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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2016, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/ptlrpc/sec_bulk.c
  33  *
  34  * Author: Eric Mei <ericm@clusterfs.com>
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_SEC
  38
  39 #include <libcfs/libcfs.h>
  40
  41 #include <obd.h>
  42 #include <obd_cksum.h>
  43 #include <obd_class.h>
  44 #include <obd_support.h>
  45 #include <lustre_net.h>
  46 #include <lustre_import.h>
  47 #include <lustre_dlm.h>
  48 #include <lustre_sec.h>
  49
  50 #include "ptlrpc_internal.h"
  51
  52 static int mult = 20 - PAGE_SHIFT;
  53 static int enc_pool_max_memory_mb;
  54 module_param(enc_pool_max_memory_mb, int, 0644);
  55 MODULE_PARM_DESC(enc_pool_max_memory_mb,
  56                  "Encoding pool max memory (MB), 1/8 of total physical memory by default");
  57
  58 /****************************************
  59  * bulk encryption page pools           *
  60  ****************************************/
  61
  62
  63 #define PTRS_PER_PAGE   (PAGE_SIZE / sizeof(void *))
  64 #define PAGES_PER_POOL  (PTRS_PER_PAGE)
  65
  66 #define IDLE_IDX_MAX            (100)
  67 #define IDLE_IDX_WEIGHT         (3)
  68
  69 #define CACHE_QUIESCENT_PERIOD  (20)
  70
  71 static struct ptlrpc_enc_page_pool {
  72         /*
  73          * constants
  74          */
  75         unsigned long    epp_max_pages;   /* maximum pages can hold, const */
  76         unsigned int     epp_max_pools;   /* number of pools, const */
  77
  78         /*
  79          * wait queue in case of not enough free pages.
  80          */
  81         wait_queue_head_t    epp_waitq;       /* waiting threads */
  82         unsigned int     epp_waitqlen;    /* wait queue length */
  83         unsigned long    epp_pages_short; /* # of pages wanted of in-q users */
  84         unsigned int     epp_growing:1;   /* during adding pages */
  85
  86         /*
  87          * indicating how idle the pools are, from 0 to MAX_IDLE_IDX
  88          * this is counted based on each time when getting pages from
  89          * the pools, not based on time. which means in case that system
  90          * is idled for a while but the idle_idx might still be low if no
  91          * activities happened in the pools.
  92          */
  93         unsigned long    epp_idle_idx;
  94
  95         /* last shrink time due to mem tight */
  96         time64_t        epp_last_shrink;
  97         time64_t        epp_last_access;
  98
  99         /*
 100          * in-pool pages bookkeeping
 101          */
 102         spinlock_t       epp_lock;         /* protect following fields */
 103         unsigned long    epp_total_pages; /* total pages in pools */
 104         unsigned long    epp_free_pages;  /* current pages available */
 105
 106         /*
 107          * statistics
 108          */
 109         unsigned long    epp_st_max_pages;      /* # of pages ever reached */
 110         unsigned int     epp_st_grows;          /* # of grows */
 111         unsigned int     epp_st_grow_fails;     /* # of add pages failures */
 112         unsigned int     epp_st_shrinks;        /* # of shrinks */
 113         unsigned long    epp_st_access;         /* # of access */
 114         unsigned long    epp_st_missings;       /* # of cache missing */
 115         unsigned long    epp_st_lowfree;        /* lowest free pages reached */
 116         unsigned int     epp_st_max_wqlen;      /* highest waitqueue length */
 117         cfs_time_t       epp_st_max_wait;       /* in jeffies */
 118         unsigned long    epp_st_outofmem;       /* # of out of mem requests */
 119         /*
 120          * pointers to pools, may be vmalloc'd
 121          */
 122         struct page    ***epp_pools;
 123 } page_pools;
 124
 125 /*
 126  * memory shrinker
 127  */
 128 static const int pools_shrinker_seeks = DEFAULT_SEEKS;
 129 static struct shrinker *pools_shrinker;
 130
 131
 132 /*
 133  * /proc/fs/lustre/sptlrpc/encrypt_page_pools
 134  */
 135 int sptlrpc_proc_enc_pool_seq_show(struct seq_file *m, void *v)
 136 {
 137         spin_lock(&page_pools.epp_lock);
 138
 139         seq_printf(m, "physical pages:          %lu\n"
 140                    "pages per pool:          %lu\n"
 141                    "max pages:               %lu\n"
 142                    "max pools:               %u\n"
 143                    "total pages:             %lu\n"
 144                    "total free:              %lu\n"
 145                    "idle index:              %lu/100\n"
 146                    "last shrink:             %lds\n"
