_pools.epp_lock);
 257         }
 258
 259         /*
 260          * if no pool access for a long time, we consider it's fully idle.
 261          * a little race here is fine.
 262          */
 263         if (unlikely(cfs_time_current_sec() - page_pools.epp_last_access >
 264                      CACHE_QUIESCENT_PERIOD)) {
 265                 spin_lock(&page_pools.epp_lock);
 266                 page_pools.epp_idle_idx = IDLE_IDX_MAX;
 267                 spin_unlock(&page_pools.epp_lock);
 268         }
 269
 270         LASSERT(page_pools.epp_idle_idx <= IDLE_IDX_MAX);
 271         return max((int) page_pools.epp_free_pages - PTLRPC_MAX_BRW_PAGES, 0) *
 272                (IDLE_IDX_MAX - page_pools.epp_idle_idx) / IDLE_IDX_MAX;
 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(cfs_page_t ***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                                         cfs_free_page(pools[i][j]);
 294                                         cleaned++;
 295                                 }
 296                         }
 297                         OBD_FREE(pools[i], CFS_PAGE_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(cfs_page_t ***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 DECLARE_MUTEX(sem_add_pages);
 396         cfs_page_t   ***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         down(&sem_add_pages);
 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], CFS_PAGE_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] = cfs_alloc_page(CFS_ALLOC_IO |
 423                                                      CFS_ALLOC_HIGH);
 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         up(&sem_add_pages);
 446         return rc;
 447 }
 448
 449 static inline void enc_pools_wakeup(void)
 450 {
 451         LASSERT_SPIN_LOCKED(&page_pools.epp_lock);
 452         LASSERT(page_pools.epp_waitqlen >= 0);
 453
 454         if (unlikely(page_pools.epp_waitqlen)) {
 455                 LASSERT(cfs_waitq_active(&page_pools.epp_waitq));
 456                 cfs_waitq_broadcast(&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         cfs_waitlink_t  waitlink;
 498         unsigned long   this_idle = -1;
 499         cfs_time_t      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 = cfs_time_current_sec();
 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(CFS_TASK_UNINT);
 546                         cfs_waitlink_init(&waitlink);
 547                         cfs_waitq_add(&page_pools.epp_waitq, &waitlink);
 548
 549                         spin_unlock(&page_pools.epp_lock);
 550                         cfs_waitq_wait(&waitlink, CFS_TASK_UNINT);
 551                         cfs_waitq_del(&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 = cfs_time_current_sec();
 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         /*
 694          * on system with huge memory but small page size, this might lead to
 695          * high-order allocation. but it's not common, and we suppose memory
 696          * be not too much fragmented at module loading time.
 697          */
 698         OBD_ALLOC(page_pools.epp_pools,
 699                   page_pools.epp_max_pools * 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(page_pools.epp_pools,
 708                  page_pools.epp_max_pools * sizeof(*page_pools.epp_pools));
 709 }
 710
 711 int sptlrpc_enc_pool_init(void)
 712 {
 713         /*
 714          * maximum capacity is 1/8 of total physical memory.
 715          * is the 1/8 a good number?
