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_FIRST_ARG int nr_to_scan,
 243                             unsigned int gfp_mask)
 244 {
 245         if (unlikely(nr_to_scan != 0)) {
 246                 cfs_spin_lock(&page_pools.epp_lock);
 247                 nr_to_scan = min(nr_to_scan, (int) page_pools.epp_free_pages -
 248                                  PTLRPC_MAX_BRW_PAGES);
 249                 if (nr_to_scan > 0) {
 250                         enc_pools_release_free_pages(nr_to_scan);
 251                         CDEBUG(D_SEC, "released %d pages, %ld left\n",
 252                                nr_to_scan, page_pools.epp_free_pages);
 253
 254                         page_pools.epp_st_shrinks++;
 255                         page_pools.epp_last_shrink = cfs_time_current_sec();
 256                 }
 257                 cfs_spin_unlock(&page_pools.epp_lock);
 258         }
 259
 260         /*
 261          * if no pool access for a long time, we consider it's fully idle.
 262          * a little race here is fine.
 263          */
 264         if (unlikely(cfs_time_current_sec() - page_pools.epp_last_access >
 265                      CACHE_QUIESCENT_PERIOD)) {
 266                 cfs_spin_lock(&page_pools.epp_lock);
 267                 page_pools.epp_idle_idx = IDLE_IDX_MAX;
 268                 cfs_spin_unlock(&page_pools.epp_lock);
 269         }
 270
 271         LASSERT(page_pools.epp_idle_idx <= IDLE_IDX_MAX);
 272         return max((int) page_pools.epp_free_pages - PTLRPC_MAX_BRW_PAGES, 0) *
 273                (IDLE_IDX_MAX - page_pools.epp_idle_idx) / IDLE_IDX_MAX;
 274 }
 275
 276 static inline
 277 int npages_to_npools(unsigned long npages)
 278 {
 279         return (int) ((npages + PAGES_PER_POOL - 1) / PAGES_PER_POOL);
 280 }
 281
 282 /*
 283  * return how many pages cleaned up.
 284  */
 285 static unsigned long enc_pools_cleanup(cfs_page_t ***pools, int npools)
 286 {
 287         unsigned long cleaned = 0;
 288         int           i, j;
 289
 290         for (i = 0; i < npools; i++) {
 291                 if (pools[i]) {
 292                         for (j = 0; j < PAGES_PER_POOL; j++) {
 293                                 if (pools[i][j]) {
 294                                         cfs_free_page(pools[i][j]);
 295                                         cleaned++;
 296                                 }
 297                         }
 298                         OBD_FREE(pools[i], CFS_PAGE_SIZE);
 299                         pools[i] = NULL;
 300                 }
 301         }
 302
 303         return cleaned;
 304 }
 305
 306 /*
 307  * merge @npools pointed by @pools which contains @npages new pages
 308  * into current pools.
 309  *
 310  * we have options to avoid most memory copy with some tricks. but we choose
 311  * the simplest way to avoid complexity. It's not frequently called.
 312  */
 313 static void enc_pools_insert(cfs_page_t ***pools, int npools, int npages)
 314 {
 315         int     freeslot;
 316         int     op_idx, np_idx, og_idx, ng_idx;
 317         int     cur_npools, end_npools;
 318
 319         LASSERT(npages > 0);
 320         LASSERT(page_pools.epp_total_pages+npages <= page_pools.epp_max_pages);
 321         LASSERT(npages_to_npools(npages) == npools);
 322         LASSERT(page_pools.epp_growing);
 323
 324         cfs_spin_lock(&page_pools.epp_lock);
 325
 326         /*
 327          * (1) fill all the free slots of current pools.
