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  */
 241 static int enc_pools_shrink(int nr_to_scan, unsigned int gfp_mask)
 242 {
 243         unsigned long   ret;
 244
 245         spin_lock(&page_pools.epp_lock);
 246
 247         if (nr_to_scan > page_pools.epp_free_pages)
 248                 nr_to_scan = page_pools.epp_free_pages;
 249
 250         if (nr_to_scan > 0) {
 251                 enc_pools_release_free_pages(nr_to_scan);
 252                 CDEBUG(D_SEC, "released %d pages, %ld left\n",
 253                        nr_to_scan, page_pools.epp_free_pages);
 254
 255                 page_pools.epp_st_shrinks++;
 256                 page_pools.epp_last_shrink = cfs_time_current_sec();
 257         }
 258
 259         /*
 260          * try to keep at least PTLRPC_MAX_BRW_PAGES pages in the pool
 261          */
 262         if (page_pools.epp_free_pages <= PTLRPC_MAX_BRW_PAGES) {
 263                 ret = 0;
 264                 goto out_unlock;
 265         }
 266
 267         /*
 268          * if no pool access for a long time, we consider it's fully idle
 269          */
 270         if (cfs_time_current_sec() - page_pools.epp_last_access >
 271             CACHE_QUIESCENT_PERIOD)
 272                 page_pools.epp_idle_idx = IDLE_IDX_MAX;
 273
 274         LASSERT(page_pools.epp_idle_idx <= IDLE_IDX_MAX);
 275         ret = (page_pools.epp_free_pages * page_pools.epp_idle_idx /
 276                IDLE_IDX_MAX);
 277         if (page_pools.epp_free_pages - ret < PTLRPC_MAX_BRW_PAGES)
 278                 ret = page_pools.epp_free_pages - PTLRPC_MAX_BRW_PAGES;
 279
 280 out_unlock:
 281         spin_unlock(&page_pools.epp_lock);
 282         return ret;
 283 }
 284
 285 static inline
 286 int npages_to_npools(unsigned long npages)
 287 {
 288         return (int) ((npages + PAGES_PER_POOL - 1) / PAGES_PER_POOL);
 289 }
 290
 291 /*
 292  * return how many pages cleaned up.
 293  */
 294 static unsigned long enc_pools_cleanup(cfs_page_t ***pools, int npools)
 295 {
 296         unsigned long cleaned = 0;
 297         int           i, j;
 298
 299         for (i = 0; i < npools; i++) {
 300                 if (pools[i]) {
 301                         for (j = 0; j < PAGES_PER_POOL; j++) {
 302                                 if (pools[i][j]) {
 303                                         cfs_free_page(pools[i][j]);
 304                                         cleaned++;
 305                                 }
 306                         }
 307                         OBD_FREE(pools[i], CFS_PAGE_SIZE);
 308                         pools[i] = NULL;
 309                 }
 310         }
 311
 312         return cleaned;
 313 }
 314
 315 /*
 316  * merge @npools pointed by @pools which contains @npages new pages
 317  * into current pools.
 318  *
 319  * we have options to avoid most memory copy with some tricks. but we choose
 320  * the simplest way to avoid complexity. It's not frequently called.
 321  */
 322 static void enc_pools_insert(cfs_page_t ***pools, int npools, int npages)
 323 {
 324         int     freeslot;
 325         int     op_idx, np_idx, og_idx, ng_idx;
 326         int     cur_npools, end_npools;
 327
 328         LASSERT(npages > 0);
 329         LASSERT(page_pools.epp_total_pages+npages <= page_pools.epp_max_pages);
 330         LASSERT(npages_to_npools(npages) == npools);
 331
 332         spin_lock(&page_pools.epp_lock);
 333
 334         /*
 335          * (1) fill all the free slots of current pools.