 147                    "last access:             %lds\n"
 148                    "max pages reached:       %lu\n"
 149                    "grows:                   %u\n"
 150                    "grows failure:           %u\n"
 151                    "shrinks:                 %u\n"
 152                    "cache access:            %lu\n"
 153                    "cache missing:           %lu\n"
 154                    "low free mark:           %lu\n"
 155                    "max waitqueue depth:     %u\n"
 156                    "max wait time:           "CFS_TIME_T"/%lu\n"
 157                    "out of mem:             %lu\n",
 158                    totalram_pages, PAGES_PER_POOL,
 159                    page_pools.epp_max_pages,
 160                    page_pools.epp_max_pools,
 161                    page_pools.epp_total_pages,
 162                    page_pools.epp_free_pages,
 163                    page_pools.epp_idle_idx,
 164                    (long)(ktime_get_seconds() - page_pools.epp_last_shrink),
 165                    (long)(ktime_get_seconds() - page_pools.epp_last_access),
 166                    page_pools.epp_st_max_pages,
 167                    page_pools.epp_st_grows,
 168                    page_pools.epp_st_grow_fails,
 169                    page_pools.epp_st_shrinks,
 170                    page_pools.epp_st_access,
 171                    page_pools.epp_st_missings,
 172                    page_pools.epp_st_lowfree,
 173                    page_pools.epp_st_max_wqlen,
 174                    page_pools.epp_st_max_wait,
 175                    msecs_to_jiffies(MSEC_PER_SEC),
 176                    page_pools.epp_st_outofmem);
 177
 178         spin_unlock(&page_pools.epp_lock);
 179         return 0;
 180 }
 181
 182 static void enc_pools_release_free_pages(long npages)
 183 {
 184         int     p_idx, g_idx;
 185         int     p_idx_max1, p_idx_max2;
 186
 187         LASSERT(npages > 0);
 188         LASSERT(npages <= page_pools.epp_free_pages);
 189         LASSERT(page_pools.epp_free_pages <= page_pools.epp_total_pages);
 190
 191         /* max pool index before the release */
 192         p_idx_max2 = (page_pools.epp_total_pages - 1) / PAGES_PER_POOL;
 193
 194         page_pools.epp_free_pages -= npages;
 195         page_pools.epp_total_pages -= npages;
 196
 197         /* max pool index after the release */
 198         p_idx_max1 = page_pools.epp_total_pages == 0 ? -1 :
 199                      ((page_pools.epp_total_pages - 1) / PAGES_PER_POOL);
 200
 201         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 202         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 203         LASSERT(page_pools.epp_pools[p_idx]);
 204
 205         while (npages--) {
 206                 LASSERT(page_pools.epp_pools[p_idx]);
 207                 LASSERT(page_pools.epp_pools[p_idx][g_idx] != NULL);
 208
 209                 __free_page(page_pools.epp_pools[p_idx][g_idx]);
 210                 page_pools.epp_pools[p_idx][g_idx] = NULL;
 211
 212                 if (++g_idx == PAGES_PER_POOL) {
 213                         p_idx++;
 214                         g_idx = 0;
 215                 }
 216         }
 217
 218         /* free unused pools */
 219         while (p_idx_max1 < p_idx_max2) {
 220                 LASSERT(page_pools.epp_pools[p_idx_max2]);
 221                 OBD_FREE(page_pools.epp_pools[p_idx_max2], PAGE_SIZE);
 222                 page_pools.epp_pools[p_idx_max2] = NULL;
 223                 p_idx_max2--;
 224         }
 225 }
 226
 227 /*
 228  * we try to keep at least PTLRPC_MAX_BRW_PAGES pages in the pool.
 229  */
 230 static unsigned long enc_pools_shrink_count(struct shrinker *s,
 231                                             struct shrink_control *sc)
 232 {
 233         /*
 234          * if no pool access for a long time, we consider it's fully idle.
 235          * a little race here is fine.
 236          */
 237         if (unlikely(ktime_get_real_seconds() - page_pools.epp_last_access >
 238                      CACHE_QUIESCENT_PERIOD)) {
 239                 spin_lock(&page_pools.epp_lock);
 240                 page_pools.epp_idle_idx = IDLE_IDX_MAX;
 241                 spin_unlock(&page_pools.epp_lock);
 242         }
 243
 244         LASSERT(page_pools.epp_idle_idx <= IDLE_IDX_MAX);
 245         return max(page_pools.epp_free_pages - PTLRPC_MAX_BRW_PAGES, 0UL) *
 246                 (IDLE_IDX_MAX - page_pools.epp_idle_idx) / IDLE_IDX_MAX;
 247 }
 248
 249 /*
 250  * we try to keep at least PTLRPC_MAX_BRW_PAGES pages in the pool.
