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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
31 * This file is part of Lustre, http://www.lustre.org/
32 * Lustre is a trademark of Sun Microsystems, Inc.
34 * lustre/ptlrpc/gss/gss_keyring.c
36 * Author: Eric Mei <ericm@clusterfs.com>
40 # define EXPORT_SYMTAB
42 #define DEBUG_SUBSYSTEM S_SEC
44 #include <linux/init.h>
45 #include <linux/module.h>
46 #include <linux/slab.h>
47 #include <linux/dcache.h>
49 #include <linux/crypto.h>
50 #include <linux/key.h>
51 #include <linux/keyctl.h>
52 #include <linux/mutex.h>
53 #include <asm/atomic.h>
55 #include <liblustre.h>
59 #include <obd_class.h>
60 #include <obd_support.h>
61 #include <lustre/lustre_idl.h>
62 #include <lustre_sec.h>
63 #include <lustre_net.h>
64 #include <lustre_import.h>
67 #include "gss_internal.h"
70 static struct ptlrpc_sec_policy gss_policy_keyring;
71 static struct ptlrpc_ctx_ops gss_keyring_ctxops;
72 static struct key_type gss_key_type;
74 static int sec_install_rctx_kr(struct ptlrpc_sec *sec,
75 struct ptlrpc_svc_ctx *svc_ctx);
78 * the timeout is only for the case that upcall child process die abnormally.
79 * in any other cases it should finally update kernel key.
81 * FIXME we'd better to incorporate the client & server side upcall timeouts
82 * into the framework of Adaptive Timeouts, but we need to figure out how to
83 * make sure that kernel knows the upcall processes is in-progress or died
86 #define KEYRING_UPCALL_TIMEOUT (obd_timeout + obd_timeout)
88 /****************************************
90 ****************************************/
92 #define DUMP_PROCESS_KEYRINGS(tsk) \
94 CWARN("DUMP PK: %s[%u,%u/%u](<-%s[%u,%u/%u]): " \
95 "a %d, t %d, p %d, s %d, u %d, us %d, df %d\n", \
96 tsk->comm, tsk->pid, tsk->uid, tsk->fsuid, \
97 tsk->parent->comm, tsk->parent->pid, \
98 tsk->parent->uid, tsk->parent->fsuid, \
99 tsk->request_key_auth ? \
100 tsk->request_key_auth->serial : 0, \
101 tsk->thread_keyring ? \
102 tsk->thread_keyring->serial : 0, \
103 tsk->signal->process_keyring ? \
104 tsk->signal->process_keyring->serial : 0, \
105 tsk->signal->session_keyring ? \
106 tsk->signal->session_keyring->serial : 0, \
107 tsk->user->uid_keyring ? \
108 tsk->user->uid_keyring->serial : 0, \
109 tsk->user->session_keyring ? \
110 tsk->user->session_keyring->serial : 0, \
115 #define DUMP_KEY(key) \
117 CWARN("DUMP KEY: %p(%d) ref %d u%u/g%u desc %s\n", \
118 key, key->serial, atomic_read(&key->usage), \
119 key->uid, key->gid, \
120 key->description ? key->description : "n/a" \
125 static inline void keyring_upcall_lock(struct gss_sec_keyring *gsec_kr)
127 #ifdef HAVE_KEYRING_UPCALL_SERIALIZED
128 cfs_mutex_lock(&gsec_kr->gsk_uc_lock);
132 static inline void keyring_upcall_unlock(struct gss_sec_keyring *gsec_kr)
134 #ifdef HAVE_KEYRING_UPCALL_SERIALIZED
135 cfs_mutex_unlock(&gsec_kr->gsk_uc_lock);
139 static inline void key_revoke_locked(struct key *key)
141 set_bit(KEY_FLAG_REVOKED, &key->flags);
144 static void ctx_upcall_timeout_kr(unsigned long data)
146 struct ptlrpc_cli_ctx *ctx = (struct ptlrpc_cli_ctx *) data;
147 struct key *key = ctx2gctx_keyring(ctx)->gck_key;
149 CWARN("ctx %p, key %p\n", ctx, key);
154 key_revoke_locked(key);
158 void ctx_start_timer_kr(struct ptlrpc_cli_ctx *ctx, long timeout)
160 struct gss_cli_ctx_keyring *gctx_kr = ctx2gctx_keyring(ctx);
161 struct timer_list *timer = gctx_kr->gck_timer;
165 CDEBUG(D_SEC, "ctx %p: start timer %lds\n", ctx, timeout);
166 timeout = timeout * CFS_HZ + cfs_time_current();
169 timer->expires = timeout;
170 timer->data = (unsigned long ) ctx;
171 timer->function = ctx_upcall_timeout_kr;
177 * caller should make sure no race with other threads