 336          */
 337         /* free slots are those left by rent pages, and the extra ones with
 338          * index >= total_pages, locate at the tail of last pool. */
 339         freeslot = page_pools.epp_total_pages % PAGES_PER_POOL;
 340         if (freeslot != 0)
 341                 freeslot = PAGES_PER_POOL - freeslot;
 342         freeslot += page_pools.epp_total_pages - page_pools.epp_free_pages;
 343
 344         op_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 345         og_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 346         np_idx = npools - 1;
 347         ng_idx = (npages - 1) % PAGES_PER_POOL;
 348
 349         while (freeslot) {
 350                 LASSERT(page_pools.epp_pools[op_idx][og_idx] == NULL);
 351                 LASSERT(pools[np_idx][ng_idx] != NULL);
 352
 353                 page_pools.epp_pools[op_idx][og_idx] = pools[np_idx][ng_idx];
 354                 pools[np_idx][ng_idx] = NULL;
 355
 356                 freeslot--;
 357
 358                 if (++og_idx == PAGES_PER_POOL) {
 359                         op_idx++;
 360                         og_idx = 0;
 361                 }
 362                 if (--ng_idx < 0) {
 363                         if (np_idx == 0)
 364                                 break;
 365                         np_idx--;
 366                         ng_idx = PAGES_PER_POOL - 1;
 367                 }
 368         }
 369
 370         /*
 371          * (2) add pools if needed.
 372          */
 373         cur_npools = (page_pools.epp_total_pages + PAGES_PER_POOL - 1) /
 374                      PAGES_PER_POOL;
 375         end_npools = (page_pools.epp_total_pages + npages + PAGES_PER_POOL -1) /
 376                      PAGES_PER_POOL;
 377         LASSERT(end_npools <= page_pools.epp_max_pools);
 378
 379         np_idx = 0;
 380         while (cur_npools < end_npools) {
 381                 LASSERT(page_pools.epp_pools[cur_npools] == NULL);
 382                 LASSERT(np_idx < npools);
 383                 LASSERT(pools[np_idx] != NULL);
 384
 385                 page_pools.epp_pools[cur_npools++] = pools[np_idx];
 386                 pools[np_idx++] = NULL;
 387         }
 388
 389         page_pools.epp_total_pages += npages;
 390         page_pools.epp_free_pages += npages;
 391         page_pools.epp_st_lowfree = page_pools.epp_free_pages;
 392
 393         if (page_pools.epp_total_pages > page_pools.epp_st_max_pages)
 394                 page_pools.epp_st_max_pages = page_pools.epp_total_pages;
 395
 396         CDEBUG(D_SEC, "add %d pages to total %lu\n", npages,
 397                page_pools.epp_total_pages);
 398
 399         spin_unlock(&page_pools.epp_lock);
 400 }
 401
 402 static int enc_pools_add_pages(int npages)
 403 {
 404         static DECLARE_MUTEX(sem_add_pages);
 405         cfs_page_t   ***pools;
 406         int             npools, alloced = 0;
 407         int             i, j, rc = -ENOMEM;
 408
 409         if (npages < PTLRPC_MAX_BRW_PAGES)
 410                 npages = PTLRPC_MAX_BRW_PAGES;
 411
 412         down(&sem_add_pages);
 413
 414         if (npages + page_pools.epp_total_pages > page_pools.epp_max_pages)
 415                 npages = page_pools.epp_max_pages - page_pools.epp_total_pages;
 416         LASSERT(npages > 0);
 417
 418         page_pools.epp_st_grows++;
 419
 420         npools = npages_to_npools(npages);
 421         OBD_ALLOC(pools, npools * sizeof(*pools));
 422         if (pools == NULL)
 423                 goto out;
 424
 425         for (i = 0; i < npools; i++) {
 426                 OBD_ALLOC(pools[i], CFS_PAGE_SIZE);
 427                 if (pools[i] == NULL)
 428                         goto out_pools;
 429
 430                 for (j = 0; j < PAGES_PER_POOL && alloced < npages; j++) {
 431                         pools[i][j] = cfs_alloc_page(CFS_ALLOC_IO |
 432                                                      CFS_ALLOC_HIGH);
 433                         if (pools[i][j] == NULL)
 434                                 goto out_pools;
 435
 436                         alloced++;
 437                 }
 438         }
 439
 440         enc_pools_insert(pools, npools, npages);
 441         CDEBUG(D_SEC, "added %d pages into pools\n", npages);