 251  */
 252 static unsigned long enc_pools_shrink_scan(struct shrinker *s,
 253                                            struct shrink_control *sc)
 254 {
 255         spin_lock(&page_pools.epp_lock);
 256         sc->nr_to_scan = min_t(unsigned long, sc->nr_to_scan,
 257                               page_pools.epp_free_pages - PTLRPC_MAX_BRW_PAGES);
 258         if (sc->nr_to_scan > 0) {
 259                 enc_pools_release_free_pages(sc->nr_to_scan);
 260                 CDEBUG(D_SEC, "released %ld pages, %ld left\n",
 261                        (long)sc->nr_to_scan, page_pools.epp_free_pages);
 262
 263                 page_pools.epp_st_shrinks++;
 264                 page_pools.epp_last_shrink = ktime_get_real_seconds();
 265         }
 266         spin_unlock(&page_pools.epp_lock);
 267
 268         /*
 269          * if no pool access for a long time, we consider it's fully idle.
 270          * a little race here is fine.
 271          */
 272         if (unlikely(ktime_get_real_seconds() - page_pools.epp_last_access >
 273                      CACHE_QUIESCENT_PERIOD)) {
 274                 spin_lock(&page_pools.epp_lock);
 275                 page_pools.epp_idle_idx = IDLE_IDX_MAX;
 276                 spin_unlock(&page_pools.epp_lock);
 277         }
 278
 279         LASSERT(page_pools.epp_idle_idx <= IDLE_IDX_MAX);
 280         return sc->nr_to_scan;
 281 }
 282
 283 #ifndef HAVE_SHRINKER_COUNT
 284 /*
 285  * could be called frequently for query (@nr_to_scan == 0).
 286  * we try to keep at least PTLRPC_MAX_BRW_PAGES pages in the pool.
 287  */
 288 static int enc_pools_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
 289 {
 290         struct shrink_control scv = {
 291                 .nr_to_scan = shrink_param(sc, nr_to_scan),
 292                 .gfp_mask   = shrink_param(sc, gfp_mask)
 293         };
 294 #if !defined(HAVE_SHRINKER_WANT_SHRINK_PTR) && !defined(HAVE_SHRINK_CONTROL)
 295         struct shrinker* shrinker = NULL;
 296 #endif
 297
 298         enc_pools_shrink_scan(shrinker, &scv);
 299
 300         return enc_pools_shrink_count(shrinker, &scv);
 301 }
 302
 303 #endif /* HAVE_SHRINKER_COUNT */
 304
 305 static inline
 306 int npages_to_npools(unsigned long npages)
 307 {
 308         return (int) ((npages + PAGES_PER_POOL - 1) / PAGES_PER_POOL);
 309 }
 310
 311 /*
 312  * return how many pages cleaned up.
 313  */
 314 static unsigned long enc_pools_cleanup(struct page ***pools, int npools)
 315 {
 316         unsigned long cleaned = 0;
 317         int           i, j;
 318
 319         for (i = 0; i < npools; i++) {
 320                 if (pools[i]) {
 321                         for (j = 0; j < PAGES_PER_POOL; j++) {
 322                                 if (pools[i][j]) {
 323                                         __free_page(pools[i][j]);
 324                                         cleaned++;
 325                                 }
 326                         }
 327                         OBD_FREE(pools[i], PAGE_SIZE);
 328                         pools[i] = NULL;
 329                 }
 330         }
 331
 332         return cleaned;
 333 }
 334
 335 /*
 336  * merge @npools pointed by @pools which contains @npages new pages
 337  * into current pools.
 338  *
 339  * we have options to avoid most memory copy with some tricks. but we choose
 340  * the simplest way to avoid complexity. It's not frequently called.
 341  */
 342 static void enc_pools_insert(struct page ***pools, int npools, int npages)
 343 {
 344         int     freeslot;
 345         int     op_idx, np_idx, og_idx, ng_idx;
 346         int     cur_npools, end_npools;
 347
 348         LASSERT(npages > 0);
 349         LASSERT(page_pools.epp_total_pages+npages <= page_pools.epp_max_pages);
 350         LASSERT(npages_to_npools(npages) == npools);
 351         LASSERT(page_pools.epp_growing);
 352
 353         spin_lock(&page_pools.epp_lock);
 354
 355         /*
 356          * (1) fill all the free slots of current pools.
 357          */
 358         /* free slots are those left by rent pages, and the extra ones with
 359          * index >= total_pages, locate at the tail of last pool. */
 360         freeslot = page_pools.epp_total_pages % PAGES_PER_POOL;
 361         if (freeslot != 0)
 362                 freeslot = PAGES_PER_POOL - freeslot;
 363         freeslot += page_pools.epp_total_pages - page_pools.epp_free_pages;
 364
 365         op_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 366         og_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 367         np_idx = npools - 1;
 368         ng_idx = (npages - 1) % PAGES_PER_POOL;
 369
 370         while (freeslot) {
 371                 LASSERT(page_pools.epp_pools[op_idx][og_idx] == NULL);
 372                 LASSERT(pools[np_idx][ng_idx] != NULL);
 373
 374                 page_pools.epp_pools[op_idx][og_idx] = pools[np_idx][ng_idx];
 375                 pools[np_idx][ng_idx] = NULL;
 376
 377                 freeslot--;
 378
 379                 if (++og_idx == PAGES_PER_POOL) {
 380                         op_idx++;
 381                         og_idx = 0;
 382                 }
 383                 if (--ng_idx < 0) {
 384                         if (np_idx == 0)
 385                                 break;
 386                         np_idx--;
 387                         ng_idx = PAGES_PER_POOL - 1;
 388                 }
 389         }
 390
 391         /*
 392          * (2) add pools if needed.
