lustre/ptlrpc/sec_bulk.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/ptlrpc/sec_bulk.c
  37  *
  38  * Author: Eric Mei <ericm@clusterfs.com>
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_SEC
  42
  43 #include <libcfs/libcfs.h>
  44 #ifndef __KERNEL__
  45 #include <liblustre.h>
  46 #include <libcfs/list.h>
  47 #else
  48 #include <linux/crypto.h>
  49 #endif
  50
  51 #include <obd.h>
  52 #include <obd_cksum.h>
  53 #include <obd_class.h>
  54 #include <obd_support.h>
  55 #include <lustre_net.h>
  56 #include <lustre_import.h>
  57 #include <lustre_dlm.h>
  58 #include <lustre_sec.h>
  59
  60 #include "ptlrpc_internal.h"
  61
  62 /****************************************
  63  * bulk encryption page pools           *
  64  ****************************************/
  65
  66 #ifdef __KERNEL__
  67
  68 #define PTRS_PER_PAGE   (PAGE_CACHE_SIZE / sizeof(void *))
  69 #define PAGES_PER_POOL  (PTRS_PER_PAGE)
  70
  71 #define IDLE_IDX_MAX            (100)
  72 #define IDLE_IDX_WEIGHT         (3)
  73
  74 #define CACHE_QUIESCENT_PERIOD  (20)
  75
  76 static struct ptlrpc_enc_page_pool {
  77         /*
  78          * constants
  79          */
  80         unsigned long    epp_max_pages;   /* maximum pages can hold, const */
  81         unsigned int     epp_max_pools;   /* number of pools, const */
  82
  83         /*
  84          * wait queue in case of not enough free pages.
  85          */
  86         wait_queue_head_t    epp_waitq;       /* waiting threads */
  87         unsigned int     epp_waitqlen;    /* wait queue length */
  88         unsigned long    epp_pages_short; /* # of pages wanted of in-q users */
  89         unsigned int     epp_growing:1;   /* during adding pages */
  90
  91         /*
  92          * indicating how idle the pools are, from 0 to MAX_IDLE_IDX
  93          * this is counted based on each time when getting pages from
  94          * the pools, not based on time. which means in case that system
  95          * is idled for a while but the idle_idx might still be low if no
  96          * activities happened in the pools.
  97          */
  98         unsigned long    epp_idle_idx;
  99
 100         /* last shrink time due to mem tight */
 101         long             epp_last_shrink;
 102         long             epp_last_access;
 103
 104         /*
 105          * in-pool pages bookkeeping
 106          */
 107         spinlock_t       epp_lock;         /* protect following fields */
 108         unsigned long    epp_total_pages; /* total pages in pools */
 109         unsigned long    epp_free_pages;  /* current pages available */
 110
 111         /*
 112          * statistics
 113          */
 114         unsigned long    epp_st_max_pages;      /* # of pages ever reached */
 115         unsigned int     epp_st_grows;          /* # of grows */
 116         unsigned int     epp_st_grow_fails;     /* # of add pages failures */
 117         unsigned int     epp_st_shrinks;        /* # of shrinks */
 118         unsigned long    epp_st_access;         /* # of access */
 119         unsigned long    epp_st_missings;       /* # of cache missing */
 120         unsigned long    epp_st_lowfree;        /* lowest free pages reached */
 121         unsigned int     epp_st_max_wqlen;      /* highest waitqueue length */
 122         cfs_time_t       epp_st_max_wait;       /* in jeffies */
 123         /*
 124          * pointers to pools
 125          */
 126         struct page    ***epp_pools;
 127 } page_pools;
 128
 129 /*
 130  * memory shrinker
 131  */
 132 const int pools_shrinker_seeks = DEFAULT_SEEKS;
 133 static struct shrinker *pools_shrinker;
 134
 135
 136 /*
 137  * /proc/fs/lustre/sptlrpc/encrypt_page_pools
 138  */
 139 int sptlrpc_proc_enc_pool_seq_show(struct seq_file *m, void *v)
 140 {
 141         int     rc;
 142
 143         spin_lock(&page_pools.epp_lock);
 144
 145         rc = seq_printf(m,
 146                       "physical pages:          %lu\n"
 147                       "pages per pool:          %lu\n"
