1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (C) 2006-2007 Cluster File Systems, Inc.
5 * Author: Eric Mei <ericm@clusterfs.com>
7 * This file is part of Lustre, http://www.lustre.org.
9 * Lustre is free software; you can redistribute it and/or
10 * modify it under the terms of version 2 of the GNU General Public
11 * License as published by the Free Software Foundation.
13 * Lustre is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 * GNU General Public License for more details.
18 * You should have received a copy of the GNU General Public License
19 * along with Lustre; if not, write to the Free Software
20 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
26 #define DEBUG_SUBSYSTEM S_SEC
28 #include <libcfs/libcfs.h>
30 #include <liblustre.h>
31 #include <libcfs/list.h>
33 #include <linux/crypto.h>
37 #include <obd_cksum.h>
38 #include <obd_class.h>
39 #include <obd_support.h>
40 #include <lustre_net.h>
41 #include <lustre_import.h>
42 #include <lustre_dlm.h>
43 #include <lustre_sec.h>
45 #include "ptlrpc_internal.h"
47 /****************************************
48 * bulk encryption page pools *
49 ****************************************/
53 #define PTRS_PER_PAGE (CFS_PAGE_SIZE / sizeof(void *))
54 #define PAGES_PER_POOL (PTRS_PER_PAGE)
56 #define IDLE_IDX_MAX (100)
57 #define IDLE_IDX_WEIGHT (3)
59 #define CACHE_QUIESCENCE_PERIOD (20)
61 static struct ptlrpc_enc_page_pool {
65 unsigned long epp_max_pages; /* maximum pages can hold, const */
66 unsigned int epp_max_pools; /* number of pools, const */
69 * wait queue in case of not enough free pages.
71 cfs_waitq_t epp_waitq; /* waiting threads */
72 unsigned int epp_waitqlen; /* wait queue length */
73 unsigned long epp_pages_short; /* # of pages wanted of in-q users */
74 unsigned int epp_growing:1; /* during adding pages */
77 * indicating how idle the pools are, from 0 to MAX_IDLE_IDX
78 * this is counted based on each time when getting pages from
79 * the pools, not based on time. which means in case that system
80 * is idled for a while but the idle_idx might still be low if no
81 * activities happened in the pools.
83 unsigned long epp_idle_idx;
85 /* last shrink time due to mem tight */
90 * in-pool pages bookkeeping
92 spinlock_t epp_lock; /* protect following fields */
93 unsigned long epp_total_pages; /* total pages in pools */
94 unsigned long epp_free_pages; /* current pages available */
99 unsigned int epp_st_grows; /* # of grows */
100 unsigned int epp_st_grow_fails; /* # of add pages failures */
101 unsigned int epp_st_shrinks; /* # of shrinks */
102 unsigned long epp_st_access; /* # of access */
103 unsigned long epp_st_missings; /* # of cache missing */
104 unsigned long epp_st_lowfree; /* lowest free pages reached */
105 unsigned long epp_st_max_wqlen; /* highest waitqueue length */
106 cfs_time_t epp_st_max_wait; /* in jeffies */
110 cfs_page_t ***epp_pools;
116 const int pools_shrinker_seeks = DEFAULT_SEEKS;
117 static struct shrinker *pools_shrinker = NULL;
121 * /proc/fs/lustre/sptlrpc/encrypt_page_pools
123 int sptlrpc_proc_read_enc_pool(char *page, char **start, off_t off, int count,
124 int *eof, void *data)
128 spin_lock(&page_pools.epp_lock);
130 rc = snprintf(page, count,
131 "physical pages: %lu\n"
132 "pages per pool: %lu\n"
137 "idle index: %lu/100\n"
138 "last shrink: %lds\n"
139 "last access: %lds\n"
141 "grows failure: %u\n"
143 "cache access: %lu\n"
144 "cache missing: %lu\n"
145 "low free mark: %lu\n"
146 "max waitqueue depth: %lu\n"
147 "max wait time: "CFS_TIME_T"/%u\n"
151 page_pools.epp_max_pages,
152 page_pools.epp_max_pools,
153 page_pools.epp_total_pages,
154 page_pools.epp_free_pages,
155 page_pools.epp_idle_idx,
156 cfs_time_current_sec() - page_pools.epp_last_shrink,
157 cfs_time_current_sec() - page_pools.epp_last_access,