 393          */
 394         cur_npools = (page_pools.epp_total_pages + PAGES_PER_POOL - 1) /
 395                      PAGES_PER_POOL;
 396         end_npools = (page_pools.epp_total_pages + npages + PAGES_PER_POOL -1) /
 397                      PAGES_PER_POOL;
 398         LASSERT(end_npools <= page_pools.epp_max_pools);
 399
 400         np_idx = 0;
 401         while (cur_npools < end_npools) {
 402                 LASSERT(page_pools.epp_pools[cur_npools] == NULL);
 403                 LASSERT(np_idx < npools);
 404                 LASSERT(pools[np_idx] != NULL);
 405
 406                 page_pools.epp_pools[cur_npools++] = pools[np_idx];
 407                 pools[np_idx++] = NULL;
 408         }
 409
 410         page_pools.epp_total_pages += npages;
 411         page_pools.epp_free_pages += npages;
 412         page_pools.epp_st_lowfree = page_pools.epp_free_pages;
 413
 414         if (page_pools.epp_total_pages > page_pools.epp_st_max_pages)
 415                 page_pools.epp_st_max_pages = page_pools.epp_total_pages;
 416
 417         CDEBUG(D_SEC, "add %d pages to total %lu\n", npages,
 418                page_pools.epp_total_pages);
 419
 420         spin_unlock(&page_pools.epp_lock);
 421 }
 422
 423 static int enc_pools_add_pages(int npages)
 424 {
 425         static DEFINE_MUTEX(add_pages_mutex);
 426         struct page   ***pools;
 427         int             npools, alloced = 0;
 428         int             i, j, rc = -ENOMEM;
 429
 430         if (npages < PTLRPC_MAX_BRW_PAGES)
 431                 npages = PTLRPC_MAX_BRW_PAGES;
 432
 433         mutex_lock(&add_pages_mutex);
 434
 435         if (npages + page_pools.epp_total_pages > page_pools.epp_max_pages)
 436                 npages = page_pools.epp_max_pages - page_pools.epp_total_pages;
 437         LASSERT(npages > 0);
 438
 439         page_pools.epp_st_grows++;
 440
 441         npools = npages_to_npools(npages);
 442         OBD_ALLOC(pools, npools * sizeof(*pools));
 443         if (pools == NULL)
 444                 goto out;
 445
 446         for (i = 0; i < npools; i++) {
 447                 OBD_ALLOC(pools[i], PAGE_SIZE);
 448                 if (pools[i] == NULL)
 449                         goto out_pools;
 450
 451                 for (j = 0; j < PAGES_PER_POOL && alloced < npages; j++) {
 452                         pools[i][j] = alloc_page(GFP_NOFS |
 453                                                  __GFP_HIGHMEM);
 454                         if (pools[i][j] == NULL)
 455                                 goto out_pools;
 456
 457                         alloced++;
 458                 }
 459         }
 460         LASSERT(alloced == npages);
 461
 462         enc_pools_insert(pools, npools, npages);
 463         CDEBUG(D_SEC, "added %d pages into pools\n", npages);
 464         rc = 0;
 465
 466 out_pools:
 467         enc_pools_cleanup(pools, npools);
 468         OBD_FREE(pools, npools * sizeof(*pools));
 469 out:
 470         if (rc) {
 471                 page_pools.epp_st_grow_fails++;
 472                 CERROR("Failed to allocate %d enc pages\n", npages);
 473         }
 474
 475         mutex_unlock(&add_pages_mutex);
 476         return rc;
 477 }
 478
 479 static inline void enc_pools_wakeup(void)
 480 {
 481         assert_spin_locked(&page_pools.epp_lock);
 482
 483         if (unlikely(page_pools.epp_waitqlen)) {
 484                 LASSERT(waitqueue_active(&page_pools.epp_waitq));
 485                 wake_up_all(&page_pools.epp_waitq);
 486         }
 487 }
 488
 489 static int enc_pools_should_grow(int page_needed, time64_t now)
 490 {
 491         /* don't grow if someone else is growing the pools right now,
 492          * or the pools has reached its full capacity
 493          */
 494         if (page_pools.epp_growing ||
 495             page_pools.epp_total_pages == page_pools.epp_max_pages)
 496                 return 0;
 497
 498         /* if total pages is not enough, we need to grow */
 499         if (page_pools.epp_total_pages < page_needed)
 500                 return 1;
 501
 502         /*
 503          * we wanted to return 0 here if there was a shrink just
 504          * happened a moment ago, but this may cause deadlock if both
 505          * client and ost live on single node.