 328          */
 329         /* free slots are those left by rent pages, and the extra ones with
 330          * index >= total_pages, locate at the tail of last pool. */
 331         freeslot = page_pools.epp_total_pages % PAGES_PER_POOL;
 332         if (freeslot != 0)
 333                 freeslot = PAGES_PER_POOL - freeslot;
 334         freeslot += page_pools.epp_total_pages - page_pools.epp_free_pages;
 335
 336         op_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 337         og_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 338         np_idx = npools - 1;
 339         ng_idx = (npages - 1) % PAGES_PER_POOL;
 340
 341         while (freeslot) {
 342                 LASSERT(page_pools.epp_pools[op_idx][og_idx] == NULL);
 343                 LASSERT(pools[np_idx][ng_idx] != NULL);
 344
 345                 page_pools.epp_pools[op_idx][og_idx] = pools[np_idx][ng_idx];
 346                 pools[np_idx][ng_idx] = NULL;
 347
 348                 freeslot--;
 349
 350                 if (++og_idx == PAGES_PER_POOL) {
 351                         op_idx++;
 352                         og_idx = 0;
 353                 }
 354                 if (--ng_idx < 0) {
 355                         if (np_idx == 0)
 356                                 break;
 357                         np_idx--;
 358                         ng_idx = PAGES_PER_POOL - 1;
 359                 }
 360         }
 361
 362         /*
 363          * (2) add pools if needed.
 364          */
 365         cur_npools = (page_pools.epp_total_pages + PAGES_PER_POOL - 1) /
 366                      PAGES_PER_POOL;
 367         end_npools = (page_pools.epp_total_pages + npages + PAGES_PER_POOL -1) /
 368                      PAGES_PER_POOL;
 369         LASSERT(end_npools <= page_pools.epp_max_pools);
 370
 371         np_idx = 0;
 372         while (cur_npools < end_npools) {
 373                 LASSERT(page_pools.epp_pools[cur_npools] == NULL);
 374                 LASSERT(np_idx < npools);
 375                 LASSERT(pools[np_idx] != NULL);
 376
 377                 page_pools.epp_pools[cur_npools++] = pools[np_idx];
 378                 pools[np_idx++] = NULL;
 379         }
 380
 381         page_pools.epp_total_pages += npages;
 382         page_pools.epp_free_pages += npages;
 383         page_pools.epp_st_lowfree = page_pools.epp_free_pages;
 384
 385         if (page_pools.epp_total_pages > page_pools.epp_st_max_pages)
 386                 page_pools.epp_st_max_pages = page_pools.epp_total_pages;
 387
 388         CDEBUG(D_SEC, "add %d pages to total %lu\n", npages,
 389                page_pools.epp_total_pages);
 390
 391         cfs_spin_unlock(&page_pools.epp_lock);
 392 }
 393
 394 static int enc_pools_add_pages(int npages)
 395 {
 396         static CFS_DECLARE_MUTEX(sem_add_pages);
 397         cfs_page_t   ***pools;
 398         int             npools, alloced = 0;
 399         int             i, j, rc = -ENOMEM;
 400
 401         if (npages < PTLRPC_MAX_BRW_PAGES)
 402                 npages = PTLRPC_MAX_BRW_PAGES;
 403
 404         cfs_down(&sem_add_pages);
 405
 406         if (npages + page_pools.epp_total_pages > page_pools.epp_max_pages)
 407                 npages = page_pools.epp_max_pages - page_pools.epp_total_pages;
 408         LASSERT(npages > 0);
 409
 410         page_pools.epp_st_grows++;
 411
 412         npools = npages_to_npools(npages);
 413         OBD_ALLOC(pools, npools * sizeof(*pools));
 414         if (pools == NULL)
 415                 goto out;
 416
 417         for (i = 0; i < npools; i++) {
 418                 OBD_ALLOC(pools[i], CFS_PAGE_SIZE);
 419                 if (pools[i] == NULL)
 420                         goto out_pools;
 421
 422                 for (j = 0; j < PAGES_PER_POOL && alloced < npages; j++) {
 423                         pools[i][j] = cfs_alloc_page(CFS_ALLOC_IO |
 424                                                      CFS_ALLOC_HIGH);
 425                         if (pools[i][j] == NULL)
 426                                 goto out_pools;
 427
 428                         alloced++;
 429                 }
 430         }
 431         LASSERT(alloced == npages);
 432
 433         enc_pools_insert(pools, npools, npages);
 434         CDEBUG(D_SEC, "added %d pages into pools\n", npages);