158 page_pools.epp_st_grows,
159 page_pools.epp_st_grow_fails,
160 page_pools.epp_st_shrinks,
161 page_pools.epp_st_access,
162 page_pools.epp_st_missings,
163 page_pools.epp_st_lowfree,
164 page_pools.epp_st_max_wqlen,
165 page_pools.epp_st_max_wait, HZ
168 spin_unlock(&page_pools.epp_lock);
172 static void enc_pools_release_free_pages(long npages)
176 LASSERT(npages <= page_pools.epp_free_pages);
178 p_idx = (page_pools.epp_free_pages - 1) / PAGES_PER_POOL;
179 g_idx = (page_pools.epp_free_pages - 1) % PAGES_PER_POOL;
180 LASSERT(page_pools.epp_pools[p_idx]);
182 page_pools.epp_free_pages -= npages;
183 page_pools.epp_total_pages -= npages;
185 while (npages-- > 0) {
186 LASSERT(page_pools.epp_pools[p_idx][g_idx] != NULL);
188 cfs_free_page(page_pools.epp_pools[p_idx][g_idx]);
189 page_pools.epp_pools[p_idx][g_idx] = NULL;
193 g_idx = PAGES_PER_POOL - 1;
195 LASSERT(page_pools.epp_pools[p_idx]);
201 * could be called frequently for query (@nr_to_scan == 0)
203 static int enc_pools_shrink(int nr_to_scan, unsigned int gfp_mask)
207 spin_lock(&page_pools.epp_lock);
210 if (nr_to_scan > page_pools.epp_free_pages)
211 nr_to_scan = page_pools.epp_free_pages;
213 enc_pools_release_free_pages(nr_to_scan);
214 CDEBUG(D_SEC, "released %d pages, %ld left\n",
215 nr_to_scan, page_pools.epp_free_pages);
217 page_pools.epp_st_shrinks++;
218 page_pools.epp_last_shrink = cfs_time_current_sec();
222 * try to keep at least PTLRPC_MAX_BRW_PAGES pages in the pool
224 if (page_pools.epp_free_pages <= PTLRPC_MAX_BRW_PAGES) {
230 * if no pool access for a long time, we consider it's fully idle
232 if (cfs_time_current_sec() - page_pools.epp_last_access >
233 CACHE_QUIESCENCE_PERIOD)
234 page_pools.epp_idle_idx = IDLE_IDX_MAX;
236 LASSERT(page_pools.epp_idle_idx <= IDLE_IDX_MAX);
237 ret = (page_pools.epp_free_pages * page_pools.epp_idle_idx /
239 if (page_pools.epp_free_pages - ret < PTLRPC_MAX_BRW_PAGES)
240 ret = page_pools.epp_free_pages - PTLRPC_MAX_BRW_PAGES;
243 spin_unlock(&page_pools.epp_lock);
248 int npages_to_npools(unsigned long npages)
250 return (int) ((npages + PAGES_PER_POOL - 1) / PAGES_PER_POOL);
254 * return how many pages cleaned up.
256 static unsigned long enc_pools_cleanup(cfs_page_t ***pools, int npools)
258 unsigned long cleaned = 0;
261 for (i = 0; i < npools; i++) {
263 for (j = 0; j < PAGES_PER_POOL; j++) {
265 cfs_free_page(pools[i][j]);
269 OBD_FREE(pools[i], CFS_PAGE_SIZE);
278 * merge @npools pointed by @pools which contains @npages new pages
279 * into current pools.
281 * we have options to avoid most memory copy with some tricks. but we choose
282 * the simplest way to avoid complexity. It's not frequently called.
284 static void enc_pools_insert(cfs_page_t ***pools, int npools, int npages)
287 int op_idx, np_idx, og_idx, ng_idx;
288 int cur_npools, end_npools;
291 LASSERT(page_pools.epp_total_pages+npages <= page_pools.epp_max_pages);
292 LASSERT(npages_to_npools(npages) == npools);
294 spin_lock(&page_pools.epp_lock);
297 * (1) fill all the free slots of current pools.
299 /* free slots are those left by rent pages, and the extra ones with
300 * index >= eep_total_pages, locate at the tail of last pool. */
301 freeslot = page_pools.epp_total_pages % PAGES_PER_POOL;
303 freeslot = PAGES_PER_POOL - freeslot;
304 freeslot += page_pools.epp_total_pages - page_pools.epp_free_pages;
306 op_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
307 og_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
309 ng_idx = (npages - 1) % PAGES_PER_POOL;
312 LASSERT(page_pools.epp_pools[op_idx][og_idx] == NULL);
313 LASSERT(pools[np_idx][ng_idx] != NULL);
315 page_pools.epp_pools[op_idx][og_idx] = pools[np_idx][ng_idx];
316 pools[np_idx][ng_idx] = NULL;
320 if (++og_idx == PAGES_PER_POOL) {
328 ng_idx = PAGES_PER_POOL - 1;
333 * (2) add pools if needed.