180 void ctx_clear_timer_kr(struct ptlrpc_cli_ctx *ctx)
182 struct gss_cli_ctx_keyring *gctx_kr = ctx2gctx_keyring(ctx);
183 struct timer_list *timer = gctx_kr->gck_timer;
188 CDEBUG(D_SEC, "ctx %p, key %p\n", ctx, gctx_kr->gck_key);
190 gctx_kr->gck_timer = NULL;
192 del_singleshot_timer_sync(timer);
198 struct ptlrpc_cli_ctx *ctx_create_kr(struct ptlrpc_sec *sec,
199 struct vfs_cred *vcred)
201 struct ptlrpc_cli_ctx *ctx;
202 struct gss_cli_ctx_keyring *gctx_kr;
204 OBD_ALLOC_PTR(gctx_kr);
208 OBD_ALLOC_PTR(gctx_kr->gck_timer);
209 if (gctx_kr->gck_timer == NULL) {
210 OBD_FREE_PTR(gctx_kr);
213 init_timer(gctx_kr->gck_timer);
215 ctx = &gctx_kr->gck_base.gc_base;
217 if (gss_cli_ctx_init_common(sec, ctx, &gss_keyring_ctxops, vcred)) {
218 OBD_FREE_PTR(gctx_kr->gck_timer);
219 OBD_FREE_PTR(gctx_kr);
223 ctx->cc_expire = cfs_time_current_sec() + KEYRING_UPCALL_TIMEOUT;
224 cfs_clear_bit(PTLRPC_CTX_NEW_BIT, &ctx->cc_flags);
225 cfs_atomic_inc(&ctx->cc_refcount); /* for the caller */
230 static void ctx_destroy_kr(struct ptlrpc_cli_ctx *ctx)
232 struct ptlrpc_sec *sec = ctx->cc_sec;
233 struct gss_cli_ctx_keyring *gctx_kr = ctx2gctx_keyring(ctx);
235 CDEBUG(D_SEC, "destroying ctx %p\n", ctx);
237 /* at this time the association with key has been broken. */
239 LASSERT(cfs_atomic_read(&sec->ps_refcount) > 0);
240 LASSERT(cfs_atomic_read(&sec->ps_nctx) > 0);
241 LASSERT(cfs_test_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags) == 0);
242 LASSERT(gctx_kr->gck_key == NULL);
244 ctx_clear_timer_kr(ctx);
245 LASSERT(gctx_kr->gck_timer == NULL);
247 if (gss_cli_ctx_fini_common(sec, ctx))
250 OBD_FREE_PTR(gctx_kr);
252 cfs_atomic_dec(&sec->ps_nctx);
253 sptlrpc_sec_put(sec);
256 static void ctx_release_kr(struct ptlrpc_cli_ctx *ctx, int sync)
261 cfs_atomic_inc(&ctx->cc_refcount);
262 sptlrpc_gc_add_ctx(ctx);
266 static void ctx_put_kr(struct ptlrpc_cli_ctx *ctx, int sync)
268 LASSERT(cfs_atomic_read(&ctx->cc_refcount) > 0);
270 if (cfs_atomic_dec_and_test(&ctx->cc_refcount))
271 ctx_release_kr(ctx, sync);
275 * key <-> ctx association and rules:
276 * - ctx might not bind with any key
277 * - key/ctx binding is protected by key semaphore (if the key present)
278 * - key and ctx each take a reference of the other
279 * - ctx enlist/unlist is protected by ctx spinlock
280 * - never enlist a ctx after it's been unlisted
281 * - whoever do enlist should also do bind, lock key before enlist:
282 * - lock key -> lock ctx -> enlist -> unlock ctx -> bind -> unlock key
283 * - whoever do unlist should also do unbind:
284 * - lock key -> lock ctx -> unlist -> unlock ctx -> unbind -> unlock key
285 * - lock ctx -> unlist -> unlock ctx -> lock key -> unbind -> unlock key
288 static inline void spin_lock_if(cfs_spinlock_t *lock, int condition)
294 static inline void spin_unlock_if(cfs_spinlock_t *lock, int condition)
297 cfs_spin_unlock(lock);
300 static void ctx_enlist_kr(struct ptlrpc_cli_ctx *ctx, int is_root, int locked)
302 struct ptlrpc_sec *sec = ctx->cc_sec;
303 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
305 LASSERT(!cfs_test_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags));
306 LASSERT(cfs_atomic_read(&ctx->cc_refcount) > 0);
308 spin_lock_if(&sec->ps_lock, !locked);
310 cfs_atomic_inc(&ctx->cc_refcount);
311 cfs_set_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags);
312 cfs_hlist_add_head(&ctx->cc_cache, &gsec_kr->gsk_clist);
314 gsec_kr->gsk_root_ctx = ctx;
316 spin_unlock_if(&sec->ps_lock, !locked);
320 * Note after this get called, caller should not access ctx again because
321 * it might have been freed, unless caller hold at least one refcount of
324 * return non-zero if we indeed unlist this ctx.