 148                       "max pages:               %lu\n"
 149                       "max pools:               %u\n"
 150                       "total pages:             %lu\n"
 151                       "total free:              %lu\n"
 152                       "idle index:              %lu/100\n"
 153                       "last shrink:             %lds\n"
 154                       "last access:             %lds\n"
 155                       "max pages reached:       %lu\n"
 156                       "grows:                   %u\n"
 157                       "grows failure:           %u\n"
 158                       "shrinks:                 %u\n"
 159                       "cache access:            %lu\n"
 160                       "cache missing:           %lu\n"
 161                       "low free mark:           %lu\n"
 162                       "max waitqueue depth:     %u\n"
 163                       "max wait time:           "CFS_TIME_T"/%u\n"
 164                       ,
 165                       totalram_pages,
 166                       PAGES_PER_POOL,
 167                       page_pools.epp_max_pages,
 168                       page_pools.epp_max_pools,
 169                       page_pools.epp_total_pages,
 170                       page_pools.epp_free_pages,
 171                       page_pools.epp_idle_idx,
 172                       cfs_time_current_sec() - page_pools.epp_last_shrink,
 173                       cfs_time_current_sec() - page_pools.epp_last_access,
 174                       page_pools.epp_st_max_pages,
 175                       page_pools.epp_st_grows,
 176                       page_pools.epp_st_grow_fails,
 177                       page_pools.epp_st_shrinks,
 178                       page_pools.epp_st_access,
 179                       page_pools.epp_st_missings,
 180                       page_pools.epp_st_lowfree,
 181                       page_pools.epp_st_max_wqlen,
 182                       page_pools.epp_st_max_wait, HZ
 183                      );
 184
 185         spin_unlock(&page_pools.epp_lock);
 186         return rc;
 187 }
 188
 189 static void enc_pools_release_free_pages(long npages)
 190 {
 191         int     p_idx, g_idx;
 192         int     p_idx_max1, p_idx_max2;
 193
 194         LASSERT(npages > 0);
 195         LASSERT(npages <= page_pools.epp_free_pages);
 196         LASSERT(page_pools.epp_free_pages <= page_pools.epp_total_pages);
 197
 198         /* max pool index before the release */
 199         p_idx_max2 = (page_pools.epp_total_pages - 1) / PAGES_PER_POOL;
 200
 201         page_pools.epp_free_pages -= npages;
 202         page_pools.epp_total_pages -= npages;
 203
 204         /* max pool index after the release */
 205         p_idx_max1 = page_pools.epp_total_pages == 0 ? -1 :
 206                      ((page_pools.epp_total_pages - 1) / PAGES_PER_POOL);
 207
 208         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 209         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 210         LASSERT(page_pools.epp_pools[p_idx]);
 211
 212         while (npages--) {
 213                 LASSERT(page_pools.epp_pools[p_idx]);
 214                 LASSERT(page_pools.epp_pools[p_idx][g_idx] != NULL);
 215
 216                 __free_page(page_pools.epp_pools[p_idx][g_idx]);
 217                 page_pools.epp_pools[p_idx][g_idx] = NULL;
 218
 219                 if (++g_idx == PAGES_PER_POOL) {
 220                         p_idx++;
 221                         g_idx = 0;
 222                 }
 223         }
 224
 225         /* free unused pools */
 226         while (p_idx_max1 < p_idx_max2) {
 227                 LASSERT(page_pools.epp_pools[p_idx_max2]);
 228                 OBD_FREE(page_pools.epp_pools[p_idx_max2], PAGE_CACHE_SIZE);
 229                 page_pools.epp_pools[p_idx_max2] = NULL;
 230                 p_idx_max2--;
 231         }
 232 }
 233
 234 /*
 235  * could be called frequently for query (@nr_to_scan == 0).
 236  * we try to keep at least PTLRPC_MAX_BRW_PAGES pages in the pool.