 442         rc = 0;
 443
 444 out_pools:
 445         enc_pools_cleanup(pools, npools);
 446         OBD_FREE(pools, npools * sizeof(*pools));
 447 out:
 448         if (rc) {
 449                 page_pools.epp_st_grow_fails++;
 450                 CERROR("Failed to allocate %d enc pages\n", npages);
 451         }
 452
 453         up(&sem_add_pages);
 454         return rc;
 455 }
 456
 457 static inline void enc_pools_wakeup(void)
 458 {
 459         LASSERT_SPIN_LOCKED(&page_pools.epp_lock);
 460         LASSERT(page_pools.epp_waitqlen >= 0);
 461
 462         if (unlikely(page_pools.epp_waitqlen)) {
 463                 LASSERT(cfs_waitq_active(&page_pools.epp_waitq));
 464                 cfs_waitq_broadcast(&page_pools.epp_waitq);
 465         }
 466 }
 467
 468 static int enc_pools_should_grow(int page_needed, long now)
 469 {
 470         /* don't grow if someone else is growing the pools right now,
 471          * or the pools has reached its full capacity
 472          */
 473         if (page_pools.epp_growing ||
 474             page_pools.epp_total_pages == page_pools.epp_max_pages)
 475                 return 0;
 476
 477         /* if total pages is not enough, we need to grow */
 478         if (page_pools.epp_total_pages < page_needed)
 479                 return 1;
 480
 481         /*
 482          * we wanted to return 0 here if there was a shrink just happened
 483          * moment ago, but this may cause deadlock if both client and ost
 484          * live on single node.
 485          */
 486 #if 0
 487         if (now - page_pools.epp_last_shrink < 2)
 488                 return 0;
 489 #endif
 490
 491         /*
 492          * here we perhaps need consider other factors like wait queue
 493          * length, idle index, etc. ?
 494          */
 495
 496         /* grow the pools in any other cases */
 497         return 1;
 498 }
 499
 500 /*
 501  * we allocate the requested pages atomically.
 502  */
 503 int sptlrpc_enc_pool_get_pages(struct ptlrpc_bulk_desc *desc)
 504 {
 505         cfs_waitlink_t  waitlink;
 506         unsigned long   this_idle = -1;
 507         cfs_time_t      tick = 0;
 508         long            now;
 509         int             p_idx, g_idx;
 510         int             i;
 511
 512         LASSERT(desc->bd_iov_count > 0);
 513         LASSERT(desc->bd_iov_count <= page_pools.epp_max_pages);
 514
 515         /* resent bulk, enc iov might have been allocated previously */
 516         if (desc->bd_enc_iov != NULL)
 517                 return 0;
 518
 519         OBD_ALLOC(desc->bd_enc_iov,
 520                   desc->bd_iov_count * sizeof(*desc->bd_enc_iov));
 521         if (desc->bd_enc_iov == NULL)
 522                 return -ENOMEM;
 523
 524         spin_lock(&page_pools.epp_lock);
 525
 526         page_pools.epp_st_access++;
 527 again:
 528         if (unlikely(page_pools.epp_free_pages < desc->bd_iov_count)) {
 529                 if (tick == 0)
 530                         tick = cfs_time_current();
 531
 532                 now = cfs_time_current_sec();
 533
 534                 page_pools.epp_st_missings++;
 535                 page_pools.epp_pages_short += desc->bd_iov_count;
 536
 537                 if (enc_pools_should_grow(desc->bd_iov_count, now)) {
 538                         page_pools.epp_growing = 1;
 539
 540                         spin_unlock(&page_pools.epp_lock);
 541                         enc_pools_add_pages(page_pools.epp_pages_short / 2);
 542                         spin_lock(&page_pools.epp_lock);
 543
 544                         page_pools.epp_growing = 0;
 545
 546                         enc_pools_wakeup();
 547                 } else {
 548                         if (++page_pools.epp_waitqlen >
 549                             page_pools.epp_st_max_wqlen)
 550                                 page_pools.epp_st_max_wqlen =
 551                                                 page_pools.epp_waitqlen;
 552
 553                         set_current_state(CFS_TASK_UNINT);
 554                         cfs_waitlink_init(&waitlink);
 555                         cfs_waitq_add(&page_pools.epp_waitq, &waitlink);
 556
 557                         spin_unlock(&page_pools.epp_lock);