326 static int ctx_unlist_kr(struct ptlrpc_cli_ctx *ctx, int locked)
328 struct ptlrpc_sec *sec = ctx->cc_sec;
329 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
331 /* if hashed bit has gone, leave the job to somebody who is doing it */
332 if (cfs_test_and_clear_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags) == 0)
335 /* drop ref inside spin lock to prevent race with other operations */
336 spin_lock_if(&sec->ps_lock, !locked);
338 if (gsec_kr->gsk_root_ctx == ctx)
339 gsec_kr->gsk_root_ctx = NULL;
340 cfs_hlist_del_init(&ctx->cc_cache);
341 cfs_atomic_dec(&ctx->cc_refcount);
343 spin_unlock_if(&sec->ps_lock, !locked);
349 * bind a key with a ctx together.
350 * caller must hold write lock of the key, as well as ref on key & ctx.
352 static void bind_key_ctx(struct key *key, struct ptlrpc_cli_ctx *ctx)
354 LASSERT(cfs_atomic_read(&ctx->cc_refcount) > 0);
355 LASSERT(atomic_read(&key->usage) > 0);
356 LASSERT(ctx2gctx_keyring(ctx)->gck_key == NULL);
357 LASSERT(key->payload.data == NULL);
359 /* at this time context may or may not in list. */
361 cfs_atomic_inc(&ctx->cc_refcount);
362 ctx2gctx_keyring(ctx)->gck_key = key;
363 key->payload.data = ctx;
367 * unbind a key and a ctx.
368 * caller must hold write lock, as well as a ref of the key.
370 static void unbind_key_ctx(struct key *key, struct ptlrpc_cli_ctx *ctx)
372 LASSERT(key->payload.data == ctx);
373 LASSERT(cfs_test_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags) == 0);
375 /* must revoke the key, or others may treat it as newly created */
376 key_revoke_locked(key);
378 key->payload.data = NULL;
379 ctx2gctx_keyring(ctx)->gck_key = NULL;
381 /* once ctx get split from key, the timer is meaningless */
382 ctx_clear_timer_kr(ctx);
389 * given a ctx, unbind with its coupled key, if any.
390 * unbind could only be called once, so we don't worry the key be released
393 static void unbind_ctx_kr(struct ptlrpc_cli_ctx *ctx)
395 struct key *key = ctx2gctx_keyring(ctx)->gck_key;
398 LASSERT(key->payload.data == ctx);
401 down_write(&key->sem);
402 unbind_key_ctx(key, ctx);
409 * given a key, unbind with its coupled ctx, if any.
410 * caller must hold write lock, as well as a ref of the key.
412 static void unbind_key_locked(struct key *key)
414 struct ptlrpc_cli_ctx *ctx = key->payload.data;
417 unbind_key_ctx(key, ctx);
421 * unlist a ctx, and unbind from coupled key
423 static void kill_ctx_kr(struct ptlrpc_cli_ctx *ctx)
425 if (ctx_unlist_kr(ctx, 0))
430 * given a key, unlist and unbind with the coupled ctx (if any).
