4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
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.
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).
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
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2015, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/ptlrpc/sec_bulk.c
34 * Author: Eric Mei <ericm@clusterfs.com>
37 #define DEBUG_SUBSYSTEM S_SEC
39 #include "../../include/linux/libcfs/libcfs.h"
41 #include "../include/obd.h"
42 #include "../include/obd_cksum.h"
43 #include "../include/obd_class.h"
44 #include "../include/obd_support.h"
45 #include "../include/lustre_net.h"
46 #include "../include/lustre_import.h"
47 #include "../include/lustre_dlm.h"
48 #include "../include/lustre_sec.h"
50 #include "ptlrpc_internal.h"
52 /****************************************
53 * bulk encryption page pools *
54 ****************************************/
56 #define POINTERS_PER_PAGE (PAGE_SIZE / sizeof(void *))
57 #define PAGES_PER_POOL (POINTERS_PER_PAGE)
59 #define IDLE_IDX_MAX (100)
60 #define IDLE_IDX_WEIGHT (3)
62 #define CACHE_QUIESCENT_PERIOD (20)
64 static struct ptlrpc_enc_page_pool {
68 unsigned long epp_max_pages; /* maximum pages can hold, const */
69 unsigned int epp_max_pools; /* number of pools, const */
72 * wait queue in case of not enough free pages.
74 wait_queue_head_t epp_waitq; /* waiting threads */
75 unsigned int epp_waitqlen; /* wait queue length */
76 unsigned long epp_pages_short; /* # of pages wanted of in-q users */
77 unsigned int epp_growing:1; /* during adding pages */
80 * indicating how idle the pools are, from 0 to MAX_IDLE_IDX
81 * this is counted based on each time when getting pages from
82 * the pools, not based on time. which means in case that system
83 * is idled for a while but the idle_idx might still be low if no
84 * activities happened in the pools.
86 unsigned long epp_idle_idx;
88 /* last shrink time due to mem tight */
89 time64_t epp_last_shrink;
90 time64_t epp_last_access;
93 * in-pool pages bookkeeping
95 spinlock_t epp_lock; /* protect following fields */
96 unsigned long epp_total_pages; /* total pages in pools */
97 unsigned long epp_free_pages; /* current pages available */
102 unsigned long epp_st_max_pages; /* # of pages ever reached */
103 unsigned int epp_st_grows; /* # of grows */
104 unsigned int epp_st_grow_fails; /* # of add pages failures */
105 unsigned int epp_st_shrinks; /* # of shrinks */
106 unsigned long epp_st_access; /* # of access */
107 unsigned long epp_st_missings; /* # of cache missing */
108 unsigned long epp_st_lowfree; /* lowest free pages reached */
109 unsigned int epp_st_max_wqlen; /* highest waitqueue length */
110 unsigned long epp_st_max_wait; /* in jiffies */
114 struct page ***epp_pools;
118 * /sys/kernel/debug/lustre/sptlrpc/encrypt_page_pools
120 int sptlrpc_proc_enc_pool_seq_show(struct seq_file *m, void *v)
122 spin_lock(&page_pools.epp_lock);
125 "physical pages: %lu\n"
126 "pages per pool: %lu\n"
131 "idle index: %lu/100\n"
132 "last shrink: %lds\n"
133 "last access: %lds\n"
134 "max pages reached: %lu\n"
136 "grows failure: %u\n"
138 "cache access: %lu\n"
139 "cache missing: %lu\n"
140 "low free mark: %lu\n"
141 "max waitqueue depth: %u\n"
142 "max wait time: %ld/%lu\n",
145 page_pools.epp_max_pages,
146 page_pools.epp_max_pools,
147 page_pools.epp_total_pages,
148 page_pools.epp_free_pages,
149 page_pools.epp_idle_idx,
150 (long)(ktime_get_seconds() - page_pools.epp_last_shrink),