335 cur_npools = (page_pools.epp_total_pages + PAGES_PER_POOL - 1) /
337 end_npools = (page_pools.epp_total_pages + npages + PAGES_PER_POOL -1) /
339 LASSERT(end_npools <= page_pools.epp_max_pools);
342 while (cur_npools < end_npools) {
343 LASSERT(page_pools.epp_pools[cur_npools] == NULL);
344 LASSERT(np_idx < npools);
345 LASSERT(pools[np_idx] != NULL);
347 page_pools.epp_pools[cur_npools++] = pools[np_idx];
348 pools[np_idx++] = NULL;
351 page_pools.epp_total_pages += npages;
352 page_pools.epp_free_pages += npages;
353 page_pools.epp_st_lowfree = page_pools.epp_free_pages;
355 CDEBUG(D_SEC, "add %d pages to total %lu\n", npages,
356 page_pools.epp_total_pages);
358 spin_unlock(&page_pools.epp_lock);
361 static int enc_pools_add_pages(int npages)
363 static DECLARE_MUTEX(sem_add_pages);
365 int npools, alloced = 0;
366 int i, j, rc = -ENOMEM;
368 if (npages < PTLRPC_MAX_BRW_PAGES)
369 npages = PTLRPC_MAX_BRW_PAGES;
371 down(&sem_add_pages);
373 if (npages + page_pools.epp_total_pages > page_pools.epp_max_pages)
374 npages = page_pools.epp_max_pages - page_pools.epp_total_pages;
377 page_pools.epp_st_grows++;
379 npools = npages_to_npools(npages);
380 OBD_ALLOC(pools, npools * sizeof(*pools));
384 for (i = 0; i < npools; i++) {
385 OBD_ALLOC(pools[i], CFS_PAGE_SIZE);
386 if (pools[i] == NULL)
389 for (j = 0; j < PAGES_PER_POOL && alloced < npages; j++) {
390 pools[i][j] = cfs_alloc_page(CFS_ALLOC_IO |
392 if (pools[i][j] == NULL)
399 enc_pools_insert(pools, npools, npages);
400 CDEBUG(D_SEC, "added %d pages into pools\n", npages);
404 enc_pools_cleanup(pools, npools);
405 OBD_FREE(pools, npools * sizeof(*pools));
408 page_pools.epp_st_grow_fails++;
409 CERROR("Failed to allocate %d enc pages\n", npages);
416 static inline void enc_pools_wakeup(void)
418 if (unlikely(page_pools.epp_waitqlen)) {
419 LASSERT(page_pools.epp_waitqlen > 0);
420 LASSERT(cfs_waitq_active(&page_pools.epp_waitq));
421 cfs_waitq_broadcast(&page_pools.epp_waitq);
425 static int enc_pools_should_grow(int page_needed, long now)
427 /* don't grow if someone else is growing the pools right now,
428 * or the pools has reached its full capacity
430 if (page_pools.epp_growing ||
431 page_pools.epp_total_pages == page_pools.epp_max_pages)
434 /* if total pages is not enough, we need to grow */
435 if (page_pools.epp_total_pages < page_needed)
438 /* if we just did a shrink due to memory tight, we'd better
439 * wait a while to grow again.
441 if (now - page_pools.epp_last_shrink < 2)
445 * here we perhaps need consider other factors like wait queue
446 * length, idle index, etc. ?
449 /* grow the pools in any other cases */
454 * we allocate the requested pages atomically.
456 int sptlrpc_enc_pool_get_pages(struct ptlrpc_bulk_desc *desc)
458 cfs_waitlink_t waitlink;
459 unsigned long this_idle = -1;
465 LASSERT(desc->bd_max_iov > 0);
466 LASSERT(desc->bd_max_iov <= page_pools.epp_max_pages);
468 /* resent bulk, enc pages might have been allocated previously */
469 if (desc->bd_enc_pages != NULL)
472 OBD_ALLOC(desc->bd_enc_pages,
473 desc->bd_max_iov * sizeof(*desc->bd_enc_pages));
474 if (desc->bd_enc_pages == NULL)
477 spin_lock(&page_pools.epp_lock);
479 page_pools.epp_st_access++;
481 if (unlikely(page_pools.epp_free_pages < desc->bd_max_iov)) {
483 tick = cfs_time_current();
485 now = cfs_time_current_sec();
487 page_pools.epp_st_missings++;
488 page_pools.epp_pages_short += desc->bd_max_iov;
490 if (enc_pools_should_grow(desc->bd_max_iov, now)) {
491 page_pools.epp_growing = 1;
493 spin_unlock(&page_pools.epp_lock);
494 enc_pools_add_pages(page_pools.epp_pages_short / 2);
495 spin_lock(&page_pools.epp_lock);
497 page_pools.epp_growing = 0;
499 if (++page_pools.epp_waitqlen >
500 page_pools.epp_st_max_wqlen)
501 page_pools.epp_st_max_wqlen =
502 page_pools.epp_waitqlen;
504 set_current_state(TASK_UNINTERRUPTIBLE);
505 cfs_waitlink_init(&waitlink);