 558                         cfs_waitq_wait(&waitlink, CFS_TASK_UNINT);
 559                         cfs_waitq_del(&page_pools.epp_waitq, &waitlink);
 560                         LASSERT(page_pools.epp_waitqlen > 0);
 561                         spin_lock(&page_pools.epp_lock);
 562                         page_pools.epp_waitqlen--;
 563                 }
 564
 565                 LASSERT(page_pools.epp_pages_short >= desc->bd_iov_count);
 566                 page_pools.epp_pages_short -= desc->bd_iov_count;
 567
 568                 this_idle = 0;
 569                 goto again;
 570         }
 571
 572         /* record max wait time */
 573         if (unlikely(tick != 0)) {
 574                 tick = cfs_time_current() - tick;
 575                 if (tick > page_pools.epp_st_max_wait)
 576                         page_pools.epp_st_max_wait = tick;
 577         }
 578
 579         /* proceed with rest of allocation */
 580         page_pools.epp_free_pages -= desc->bd_iov_count;
 581
 582         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 583         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 584
 585         for (i = 0; i < desc->bd_iov_count; i++) {
 586                 LASSERT(page_pools.epp_pools[p_idx][g_idx] != NULL);
 587                 desc->bd_enc_iov[i].kiov_page =
 588                                         page_pools.epp_pools[p_idx][g_idx];
 589                 page_pools.epp_pools[p_idx][g_idx] = NULL;
 590
 591                 if (++g_idx == PAGES_PER_POOL) {
 592                         p_idx++;
 593                         g_idx = 0;
 594                 }
 595         }
 596
 597         if (page_pools.epp_free_pages < page_pools.epp_st_lowfree)
 598                 page_pools.epp_st_lowfree = page_pools.epp_free_pages;
 599
 600         /*
 601          * new idle index = (old * weight + new) / (weight + 1)
 602          */
 603         if (this_idle == -1) {
 604                 this_idle = page_pools.epp_free_pages * IDLE_IDX_MAX /
 605                             page_pools.epp_total_pages;
 606         }
 607         page_pools.epp_idle_idx = (page_pools.epp_idle_idx * IDLE_IDX_WEIGHT +
 608                                    this_idle) /
 609                                   (IDLE_IDX_WEIGHT + 1);
 610
 611         page_pools.epp_last_access = cfs_time_current_sec();
 612
 613         spin_unlock(&page_pools.epp_lock);
 614         return 0;
 615 }
 616 EXPORT_SYMBOL(sptlrpc_enc_pool_get_pages);
 617
 618 void sptlrpc_enc_pool_put_pages(struct ptlrpc_bulk_desc *desc)
 619 {
 620         int     p_idx, g_idx;
 621         int     i;
 622
 623         if (desc->bd_enc_iov == NULL)
 624                 return;
 625
 626         LASSERT(desc->bd_iov_count > 0);
 627
 628         spin_lock(&page_pools.epp_lock);
 629
 630         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 631         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 632
 633         LASSERT(page_pools.epp_free_pages + desc->bd_iov_count <=
 634                 page_pools.epp_total_pages);
 635         LASSERT(page_pools.epp_pools[p_idx]);
 636
 637         for (i = 0; i < desc->bd_iov_count; i++) {
 638                 LASSERT(desc->bd_enc_iov[i].kiov_page != NULL);
 639                 LASSERT(g_idx != 0 || page_pools.epp_pools[p_idx]);
 640                 LASSERT(page_pools.epp_pools[p_idx][g_idx] == NULL);
 641
 642                 page_pools.epp_pools[p_idx][g_idx] =
 643                                         desc->bd_enc_iov[i].kiov_page;
 644
 645                 if (++g_idx == PAGES_PER_POOL) {
 646                         p_idx++;
 647                         g_idx = 0;
 648                 }
 649         }
 650
 651         page_pools.epp_free_pages += desc->bd_iov_count;
 652
 653         enc_pools_wakeup();
 654
 655         spin_unlock(&page_pools.epp_lock);
 656
 657         OBD_FREE(desc->bd_enc_iov,
 658                  desc->bd_iov_count * sizeof(*desc->bd_enc_iov));
 659         desc->bd_enc_iov = NULL;
 660 }
 661 EXPORT_SYMBOL(sptlrpc_enc_pool_put_pages);
 662
 663 /*
 664  * we don't do much stuff for add_user/del_user anymore, except adding some
 665  * initial pages in add_user() if current pools are empty, rest would be
 666  * handled by the pools's self-adaption.