 435         rc = 0;
 436
 437 out_pools:
 438         enc_pools_cleanup(pools, npools);
 439         OBD_FREE(pools, npools * sizeof(*pools));
 440 out:
 441         if (rc) {
 442                 page_pools.epp_st_grow_fails++;
 443                 CERROR("Failed to allocate %d enc pages\n", npages);
 444         }
 445
 446         cfs_up(&sem_add_pages);
 447         return rc;
 448 }
 449
 450 static inline void enc_pools_wakeup(void)
 451 {
 452         LASSERT_SPIN_LOCKED(&page_pools.epp_lock);
 453         LASSERT(page_pools.epp_waitqlen >= 0);
 454
 455         if (unlikely(page_pools.epp_waitqlen)) {
 456                 LASSERT(cfs_waitq_active(&page_pools.epp_waitq));
 457                 cfs_waitq_broadcast(&page_pools.epp_waitq);
 458         }
 459 }
 460
 461 static int enc_pools_should_grow(int page_needed, long now)
 462 {
 463         /* don't grow if someone else is growing the pools right now,
 464          * or the pools has reached its full capacity
 465          */
 466         if (page_pools.epp_growing ||
 467             page_pools.epp_total_pages == page_pools.epp_max_pages)
 468                 return 0;
 469
 470         /* if total pages is not enough, we need to grow */
 471         if (page_pools.epp_total_pages < page_needed)
 472                 return 1;
 473
 474         /*
 475          * we wanted to return 0 here if there was a shrink just happened
 476          * moment ago, but this may cause deadlock if both client and ost
 477          * live on single node.
 478          */
 479 #if 0
 480         if (now - page_pools.epp_last_shrink < 2)
 481                 return 0;
 482 #endif
 483
 484         /*
 485          * here we perhaps need consider other factors like wait queue
 486          * length, idle index, etc. ?
 487          */
 488
 489         /* grow the pools in any other cases */
 490         return 1;
 491 }
 492
 493 /*
 494  * we allocate the requested pages atomically.
 495  */
 496 int sptlrpc_enc_pool_get_pages(struct ptlrpc_bulk_desc *desc)
 497 {
 498         cfs_waitlink_t  waitlink;
 499         unsigned long   this_idle = -1;
 500         cfs_time_t      tick = 0;
 501         long            now;
 502         int             p_idx, g_idx;
 503         int             i;
 504
 505         LASSERT(desc->bd_iov_count > 0);
 506         LASSERT(desc->bd_iov_count <= page_pools.epp_max_pages);
 507
 508         /* resent bulk, enc iov might have been allocated previously */
 509         if (desc->bd_enc_iov != NULL)
 510                 return 0;
 511
 512         OBD_ALLOC(desc->bd_enc_iov,
 513                   desc->bd_iov_count * sizeof(*desc->bd_enc_iov));
 514         if (desc->bd_enc_iov == NULL)
 515                 return -ENOMEM;
 516
 517         cfs_spin_lock(&page_pools.epp_lock);
 518
 519         page_pools.epp_st_access++;
 520 again:
 521         if (unlikely(page_pools.epp_free_pages < desc->bd_iov_count)) {
 522                 if (tick == 0)
 523                         tick = cfs_time_current();
 524
 525                 now = cfs_time_current_sec();
 526
 527                 page_pools.epp_st_missings++;
 528                 page_pools.epp_pages_short += desc->bd_iov_count;
 529
 530                 if (enc_pools_should_grow(desc->bd_iov_count, now)) {
 531                         page_pools.epp_growing = 1;
 532
 533                         cfs_spin_unlock(&page_pools.epp_lock);
 534                         enc_pools_add_pages(page_pools.epp_pages_short / 2);
 535                         cfs_spin_lock(&page_pools.epp_lock);
 536
 537                         page_pools.epp_growing = 0;
 538
 539                         enc_pools_wakeup();
 540                 } else {
 541                         if (++page_pools.epp_waitqlen >
 542                             page_pools.epp_st_max_wqlen)
 543                                 page_pools.epp_st_max_wqlen =
 544                                                 page_pools.epp_waitqlen;
 545
 546                         cfs_set_current_state(CFS_TASK_UNINT);
 547                         cfs_waitlink_init(&waitlink);
 548                         cfs_waitq_add(&page_pools.epp_waitq, &waitlink);
 549
 550                         cfs_spin_unlock(&page_pools.epp_lock);