 716          */
 717         page_pools.epp_max_pages = num_physpages / 8;
 718         page_pools.epp_max_pools = npages_to_npools(page_pools.epp_max_pages);
 719
 720         cfs_waitq_init(&page_pools.epp_waitq);
 721         page_pools.epp_waitqlen = 0;
 722         page_pools.epp_pages_short = 0;
 723
 724         page_pools.epp_growing = 0;
 725
 726         page_pools.epp_idle_idx = 0;
 727         page_pools.epp_last_shrink = cfs_time_current_sec();
 728         page_pools.epp_last_access = cfs_time_current_sec();
 729
 730         spin_lock_init(&page_pools.epp_lock);
 731         page_pools.epp_total_pages = 0;
 732         page_pools.epp_free_pages = 0;
 733
 734         page_pools.epp_st_max_pages = 0;
 735         page_pools.epp_st_grows = 0;
 736         page_pools.epp_st_grow_fails = 0;
 737         page_pools.epp_st_shrinks = 0;
 738         page_pools.epp_st_access = 0;
 739         page_pools.epp_st_missings = 0;
 740         page_pools.epp_st_lowfree = 0;
 741         page_pools.epp_st_max_wqlen = 0;
 742         page_pools.epp_st_max_wait = 0;
 743
 744         enc_pools_alloc();
 745         if (page_pools.epp_pools == NULL)
 746                 return -ENOMEM;
 747
 748         pools_shrinker = set_shrinker(pools_shrinker_seeks, enc_pools_shrink);
 749         if (pools_shrinker == NULL) {
 750                 enc_pools_free();
 751                 return -ENOMEM;
 752         }
 753
 754         return 0;
 755 }
 756
 757 void sptlrpc_enc_pool_fini(void)
 758 {
 759         unsigned long cleaned, npools;
 760
 761         LASSERT(pools_shrinker);
 762         LASSERT(page_pools.epp_pools);
 763         LASSERT(page_pools.epp_total_pages == page_pools.epp_free_pages);
 764
 765         remove_shrinker(pools_shrinker);
 766
 767         npools = npages_to_npools(page_pools.epp_total_pages);
 768         cleaned = enc_pools_cleanup(page_pools.epp_pools, npools);
 769         LASSERT(cleaned == page_pools.epp_total_pages);
 770
 771         enc_pools_free();
 772
 773         if (page_pools.epp_st_access > 0) {
 774                 CWARN("max pages %lu, grows %u, grow fails %u, shrinks %u, "
 775                       "access %lu, missing %lu, max qlen %u, max wait "
 776                       CFS_TIME_T"/%d\n",
 777                       page_pools.epp_st_max_pages, page_pools.epp_st_grows,
 778                       page_pools.epp_st_grow_fails,
 779                       page_pools.epp_st_shrinks, page_pools.epp_st_access,
 780                       page_pools.epp_st_missings, page_pools.epp_st_max_wqlen,
 781                       page_pools.epp_st_max_wait, HZ);
 782         }
 783 }
 784
 785 #else /* !__KERNEL__ */
 786
 787 int sptlrpc_enc_pool_get_pages(struct ptlrpc_bulk_desc *desc)
 788 {
 789         return 0;
 790 }
 791
 792 void sptlrpc_enc_pool_put_pages(struct ptlrpc_bulk_desc *desc)
 793 {
 794 }
 795
 796 int sptlrpc_enc_pool_init(void)
 797 {
 798         return 0;
 799 }
 800
 801 void sptlrpc_enc_pool_fini(void)
 802 {
 803 }
 804 #endif
 805
 806 /****************************************
 807  * Helpers to assist policy modules to  *
 808  * implement checksum funcationality    *
 809  ****************************************/
 810
 811 static struct sptlrpc_hash_type hash_types[] = {
 812         [BULK_HASH_ALG_NULL]    = { "null",     "null",         0 },
 813         [BULK_HASH_ALG_ADLER32] = { "adler32",  "adler32",      4 },
 814         [BULK_HASH_ALG_CRC32]   = { "crc32",    "crc32",        4 },
 815         [BULK_HASH_ALG_MD5]     = { "md5",      "md5",          16 },
 816         [BULK_HASH_ALG_SHA1]    = { "sha1",     "sha1",         20 },
 817         [BULK_HASH_ALG_SHA256]  = { "sha256",   "sha256",       32 },
 818         [BULK_HASH_ALG_SHA384]  = { "sha384",   "sha384",       48 },