431 * caller must hold write lock, as well as a ref of the key.
433 static void kill_key_locked(struct key *key)
435 struct ptlrpc_cli_ctx *ctx = key->payload.data;
437 if (ctx && ctx_unlist_kr(ctx, 0))
438 unbind_key_locked(key);
442 * caller should hold one ref on contexts in freelist.
444 static void dispose_ctx_list_kr(cfs_hlist_head_t *freelist)
446 cfs_hlist_node_t *pos, *next;
447 struct ptlrpc_cli_ctx *ctx;
448 struct gss_cli_ctx *gctx;
450 cfs_hlist_for_each_entry_safe(ctx, pos, next, freelist, cc_cache) {
451 cfs_hlist_del_init(&ctx->cc_cache);
453 /* reverse ctx: update current seq to buddy svcctx if exist.
454 * ideally this should be done at gss_cli_ctx_finalize(), but
455 * the ctx destroy could be delayed by:
456 * 1) ctx still has reference;
457 * 2) ctx destroy is asynchronous;
458 * and reverse import call inval_all_ctx() require this be done
459 *_immediately_ otherwise newly created reverse ctx might copy
460 * the very old sequence number from svcctx. */
461 gctx = ctx2gctx(ctx);
462 if (!rawobj_empty(&gctx->gc_svc_handle) &&
463 sec_is_reverse(gctx->gc_base.cc_sec)) {
464 gss_svc_upcall_update_sequence(&gctx->gc_svc_handle,
465 (__u32) cfs_atomic_read(&gctx->gc_seq));
468 /* we need to wakeup waiting reqs here. the context might
469 * be forced released before upcall finished, then the
470 * late-arrived downcall can't find the ctx even. */
471 sptlrpc_cli_ctx_wakeup(ctx);
479 * lookup a root context directly in a sec, return root ctx with a
480 * reference taken or NULL.
483 struct ptlrpc_cli_ctx * sec_lookup_root_ctx_kr(struct ptlrpc_sec *sec)
485 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
486 struct ptlrpc_cli_ctx *ctx = NULL;
488 cfs_spin_lock(&sec->ps_lock);
490 ctx = gsec_kr->gsk_root_ctx;
492 if (ctx == NULL && unlikely(sec_is_reverse(sec))) {
493 cfs_hlist_node_t *node;
494 struct ptlrpc_cli_ctx *tmp;
496 /* reverse ctx, search root ctx in list, choose the one
497 * with shortest expire time, which is most possibly have
498 * an established peer ctx at client side. */
499 cfs_hlist_for_each_entry(tmp, node, &gsec_kr->gsk_clist,
501 if (ctx == NULL || ctx->cc_expire == 0 ||
502 ctx->cc_expire > tmp->cc_expire) {
504 /* promote to be root_ctx */
505 gsec_kr->gsk_root_ctx = ctx;
511 LASSERT(cfs_atomic_read(&ctx->cc_refcount) > 0);
512 LASSERT(!cfs_hlist_empty(&gsec_kr->gsk_clist));
513 cfs_atomic_inc(&ctx->cc_refcount);
516 cfs_spin_unlock(&sec->ps_lock);
521 #define RVS_CTX_EXPIRE_NICE (10)
524 void rvs_sec_install_root_ctx_kr(struct ptlrpc_sec *sec,
525 struct ptlrpc_cli_ctx *new_ctx,
528 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
529 cfs_hlist_node_t *hnode;
530 struct ptlrpc_cli_ctx *ctx;
534 LASSERT(sec_is_reverse(sec));
536 cfs_spin_lock(&sec->ps_lock);
538 now = cfs_time_current_sec();
540 /* set all existing ctxs short expiry */
541 cfs_hlist_for_each_entry(ctx, hnode, &gsec_kr->gsk_clist, cc_cache) {
542 if (ctx->cc_expire > now + RVS_CTX_EXPIRE_NICE) {
543 ctx->cc_early_expire = 1;
544 ctx->cc_expire = now + RVS_CTX_EXPIRE_NICE;
548 /* if there's root_ctx there, instead obsolete the current
549 * immediately, we leave it continue operating for a little while.