 551                         cfs_waitq_wait(&waitlink, CFS_TASK_UNINT);
 552                         cfs_waitq_del(&page_pools.epp_waitq, &waitlink);
 553                         LASSERT(page_pools.epp_waitqlen > 0);
 554                         cfs_spin_lock(&page_pools.epp_lock);
 555                         page_pools.epp_waitqlen--;
 556                 }
 557
 558                 LASSERT(page_pools.epp_pages_short >= desc->bd_iov_count);
 559                 page_pools.epp_pages_short -= desc->bd_iov_count;
 560
 561                 this_idle = 0;
 562                 goto again;
 563         }
 564
 565         /* record max wait time */
 566         if (unlikely(tick != 0)) {
 567                 tick = cfs_time_current() - tick;
 568                 if (tick > page_pools.epp_st_max_wait)
 569                         page_pools.epp_st_max_wait = tick;
 570         }
 571
 572         /* proceed with rest of allocation */
 573         page_pools.epp_free_pages -= desc->bd_iov_count;
 574
 575         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 576         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 577
 578         for (i = 0; i < desc->bd_iov_count; i++) {
 579                 LASSERT(page_pools.epp_pools[p_idx][g_idx] != NULL);
 580                 desc->bd_enc_iov[i].kiov_page =
 581                                         page_pools.epp_pools[p_idx][g_idx];
 582                 page_pools.epp_pools[p_idx][g_idx] = NULL;
 583
 584                 if (++g_idx == PAGES_PER_POOL) {
 585                         p_idx++;
 586                         g_idx = 0;
 587                 }
 588         }
 589
 590         if (page_pools.epp_free_pages < page_pools.epp_st_lowfree)
 591                 page_pools.epp_st_lowfree = page_pools.epp_free_pages;
 592
 593         /*
 594          * new idle index = (old * weight + new) / (weight + 1)
 595          */
 596         if (this_idle == -1) {
 597                 this_idle = page_pools.epp_free_pages * IDLE_IDX_MAX /
 598                             page_pools.epp_total_pages;
 599         }
 600         page_pools.epp_idle_idx = (page_pools.epp_idle_idx * IDLE_IDX_WEIGHT +
 601                                    this_idle) /
 602                                   (IDLE_IDX_WEIGHT + 1);
 603
 604         page_pools.epp_last_access = cfs_time_current_sec();
 605
 606         cfs_spin_unlock(&page_pools.epp_lock);
 607         return 0;
 608 }
 609 EXPORT_SYMBOL(sptlrpc_enc_pool_get_pages);
 610
 611 void sptlrpc_enc_pool_put_pages(struct ptlrpc_bulk_desc *desc)
 612 {
 613         int     p_idx, g_idx;
 614         int     i;
 615
 616         if (desc->bd_enc_iov == NULL)
 617                 return;
 618
 619         LASSERT(desc->bd_iov_count > 0);
 620
 621         cfs_spin_lock(&page_pools.epp_lock);
 622
 623         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 624         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 625
 626         LASSERT(page_pools.epp_free_pages + desc->bd_iov_count <=
 627                 page_pools.epp_total_pages);
 628         LASSERT(page_pools.epp_pools[p_idx]);
 629
 630         for (i = 0; i < desc->bd_iov_count; i++) {
 631                 LASSERT(desc->bd_enc_iov[i].kiov_page != NULL);
 632                 LASSERT(g_idx != 0 || page_pools.epp_pools[p_idx]);
 633                 LASSERT(page_pools.epp_pools[p_idx][g_idx] == NULL);
 634
 635                 page_pools.epp_pools[p_idx][g_idx] =
 636                                         desc->bd_enc_iov[i].kiov_page;
 637
 638                 if (++g_idx == PAGES_PER_POOL) {
 639                         p_idx++;
 640                         g_idx = 0;
 641                 }
 642         }
 643
 644         page_pools.epp_free_pages += desc->bd_iov_count;
 645
 646         enc_pools_wakeup();
 647
 648         cfs_spin_unlock(&page_pools.epp_lock);
 649
 650         OBD_FREE(desc->bd_enc_iov,
 651                  desc->bd_iov_count * sizeof(*desc->bd_enc_iov));
 652         desc->bd_enc_iov = NULL;
 653 }
 654 EXPORT_SYMBOL(sptlrpc_enc_pool_put_pages);
 655
 656 /*
 657  * we don't do much stuff for add_user/del_user anymore, except adding some
 658  * initial pages in add_user() if current pools are empty, rest would be
 659  * handled by the pools's self-adaption.