151 (long)(ktime_get_seconds() - page_pools.epp_last_access),
152 page_pools.epp_st_max_pages,
153 page_pools.epp_st_grows,
154 page_pools.epp_st_grow_fails,
155 page_pools.epp_st_shrinks,
156 page_pools.epp_st_access,
157 page_pools.epp_st_missings,
158 page_pools.epp_st_lowfree,
159 page_pools.epp_st_max_wqlen,
160 page_pools.epp_st_max_wait,
161 msecs_to_jiffies(MSEC_PER_SEC));
163 spin_unlock(&page_pools.epp_lock);
168 static void enc_pools_release_free_pages(long npages)
171 int p_idx_max1, p_idx_max2;
174 LASSERT(npages <= page_pools.epp_free_pages);
175 LASSERT(page_pools.epp_free_pages <= page_pools.epp_total_pages);
177 /* max pool index before the release */
178 p_idx_max2 = (page_pools.epp_total_pages - 1) / PAGES_PER_POOL;
180 page_pools.epp_free_pages -= npages;
181 page_pools.epp_total_pages -= npages;
183 /* max pool index after the release */
184 p_idx_max1 = page_pools.epp_total_pages == 0 ? -1 :
185 ((page_pools.epp_total_pages - 1) / PAGES_PER_POOL);
187 p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
188 g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
189 LASSERT(page_pools.epp_pools[p_idx]);
192 LASSERT(page_pools.epp_pools[p_idx]);
193 LASSERT(page_pools.epp_pools[p_idx][g_idx]);
195 __free_page(page_pools.epp_pools[p_idx][g_idx]);
196 page_pools.epp_pools[p_idx][g_idx] = NULL;
198 if (++g_idx == PAGES_PER_POOL) {
204 /* free unused pools */
205 while (p_idx_max1 < p_idx_max2) {
206 LASSERT(page_pools.epp_pools[p_idx_max2]);
207 kfree(page_pools.epp_pools[p_idx_max2]);
208 page_pools.epp_pools[p_idx_max2] = NULL;
214 * we try to keep at least PTLRPC_MAX_BRW_PAGES pages in the pool.
216 static unsigned long enc_pools_shrink_count(struct shrinker *s,
217 struct shrink_control *sc)
220 * if no pool access for a long time, we consider it's fully idle.
221 * a little race here is fine.
223 if (unlikely(ktime_get_seconds() - page_pools.epp_last_access >
224 CACHE_QUIESCENT_PERIOD)) {
225 spin_lock(&page_pools.epp_lock);
226 page_pools.epp_idle_idx = IDLE_IDX_MAX;
227 spin_unlock(&page_pools.epp_lock);
230 LASSERT(page_pools.epp_idle_idx <= IDLE_IDX_MAX);
231 return max((int)page_pools.epp_free_pages - PTLRPC_MAX_BRW_PAGES, 0) *
232 (IDLE_IDX_MAX - page_pools.epp_idle_idx) / IDLE_IDX_MAX;
236 * we try to keep at least PTLRPC_MAX_BRW_PAGES pages in the pool.
238 static unsigned long enc_pools_shrink_scan(struct shrinker *s,
239 struct shrink_control *sc)
241 spin_lock(&page_pools.epp_lock);
242 sc->nr_to_scan = min_t(unsigned long, sc->nr_to_scan,
243 page_pools.epp_free_pages - PTLRPC_MAX_BRW_PAGES);
244 if (sc->nr_to_scan > 0) {
245 enc_pools_release_free_pages(sc->nr_to_scan);
246 CDEBUG(D_SEC, "released %ld pages, %ld left\n",
247 (long)sc->nr_to_scan, page_pools.epp_free_pages);
249 page_pools.epp_st_shrinks++;
250 page_pools.epp_last_shrink = ktime_get_seconds();
252 spin_unlock(&page_pools.epp_lock);
255 * if no pool access for a long time, we consider it's fully idle.
256 * a little race here is fine.
258 if (unlikely(ktime_get_seconds() - page_pools.epp_last_access >
259 CACHE_QUIESCENT_PERIOD)) {
260 spin_lock(&page_pools.epp_lock);
261 page_pools.epp_idle_idx = IDLE_IDX_MAX;
262 spin_unlock(&page_pools.epp_lock);
265 LASSERT(page_pools.epp_idle_idx <= IDLE_IDX_MAX);
266 return sc->nr_to_scan;
270 int npages_to_npools(unsigned long npages)
272 return (int)((npages + PAGES_PER_POOL - 1) / PAGES_PER_POOL);
276 * return how many pages cleaned up.