 237  */
 238 static int enc_pools_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
 239 {
 240         if (unlikely(shrink_param(sc, nr_to_scan) != 0)) {
 241                 spin_lock(&page_pools.epp_lock);
 242                 shrink_param(sc, nr_to_scan) = min_t(unsigned long,
 243                                                    shrink_param(sc, nr_to_scan),
 244                                                    page_pools.epp_free_pages -
 245                                                    PTLRPC_MAX_BRW_PAGES);
 246                 if (shrink_param(sc, nr_to_scan) > 0) {
 247                         enc_pools_release_free_pages(shrink_param(sc,
 248                                                                   nr_to_scan));
 249                         CDEBUG(D_SEC, "released %ld pages, %ld left\n",
 250                                (long)shrink_param(sc, nr_to_scan),
 251                                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(struct page ***pools, int npools)
 285 {
 286         unsigned long cleaned = 0;
 287         int           i, j;
 288
 289         for (i = 0; i < npools; i++) {
 290                 if (pools[i]) {
 291                         for (j = 0; j < PAGES_PER_POOL; j++) {
 292                                 if (pools[i][j]) {
 293                                         __free_page(pools[i][j]);
 294                                         cleaned++;
 295                                 }
 296                         }
 297                         OBD_FREE(pools[i], PAGE_CACHE_SIZE);
 298                         pools[i] = NULL;
 299                 }
 300         }
 301
 302         return cleaned;
 303 }
 304
 305 /*
 306  * merge @npools pointed by @pools which contains @npages new pages
 307  * into current pools.
 308  *
 309  * we have options to avoid most memory copy with some tricks. but we choose
 310  * the simplest way to avoid complexity. It's not frequently called.
 311  */
 312 static void enc_pools_insert(struct page ***pools, int npools, int npages)
 313 {
 314         int     freeslot;
 315         int     op_idx, np_idx, og_idx, ng_idx;
 316         int     cur_npools, end_npools;
 317
 318         LASSERT(npages > 0);
 319         LASSERT(page_pools.epp_total_pages+npages <= page_pools.epp_max_pages);
 320         LASSERT(npages_to_npools(npages) == npools);
 321         LASSERT(page_pools.epp_growing);
 322
 323         spin_lock(&page_pools.epp_lock);
 324
 325         /*
 326          * (1) fill all the free slots of current pools.
 327          */
 328         /* free slots are those left by rent pages, and the extra ones with
 329          * index >= total_pages, locate at the tail of last pool. */
 330         freeslot = page_pools.epp_total_pages % PAGES_PER_POOL;
 331         if (freeslot != 0)
 332                 freeslot = PAGES_PER_POOL - freeslot;
 333         freeslot += page_pools.epp_total_pages - page_pools.epp_free_pages;
 334
 335         op_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 336         og_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 337         np_idx = npools - 1;
 338         ng_idx = (npages - 1) % PAGES_PER_POOL;
 339
 340         while (freeslot) {
 341                 LASSERT(page_pools.epp_pools[op_idx][og_idx] == NULL);
 342                 LASSERT(pools[np_idx][ng_idx] != NULL);
 343
 344                 page_pools.epp_pools[op_idx][og_idx] = pools[np_idx][ng_idx];
 345                 pools[np_idx][ng_idx] = NULL;
 346
 347                 freeslot--;
 348
 349                 if (++og_idx == PAGES_PER_POOL) {
 350                         op_idx++;
 351                         og_idx = 0;
 352                 }
 353                 if (--ng_idx < 0) {
 354                         if (np_idx == 0)
 355                                 break;
 356                         np_idx--;
 357                         ng_idx = PAGES_PER_POOL - 1;
 358                 }
 359         }
 360
 361         /*
 362          * (2) add pools if needed.