 660  */
 661 int sptlrpc_enc_pool_add_user(void)
 662 {
 663         int     need_grow = 0;
 664
 665         cfs_spin_lock(&page_pools.epp_lock);
 666         if (page_pools.epp_growing == 0 && page_pools.epp_total_pages == 0) {
 667                 page_pools.epp_growing = 1;
 668                 need_grow = 1;
 669         }
 670         cfs_spin_unlock(&page_pools.epp_lock);
 671
 672         if (need_grow) {
 673                 enc_pools_add_pages(PTLRPC_MAX_BRW_PAGES +
 674                                     PTLRPC_MAX_BRW_PAGES);
 675
 676                 cfs_spin_lock(&page_pools.epp_lock);
 677                 page_pools.epp_growing = 0;
 678                 enc_pools_wakeup();
 679                 cfs_spin_unlock(&page_pools.epp_lock);
 680         }
 681         return 0;
 682 }
 683 EXPORT_SYMBOL(sptlrpc_enc_pool_add_user);
 684
 685 int sptlrpc_enc_pool_del_user(void)
 686 {
 687         return 0;
 688 }
 689 EXPORT_SYMBOL(sptlrpc_enc_pool_del_user);
 690
 691 static inline void enc_pools_alloc(void)
 692 {
 693         LASSERT(page_pools.epp_max_pools);
 694         OBD_ALLOC_LARGE(page_pools.epp_pools,
 695                         page_pools.epp_max_pools *
 696                         sizeof(*page_pools.epp_pools));
 697 }
 698
 699 static inline void enc_pools_free(void)
 700 {
 701         LASSERT(page_pools.epp_max_pools);
 702         LASSERT(page_pools.epp_pools);
 703
 704         OBD_FREE_LARGE(page_pools.epp_pools,
 705                        page_pools.epp_max_pools *
 706                        sizeof(*page_pools.epp_pools));
 707 }
 708
 709 int sptlrpc_enc_pool_init(void)
 710 {
 711         /*
 712          * maximum capacity is 1/8 of total physical memory.
 713          * is the 1/8 a good number?