550 * hopefully when the first backward rpc with newest ctx send out,
551 * the client side already have the peer ctx well established. */
552 ctx_enlist_kr(new_ctx, gsec_kr->gsk_root_ctx ? 0 : 1, 1);
555 bind_key_ctx(key, new_ctx);
557 cfs_spin_unlock(&sec->ps_lock);
560 static void construct_key_desc(void *buf, int bufsize,
561 struct ptlrpc_sec *sec, uid_t uid)
563 snprintf(buf, bufsize, "%d@%x", uid, sec->ps_id);
564 ((char *)buf)[bufsize - 1] = '\0';
567 /****************************************
569 ****************************************/
572 struct ptlrpc_sec * gss_sec_create_kr(struct obd_import *imp,
573 struct ptlrpc_svc_ctx *svcctx,
574 struct sptlrpc_flavor *sf)
576 struct gss_sec_keyring *gsec_kr;
579 OBD_ALLOC(gsec_kr, sizeof(*gsec_kr));
583 CFS_INIT_HLIST_HEAD(&gsec_kr->gsk_clist);
584 gsec_kr->gsk_root_ctx = NULL;
585 cfs_mutex_init(&gsec_kr->gsk_root_uc_lock);
586 #ifdef HAVE_KEYRING_UPCALL_SERIALIZED
587 cfs_mutex_init(&gsec_kr->gsk_uc_lock);
590 if (gss_sec_create_common(&gsec_kr->gsk_base, &gss_policy_keyring,
594 if (svcctx != NULL &&
595 sec_install_rctx_kr(&gsec_kr->gsk_base.gs_base, svcctx)) {
596 gss_sec_destroy_common(&gsec_kr->gsk_base);
600 RETURN(&gsec_kr->gsk_base.gs_base);
603 OBD_FREE(gsec_kr, sizeof(*gsec_kr));
608 void gss_sec_destroy_kr(struct ptlrpc_sec *sec)
610 struct gss_sec *gsec = sec2gsec(sec);
611 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
613 CDEBUG(D_SEC, "destroy %s@%p\n", sec->ps_policy->sp_name, sec);
615 LASSERT(cfs_hlist_empty(&gsec_kr->gsk_clist));
616 LASSERT(gsec_kr->gsk_root_ctx == NULL);
618 gss_sec_destroy_common(gsec);
620 OBD_FREE(gsec_kr, sizeof(*gsec_kr));
623 static inline int user_is_root(struct ptlrpc_sec *sec, struct vfs_cred *vcred)
625 /* except the ROOTONLY flag, treat it as root user only if real uid
626 * is 0, euid/fsuid being 0 are handled as setuid scenarios */
627 if (sec_is_rootonly(sec) || (vcred->vc_uid == 0))
634 * unlink request key from it's ring, which is linked during request_key().
635 * sadly, we have to 'guess' which keyring it's linked to.
637 * FIXME this code is fragile, depend on how request_key_link() is implemented.
639 static void request_key_unlink(struct key *key)
641 struct task_struct *tsk = current;
644 switch (tsk->jit_keyring) {
645 case KEY_REQKEY_DEFL_DEFAULT:
646 case KEY_REQKEY_DEFL_THREAD_KEYRING:
647 ring = key_get(tsk->thread_keyring);
650 case KEY_REQKEY_DEFL_PROCESS_KEYRING:
651 ring = key_get(tsk->signal->process_keyring);
654 case KEY_REQKEY_DEFL_SESSION_KEYRING:
656 ring = key_get(rcu_dereference(tsk->signal->session_keyring));
660 case KEY_REQKEY_DEFL_USER_SESSION_KEYRING:
661 ring = key_get(tsk->user->session_keyring);
663 case KEY_REQKEY_DEFL_USER_KEYRING:
664 ring = key_get(tsk->user->uid_keyring);
666 case KEY_REQKEY_DEFL_GROUP_KEYRING:
672 key_unlink(ring, key);
677 struct ptlrpc_cli_ctx * gss_sec_lookup_ctx_kr(struct ptlrpc_sec *sec,
678 struct vfs_cred *vcred,
679 int create, int remove_dead)
681 struct obd_import *imp = sec->ps_import;
682 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
683 struct ptlrpc_cli_ctx *ctx = NULL;
684 unsigned int is_root = 0, create_new = 0;
692 LASSERT(imp != NULL);
694 is_root = user_is_root(sec, vcred);
696 /* a little bit optimization for root context */
698 ctx = sec_lookup_root_ctx_kr(sec);
700 * Only lookup directly for REVERSE sec, which should
703 if (ctx || sec_is_reverse(sec))
707 LASSERT(create != 0);
709 /* for root context, obtain lock and check again, this time hold
710 * the root upcall lock, make sure nobody else populated new root
711 * context after last check. */
713 cfs_mutex_lock(&gsec_kr->gsk_root_uc_lock);
715 ctx = sec_lookup_root_ctx_kr(sec);
719 /* update reverse handle for root user */
720 sec2gsec(sec)->gs_rvs_hdl = gss_get_next_ctx_index();
722 switch (sec->ps_part) {
741 /* in case of setuid, key will be constructed as owner of fsuid/fsgid,
742 * but we do authentication based on real uid/gid. the key permission
743 * bits will be exactly as POS_ALL, so only processes who subscribed
744 * this key could have the access, although the quota might be counted
745 * on others (fsuid/fsgid).