506 cfs_waitq_add(&page_pools.epp_waitq, &waitlink);
508 spin_unlock(&page_pools.epp_lock);
510 spin_lock(&page_pools.epp_lock);
512 LASSERT(page_pools.epp_waitqlen > 0);
513 page_pools.epp_waitqlen--;
516 LASSERT(page_pools.epp_pages_short >= desc->bd_max_iov);
517 page_pools.epp_pages_short -= desc->bd_max_iov;
523 /* record max wait time */
524 if (unlikely(tick != 0)) {
525 tick = cfs_time_current() - tick;
526 if (tick > page_pools.epp_st_max_wait)
527 page_pools.epp_st_max_wait = tick;
530 /* proceed with rest of allocation */
531 page_pools.epp_free_pages -= desc->bd_max_iov;
533 p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
534 g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
536 for (i = 0; i < desc->bd_max_iov; i++) {
537 LASSERT(page_pools.epp_pools[p_idx][g_idx] != NULL);
538 desc->bd_enc_pages[i] = page_pools.epp_pools[p_idx][g_idx];
539 page_pools.epp_pools[p_idx][g_idx] = NULL;
541 if (++g_idx == PAGES_PER_POOL) {
547 if (page_pools.epp_free_pages < page_pools.epp_st_lowfree)
548 page_pools.epp_st_lowfree = page_pools.epp_free_pages;
551 * new idle index = (old * weight + new) / (weight + 1)
553 if (this_idle == -1) {
554 this_idle = page_pools.epp_free_pages * IDLE_IDX_MAX /
555 page_pools.epp_total_pages;
557 page_pools.epp_idle_idx = (page_pools.epp_idle_idx * IDLE_IDX_WEIGHT +
559 (IDLE_IDX_WEIGHT + 1);
561 page_pools.epp_last_access = cfs_time_current_sec();
563 spin_unlock(&page_pools.epp_lock);
566 EXPORT_SYMBOL(sptlrpc_enc_pool_get_pages);
568 void sptlrpc_enc_pool_put_pages(struct ptlrpc_bulk_desc *desc)
573 if (desc->bd_enc_pages == NULL)
575 if (desc->bd_max_iov == 0)
578 spin_lock(&page_pools.epp_lock);
580 p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
581 g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
583 LASSERT(page_pools.epp_free_pages + desc->bd_max_iov <=
584 page_pools.epp_total_pages);
585 LASSERT(page_pools.epp_pools[p_idx]);
587 for (i = 0; i < desc->bd_max_iov; i++) {
588 LASSERT(desc->bd_enc_pages[i] != NULL);
589 LASSERT(g_idx != 0 || page_pools.epp_pools[p_idx]);
590 LASSERT(page_pools.epp_pools[p_idx][g_idx] == NULL);
592 page_pools.epp_pools[p_idx][g_idx] = desc->bd_enc_pages[i];
594 if (++g_idx == PAGES_PER_POOL) {
600 page_pools.epp_free_pages += desc->bd_max_iov;
604 spin_unlock(&page_pools.epp_lock);
606 OBD_FREE(desc->bd_enc_pages,
607 desc->bd_max_iov * sizeof(*desc->bd_enc_pages));
608 desc->bd_enc_pages = NULL;
610 EXPORT_SYMBOL(sptlrpc_enc_pool_put_pages);
613 * we don't do much stuff for add_user/del_user anymore, except adding some
614 * initial pages in add_user() if current pools are empty, rest would be
615 * handled by the pools's self-adaption.
617 int sptlrpc_enc_pool_add_user(void)
621 spin_lock(&page_pools.epp_lock);
622 if (page_pools.epp_growing == 0 && page_pools.epp_total_pages == 0) {
623 page_pools.epp_growing = 1;
626 spin_unlock(&page_pools.epp_lock);
629 enc_pools_add_pages(PTLRPC_MAX_BRW_PAGES);
631 spin_lock(&page_pools.epp_lock);
632 page_pools.epp_growing = 0;
634 spin_unlock(&page_pools.epp_lock);
638 EXPORT_SYMBOL(sptlrpc_enc_pool_add_user);
640 int sptlrpc_enc_pool_del_user(void)
644 EXPORT_SYMBOL(sptlrpc_enc_pool_del_user);
646 static inline void enc_pools_alloc(void)
648 LASSERT(page_pools.epp_max_pools);
650 * on system with huge memory but small page size, this might lead to
651 * high-order allocation. but it's not common, and we suppose memory
652 * be not too much fragmented at module loading time.
654 OBD_ALLOC(page_pools.epp_pools,
655 page_pools.epp_max_pools * sizeof(*page_pools.epp_pools));
658 static inline void enc_pools_free(void)
660 LASSERT(page_pools.epp_max_pools);
661 LASSERT(page_pools.epp_pools);
663 OBD_FREE(page_pools.epp_pools,
664 page_pools.epp_max_pools * sizeof(*page_pools.epp_pools));
667 int sptlrpc_enc_pool_init(void)
670 * maximum capacity is 1/8 of total physical memory.
671 * is the 1/8 a good number?