278 static unsigned long enc_pools_cleanup(struct page ***pools, int npools)
280 unsigned long cleaned = 0;
283 for (i = 0; i < npools; i++) {
285 for (j = 0; j < PAGES_PER_POOL; j++) {
287 __free_page(pools[i][j]);
299 static inline void enc_pools_wakeup(void)
301 assert_spin_locked(&page_pools.epp_lock);
303 if (unlikely(page_pools.epp_waitqlen)) {
304 LASSERT(waitqueue_active(&page_pools.epp_waitq));
305 wake_up_all(&page_pools.epp_waitq);
309 void sptlrpc_enc_pool_put_pages(struct ptlrpc_bulk_desc *desc)
314 if (!desc->bd_enc_iov)
317 LASSERT(desc->bd_iov_count > 0);
319 spin_lock(&page_pools.epp_lock);
321 p_idx = page_pools.epp_free_pages / PAGES_PER_POOL;
322 g_idx = page_pools.epp_free_pages % PAGES_PER_POOL;
324 LASSERT(page_pools.epp_free_pages + desc->bd_iov_count <=
325 page_pools.epp_total_pages);
326 LASSERT(page_pools.epp_pools[p_idx]);
328 for (i = 0; i < desc->bd_iov_count; i++) {
329 LASSERT(desc->bd_enc_iov[i].bv_page);
330 LASSERT(g_idx != 0 || page_pools.epp_pools[p_idx]);
331 LASSERT(!page_pools.epp_pools[p_idx][g_idx]);
333 page_pools.epp_pools[p_idx][g_idx] =
334 desc->bd_enc_iov[i].bv_page;
336 if (++g_idx == PAGES_PER_POOL) {
342 page_pools.epp_free_pages += desc->bd_iov_count;
346 spin_unlock(&page_pools.epp_lock);
348 kfree(desc->bd_enc_iov);
349 desc->bd_enc_iov = NULL;
352 static inline void enc_pools_alloc(void)
354 LASSERT(page_pools.epp_max_pools);
355 page_pools.epp_pools =
356 libcfs_kvzalloc(page_pools.epp_max_pools *
357 sizeof(*page_pools.epp_pools),
361 static inline void enc_pools_free(void)
363 LASSERT(page_pools.epp_max_pools);
364 LASSERT(page_pools.epp_pools);
366 kvfree(page_pools.epp_pools);
369 static struct shrinker pools_shrinker = {
370 .count_objects = enc_pools_shrink_count,
371 .scan_objects = enc_pools_shrink_scan,
372 .seeks = DEFAULT_SEEKS,
375 int sptlrpc_enc_pool_init(void)
378 * maximum capacity is 1/8 of total physical memory.
379 * is the 1/8 a good number?
381 page_pools.epp_max_pages = totalram_pages / 8;
382 page_pools.epp_max_pools = npages_to_npools(page_pools.epp_max_pages);
384 init_waitqueue_head(&page_pools.epp_waitq);
385 page_pools.epp_waitqlen = 0;
386 page_pools.epp_pages_short = 0;
388 page_pools.epp_growing = 0;
390 page_pools.epp_idle_idx = 0;
391 page_pools.epp_last_shrink = ktime_get_seconds();
392 page_pools.epp_last_access = ktime_get_seconds();
394 spin_lock_init(&page_pools.epp_lock);
395 page_pools.epp_total_pages = 0;
396 page_pools.epp_free_pages = 0;
398 page_pools.epp_st_max_pages = 0;
399 page_pools.epp_st_grows = 0;
400 page_pools.epp_st_grow_fails = 0;
401 page_pools.epp_st_shrinks = 0;
402 page_pools.epp_st_access = 0;
403 page_pools.epp_st_missings = 0;
404 page_pools.epp_st_lowfree = 0;
405 page_pools.epp_st_max_wqlen = 0;
406 page_pools.epp_st_max_wait = 0;
409 if (!page_pools.epp_pools)
412 register_shrinker(&pools_shrinker);
417 void sptlrpc_enc_pool_fini(void)
419 unsigned long cleaned, npools;
421 LASSERT(page_pools.epp_pools);
422 LASSERT(page_pools.epp_total_pages == page_pools.epp_free_pages);