 363          */
 364         cur_npools = (page_pools.epp_total_pages + PAGES_PER_POOL - 1) /
 365                      PAGES_PER_POOL;
 366         end_npools = (page_pools.epp_total_pages + npages + PAGES_PER_POOL -1) /
 367                      PAGES_PER_POOL;
 368         LASSERT(end_npools <= page_pools.epp_max_pools);
 369
 370         np_idx = 0;
 371         while (cur_npools < end_npools) {
 372                 LASSERT(page_pools.epp_pools[cur_npools] == NULL);
 373                 LASSERT(np_idx < npools);
 374                 LASSERT(pools[np_idx] != NULL);
 375
 376                 page_pools.epp_pools[cur_npools++] = pools[np_idx];
 377                 pools[np_idx++] = NULL;
 378         }
 379
 380         page_pools.epp_total_pages += npages;
 381         page_pools.epp_free_pages += npages;
 382         page_pools.epp_st_lowfree = page_pools.epp_free_pages;
 383
 384         if (page_pools.epp_total_pages > page_pools.epp_st_max_pages)
 385                 page_pools.epp_st_max_pages = page_pools.epp_total_pages;
 386
 387         CDEBUG(D_SEC, "add %d pages to total %lu\n", npages,
 388                page_pools.epp_total_pages);
 389
 390         spin_unlock(&page_pools.epp_lock);
 391 }
 392
 393 static int enc_pools_add_pages(int npages)
 394 {
 395         static DEFINE_MUTEX(add_pages_mutex);
 396         struct page   ***pools;
 397         int             npools, alloced = 0;
 398         int             i, j, rc = -ENOMEM;
 399
 400         if (npages < PTLRPC_MAX_BRW_PAGES)
 401                 npages = PTLRPC_MAX_BRW_PAGES;
 402
 403         mutex_lock(&add_pages_mutex);
 404
 405         if (npages + page_pools.epp_total_pages > page_pools.epp_max_pages)
 406                 npages = page_pools.epp_max_pages - page_pools.epp_total_pages;
 407         LASSERT(npages > 0);
 408
 409         page_pools.epp_st_grows++;
 410
 411         npools = npages_to_npools(npages);
 412         OBD_ALLOC(pools, npools * sizeof(*pools));
 413         if (pools == NULL)
 414                 goto out;
 415
 416         for (i = 0; i < npools; i++) {
 417                 OBD_ALLOC(pools[i], PAGE_CACHE_SIZE);
 418                 if (pools[i] == NULL)
 419                         goto out_pools;
 420
 421                 for (j = 0; j < PAGES_PER_POOL && alloced < npages; j++) {
 422                         pools[i][j] = alloc_page(GFP_NOFS |
 423                                                  __GFP_HIGHMEM);
 424                         if (pools[i][j] == NULL)
 425                                 goto out_pools;
 426
 427                         alloced++;
 428                 }
 429         }
 430         LASSERT(alloced == npages);
 431
 432         enc_pools_insert(pools, npools, npages);
 433         CDEBUG(D_SEC, "added %d pages into pools\n", npages);
 434         rc = 0;
 435
 436 out_pools:
 437         enc_pools_cleanup(pools, npools);
 438         OBD_FREE(pools, npools * sizeof(*pools));
 439 out:
 440         if (rc) {
 441                 page_pools.epp_st_grow_fails++;
 442                 CERROR("Failed to allocate %d enc pages\n", npages);
 443         }
 444
 445         mutex_unlock(&add_pages_mutex);
 446         return rc;
 447 }
 448
 449 static inline void enc_pools_wakeup(void)
 450 {
 451         LASSERT(spin_is_locked(&page_pools.epp_lock));
 452
 453         if (unlikely(page_pools.epp_waitqlen)) {
 454                 LASSERT(waitqueue_active(&page_pools.epp_waitq));
 455                 wake_up_all(&page_pools.epp_waitq);
 456         }
 457 }
 458
 459 static int enc_pools_should_grow(int page_needed, long now)
 460 {
 461         /* don't grow if someone else is growing the pools right now,
 462          * or the pools has reached its full capacity
 463          */
 464         if (page_pools.epp_growing ||
 465             page_pools.epp_total_pages == page_pools.epp_max_pages)
 466                 return 0;
 467
 468         /* if total pages is not enough, we need to grow */
 469         if (page_pools.epp_total_pages < page_needed)
 470                 return 1;
 471
 472         /*
 473          * we wanted to return 0 here if there was a shrink just happened
 474          * moment ago, but this may cause deadlock if both client and ost
 475          * live on single node.
 476          */
 477 #if 0
 478         if (now - page_pools.epp_last_shrink < 2)
 479                 return 0;
 480 #endif
 481
 482         /*
 483          * here we perhaps need consider other factors like wait queue
 484          * length, idle index, etc. ?
 485          */
 486
 487         /* grow the pools in any other cases */
 488         return 1;
 489 }
 490
 491 /*
 492  * we allocate the requested pages atomically.