673 page_pools.epp_max_pages = num_physpages / 8;
674 page_pools.epp_max_pools = npages_to_npools(page_pools.epp_max_pages);
676 cfs_waitq_init(&page_pools.epp_waitq);
677 page_pools.epp_waitqlen = 0;
678 page_pools.epp_pages_short = 0;
680 page_pools.epp_growing = 0;
682 page_pools.epp_idle_idx = 0;
683 page_pools.epp_last_shrink = cfs_time_current_sec();
684 page_pools.epp_last_access = cfs_time_current_sec();
686 spin_lock_init(&page_pools.epp_lock);
687 page_pools.epp_total_pages = 0;
688 page_pools.epp_free_pages = 0;
690 page_pools.epp_st_grows = 0;
691 page_pools.epp_st_grow_fails = 0;
692 page_pools.epp_st_shrinks = 0;
693 page_pools.epp_st_access = 0;
694 page_pools.epp_st_missings = 0;
695 page_pools.epp_st_lowfree = 0;
696 page_pools.epp_st_max_wqlen = 0;
697 page_pools.epp_st_max_wait = 0;
700 if (page_pools.epp_pools == NULL)
703 pools_shrinker = set_shrinker(pools_shrinker_seeks, enc_pools_shrink);
704 if (pools_shrinker == NULL) {
712 void sptlrpc_enc_pool_fini(void)
714 unsigned long cleaned, npools;
716 LASSERT(pools_shrinker);
717 LASSERT(page_pools.epp_pools);
718 LASSERT(page_pools.epp_total_pages == page_pools.epp_free_pages);
720 remove_shrinker(pools_shrinker);
722 npools = npages_to_npools(page_pools.epp_total_pages);
723 cleaned = enc_pools_cleanup(page_pools.epp_pools, npools);
724 LASSERT(cleaned == page_pools.epp_total_pages);
729 #else /* !__KERNEL__ */
731 int sptlrpc_enc_pool_get_pages(struct ptlrpc_bulk_desc *desc)
736 void sptlrpc_enc_pool_put_pages(struct ptlrpc_bulk_desc *desc)
740 int sptlrpc_enc_pool_init(void)
745 void sptlrpc_enc_pool_fini(void)
750 /****************************************
751 * Helpers to assist policy modules to *
752 * implement checksum funcationality *
753 ****************************************/
755 static struct sptlrpc_hash_type hash_types[] = {
756 [BULK_HASH_ALG_NULL] = { "null", "null", 0 },
757 [BULK_HASH_ALG_ADLER32] = { "adler32", "adler32", 4 },
758 [BULK_HASH_ALG_CRC32] = { "crc32", "crc32", 4 },
759 [BULK_HASH_ALG_MD5] = { "md5", "md5", 16 },
760 [BULK_HASH_ALG_SHA1] = { "sha1", "sha1", 20 },
761 [BULK_HASH_ALG_SHA256] = { "sha256", "sha256", 32 },
762 [BULK_HASH_ALG_SHA384] = { "sha384", "sha384", 48 },
763 [BULK_HASH_ALG_SHA512] = { "sha512", "sha512", 64 },
764 [BULK_HASH_ALG_WP256] = { "wp256", "wp256", 32 },
765 [BULK_HASH_ALG_WP384] = { "wp384", "wp384", 48 },
766 [BULK_HASH_ALG_WP512] = { "wp512", "wp512", 64 },
769 const struct sptlrpc_hash_type *sptlrpc_get_hash_type(__u8 hash_alg)
771 struct sptlrpc_hash_type *ht;
773 if (hash_alg < BULK_HASH_ALG_MAX) {
774 ht = &hash_types[hash_alg];
775 if (ht->sht_tfm_name)
780 EXPORT_SYMBOL(sptlrpc_get_hash_type);
782 const char * sptlrpc_get_hash_name(__u8 hash_alg)
784 const struct sptlrpc_hash_type *ht;
786 ht = sptlrpc_get_hash_type(hash_alg);
792 EXPORT_SYMBOL(sptlrpc_get_hash_name);
794 int bulk_sec_desc_size(__u8 hash_alg, int request, int read)
796 int size = sizeof(struct ptlrpc_bulk_sec_desc);
798 LASSERT(hash_alg < BULK_HASH_ALG_MAX);
800 /* read request don't need extra data */
801 if (!(read && request))
802 size += hash_types[hash_alg].sht_size;
806 EXPORT_SYMBOL(bulk_sec_desc_size);
808 int bulk_sec_desc_unpack(struct lustre_msg *msg, int offset)
810 struct ptlrpc_bulk_sec_desc *bsd;
811 int size = msg->lm_buflens[offset];
813 bsd = lustre_msg_buf(msg, offset, sizeof(*bsd));
815 CERROR("Invalid bulk sec desc: size %d\n", size);