 667  */
 668 int sptlrpc_enc_pool_add_user(void)
 669 {
 670         int     need_grow = 0;
 671
 672         spin_lock(&page_pools.epp_lock);
 673         if (page_pools.epp_growing == 0 && page_pools.epp_total_pages == 0) {
 674                 page_pools.epp_growing = 1;
 675                 need_grow = 1;
 676         }
 677         spin_unlock(&page_pools.epp_lock);
 678
 679         if (need_grow) {
 680                 enc_pools_add_pages(PTLRPC_MAX_BRW_PAGES +
 681                                     PTLRPC_MAX_BRW_PAGES);
 682
 683                 spin_lock(&page_pools.epp_lock);
 684                 page_pools.epp_growing = 0;
 685                 enc_pools_wakeup();
 686                 spin_unlock(&page_pools.epp_lock);
 687         }
 688         return 0;
 689 }
 690 EXPORT_SYMBOL(sptlrpc_enc_pool_add_user);
 691
 692 int sptlrpc_enc_pool_del_user(void)
 693 {
 694         return 0;
 695 }
 696 EXPORT_SYMBOL(sptlrpc_enc_pool_del_user);
 697
 698 static inline void enc_pools_alloc(void)
 699 {
 700         LASSERT(page_pools.epp_max_pools);
 701         /*
 702          * on system with huge memory but small page size, this might lead to
 703          * high-order allocation. but it's not common, and we suppose memory
 704          * be not too much fragmented at module loading time.
 705          */
 706         OBD_ALLOC(page_pools.epp_pools,
 707                   page_pools.epp_max_pools * sizeof(*page_pools.epp_pools));
 708 }
 709
 710 static inline void enc_pools_free(void)
 711 {
 712         LASSERT(page_pools.epp_max_pools);
 713         LASSERT(page_pools.epp_pools);
 714
 715         OBD_FREE(page_pools.epp_pools,
 716                  page_pools.epp_max_pools * sizeof(*page_pools.epp_pools));
 717 }
 718
 719 int sptlrpc_enc_pool_init(void)
 720 {
 721         /*
 722          * maximum capacity is 1/8 of total physical memory.
 723          * is the 1/8 a good number?