424 unregister_shrinker(&pools_shrinker);
426 npools = npages_to_npools(page_pools.epp_total_pages);
427 cleaned = enc_pools_cleanup(page_pools.epp_pools, npools);
428 LASSERT(cleaned == page_pools.epp_total_pages);
432 if (page_pools.epp_st_access > 0) {
434 "max pages %lu, grows %u, grow fails %u, shrinks %u, access %lu, missing %lu, max qlen %u, max wait %ld/%ld\n",
435 page_pools.epp_st_max_pages, page_pools.epp_st_grows,
436 page_pools.epp_st_grow_fails,
437 page_pools.epp_st_shrinks, page_pools.epp_st_access,
438 page_pools.epp_st_missings, page_pools.epp_st_max_wqlen,
439 page_pools.epp_st_max_wait,
440 msecs_to_jiffies(MSEC_PER_SEC));
444 static int cfs_hash_alg_id[] = {
445 [BULK_HASH_ALG_NULL] = CFS_HASH_ALG_NULL,
446 [BULK_HASH_ALG_ADLER32] = CFS_HASH_ALG_ADLER32,
447 [BULK_HASH_ALG_CRC32] = CFS_HASH_ALG_CRC32,
448 [BULK_HASH_ALG_MD5] = CFS_HASH_ALG_MD5,
449 [BULK_HASH_ALG_SHA1] = CFS_HASH_ALG_SHA1,
450 [BULK_HASH_ALG_SHA256] = CFS_HASH_ALG_SHA256,
451 [BULK_HASH_ALG_SHA384] = CFS_HASH_ALG_SHA384,
452 [BULK_HASH_ALG_SHA512] = CFS_HASH_ALG_SHA512,
455 const char *sptlrpc_get_hash_name(__u8 hash_alg)
457 return cfs_crypto_hash_name(cfs_hash_alg_id[hash_alg]);
460 __u8 sptlrpc_get_hash_alg(const char *algname)
462 return cfs_crypto_hash_alg(algname);
465 int bulk_sec_desc_unpack(struct lustre_msg *msg, int offset, int swabbed)
467 struct ptlrpc_bulk_sec_desc *bsd;
468 int size = msg->lm_buflens[offset];
470 bsd = lustre_msg_buf(msg, offset, sizeof(*bsd));
472 CERROR("Invalid bulk sec desc: size %d\n", size);
477 __swab32s(&bsd->bsd_nob);
479 if (unlikely(bsd->bsd_version != 0)) {
480 CERROR("Unexpected version %u\n", bsd->bsd_version);
484 if (unlikely(bsd->bsd_type >= SPTLRPC_BULK_MAX)) {
485 CERROR("Invalid type %u\n", bsd->bsd_type);
489 /* FIXME more sanity check here */
491 if (unlikely(bsd->bsd_svc != SPTLRPC_BULK_SVC_NULL &&
492 bsd->bsd_svc != SPTLRPC_BULK_SVC_INTG &&
493 bsd->bsd_svc != SPTLRPC_BULK_SVC_PRIV)) {
494 CERROR("Invalid svc %u\n", bsd->bsd_svc);
500 EXPORT_SYMBOL(bulk_sec_desc_unpack);
502 int sptlrpc_get_bulk_checksum(struct ptlrpc_bulk_desc *desc, __u8 alg,
503 void *buf, int buflen)
505 struct cfs_crypto_hash_desc *hdesc;
507 unsigned int bufsize;
510 LASSERT(alg > BULK_HASH_ALG_NULL && alg < BULK_HASH_ALG_MAX);
511 LASSERT(buflen >= 4);
513 hdesc = cfs_crypto_hash_init(cfs_hash_alg_id[alg], NULL, 0);
515 CERROR("Unable to initialize checksum hash %s\n",
516 cfs_crypto_hash_name(cfs_hash_alg_id[alg]));
517 return PTR_ERR(hdesc);
520 hashsize = cfs_crypto_hash_digestsize(cfs_hash_alg_id[alg]);
522 for (i = 0; i < desc->bd_iov_count; i++) {
523 cfs_crypto_hash_update_page(hdesc, desc->bd_iov[i].bv_page,
524 desc->bd_iov[i].bv_offset &
526 desc->bd_iov[i].bv_len);
529 if (hashsize > buflen) {
530 unsigned char hashbuf[CFS_CRYPTO_HASH_DIGESTSIZE_MAX];
532 bufsize = sizeof(hashbuf);
533 LASSERTF(bufsize >= hashsize, "bufsize = %u < hashsize %u\n",
535 err = cfs_crypto_hash_final(hdesc, hashbuf, &bufsize);
536 memcpy(buf, hashbuf, buflen);
539 err = cfs_crypto_hash_final(hdesc, buf, &bufsize);