819 /* nothing to swab */
821 if (unlikely(bsd->bsd_version != 0)) {
822 CERROR("Unexpected version %u\n", bsd->bsd_version);
826 if (unlikely(bsd->bsd_flags != 0)) {
827 CERROR("Unexpected flags %x\n", bsd->bsd_flags);
831 if (unlikely(!sptlrpc_get_hash_type(bsd->bsd_hash_alg))) {
832 CERROR("Unsupported checksum algorithm %u\n",
837 if (unlikely(!sptlrpc_get_ciph_type(bsd->bsd_ciph_alg))) {
838 CERROR("Unsupported cipher algorithm %u\n",
843 if (unlikely(size > sizeof(*bsd)) &&
844 size < sizeof(*bsd) + hash_types[bsd->bsd_hash_alg].sht_size) {
845 CERROR("Mal-formed checksum data: csum alg %u, size %d\n",
846 bsd->bsd_hash_alg, size);
852 EXPORT_SYMBOL(bulk_sec_desc_unpack);
857 static int do_bulk_checksum_adler32(struct ptlrpc_bulk_desc *desc, void *buf)
865 for (i = 0; i < desc->bd_iov_count; i++) {
866 page = desc->bd_iov[i].kiov_page;
867 off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK;
868 ptr = cfs_kmap(page) + off;
869 len = desc->bd_iov[i].kiov_len;
871 adler32 = adler32(adler32, ptr, len);
876 adler32 = cpu_to_le32(adler32);
877 memcpy(buf, &adler32, sizeof(adler32));
882 static int do_bulk_checksum_crc32(struct ptlrpc_bulk_desc *desc, void *buf)
890 for (i = 0; i < desc->bd_iov_count; i++) {
891 page = desc->bd_iov[i].kiov_page;
892 off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK;
893 ptr = cfs_kmap(page) + off;
894 len = desc->bd_iov[i].kiov_len;
896 crc32 = crc32_le(crc32, ptr, len);
901 crc32 = cpu_to_le32(crc32);
902 memcpy(buf, &crc32, sizeof(crc32));
906 static int do_bulk_checksum(struct ptlrpc_bulk_desc *desc, __u32 alg, void *buf)
908 struct hash_desc hdesc;
909 struct scatterlist *sl;
910 int i, rc = 0, bytes = 0;
912 LASSERT(alg > BULK_HASH_ALG_NULL &&
913 alg < BULK_HASH_ALG_MAX);
916 case BULK_HASH_ALG_ADLER32:
918 return do_bulk_checksum_adler32(desc, buf);
920 CERROR("Adler32 not supported\n");
923 case BULK_HASH_ALG_CRC32:
924 return do_bulk_checksum_crc32(desc, buf);
927 hdesc.tfm = ll_crypto_alloc_hash(hash_types[alg].sht_tfm_name, 0, 0);
928 if (hdesc.tfm == NULL) {
929 CERROR("Unable to allocate TFM %s\n", hash_types[alg].sht_name);
934 OBD_ALLOC(sl, sizeof(*sl) * desc->bd_iov_count);
940 for (i = 0; i < desc->bd_iov_count; i++) {
941 sl[i].page = desc->bd_iov[i].kiov_page;
942 sl[i].offset = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK;
943 sl[i].length = desc->bd_iov[i].kiov_len;
944 bytes += desc->bd_iov[i].kiov_len;
947 ll_crypto_hash_init(&hdesc);
948 ll_crypto_hash_update(&hdesc, sl, bytes);
949 ll_crypto_hash_final(&hdesc, buf);
951 OBD_FREE(sl, sizeof(*sl) * desc->bd_iov_count);
954 ll_crypto_free_hash(hdesc.tfm);
958 #else /* !__KERNEL__ */
960 static int do_bulk_checksum(struct ptlrpc_bulk_desc *desc, __u32 alg, void *buf)
965 LASSERT(alg == BULK_HASH_ALG_ADLER32 || alg == BULK_HASH_ALG_CRC32);
967 if (alg == BULK_HASH_ALG_ADLER32)
972 for (i = 0; i < desc->bd_iov_count; i++) {
973 char *ptr = desc->bd_iov[i].iov_base;
974 int len = desc->bd_iov[i].iov_len;
977 case BULK_HASH_ALG_ADLER32:
979 csum32 = adler32(csum32, ptr, len);
981 CERROR("Adler32 not supported\n");
985 case BULK_HASH_ALG_CRC32:
986 csum32 = crc32_le(csum32, ptr, len);
991 csum32 = cpu_to_le32(csum32);
992 memcpy(buf, &csum32, sizeof(csum32));
996 #endif /* __KERNEL__ */
999 * perform algorithm @alg checksum on @desc, store result in @buf.
1000 * if anything goes wrong, leave 'alg' be BULK_HASH_ALG_NULL.