 493  */
 494 int sptlrpc_enc_pool_get_pages(struct ptlrpc_bulk_desc *desc)
 495 {
 496         wait_queue_t  waitlink;
 497         unsigned long   this_idle = -1;
 498         cfs_time_t      tick = 0;
 499         long            now;
 500         int             p_idx, g_idx;
 501         int             i;
 502
 503         LASSERT(desc->bd_iov_count > 0);
 504         LASSERT(desc->bd_iov_count <= page_pools.epp_max_pages);
 505
 506         /* resent bulk, enc iov might have been allocated previously */
 507         if (desc->bd_enc_iov != NULL)
 508                 return 0;
 509
 510         OBD_ALLOC(desc->bd_enc_iov,
 511                   desc->bd_iov_count * sizeof(*desc->bd_enc_iov));
 512         if (desc->bd_enc_iov == NULL)
 513                 return -ENOMEM;
 514
 515         spin_lock(&page_pools.epp_lock);
 516
 517         page_pools.epp_st_access++;
 518 again:
 519         if (unlikely(page_pools.epp_free_pages < desc->bd_iov_count)) {
 520                 if (tick == 0)
 521                         tick = cfs_time_current();
 522
 523                 now = cfs_time_current_sec();
 524
 525                 page_pools.epp_st_missings++;
 526                 page_pools.epp_pages_short += desc->bd_iov_count;
 527
 528                 if (enc_pools_should_grow(desc->bd_iov_count, now)) {
 529                         page_pools.epp_growing = 1;
 530
 531                         spin_unlock(&page_pools.epp_lock);
 532                         enc_pools_add_pages(page_pools.epp_pages_short / 2);
 533                         spin_lock(&page_pools.epp_lock);
 534
 535                         page_pools.epp_growing = 0;
 536
 537                         enc_pools_wakeup();
 538                 } else {
 539                         if (++page_pools.epp_waitqlen >
 540                             page_pools.epp_st_max_wqlen)
 541                                 page_pools.epp_st_max_wqlen =
 542                                                 page_pools.epp_waitqlen;
 543
 544                         set_current_state(TASK_UNINTERRUPTIBLE);
 545                         init_waitqueue_entry_current(&waitlink);
 546                         add_wait_queue(&page_pools.epp_waitq, &waitlink);
 547
 548                         spin_unlock(&page_pools.epp_lock);
 549                         waitq_wait(&waitlink, TASK_UNINTERRUPTIBLE);
 550                         remove_wait_queue(&page_pools.epp_waitq, &waitlink);
 551                         LASSERT(page_pools.epp_waitqlen > 0);
 552                         spin_lock(&page_pools.epp_lock);
 553                         page_pools.epp_waitqlen--;
 554                 }
 555
 556                 LASSERT(page_pools.epp_pages_short >= desc->bd_iov_count);
 557                 page_pools.epp_pages_short -= desc->bd_iov_count;
 558
 559                 this_idle = 0;
 560                 goto again;
 561         }
 562
 563         /* record max wait time */
 564         if (unlikely(tick != 0)) {
 565                 tick = cfs_time_current() - tick;
 566                 if (tick > page_pools.epp_st_max_wait)
 567                         page_pools.epp_st_max_wait = tick;
 568         }
 569
 570         /* proceed with rest of allocation */
 571         page_pools.epp_free_pages -= desc->bd_iov_count;
 572
 573         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 574         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 575
 576         for (i = 0; i < desc->bd_iov_count; i++) {
 577                 LASSERT(page_pools.epp_pools[p_idx][g_idx] != NULL);
 578                 desc->bd_enc_iov[i].kiov_page =
 579                                         page_pools.epp_pools[p_idx][g_idx];
 580                 page_pools.epp_pools[p_idx][g_idx] = NULL;
 581
 582                 if (++g_idx == PAGES_PER_POOL) {
 583                         p_idx++;
 584                         g_idx = 0;
 585                 }
 586         }
 587
 588         if (page_pools.epp_free_pages < page_pools.epp_st_lowfree)
 589                 page_pools.epp_st_lowfree = page_pools.epp_free_pages;
 590
 591         /*
 592          * new idle index = (old * weight + new) / (weight + 1)