 506          */
 507
 508         /*
 509          * here we perhaps need consider other factors like wait queue
 510          * length, idle index, etc. ?
 511          */
 512
 513         /* grow the pools in any other cases */
 514         return 1;
 515 }
 516
 517 /*
 518  * Export the number of free pages in the pool
 519  */
 520 int get_free_pages_in_pool(void)
 521 {
 522         return page_pools.epp_free_pages;
 523 }
 524 EXPORT_SYMBOL(get_free_pages_in_pool);
 525
 526 /*
 527  * Let outside world know if enc_pool full capacity is reached
 528  */
 529 int pool_is_at_full_capacity(void)
 530 {
 531         return (page_pools.epp_total_pages == page_pools.epp_max_pages);
 532 }
 533 EXPORT_SYMBOL(pool_is_at_full_capacity);
 534
 535 /*
 536  * we allocate the requested pages atomically.
 537  */
 538 int sptlrpc_enc_pool_get_pages(struct ptlrpc_bulk_desc *desc)
 539 {
 540         wait_queue_t  waitlink;
 541         unsigned long   this_idle = -1;
 542         cfs_time_t      tick = 0;
 543         long            now;
 544         int             p_idx, g_idx;
 545         int             i;
 546
 547         LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
 548         LASSERT(desc->bd_iov_count > 0);
 549         LASSERT(desc->bd_iov_count <= page_pools.epp_max_pages);
 550
 551         /* resent bulk, enc iov might have been allocated previously */
 552         if (GET_ENC_KIOV(desc) != NULL)
 553                 return 0;
 554
 555         OBD_ALLOC_LARGE(GET_ENC_KIOV(desc),
 556                   desc->bd_iov_count * sizeof(*GET_ENC_KIOV(desc)));
 557         if (GET_ENC_KIOV(desc) == NULL)
 558                 return -ENOMEM;
 559
 560         spin_lock(&page_pools.epp_lock);
 561
 562         page_pools.epp_st_access++;
 563 again:
 564         if (unlikely(page_pools.epp_free_pages < desc->bd_iov_count)) {
 565                 if (tick == 0)
 566                         tick = cfs_time_current();
 567
 568                 now = ktime_get_real_seconds();
 569
 570                 page_pools.epp_st_missings++;
 571                 page_pools.epp_pages_short += desc->bd_iov_count;
 572
 573                 if (enc_pools_should_grow(desc->bd_iov_count, now)) {
 574                         page_pools.epp_growing = 1;
 575
 576                         spin_unlock(&page_pools.epp_lock);
 577                         enc_pools_add_pages(page_pools.epp_pages_short / 2);
 578                         spin_lock(&page_pools.epp_lock);
 579
 580                         page_pools.epp_growing = 0;
 581
 582                         enc_pools_wakeup();
 583                 } else {
 584                         if (page_pools.epp_growing) {
 585                                 if (++page_pools.epp_waitqlen >
 586                                     page_pools.epp_st_max_wqlen)
 587                                         page_pools.epp_st_max_wqlen =
 588                                                         page_pools.epp_waitqlen;
 589
 590                                 set_current_state(TASK_UNINTERRUPTIBLE);
 591                                 init_waitqueue_entry(&waitlink, current);
 592                                 add_wait_queue(&page_pools.epp_waitq,
 593                                                &waitlink);
 594
 595                                 spin_unlock(&page_pools.epp_lock);
 596                                 schedule();
 597                                 remove_wait_queue(&page_pools.epp_waitq,
 598                                                   &waitlink);
 599                                 LASSERT(page_pools.epp_waitqlen > 0);
 600                                 spin_lock(&page_pools.epp_lock);
 601                                 page_pools.epp_waitqlen--;
 602                         } else {
 603                                 /* ptlrpcd thread should not sleep in that case,
 604                                  * or deadlock may occur!