1003 int generate_bulk_csum(struct ptlrpc_bulk_desc *desc, __u32 alg,
1004 struct ptlrpc_bulk_sec_desc *bsd, int bsdsize)
1009 LASSERT(alg < BULK_HASH_ALG_MAX);
1011 bsd->bsd_hash_alg = BULK_HASH_ALG_NULL;
1013 if (alg == BULK_HASH_ALG_NULL)
1016 LASSERT(bsdsize >= sizeof(*bsd) + hash_types[alg].sht_size);
1018 rc = do_bulk_checksum(desc, alg, bsd->bsd_csum);
1020 bsd->bsd_hash_alg = alg;
1026 int verify_bulk_csum(struct ptlrpc_bulk_desc *desc, int read,
1027 struct ptlrpc_bulk_sec_desc *bsdv, int bsdvsize,
1028 struct ptlrpc_bulk_sec_desc *bsdr, int bsdrsize)
1032 int csum_size, rc = 0;
1035 LASSERT(bsdv->bsd_hash_alg < BULK_HASH_ALG_MAX);
1038 bsdr->bsd_hash_alg = BULK_HASH_ALG_NULL;
1040 if (bsdv->bsd_hash_alg == BULK_HASH_ALG_NULL)
1043 /* for all supported algorithms */
1044 csum_size = hash_types[bsdv->bsd_hash_alg].sht_size;
1046 if (bsdvsize < sizeof(*bsdv) + csum_size) {
1047 CERROR("verifier size %d too small, require %d\n",
1048 bsdvsize, (int) sizeof(*bsdv) + csum_size);
1053 LASSERT(bsdrsize >= sizeof(*bsdr) + csum_size);
1054 csum_p = (char *) bsdr->bsd_csum;
1056 OBD_ALLOC(buf, csum_size);
1062 rc = do_bulk_checksum(desc, bsdv->bsd_hash_alg, csum_p);
1064 if (memcmp(bsdv->bsd_csum, csum_p, csum_size)) {
1065 CERROR("BAD %s CHECKSUM (%s), data mutated during "
1066 "transfer!\n", read ? "READ" : "WRITE",
1067 hash_types[bsdv->bsd_hash_alg].sht_name);
1070 CDEBUG(D_SEC, "bulk %s checksum (%s) verified\n",
1071 read ? "read" : "write",
1072 hash_types[bsdv->bsd_hash_alg].sht_name);
1076 bsdr->bsd_hash_alg = bsdv->bsd_hash_alg;
1077 memcpy(bsdr->bsd_csum, csum_p, csum_size);
1080 OBD_FREE(buf, csum_size);
1086 int bulk_csum_cli_request(struct ptlrpc_bulk_desc *desc, int read,
1087 __u32 alg, struct lustre_msg *rmsg, int roff)
1089 struct ptlrpc_bulk_sec_desc *bsdr;
1092 rsize = rmsg->lm_buflens[roff];
1093 bsdr = lustre_msg_buf(rmsg, roff, sizeof(*bsdr));
1096 LASSERT(rsize >= sizeof(*bsdr));
1097 LASSERT(alg < BULK_HASH_ALG_MAX);
1100 bsdr->bsd_hash_alg = alg;
1102 rc = generate_bulk_csum(desc, alg, bsdr, rsize);
1104 CERROR("bulk write: client failed to compute "
1105 "checksum: %d\n", rc);
1107 /* For sending we only compute the wrong checksum instead
1108 * of corrupting the data so it is still correct on a redo */
1109 if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND) &&
1110 bsdr->bsd_hash_alg != BULK_HASH_ALG_NULL)
1111 bsdr->bsd_csum[0] ^= 0x1;
1116 EXPORT_SYMBOL(bulk_csum_cli_request);
1118 int bulk_csum_cli_reply(struct ptlrpc_bulk_desc *desc, int read,
1119 struct lustre_msg *rmsg, int roff,
1120 struct lustre_msg *vmsg, int voff)
1122 struct ptlrpc_bulk_sec_desc *bsdv, *bsdr;
1125 rsize = rmsg->lm_buflens[roff];
1126 vsize = vmsg->lm_buflens[voff];
1127 bsdr = lustre_msg_buf(rmsg, roff, 0);
1128 bsdv = lustre_msg_buf(vmsg, voff, 0);
1130 if (bsdv == NULL || vsize < sizeof(*bsdv)) {
1131 CERROR("Invalid checksum verifier from server: size %d\n",
1137 LASSERT(rsize >= sizeof(*bsdr));
1138 LASSERT(vsize >= sizeof(*bsdv));
1140 if (bsdr->bsd_hash_alg != bsdv->bsd_hash_alg) {
1141 CERROR("bulk %s: checksum algorithm mismatch: client request "
1142 "%s but server reply with %s. try to use the new one "
1143 "for checksum verification\n",
1144 read ? "read" : "write",
1145 hash_types[bsdr->bsd_hash_alg].sht_name,
1146 hash_types[bsdv->bsd_hash_alg].sht_name);
1150 return verify_bulk_csum(desc, 1, bsdv, vsize, NULL, 0);
1152 char *cli, *srv, *new = NULL;
1153 int csum_size = hash_types[bsdr->bsd_hash_alg].sht_size;
1155 LASSERT(bsdr->bsd_hash_alg < BULK_HASH_ALG_MAX);
1156 if (bsdr->bsd_hash_alg == BULK_HASH_ALG_NULL)
1159 if (vsize < sizeof(*bsdv) + csum_size) {
1160 CERROR("verifier size %d too small, require %d\n",