 593          */
 594         if (this_idle == -1) {
 595                 this_idle = page_pools.epp_free_pages * IDLE_IDX_MAX /
 596                             page_pools.epp_total_pages;
 597         }
 598         page_pools.epp_idle_idx = (page_pools.epp_idle_idx * IDLE_IDX_WEIGHT +
 599                                    this_idle) /
 600                                   (IDLE_IDX_WEIGHT + 1);
 601
 602         page_pools.epp_last_access = cfs_time_current_sec();
 603
 604         spin_unlock(&page_pools.epp_lock);
 605         return 0;
 606 }
 607 EXPORT_SYMBOL(sptlrpc_enc_pool_get_pages);
 608
 609 void sptlrpc_enc_pool_put_pages(struct ptlrpc_bulk_desc *desc)
 610 {
 611         int     p_idx, g_idx;
 612         int     i;
 613
 614         if (desc->bd_enc_iov == NULL)
 615                 return;
 616
 617         LASSERT(desc->bd_iov_count > 0);
 618
 619         spin_lock(&page_pools.epp_lock);
 620
 621         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 622         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 623
 624         LASSERT(page_pools.epp_free_pages + desc->bd_iov_count <=
 625                 page_pools.epp_total_pages);
 626         LASSERT(page_pools.epp_pools[p_idx]);
 627
 628         for (i = 0; i < desc->bd_iov_count; i++) {
 629                 LASSERT(desc->bd_enc_iov[i].kiov_page != NULL);
 630                 LASSERT(g_idx != 0 || page_pools.epp_pools[p_idx]);
 631                 LASSERT(page_pools.epp_pools[p_idx][g_idx] == NULL);
 632
 633                 page_pools.epp_pools[p_idx][g_idx] =
 634                                         desc->bd_enc_iov[i].kiov_page;
 635
 636                 if (++g_idx == PAGES_PER_POOL) {
 637                         p_idx++;
 638                         g_idx = 0;
 639                 }
 640         }
 641
 642         page_pools.epp_free_pages += desc->bd_iov_count;
 643
 644         enc_pools_wakeup();
 645
 646         spin_unlock(&page_pools.epp_lock);
 647
 648         OBD_FREE(desc->bd_enc_iov,
 649                  desc->bd_iov_count * sizeof(*desc->bd_enc_iov));
 650         desc->bd_enc_iov = NULL;
 651 }
 652 EXPORT_SYMBOL(sptlrpc_enc_pool_put_pages);
 653
 654 /*
 655  * we don't do much stuff for add_user/del_user anymore, except adding some
 656  * initial pages in add_user() if current pools are empty, rest would be
 657  * handled by the pools's self-adaption.
 658  */
 659 int sptlrpc_enc_pool_add_user(void)
 660 {
 661         int     need_grow = 0;
 662
 663         spin_lock(&page_pools.epp_lock);
 664         if (page_pools.epp_growing == 0 && page_pools.epp_total_pages == 0) {
 665                 page_pools.epp_growing = 1;
 666                 need_grow = 1;
 667         }
 668         spin_unlock(&page_pools.epp_lock);
 669
 670         if (need_grow) {
 671                 enc_pools_add_pages(PTLRPC_MAX_BRW_PAGES +
 672                                     PTLRPC_MAX_BRW_PAGES);
 673
 674                 spin_lock(&page_pools.epp_lock);
 675                 page_pools.epp_growing = 0;
 676                 enc_pools_wakeup();
 677                 spin_unlock(&page_pools.epp_lock);
 678         }
 679         return 0;
 680 }
 681 EXPORT_SYMBOL(sptlrpc_enc_pool_add_user);
 682
 683 int sptlrpc_enc_pool_del_user(void)
 684 {
 685         return 0;
 686 }
 687 EXPORT_SYMBOL(sptlrpc_enc_pool_del_user);
 688
 689 static inline void enc_pools_alloc(void)
 690 {
 691         LASSERT(page_pools.epp_max_pools);
 692         OBD_ALLOC_LARGE(page_pools.epp_pools,
 693                         page_pools.epp_max_pools *
 694                         sizeof(*page_pools.epp_pools));
 695 }
 696
 697 static inline void enc_pools_free(void)
 698 {
 699         LASSERT(page_pools.epp_max_pools);
 700         LASSERT(page_pools.epp_pools);
 701
 702         OBD_FREE_LARGE(page_pools.epp_pools,
 703                        page_pools.epp_max_pools *
 704                        sizeof(*page_pools.epp_pools));
 705 }
 706
 707 int sptlrpc_enc_pool_init(void)
 708 {
 709         /*
 710          * maximum capacity is 1/8 of total physical memory.
 711          * is the 1/8 a good number?