 724          */
 725         page_pools.epp_max_pages = num_physpages / 8;
 726         page_pools.epp_max_pools = npages_to_npools(page_pools.epp_max_pages);
 727
 728         cfs_waitq_init(&page_pools.epp_waitq);
 729         page_pools.epp_waitqlen = 0;
 730         page_pools.epp_pages_short = 0;
 731
 732         page_pools.epp_growing = 0;
 733
 734         page_pools.epp_idle_idx = 0;
 735         page_pools.epp_last_shrink = cfs_time_current_sec();
 736         page_pools.epp_last_access = cfs_time_current_sec();
 737
 738         spin_lock_init(&page_pools.epp_lock);
 739         page_pools.epp_total_pages = 0;
 740         page_pools.epp_free_pages = 0;
 741
 742         page_pools.epp_st_max_pages = 0;
 743         page_pools.epp_st_grows = 0;
 744         page_pools.epp_st_grow_fails = 0;
 745         page_pools.epp_st_shrinks = 0;
 746         page_pools.epp_st_access = 0;
 747         page_pools.epp_st_missings = 0;
 748         page_pools.epp_st_lowfree = 0;
 749         page_pools.epp_st_max_wqlen = 0;
 750         page_pools.epp_st_max_wait = 0;
 751
 752         enc_pools_alloc();
 753         if (page_pools.epp_pools == NULL)
 754                 return -ENOMEM;
 755
 756         pools_shrinker = set_shrinker(pools_shrinker_seeks, enc_pools_shrink);
 757         if (pools_shrinker == NULL) {
 758                 enc_pools_free();
 759                 return -ENOMEM;
 760         }
 761
 762         return 0;
 763 }
 764
 765 void sptlrpc_enc_pool_fini(void)
 766 {
 767         unsigned long cleaned, npools;
 768
 769         LASSERT(pools_shrinker);
 770         LASSERT(page_pools.epp_pools);
 771         LASSERT(page_pools.epp_total_pages == page_pools.epp_free_pages);
 772
 773         remove_shrinker(pools_shrinker);
 774
 775         npools = npages_to_npools(page_pools.epp_total_pages);
 776         cleaned = enc_pools_cleanup(page_pools.epp_pools, npools);
 777         LASSERT(cleaned == page_pools.epp_total_pages);
 778
 779         enc_pools_free();
 780
 781         if (page_pools.epp_st_access > 0) {
 782                 CWARN("max pages %lu, grows %u, grow fails %u, shrinks %u, "
 783                       "access %lu, missing %lu, max qlen %u, max wait "
 784                       CFS_TIME_T"/%d\n",
 785                       page_pools.epp_st_max_pages, page_pools.epp_st_grows,
 786                       page_pools.epp_st_grow_fails,
 787                       page_pools.epp_st_shrinks, page_pools.epp_st_access,
 788                       page_pools.epp_st_missings, page_pools.epp_st_max_wqlen,
 789                       page_pools.epp_st_max_wait, HZ);
 790         }
 791 }
 792
 793 #else /* !__KERNEL__ */
 794
 795 int sptlrpc_enc_pool_get_pages(struct ptlrpc_bulk_desc *desc)
 796 {
 797         return 0;
 798 }
 799
 800 void sptlrpc_enc_pool_put_pages(struct ptlrpc_bulk_desc *desc)
 801 {
 802 }
 803
 804 int sptlrpc_enc_pool_init(void)
 805 {
 806         return 0;
 807 }
 808
 809 void sptlrpc_enc_pool_fini(void)
 810 {
 811 }
 812 #endif
 813
 814 /****************************************
 815  * Helpers to assist policy modules to  *
 816  * implement checksum funcationality    *
 817  ****************************************/
 818
 819 static struct sptlrpc_hash_type hash_types[] = {
 820         [BULK_HASH_ALG_NULL]    = { "null",     "null",         0 },
 821         [BULK_HASH_ALG_ADLER32] = { "adler32",  "adler32",      4 },
 822         [BULK_HASH_ALG_CRC32]   = { "crc32",    "crc32",        4 },
 823         [BULK_HASH_ALG_MD5]     = { "md5",      "md5",          16 },
 824         [BULK_HASH_ALG_SHA1]    = { "sha1",     "sha1",         20 },
 825         [BULK_HASH_ALG_SHA256]  = { "sha256",   "sha256",       32 },
 826         [BULK_HASH_ALG_SHA384]  = { "sha384",   "sha384",       48 },