 605                                  * Instead, return -ENOMEM so that upper layers
 606                                  * will put request back in queue. */
 607                                 page_pools.epp_st_outofmem++;
 608                                 spin_unlock(&page_pools.epp_lock);
 609                                 OBD_FREE_LARGE(GET_ENC_KIOV(desc),
 610                                                desc->bd_iov_count *
 611                                                 sizeof(*GET_ENC_KIOV(desc)));
 612                                 GET_ENC_KIOV(desc) = NULL;
 613                                 return -ENOMEM;
 614                         }
 615                 }
 616
 617                 LASSERT(page_pools.epp_pages_short >= desc->bd_iov_count);
 618                 page_pools.epp_pages_short -= desc->bd_iov_count;
 619
 620                 this_idle = 0;
 621                 goto again;
 622         }
 623
 624         /* record max wait time */
 625         if (unlikely(tick != 0)) {
 626                 tick = cfs_time_current() - tick;
 627                 if (tick > page_pools.epp_st_max_wait)
 628                         page_pools.epp_st_max_wait = tick;
 629         }
 630
 631         /* proceed with rest of allocation */
 632         page_pools.epp_free_pages -= desc->bd_iov_count;
 633
 634         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 635         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 636
 637         for (i = 0; i < desc->bd_iov_count; i++) {
 638                 LASSERT(page_pools.epp_pools[p_idx][g_idx] != NULL);
 639                 BD_GET_ENC_KIOV(desc, i).kiov_page =
 640                        page_pools.epp_pools[p_idx][g_idx];
 641                 page_pools.epp_pools[p_idx][g_idx] = NULL;
 642
 643                 if (++g_idx == PAGES_PER_POOL) {
 644                         p_idx++;
 645                         g_idx = 0;
 646                 }
 647         }
 648
 649         if (page_pools.epp_free_pages < page_pools.epp_st_lowfree)
 650                 page_pools.epp_st_lowfree = page_pools.epp_free_pages;
 651
 652         /*
 653          * new idle index = (old * weight + new) / (weight + 1)
 654          */
 655         if (this_idle == -1) {
 656                 this_idle = page_pools.epp_free_pages * IDLE_IDX_MAX /
 657                             page_pools.epp_total_pages;
 658         }
 659         page_pools.epp_idle_idx = (page_pools.epp_idle_idx * IDLE_IDX_WEIGHT +
 660                                    this_idle) /
 661                                   (IDLE_IDX_WEIGHT + 1);
 662
 663         page_pools.epp_last_access = ktime_get_real_seconds();
 664
 665         spin_unlock(&page_pools.epp_lock);
 666         return 0;
 667 }
 668 EXPORT_SYMBOL(sptlrpc_enc_pool_get_pages);
 669
 670 void sptlrpc_enc_pool_put_pages(struct ptlrpc_bulk_desc *desc)
 671 {
 672         int     p_idx, g_idx;
 673         int     i;
 674
 675         LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
 676
 677         if (GET_ENC_KIOV(desc) == NULL)
 678                 return;
 679
 680         LASSERT(desc->bd_iov_count > 0);
 681
 682         spin_lock(&page_pools.epp_lock);
 683
 684         p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
 685         g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
 686
 687         LASSERT(page_pools.epp_free_pages + desc->bd_iov_count <=
 688                 page_pools.epp_total_pages);
 689         LASSERT(page_pools.epp_pools[p_idx]);
 690
 691         for (i = 0; i < desc->bd_iov_count; i++) {
 692                 LASSERT(BD_GET_ENC_KIOV(desc, i).kiov_page != NULL);
 693                 LASSERT(g_idx != 0 || page_pools.epp_pools[p_idx]);
 694                 LASSERT(page_pools.epp_pools[p_idx][g_idx] == NULL);
 695
 696                 page_pools.epp_pools[p_idx][g_idx] =
 697                         BD_GET_ENC_KIOV(desc, i).kiov_page;
 698
 699                 if (++g_idx == PAGES_PER_POOL) {
 700                         p_idx++;
 701                         g_idx = 0;
 702                 }
 703         }
 704
 705         page_pools.epp_free_pages += desc->bd_iov_count;
 706
 707         enc_pools_wakeup();
 708
 709         spin_unlock(&page_pools.epp_lock);
 710
 711         OBD_FREE_LARGE(GET_ENC_KIOV(desc),
 712                  desc->bd_iov_count * sizeof(*GET_ENC_KIOV(desc)));
 713         GET_ENC_KIOV(desc) = NULL;
 714 }
 715
 716 /*
 717  * we don't do much stuff for add_user/del_user anymore, except adding some
 718  * initial pages in add_user() if current pools are empty, rest would be
 719  * handled by the pools's self-adaption.