 714          */
 715         page_pools.epp_max_pages = cfs_num_physpages / 8;
 716         page_pools.epp_max_pools = npages_to_npools(page_pools.epp_max_pages);
 717
 718         cfs_waitq_init(&page_pools.epp_waitq);
 719         page_pools.epp_waitqlen = 0;
 720         page_pools.epp_pages_short = 0;
 721
 722         page_pools.epp_growing = 0;
 723
 724         page_pools.epp_idle_idx = 0;
 725         page_pools.epp_last_shrink = cfs_time_current_sec();
 726         page_pools.epp_last_access = cfs_time_current_sec();
 727
 728         cfs_spin_lock_init(&page_pools.epp_lock);
 729         page_pools.epp_total_pages = 0;
 730         page_pools.epp_free_pages = 0;
 731
 732         page_pools.epp_st_max_pages = 0;
 733         page_pools.epp_st_grows = 0;
 734         page_pools.epp_st_grow_fails = 0;
 735         page_pools.epp_st_shrinks = 0;
 736         page_pools.epp_st_access = 0;
 737         page_pools.epp_st_missings = 0;
 738         page_pools.epp_st_lowfree = 0;
 739         page_pools.epp_st_max_wqlen = 0;
 740         page_pools.epp_st_max_wait = 0;
 741
 742         enc_pools_alloc();
 743         if (page_pools.epp_pools == NULL)
 744                 return -ENOMEM;
 745
 746         pools_shrinker = cfs_set_shrinker(pools_shrinker_seeks,
 747                                           enc_pools_shrink);
 748         if (pools_shrinker == NULL) {
 749                 enc_pools_free();
 750                 return -ENOMEM;
 751         }
 752
 753         return 0;
 754 }
 755
 756 void sptlrpc_enc_pool_fini(void)
 757 {
 758         unsigned long cleaned, npools;
 759
 760         LASSERT(pools_shrinker);
 761         LASSERT(page_pools.epp_pools);
 762         LASSERT(page_pools.epp_total_pages == page_pools.epp_free_pages);
 763
 764         cfs_remove_shrinker(pools_shrinker);
 765
 766         npools = npages_to_npools(page_pools.epp_total_pages);
 767         cleaned = enc_pools_cleanup(page_pools.epp_pools, npools);
 768         LASSERT(cleaned == page_pools.epp_total_pages);
 769
 770         enc_pools_free();
 771
 772         if (page_pools.epp_st_access > 0) {
 773                 CWARN("max pages %lu, grows %u, grow fails %u, shrinks %u, "
 774                       "access %lu, missing %lu, max qlen %u, max wait "
 775                       CFS_TIME_T"/%d\n",
 776                       page_pools.epp_st_max_pages, page_pools.epp_st_grows,
 777                       page_pools.epp_st_grow_fails,
 778                       page_pools.epp_st_shrinks, page_pools.epp_st_access,
 779                       page_pools.epp_st_missings, page_pools.epp_st_max_wqlen,
 780                       page_pools.epp_st_max_wait, CFS_HZ);
 781         }
 782 }
 783
 784 #else /* !__KERNEL__ */
 785
 786 int sptlrpc_enc_pool_get_pages(struct ptlrpc_bulk_desc *desc)
 787 {
 788         return 0;
 789 }
 790
 791 void sptlrpc_enc_pool_put_pages(struct ptlrpc_bulk_desc *desc)
 792 {
 793 }
 794
 795 int sptlrpc_enc_pool_init(void)
 796 {
 797         return 0;
 798 }
 799
 800 void sptlrpc_enc_pool_fini(void)
 801 {
 802 }
 803 #endif
 804
 805 /****************************************
 806  * Helpers to assist policy modules to  *
 807  * implement checksum funcationality    *
 808  ****************************************/
 809
 810 static struct sptlrpc_hash_type hash_types[] = {
 811         [BULK_HASH_ALG_NULL]    = { "null",     "null",         0 },
 812         [BULK_HASH_ALG_ADLER32] = { "adler32",  "adler32",      4 },
 813         [BULK_HASH_ALG_CRC32]   = { "crc32",    "crc32",        4 },
 814         [BULK_HASH_ALG_MD5]     = { "md5",      "md5",          16 },
 815         [BULK_HASH_ALG_SHA1]    = { "sha1",     "sha1",         20 },
 816         [BULK_HASH_ALG_SHA256]  = { "sha256",   "sha256",       32 },