 827         [BULK_HASH_ALG_SHA512]  = { "sha512",   "sha512",       64 },
 828 };
 829
 830 const struct sptlrpc_hash_type *sptlrpc_get_hash_type(__u8 hash_alg)
 831 {
 832         struct sptlrpc_hash_type *ht;
 833
 834         if (hash_alg < BULK_HASH_ALG_MAX) {
 835                 ht = &hash_types[hash_alg];
 836                 if (ht->sht_tfm_name)
 837                         return ht;
 838         }
 839         return NULL;
 840 }
 841 EXPORT_SYMBOL(sptlrpc_get_hash_type);
 842
 843 const char * sptlrpc_get_hash_name(__u8 hash_alg)
 844 {
 845         const struct sptlrpc_hash_type *ht;
 846
 847         ht = sptlrpc_get_hash_type(hash_alg);
 848         if (ht)
 849                 return ht->sht_name;
 850         else
 851                 return "unknown";
 852 }
 853 EXPORT_SYMBOL(sptlrpc_get_hash_name);
 854
 855 __u8 sptlrpc_get_hash_alg(const char *algname)
 856 {
 857         int     i;
 858
 859         for (i = 0; i < BULK_HASH_ALG_MAX; i++)
 860                 if (!strcmp(hash_types[i].sht_name, algname))
 861                         break;
 862         return i;
 863 }
 864 EXPORT_SYMBOL(sptlrpc_get_hash_alg);
 865
 866 int bulk_sec_desc_unpack(struct lustre_msg *msg, int offset)
 867 {
 868         struct ptlrpc_bulk_sec_desc *bsd;
 869         int                          size = msg->lm_buflens[offset];
 870
 871         bsd = lustre_msg_buf(msg, offset, sizeof(*bsd));
 872         if (bsd == NULL) {
 873                 CERROR("Invalid bulk sec desc: size %d\n", size);
 874                 return -EINVAL;
 875         }
 876
 877         if (lustre_msg_swabbed(msg)) {
 878                 __swab32s(&bsd->bsd_nob);
 879         }
 880
 881         if (unlikely(bsd->bsd_version != 0)) {
 882                 CERROR("Unexpected version %u\n", bsd->bsd_version);
 883                 return -EPROTO;
 884         }
 885
 886         if (unlikely(bsd->bsd_type >= SPTLRPC_BULK_MAX)) {
 887                 CERROR("Invalid type %u\n", bsd->bsd_type);
 888                 return -EPROTO;
 889         }
 890
 891         /* FIXME more sanity check here */
 892
 893         if (unlikely(bsd->bsd_svc != SPTLRPC_BULK_SVC_NULL &&
 894                      bsd->bsd_svc != SPTLRPC_BULK_SVC_INTG &&
 895                      bsd->bsd_svc != SPTLRPC_BULK_SVC_PRIV)) {
 896                 CERROR("Invalid svc %u\n", bsd->bsd_svc);
 897                 return -EPROTO;
 898         }
 899
 900         return 0;
 901 }
 902 EXPORT_SYMBOL(bulk_sec_desc_unpack);
 903
 904 #ifdef __KERNEL__
 905
 906 #ifdef HAVE_ADLER
 907 static int do_bulk_checksum_adler32(struct ptlrpc_bulk_desc *desc, void *buf)
 908 {
 909         struct page    *page;
 910         int             off;
 911         char           *ptr;
 912         __u32           adler32 = 1;
 913         int             len, i;
 914
 915         for (i = 0; i < desc->bd_iov_count; i++) {
 916                 page = desc->bd_iov[i].kiov_page;
 917                 off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK;
 918                 ptr = cfs_kmap(page) + off;
 919                 len = desc->bd_iov[i].kiov_len;
 920
 921                 adler32 = adler32(adler32, ptr, len);
 922
 923                 cfs_kunmap(page);
 924         }
 925
 926         adler32 = cpu_to_le32(adler32);
 927         memcpy(buf, &adler32, sizeof(adler32));
 928         return 0;
 929 }
 930 #endif
 931
 932 static int do_bulk_checksum_crc32(struct ptlrpc_bulk_desc *desc, void *buf)
 933 {
 934         struct page    *page;
 935         int             off;
 936         char           *ptr;
 937         __u32           crc32 = ~0;
 938         int             len, i;
 939
 940         for (i = 0; i < desc->bd_iov_count; i++) {