 720  */
 721 int sptlrpc_enc_pool_add_user(void)
 722 {
 723         int     need_grow = 0;
 724
 725         spin_lock(&page_pools.epp_lock);
 726         if (page_pools.epp_growing == 0 && page_pools.epp_total_pages == 0) {
 727                 page_pools.epp_growing = 1;
 728                 need_grow = 1;
 729         }
 730         spin_unlock(&page_pools.epp_lock);
 731
 732         if (need_grow) {
 733                 enc_pools_add_pages(PTLRPC_MAX_BRW_PAGES +
 734                                     PTLRPC_MAX_BRW_PAGES);
 735
 736                 spin_lock(&page_pools.epp_lock);
 737                 page_pools.epp_growing = 0;
 738                 enc_pools_wakeup();
 739                 spin_unlock(&page_pools.epp_lock);
 740         }
 741         return 0;
 742 }
 743 EXPORT_SYMBOL(sptlrpc_enc_pool_add_user);
 744
 745 int sptlrpc_enc_pool_del_user(void)
 746 {
 747         return 0;
 748 }
 749 EXPORT_SYMBOL(sptlrpc_enc_pool_del_user);
 750
 751 static inline void enc_pools_alloc(void)
 752 {
 753         LASSERT(page_pools.epp_max_pools);
 754         OBD_ALLOC_LARGE(page_pools.epp_pools,
 755                         page_pools.epp_max_pools *
 756                         sizeof(*page_pools.epp_pools));
 757 }
 758
 759 static inline void enc_pools_free(void)
 760 {
 761         LASSERT(page_pools.epp_max_pools);
 762         LASSERT(page_pools.epp_pools);
 763
 764         OBD_FREE_LARGE(page_pools.epp_pools,
 765                        page_pools.epp_max_pools *
 766                        sizeof(*page_pools.epp_pools));
 767 }
 768
 769 int sptlrpc_enc_pool_init(void)
 770 {
 771         DEF_SHRINKER_VAR(shvar, enc_pools_shrink,
 772                          enc_pools_shrink_count, enc_pools_shrink_scan);
 773
 774         page_pools.epp_max_pages = totalram_pages / 8;
 775         if (enc_pool_max_memory_mb > 0 &&
 776             enc_pool_max_memory_mb <= (totalram_pages >> mult))
 777                 page_pools.epp_max_pages = enc_pool_max_memory_mb << mult;
 778
 779         page_pools.epp_max_pools = npages_to_npools(page_pools.epp_max_pages);
 780
 781         init_waitqueue_head(&page_pools.epp_waitq);
 782         page_pools.epp_waitqlen = 0;
 783         page_pools.epp_pages_short = 0;
 784
 785         page_pools.epp_growing = 0;
 786
 787         page_pools.epp_idle_idx = 0;
 788         page_pools.epp_last_shrink = ktime_get_real_seconds();
 789         page_pools.epp_last_access = ktime_get_real_seconds();
 790
 791         spin_lock_init(&page_pools.epp_lock);
 792         page_pools.epp_total_pages = 0;
 793         page_pools.epp_free_pages = 0;
 794
 795         page_pools.epp_st_max_pages = 0;
 796         page_pools.epp_st_grows = 0;
 797         page_pools.epp_st_grow_fails = 0;
 798         page_pools.epp_st_shrinks = 0;
 799         page_pools.epp_st_access = 0;
 800         page_pools.epp_st_missings = 0;
 801         page_pools.epp_st_lowfree = 0;
 802         page_pools.epp_st_max_wqlen = 0;
 803         page_pools.epp_st_max_wait = 0;
 804         page_pools.epp_st_outofmem = 0;
 805
 806         enc_pools_alloc();
 807         if (page_pools.epp_pools == NULL)
 808                 return -ENOMEM;
 809
 810         pools_shrinker = set_shrinker(pools_shrinker_seeks, &shvar);
 811         if (pools_shrinker == NULL) {
 812                 enc_pools_free();
 813                 return -ENOMEM;
 814         }
 815
 816         return 0;
 817 }
 818
 819 void sptlrpc_enc_pool_fini(void)
 820 {
 821         unsigned long cleaned, npools;
 822
 823         LASSERT(pools_shrinker);
 824         LASSERT(page_pools.epp_pools);
 825         LASSERT(page_pools.epp_total_pages == page_pools.epp_free_pages);
 826
 827         remove_shrinker(pools_shrinker);
 828
 829         npools = npages_to_npools(page_pools.epp_total_pages);
 830         cleaned = enc_pools_cleanup(page_pools.epp_pools, npools);
 831         LASSERT(cleaned == page_pools.epp_total_pages);
 832
 833         enc_pools_free();
 834
 835         if (page_pools.epp_st_access > 0) {
 836                 CDEBUG(D_SEC,
 837                        "max pages %lu, grows %u, grow fails %u, shrinks %u, "
 838                        "access %lu, missing %lu, max qlen %u, max wait "
 839                        CFS_TIME_T"/%lu, out of mem %lu\n",
 840                        page_pools.epp_st_max_pages, page_pools.epp_st_grows,
 841                        page_pools.epp_st_grow_fails,
 842                        page_pools.epp_st_shrinks, page_pools.epp_st_access,
 843                        page_pools.epp_st_missings, page_pools.epp_st_max_wqlen,