 819         [BULK_HASH_ALG_SHA512]  = { "sha512",   "sha512",       64 },
 820 };
 821
 822 const struct sptlrpc_hash_type *sptlrpc_get_hash_type(__u8 hash_alg)
 823 {
 824         struct sptlrpc_hash_type *ht;
 825
 826         if (hash_alg < BULK_HASH_ALG_MAX) {
 827                 ht = &hash_types[hash_alg];
 828                 if (ht->sht_tfm_name)
 829                         return ht;
 830         }
 831         return NULL;
 832 }
 833 EXPORT_SYMBOL(sptlrpc_get_hash_type);
 834
 835 const char * sptlrpc_get_hash_name(__u8 hash_alg)
 836 {
 837         const struct sptlrpc_hash_type *ht;
 838
 839         ht = sptlrpc_get_hash_type(hash_alg);
 840         if (ht)
 841                 return ht->sht_name;
 842         else
 843                 return "unknown";
 844 }
 845 EXPORT_SYMBOL(sptlrpc_get_hash_name);
 846
 847 __u8 sptlrpc_get_hash_alg(const char *algname)
 848 {
 849         int     i;
 850
 851         for (i = 0; i < BULK_HASH_ALG_MAX; i++)
 852                 if (!strcmp(hash_types[i].sht_name, algname))
 853                         break;
 854         return i;
 855 }
 856 EXPORT_SYMBOL(sptlrpc_get_hash_alg);
 857
 858 int bulk_sec_desc_unpack(struct lustre_msg *msg, int offset, int swabbed)
 859 {
 860         struct ptlrpc_bulk_sec_desc *bsd;
 861         int                          size = msg->lm_buflens[offset];
 862
 863         bsd = lustre_msg_buf(msg, offset, sizeof(*bsd));
 864         if (bsd == NULL) {
 865                 CERROR("Invalid bulk sec desc: size %d\n", size);
 866                 return -EINVAL;
 867         }
 868
 869         if (swabbed) {
 870                 __swab32s(&bsd->bsd_nob);
 871         }
 872
 873         if (unlikely(bsd->bsd_version != 0)) {
 874                 CERROR("Unexpected version %u\n", bsd->bsd_version);
 875                 return -EPROTO;
 876         }
 877
 878         if (unlikely(bsd->bsd_type >= SPTLRPC_BULK_MAX)) {
 879                 CERROR("Invalid type %u\n", bsd->bsd_type);
 880                 return -EPROTO;
 881         }
 882
 883         /* FIXME more sanity check here */
 884
 885         if (unlikely(bsd->bsd_svc != SPTLRPC_BULK_SVC_NULL &&
 886                      bsd->bsd_svc != SPTLRPC_BULK_SVC_INTG &&
 887                      bsd->bsd_svc != SPTLRPC_BULK_SVC_PRIV)) {
 888                 CERROR("Invalid svc %u\n", bsd->bsd_svc);
 889                 return -EPROTO;
 890         }
 891
 892         return 0;
 893 }
 894 EXPORT_SYMBOL(bulk_sec_desc_unpack);
 895
 896 #ifdef __KERNEL__
 897
 898 #ifdef HAVE_ADLER
 899 static int do_bulk_checksum_adler32(struct ptlrpc_bulk_desc *desc, void *buf)
 900 {
 901         struct page    *page;
 902         int             off;
 903         char           *ptr;
 904         __u32           adler32 = 1;
 905         int             len, i;
 906
 907         for (i = 0; i < desc->bd_iov_count; i++) {
 908                 page = desc->bd_iov[i].kiov_page;
 909                 off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK;
 910                 ptr = cfs_kmap(page) + off;
 911                 len = desc->bd_iov[i].kiov_len;
 912
 913                 adler32 = adler32(adler32, ptr, len);
 914
 915                 cfs_kunmap(page);
 916         }
 917
 918         adler32 = cpu_to_le32(adler32);
 919         memcpy(buf, &adler32, sizeof(adler32));
 920         return 0;
 921 }
 922 #endif
 923
 924 static int do_bulk_checksum_crc32(struct ptlrpc_bulk_desc *desc, void *buf)
 925 {
 926         struct page    *page;
 927         int             off;
 928         char           *ptr;
 929         __u32           crc32 = ~0;
 930         int             len, i;
 931
 932         for (i = 0; i < desc->bd_iov_count; i++) {