1161 vsize, (int) sizeof(*bsdv) + csum_size);
1165 cli = (char *) (bsdr + 1);
1166 srv = (char *) (bsdv + 1);
1168 if (!memcmp(cli, srv, csum_size)) {
1169 /* checksum confirmed */
1170 CDEBUG(D_SEC, "bulk write checksum (%s) confirmed\n",
1171 hash_types[bsdr->bsd_hash_alg].sht_name);
1175 /* checksum mismatch, re-compute a new one and compare with
1176 * others, give out proper warnings. */
1177 OBD_ALLOC(new, csum_size);
1181 do_bulk_checksum(desc, bsdr->bsd_hash_alg, new);
1183 if (!memcmp(new, srv, csum_size)) {
1184 CERROR("BAD WRITE CHECKSUM (%s): pages were mutated "
1185 "on the client after we checksummed them\n",
1186 hash_types[bsdr->bsd_hash_alg].sht_name);
1187 } else if (!memcmp(new, cli, csum_size)) {
1188 CERROR("BAD WRITE CHECKSUM (%s): pages were mutated "
1190 hash_types[bsdr->bsd_hash_alg].sht_name);
1192 CERROR("BAD WRITE CHECKSUM (%s): pages were mutated "
1193 "in transit, and the current page contents "
1194 "don't match the originals and what the server "
1196 hash_types[bsdr->bsd_hash_alg].sht_name);
1198 OBD_FREE(new, csum_size);
1203 EXPORT_SYMBOL(bulk_csum_cli_reply);
1206 static void corrupt_bulk_data(struct ptlrpc_bulk_desc *desc)
1209 unsigned int off, i;
1211 for (i = 0; i < desc->bd_iov_count; i++) {
1212 if (desc->bd_iov[i].kiov_len == 0)
1215 ptr = cfs_kmap(desc->bd_iov[i].kiov_page);
1216 off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK;
1218 cfs_kunmap(desc->bd_iov[i].kiov_page);
1223 static void corrupt_bulk_data(struct ptlrpc_bulk_desc *desc)
1226 #endif /* __KERNEL__ */
1228 int bulk_csum_svc(struct ptlrpc_bulk_desc *desc, int read,
1229 struct ptlrpc_bulk_sec_desc *bsdv, int vsize,
1230 struct ptlrpc_bulk_sec_desc *bsdr, int rsize)
1234 LASSERT(vsize >= sizeof(*bsdv));
1235 LASSERT(rsize >= sizeof(*bsdr));
1236 LASSERT(bsdv && bsdr);
1239 rc = generate_bulk_csum(desc, bsdv->bsd_hash_alg, bsdr, rsize);
1241 CERROR("bulk read: server failed to generate %s "
1243 hash_types[bsdv->bsd_hash_alg].sht_name, rc);
1245 /* corrupt the data after we compute the checksum, to
1246 * simulate an OST->client data error */
1247 if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE))
1248 corrupt_bulk_data(desc);
1250 rc = verify_bulk_csum(desc, 0, bsdv, vsize, bsdr, rsize);
1255 EXPORT_SYMBOL(bulk_csum_svc);
1257 /****************************************
1258 * Helpers to assist policy modules to *
1259 * implement encryption funcationality *
1260 ****************************************/
1264 #define CRYPTO_TFM_MODE_ECB (0)
1265 #define CRYPTO_TFM_MODE_CBC (1)
1268 static struct sptlrpc_ciph_type cipher_types[] = {
1269 [BULK_CIPH_ALG_NULL] = {
1270 "null", "null", 0, 0, 0
1272 [BULK_CIPH_ALG_ARC4] = {
1273 "arc4", "ecb(arc4)", 0, 0, 16
1275 [BULK_CIPH_ALG_AES128] = {
1276 "aes128", "cbc(aes)", 0, 16, 16
1278 [BULK_CIPH_ALG_AES192] = {
1279 "aes192", "cbc(aes)", 0, 16, 24
1281 [BULK_CIPH_ALG_AES256] = {
1282 "aes256", "cbc(aes)", 0, 16, 32
1284 [BULK_CIPH_ALG_CAST128] = {
1285 "cast128", "cbc(cast5)", 0, 8, 16
1287 [BULK_CIPH_ALG_CAST256] = {
1288 "cast256", "cbc(cast6)", 0, 16, 32
1290 [BULK_CIPH_ALG_TWOFISH128] = {
1291 "twofish128", "cbc(twofish)", 0, 16, 16
1293 [BULK_CIPH_ALG_TWOFISH256] = {
1294 "twofish256", "cbc(twofish)", 0, 16, 32
1298 const struct sptlrpc_ciph_type *sptlrpc_get_ciph_type(__u8 ciph_alg)
1300 struct sptlrpc_ciph_type *ct;
1302 if (ciph_alg < BULK_CIPH_ALG_MAX) {
1303 ct = &cipher_types[ciph_alg];
1304 if (ct->sct_tfm_name)
1309 EXPORT_SYMBOL(sptlrpc_get_ciph_type);
1311 const char *sptlrpc_get_ciph_name(__u8 ciph_alg)
1313 const struct sptlrpc_ciph_type *ct;
1315 ct = sptlrpc_get_ciph_type(ciph_alg);
1317 return ct->sct_name;
1321 EXPORT_SYMBOL(sptlrpc_get_ciph_name);