 844                        page_pools.epp_st_max_wait,
 845                        msecs_to_jiffies(MSEC_PER_SEC),
 846                        page_pools.epp_st_outofmem);
 847         }
 848 }
 849
 850
 851 static int cfs_hash_alg_id[] = {
 852         [BULK_HASH_ALG_NULL]    = CFS_HASH_ALG_NULL,
 853         [BULK_HASH_ALG_ADLER32] = CFS_HASH_ALG_ADLER32,
 854         [BULK_HASH_ALG_CRC32]   = CFS_HASH_ALG_CRC32,
 855         [BULK_HASH_ALG_MD5]     = CFS_HASH_ALG_MD5,
 856         [BULK_HASH_ALG_SHA1]    = CFS_HASH_ALG_SHA1,
 857         [BULK_HASH_ALG_SHA256]  = CFS_HASH_ALG_SHA256,
 858         [BULK_HASH_ALG_SHA384]  = CFS_HASH_ALG_SHA384,
 859         [BULK_HASH_ALG_SHA512]  = CFS_HASH_ALG_SHA512,
 860 };
 861 const char * sptlrpc_get_hash_name(__u8 hash_alg)
 862 {
 863         return cfs_crypto_hash_name(cfs_hash_alg_id[hash_alg]);
 864 }
 865
 866 __u8 sptlrpc_get_hash_alg(const char *algname)
 867 {
 868         return cfs_crypto_hash_alg(algname);
 869 }
 870
 871 int bulk_sec_desc_unpack(struct lustre_msg *msg, int offset, int swabbed)
 872 {
 873         struct ptlrpc_bulk_sec_desc *bsd;
 874         int                          size = msg->lm_buflens[offset];
 875
 876         bsd = lustre_msg_buf(msg, offset, sizeof(*bsd));
 877         if (bsd == NULL) {
 878                 CERROR("Invalid bulk sec desc: size %d\n", size);
 879                 return -EINVAL;
 880         }
 881
 882         if (swabbed) {
 883                 __swab32s(&bsd->bsd_nob);
 884         }
 885
 886         if (unlikely(bsd->bsd_version != 0)) {
 887                 CERROR("Unexpected version %u\n", bsd->bsd_version);
 888                 return -EPROTO;
 889         }
 890
 891         if (unlikely(bsd->bsd_type >= SPTLRPC_BULK_MAX)) {
 892                 CERROR("Invalid type %u\n", bsd->bsd_type);
 893                 return -EPROTO;
 894         }
 895
 896         /* FIXME more sanity check here */
 897
 898         if (unlikely(bsd->bsd_svc != SPTLRPC_BULK_SVC_NULL &&
 899                      bsd->bsd_svc != SPTLRPC_BULK_SVC_INTG &&
 900                      bsd->bsd_svc != SPTLRPC_BULK_SVC_PRIV)) {
 901                 CERROR("Invalid svc %u\n", bsd->bsd_svc);
 902                 return -EPROTO;
 903         }
 904
 905         return 0;
 906 }
 907 EXPORT_SYMBOL(bulk_sec_desc_unpack);
 908
 909 /*
 910  * Compute the checksum of an RPC buffer payload.  If the return \a buflen
 911  * is not large enough, truncate the result to fit so that it is possible
 912  * to use a hash function with a large hash space, but only use a part of
 913  * the resulting hash.
 914  */
 915 int sptlrpc_get_bulk_checksum(struct ptlrpc_bulk_desc *desc, __u8 alg,
 916                               void *buf, int buflen)
 917 {
 918         struct cfs_crypto_hash_desc     *hdesc;
 919         int                             hashsize;
 920         unsigned int                    bufsize;
 921         int                             i, err;
 922
 923         LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
 924         LASSERT(alg > BULK_HASH_ALG_NULL && alg < BULK_HASH_ALG_MAX);
 925         LASSERT(buflen >= 4);
 926
 927         hdesc = cfs_crypto_hash_init(cfs_hash_alg_id[alg], NULL, 0);
 928         if (IS_ERR(hdesc)) {
 929                 CERROR("Unable to initialize checksum hash %s\n",
 930                        cfs_crypto_hash_name(cfs_hash_alg_id[alg]));
 931                 return PTR_ERR(hdesc);
 932         }
 933
 934         hashsize = cfs_crypto_hash_digestsize(cfs_hash_alg_id[alg]);
 935
 936         for (i = 0; i < desc->bd_iov_count; i++) {
 937                 cfs_crypto_hash_update_page(hdesc,
 938                                   BD_GET_KIOV(desc, i).kiov_page,
 939                                   BD_GET_KIOV(desc, i).kiov_offset &
 940                                               ~PAGE_MASK,
 941                                   BD_GET_KIOV(desc, i).kiov_len);
 942         }
 943
 944         if (hashsize > buflen) {
 945                 unsigned char hashbuf[CFS_CRYPTO_HASH_DIGESTSIZE_MAX];
 946
 947                 bufsize = sizeof(hashbuf);
 948                 LASSERTF(bufsize >= hashsize, "bufsize = %u < hashsize %u\n",
 949                          bufsize, hashsize);
 950                 err = cfs_crypto_hash_final(hdesc, hashbuf, &bufsize);
 951                 memcpy(buf, hashbuf, buflen);
 952         } else {
 953                 bufsize = buflen;
 954                 err = cfs_crypto_hash_final(hdesc, buf, &bufsize);
 955         }
 956
 957         return err;
 958 }