 817         [BULK_HASH_ALG_SHA384]  = { "sha384",   "sha384",       48 },
 818         [BULK_HASH_ALG_SHA512]  = { "sha512",   "sha512",       64 },
 819 };
 820
 821 const struct sptlrpc_hash_type *sptlrpc_get_hash_type(__u8 hash_alg)
 822 {
 823         struct sptlrpc_hash_type *ht;
 824
 825         if (hash_alg < BULK_HASH_ALG_MAX) {
 826                 ht = &hash_types[hash_alg];
 827                 if (ht->sht_tfm_name)
 828                         return ht;
 829         }
 830         return NULL;
 831 }
 832 EXPORT_SYMBOL(sptlrpc_get_hash_type);
 833
 834 const char * sptlrpc_get_hash_name(__u8 hash_alg)
 835 {
 836         const struct sptlrpc_hash_type *ht;
 837
 838         ht = sptlrpc_get_hash_type(hash_alg);
 839         if (ht)
 840                 return ht->sht_name;
 841         else
 842                 return "unknown";
 843 }
 844 EXPORT_SYMBOL(sptlrpc_get_hash_name);
 845
 846 __u8 sptlrpc_get_hash_alg(const char *algname)
 847 {
 848         int     i;
 849
 850         for (i = 0; i < BULK_HASH_ALG_MAX; i++)
 851                 if (!strcmp(hash_types[i].sht_name, algname))
 852                         break;
 853         return i;
 854 }
 855 EXPORT_SYMBOL(sptlrpc_get_hash_alg);
 856
 857 int bulk_sec_desc_unpack(struct lustre_msg *msg, int offset, int swabbed)
 858 {
 859         struct ptlrpc_bulk_sec_desc *bsd;
 860         int                          size = msg->lm_buflens[offset];
 861
 862         bsd = lustre_msg_buf(msg, offset, sizeof(*bsd));
 863         if (bsd == NULL) {
 864                 CERROR("Invalid bulk sec desc: size %d\n", size);
 865                 return -EINVAL;
 866         }
 867
 868         if (swabbed) {
 869                 __swab32s(&bsd->bsd_nob);
 870         }
 871
 872         if (unlikely(bsd->bsd_version != 0)) {
 873                 CERROR("Unexpected version %u\n", bsd->bsd_version);
 874                 return -EPROTO;
 875         }
 876
 877         if (unlikely(bsd->bsd_type >= SPTLRPC_BULK_MAX)) {
 878                 CERROR("Invalid type %u\n", bsd->bsd_type);
 879                 return -EPROTO;
 880         }
 881
 882         /* FIXME more sanity check here */
 883
 884         if (unlikely(bsd->bsd_svc != SPTLRPC_BULK_SVC_NULL &&
 885                      bsd->bsd_svc != SPTLRPC_BULK_SVC_INTG &&
 886                      bsd->bsd_svc != SPTLRPC_BULK_SVC_PRIV)) {
 887                 CERROR("Invalid svc %u\n", bsd->bsd_svc);
 888                 return -EPROTO;
 889         }
 890
 891         return 0;
 892 }
 893 EXPORT_SYMBOL(bulk_sec_desc_unpack);
 894
 895 #ifdef __KERNEL__
 896
 897 #ifdef HAVE_ADLER
 898 static int do_bulk_checksum_adler32(struct ptlrpc_bulk_desc *desc, void *buf)
 899 {
 900         struct page    *page;
 901         int             off;
 902         char           *ptr;
 903         __u32           adler32 = 1;
 904         int             len, i;
 905
 906         for (i = 0; i < desc->bd_iov_count; i++) {
 907                 page = desc->bd_iov[i].kiov_page;
 908                 off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK;
 909                 ptr = cfs_kmap(page) + off;
 910                 len = desc->bd_iov[i].kiov_len;
 911
 912                 adler32 = adler32(adler32, ptr, len);
 913
 914                 cfs_kunmap(page);
 915         }
 916
 917         adler32 = cpu_to_le32(adler32);
 918         memcpy(buf, &adler32, sizeof(adler32));
 919         return 0;
 920 }
 921 #endif
 922
 923 static int do_bulk_checksum_crc32(struct ptlrpc_bulk_desc *desc, void *buf)
 924 {
 925         struct page    *page;
 926         int             off;
 927         char           *ptr;
 928         __u32           crc32 = ~0;
 929         int             len, i;
 930
 931         for (i = 0; i < desc->bd_iov_count; i++) {