 941                 page = desc->bd_iov[i].kiov_page;
 942                 off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK;
 943                 ptr = cfs_kmap(page) + off;
 944                 len = desc->bd_iov[i].kiov_len;
 945
 946                 crc32 = crc32_le(crc32, ptr, len);
 947
 948                 cfs_kunmap(page);
 949         }
 950
 951         crc32 = cpu_to_le32(crc32);
 952         memcpy(buf, &crc32, sizeof(crc32));
 953         return 0;
 954 }
 955
 956 int sptlrpc_get_bulk_checksum(struct ptlrpc_bulk_desc *desc, __u8 alg,
 957                               void *buf, int buflen)
 958 {
 959         struct hash_desc    hdesc;
 960         int                 hashsize;
 961         char                hashbuf[64];
 962         struct scatterlist  sl;
 963         int                 i;
 964
 965         LASSERT(alg > BULK_HASH_ALG_NULL && alg < BULK_HASH_ALG_MAX);
 966         LASSERT(buflen >= 4);
 967
 968         switch (alg) {
 969         case BULK_HASH_ALG_ADLER32:
 970 #ifdef HAVE_ADLER
 971                 return do_bulk_checksum_adler32(desc, buf);
 972 #else
 973                 CERROR("Adler32 not supported\n");
 974                 return -EINVAL;
 975 #endif
 976         case BULK_HASH_ALG_CRC32:
 977                 return do_bulk_checksum_crc32(desc, buf);
 978         }
 979
 980         hdesc.tfm = ll_crypto_alloc_hash(hash_types[alg].sht_tfm_name, 0, 0);
 981         if (hdesc.tfm == NULL) {
 982                 CERROR("Unable to allocate TFM %s\n", hash_types[alg].sht_name);
 983                 return -ENOMEM;
 984         }
 985
 986         hdesc.flags = 0;
 987         ll_crypto_hash_init(&hdesc);
 988
 989         hashsize = ll_crypto_hash_digestsize(hdesc.tfm);
 990
 991         for (i = 0; i < desc->bd_iov_count; i++) {
 992                 sl.page = desc->bd_iov[i].kiov_page;
 993                 sl.offset = desc->bd_iov[i].kiov_offset;
 994                 sl.length = desc->bd_iov[i].kiov_len;
 995                 ll_crypto_hash_update(&hdesc, &sl, sl.length);
 996         }
 997
 998         if (hashsize > buflen) {
 999                 ll_crypto_hash_final(&hdesc, hashbuf);
1000                 memcpy(buf, hashbuf, buflen);
1001         } else {
1002                 ll_crypto_hash_final(&hdesc, buf);
1003         }
1004
1005         ll_crypto_free_hash(hdesc.tfm);
1006         return 0;
1007 }
1008 EXPORT_SYMBOL(sptlrpc_get_bulk_checksum);
1009
1010 #else /* !__KERNEL__ */
1011
1012 int sptlrpc_get_bulk_checksum(struct ptlrpc_bulk_desc *desc, __u8 alg,
1013                               void *buf, int buflen)
1014 {
1015         __u32   csum32;
1016         int     i;
1017
1018         LASSERT(alg == BULK_HASH_ALG_ADLER32 || alg == BULK_HASH_ALG_CRC32);
1019
1020         if (alg == BULK_HASH_ALG_ADLER32)
1021                 csum32 = 1;
1022         else
1023                 csum32 = ~0;
1024
1025         for (i = 0; i < desc->bd_iov_count; i++) {
1026                 unsigned char *ptr = desc->bd_iov[i].iov_base;
1027                 int len = desc->bd_iov[i].iov_len;
1028
1029                 switch (alg) {
1030                 case BULK_HASH_ALG_ADLER32:
1031 #ifdef HAVE_ADLER
1032                         csum32 = adler32(csum32, ptr, len);
1033 #else
1034                         CERROR("Adler32 not supported\n");
1035                         return -EINVAL;
1036 #endif
1037                         break;
1038                 case BULK_HASH_ALG_CRC32:
1039                         csum32 = crc32_le(csum32, ptr, len);
1040                         break;
1041                 }
1042         }
1043
1044         csum32 = cpu_to_le32(csum32);
1045         memcpy(buf, &csum32, sizeof(csum32));
1046         return 0;
1047 }
1048
1049 #endif /* __KERNEL__ */