 712          */
 713         page_pools.epp_max_pages = totalram_pages / 8;
 714         page_pools.epp_max_pools = npages_to_npools(page_pools.epp_max_pages);
 715
 716         init_waitqueue_head(&page_pools.epp_waitq);
 717         page_pools.epp_waitqlen = 0;
 718         page_pools.epp_pages_short = 0;
 719
 720         page_pools.epp_growing = 0;
 721
 722         page_pools.epp_idle_idx = 0;
 723         page_pools.epp_last_shrink = cfs_time_current_sec();
 724         page_pools.epp_last_access = cfs_time_current_sec();
 725
 726         spin_lock_init(&page_pools.epp_lock);
 727         page_pools.epp_total_pages = 0;
 728         page_pools.epp_free_pages = 0;
 729
 730         page_pools.epp_st_max_pages = 0;
 731         page_pools.epp_st_grows = 0;
 732         page_pools.epp_st_grow_fails = 0;
 733         page_pools.epp_st_shrinks = 0;
 734         page_pools.epp_st_access = 0;
 735         page_pools.epp_st_missings = 0;
 736         page_pools.epp_st_lowfree = 0;
 737         page_pools.epp_st_max_wqlen = 0;
 738         page_pools.epp_st_max_wait = 0;
 739
 740         enc_pools_alloc();
 741         if (page_pools.epp_pools == NULL)
 742                 return -ENOMEM;
 743
 744         pools_shrinker = set_shrinker(pools_shrinker_seeks,
 745                                           enc_pools_shrink);
 746         if (pools_shrinker == NULL) {
 747                 enc_pools_free();
 748                 return -ENOMEM;
 749         }
 750
 751         return 0;
 752 }
 753
 754 void sptlrpc_enc_pool_fini(void)
 755 {
 756         unsigned long cleaned, npools;
 757
 758         LASSERT(pools_shrinker);
 759         LASSERT(page_pools.epp_pools);
 760         LASSERT(page_pools.epp_total_pages == page_pools.epp_free_pages);
 761
 762         remove_shrinker(pools_shrinker);
 763
 764         npools = npages_to_npools(page_pools.epp_total_pages);
 765         cleaned = enc_pools_cleanup(page_pools.epp_pools, npools);
 766         LASSERT(cleaned == page_pools.epp_total_pages);
 767
 768         enc_pools_free();
 769
 770         if (page_pools.epp_st_access > 0) {
 771                 CDEBUG(D_SEC,
 772                        "max pages %lu, grows %u, grow fails %u, shrinks %u, "
 773                        "access %lu, missing %lu, max qlen %u, max wait "
 774                        CFS_TIME_T"/%d\n",
 775                        page_pools.epp_st_max_pages, page_pools.epp_st_grows,
 776                        page_pools.epp_st_grow_fails,
 777                        page_pools.epp_st_shrinks, page_pools.epp_st_access,
 778                        page_pools.epp_st_missings, page_pools.epp_st_max_wqlen,
 779                        page_pools.epp_st_max_wait, HZ);
 780         }
 781 }
 782
 783 #else /* !__KERNEL__ */
 784
 785 int sptlrpc_enc_pool_get_pages(struct ptlrpc_bulk_desc *desc)
 786 {
 787         return 0;
 788 }
 789
 790 void sptlrpc_enc_pool_put_pages(struct ptlrpc_bulk_desc *desc)
 791 {
 792 }
 793
 794 int sptlrpc_enc_pool_init(void)
 795 {
 796         return 0;
 797 }
 798
 799 void sptlrpc_enc_pool_fini(void)
 800 {
 801 }
 802 #endif
 803
 804 static int cfs_hash_alg_id[] = {
 805         [BULK_HASH_ALG_NULL]    = CFS_HASH_ALG_NULL,
 806         [BULK_HASH_ALG_ADLER32] = CFS_HASH_ALG_ADLER32,
 807         [BULK_HASH_ALG_CRC32]   = CFS_HASH_ALG_CRC32,
 808         [BULK_HASH_ALG_MD5]     = CFS_HASH_ALG_MD5,
 809         [BULK_HASH_ALG_SHA1]    = CFS_HASH_ALG_SHA1,
 810         [BULK_HASH_ALG_SHA256]  = CFS_HASH_ALG_SHA256,
 811         [BULK_HASH_ALG_SHA384]  = CFS_HASH_ALG_SHA384,
 812         [BULK_HASH_ALG_SHA512]  = CFS_HASH_ALG_SHA512,
 813 };
 814 const char * sptlrpc_get_hash_name(__u8 hash_alg)
 815 {