 932                 page = desc->bd_iov[i].kiov_page;
 933                 off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK;
 934                 ptr = cfs_kmap(page) + off;
 935                 len = desc->bd_iov[i].kiov_len;
 936
 937                 crc32 = crc32_le(crc32, ptr, len);
 938
 939                 cfs_kunmap(page);
 940         }
 941
 942         crc32 = cpu_to_le32(crc32);
 943         memcpy(buf, &crc32, sizeof(crc32));
 944         return 0;
 945 }
 946
 947 int sptlrpc_get_bulk_checksum(struct ptlrpc_bulk_desc *desc, __u8 alg,
 948                               void *buf, int buflen)
 949 {
 950         struct hash_desc    hdesc;
 951         int                 hashsize;
 952         char                hashbuf[64];
 953         struct scatterlist  sl;
 954         int                 i;
 955
 956         LASSERT(alg > BULK_HASH_ALG_NULL && alg < BULK_HASH_ALG_MAX);
 957         LASSERT(buflen >= 4);
 958
 959         switch (alg) {
 960         case BULK_HASH_ALG_ADLER32:
 961 #ifdef HAVE_ADLER
 962                 return do_bulk_checksum_adler32(desc, buf);
 963 #else
 964                 CERROR("Adler32 not supported\n");
 965                 return -EINVAL;
 966 #endif
 967         case BULK_HASH_ALG_CRC32:
 968                 return do_bulk_checksum_crc32(desc, buf);
 969         }
 970
 971         hdesc.tfm = ll_crypto_alloc_hash(hash_types[alg].sht_tfm_name, 0, 0);
 972         if (hdesc.tfm == NULL) {
 973                 CERROR("Unable to allocate TFM %s\n", hash_types[alg].sht_name);
 974                 return -ENOMEM;
 975         }
 976
 977         hdesc.flags = 0;
 978         ll_crypto_hash_init(&hdesc);
 979
 980         hashsize = ll_crypto_hash_digestsize(hdesc.tfm);
 981
 982         for (i = 0; i < desc->bd_iov_count; i++) {
 983                 sg_set_page(&sl, desc->bd_iov[i].kiov_page,
 984                              desc->bd_iov[i].kiov_len,
 985                              desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK);
 986                 ll_crypto_hash_update(&hdesc, &sl, sl.length);
 987         }
 988
 989         if (hashsize > buflen) {
 990                 ll_crypto_hash_final(&hdesc, hashbuf);
 991                 memcpy(buf, hashbuf, buflen);
 992         } else {
 993                 ll_crypto_hash_final(&hdesc, buf);
 994         }
 995
 996         ll_crypto_free_hash(hdesc.tfm);
 997         return 0;
 998 }
 999 EXPORT_SYMBOL(sptlrpc_get_bulk_checksum);
1000
1001 #else /* !__KERNEL__ */
1002
1003 int sptlrpc_get_bulk_checksum(struct ptlrpc_bulk_desc *desc, __u8 alg,
1004                               void *buf, int buflen)
1005 {
1006         __u32   csum32;
1007         int     i;
1008
1009         LASSERT(alg == BULK_HASH_ALG_ADLER32 || alg == BULK_HASH_ALG_CRC32);
1010
1011         if (alg == BULK_HASH_ALG_ADLER32)
1012                 csum32 = 1;
1013         else
1014                 csum32 = ~0;
1015
1016         for (i = 0; i < desc->bd_iov_count; i++) {
1017                 unsigned char *ptr = desc->bd_iov[i].iov_base;
1018                 int len = desc->bd_iov[i].iov_len;
1019
1020                 switch (alg) {
1021                 case BULK_HASH_ALG_ADLER32:
1022 #ifdef HAVE_ADLER
1023                         csum32 = adler32(csum32, ptr, len);
1024 #else
1025                         CERROR("Adler32 not supported\n");
1026                         return -EINVAL;
1027 #endif
1028                         break;
1029                 case BULK_HASH_ALG_CRC32:
1030                         csum32 = crc32_le(csum32, ptr, len);
1031                         break;
1032                 }
1033         }
1034
1035         csum32 = cpu_to_le32(csum32);
1036         memcpy(buf, &csum32, sizeof(csum32));
1037         return 0;
1038 }
1039
1040 #endif /* __KERNEL__ */