 933                 page = desc->bd_iov[i].kiov_page;
 934                 off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK;
 935                 ptr = cfs_kmap(page) + off;
 936                 len = desc->bd_iov[i].kiov_len;
 937
 938                 crc32 = crc32_le(crc32, ptr, len);
 939
 940                 cfs_kunmap(page);
 941         }
 942
 943         crc32 = cpu_to_le32(crc32);
 944         memcpy(buf, &crc32, sizeof(crc32));
 945         return 0;
 946 }
 947
 948 int sptlrpc_get_bulk_checksum(struct ptlrpc_bulk_desc *desc, __u8 alg,
 949                               void *buf, int buflen)
 950 {
 951         struct hash_desc    hdesc;
 952         int                 hashsize;
 953         char                hashbuf[64];
 954         struct scatterlist  sl;
 955         int                 i;
 956
 957         LASSERT(alg > BULK_HASH_ALG_NULL && alg < BULK_HASH_ALG_MAX);
 958         LASSERT(buflen >= 4);
 959
 960         switch (alg) {
 961         case BULK_HASH_ALG_ADLER32:
 962 #ifdef HAVE_ADLER
 963                 return do_bulk_checksum_adler32(desc, buf);
 964 #else
 965                 CERROR("Adler32 not supported\n");
 966                 return -EINVAL;
 967 #endif
 968         case BULK_HASH_ALG_CRC32:
 969                 return do_bulk_checksum_crc32(desc, buf);
 970         }
 971
 972         hdesc.tfm = ll_crypto_alloc_hash(hash_types[alg].sht_tfm_name, 0, 0);
 973         if (hdesc.tfm == NULL) {
 974                 CERROR("Unable to allocate TFM %s\n", hash_types[alg].sht_name);
 975                 return -ENOMEM;
 976         }
 977
 978         hdesc.flags = 0;
 979         ll_crypto_hash_init(&hdesc);
 980
 981         hashsize = ll_crypto_hash_digestsize(hdesc.tfm);
 982
 983         for (i = 0; i < desc->bd_iov_count; i++) {
 984                 sl.page = desc->bd_iov[i].kiov_page;
 985                 sl.offset = desc->bd_iov[i].kiov_offset;
 986                 sl.length = desc->bd_iov[i].kiov_len;
 987                 ll_crypto_hash_update(&hdesc, &sl, sl.length);
 988         }
 989
 990         if (hashsize > buflen) {
 991                 ll_crypto_hash_final(&hdesc, hashbuf);
 992                 memcpy(buf, hashbuf, buflen);
 993         } else {
 994                 ll_crypto_hash_final(&hdesc, buf);
 995         }
 996
 997         ll_crypto_free_hash(hdesc.tfm);
 998         return 0;
 999 }
1000 EXPORT_SYMBOL(sptlrpc_get_bulk_checksum);
1001
1002 #else /* !__KERNEL__ */
1003
1004 int sptlrpc_get_bulk_checksum(struct ptlrpc_bulk_desc *desc, __u8 alg,
1005                               void *buf, int buflen)
1006 {
1007         __u32   csum32;
1008         int     i;
1009
1010         LASSERT(alg == BULK_HASH_ALG_ADLER32 || alg == BULK_HASH_ALG_CRC32);
1011
1012         if (alg == BULK_HASH_ALG_ADLER32)
1013                 csum32 = 1;
1014         else
1015                 csum32 = ~0;
1016
1017         for (i = 0; i < desc->bd_iov_count; i++) {
1018                 unsigned char *ptr = desc->bd_iov[i].iov_base;
1019                 int len = desc->bd_iov[i].iov_len;
1020
1021                 switch (alg) {
1022                 case BULK_HASH_ALG_ADLER32:
1023 #ifdef HAVE_ADLER
1024                         csum32 = adler32(csum32, ptr, len);
1025 #else
1026                         CERROR("Adler32 not supported\n");
1027                         return -EINVAL;
1028 #endif
1029                         break;
1030                 case BULK_HASH_ALG_CRC32:
1031                         csum32 = crc32_le(csum32, ptr, len);
1032                         break;
1033                 }
1034         }
1035
1036         csum32 = cpu_to_le32(csum32);
1037         memcpy(buf, &csum32, sizeof(csum32));
1038         return 0;
1039 }
1040
1041 #endif /* __KERNEL__ */