 816         return cfs_crypto_hash_name(cfs_hash_alg_id[hash_alg]);
 817 }
 818 EXPORT_SYMBOL(sptlrpc_get_hash_name);
 819
 820 __u8 sptlrpc_get_hash_alg(const char *algname)
 821 {
 822         return cfs_crypto_hash_alg(algname);
 823 }
 824 EXPORT_SYMBOL(sptlrpc_get_hash_alg);
 825
 826 int bulk_sec_desc_unpack(struct lustre_msg *msg, int offset, int swabbed)
 827 {
 828         struct ptlrpc_bulk_sec_desc *bsd;
 829         int                          size = msg->lm_buflens[offset];
 830
 831         bsd = lustre_msg_buf(msg, offset, sizeof(*bsd));
 832         if (bsd == NULL) {
 833                 CERROR("Invalid bulk sec desc: size %d\n", size);
 834                 return -EINVAL;
 835         }
 836
 837         if (swabbed) {
 838                 __swab32s(&bsd->bsd_nob);
 839         }
 840
 841         if (unlikely(bsd->bsd_version != 0)) {
 842                 CERROR("Unexpected version %u\n", bsd->bsd_version);
 843                 return -EPROTO;
 844         }
 845
 846         if (unlikely(bsd->bsd_type >= SPTLRPC_BULK_MAX)) {
 847                 CERROR("Invalid type %u\n", bsd->bsd_type);
 848                 return -EPROTO;
 849         }
 850
 851         /* FIXME more sanity check here */
 852
 853         if (unlikely(bsd->bsd_svc != SPTLRPC_BULK_SVC_NULL &&
 854                      bsd->bsd_svc != SPTLRPC_BULK_SVC_INTG &&
 855                      bsd->bsd_svc != SPTLRPC_BULK_SVC_PRIV)) {
 856                 CERROR("Invalid svc %u\n", bsd->bsd_svc);
 857                 return -EPROTO;
 858         }
 859
 860         return 0;
 861 }
 862 EXPORT_SYMBOL(bulk_sec_desc_unpack);
 863
 864 int sptlrpc_get_bulk_checksum(struct ptlrpc_bulk_desc *desc, __u8 alg,
 865                               void *buf, int buflen)
 866 {
 867         struct cfs_crypto_hash_desc     *hdesc;
 868         int                             hashsize;
 869         char                            hashbuf[64];
 870         unsigned int                    bufsize;
 871         int                             i, err;
 872
 873         LASSERT(alg > BULK_HASH_ALG_NULL && alg < BULK_HASH_ALG_MAX);
 874         LASSERT(buflen >= 4);
 875
 876         hdesc = cfs_crypto_hash_init(cfs_hash_alg_id[alg], NULL, 0);
 877         if (IS_ERR(hdesc)) {
 878                 CERROR("Unable to initialize checksum hash %s\n",
 879                        cfs_crypto_hash_name(cfs_hash_alg_id[alg]));
 880                 return PTR_ERR(hdesc);
 881         }
 882
 883         hashsize = cfs_crypto_hash_digestsize(cfs_hash_alg_id[alg]);
 884
 885         for (i = 0; i < desc->bd_iov_count; i++) {
 886 #ifdef __KERNEL__
 887                 cfs_crypto_hash_update_page(hdesc, desc->bd_iov[i].kiov_page,
 888                                   desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK,
 889                                   desc->bd_iov[i].kiov_len);
 890 #else
 891                 cfs_crypto_hash_update(hdesc, desc->bd_iov[i].iov_base,
 892                                   desc->bd_iov[i].iov_len);
 893 #endif
 894         }
 895         if (hashsize > buflen) {
 896                 bufsize = sizeof(hashbuf);
 897                 err = cfs_crypto_hash_final(hdesc, (unsigned char *)hashbuf,
 898                                             &bufsize);
 899                 memcpy(buf, hashbuf, buflen);
 900         } else {
 901                 bufsize = buflen;
 902                 err = cfs_crypto_hash_final(hdesc, (unsigned char *)buf,
 903                                             &bufsize);
 904         }
 905
 906         if (err)
 907                 cfs_crypto_hash_final(hdesc, NULL, NULL);
 908         return err;
 909 }
 910 EXPORT_SYMBOL(sptlrpc_get_bulk_checksum);
 911
 912