747 * keyring will use fsuid/fsgid as upcall parameters, so we have to
748 * encode real uid/gid into callout info.
751 construct_key_desc(desc, sizeof(desc), sec, vcred->vc_uid);
753 /* callout info format:
754 * secid:mech:uid:gid:flags:svc_type:peer_nid:target_uuid
756 coinfo_size = sizeof(struct obd_uuid) + MAX_OBD_NAME + 64;
757 OBD_ALLOC(coinfo, coinfo_size);
761 snprintf(coinfo, coinfo_size, "%d:%s:%u:%u:%s:%d:"LPX64":%s",
762 sec->ps_id, sec2gsec(sec)->gs_mech->gm_name,
763 vcred->vc_uid, vcred->vc_gid,
764 co_flags, import_to_gss_svc(imp),
765 imp->imp_connection->c_peer.nid, imp->imp_obd->obd_name);
767 CDEBUG(D_SEC, "requesting key for %s\n", desc);
769 keyring_upcall_lock(gsec_kr);
770 key = request_key(&gss_key_type, desc, coinfo);
771 keyring_upcall_unlock(gsec_kr);
773 OBD_FREE(coinfo, coinfo_size);
776 CERROR("failed request key: %ld\n", PTR_ERR(key));
779 CDEBUG(D_SEC, "obtained key %08x for %s\n", key->serial, desc);
781 /* once payload.data was pointed to a ctx, it never changes until
782 * we de-associate them; but parallel request_key() may return
783 * a key with payload.data == NULL at the same time. so we still
784 * need wirtelock of key->sem to serialize them. */
785 down_write(&key->sem);
787 if (likely(key->payload.data != NULL)) {
788 ctx = key->payload.data;
790 LASSERT(cfs_atomic_read(&ctx->cc_refcount) >= 1);
791 LASSERT(ctx2gctx_keyring(ctx)->gck_key == key);
792 LASSERT(atomic_read(&key->usage) >= 2);
794 /* simply take a ref and return. it's upper layer's
795 * responsibility to detect & replace dead ctx. */
796 cfs_atomic_inc(&ctx->cc_refcount);
798 /* pre initialization with a cli_ctx. this can't be done in
799 * key_instantiate() because we'v no enough information
801 ctx = ctx_create_kr(sec, vcred);
803 ctx_enlist_kr(ctx, is_root, 0);
804 bind_key_ctx(key, ctx);
806 ctx_start_timer_kr(ctx, KEYRING_UPCALL_TIMEOUT);
808 CDEBUG(D_SEC, "installed key %p <-> ctx %p (sec %p)\n",
811 /* we'd prefer to call key_revoke(), but we more like
812 * to revoke it within this key->sem locked period. */
813 key_revoke_locked(key);
821 if (is_root && create_new)
822 request_key_unlink(key);
827 cfs_mutex_unlock(&gsec_kr->gsk_root_uc_lock);
832 void gss_sec_release_ctx_kr(struct ptlrpc_sec *sec,
833 struct ptlrpc_cli_ctx *ctx,
836 LASSERT(cfs_atomic_read(&sec->ps_refcount) > 0);
837 LASSERT(cfs_atomic_read(&ctx->cc_refcount) == 0);
838 ctx_release_kr(ctx, sync);
842 * flush context of normal user, we must resort to keyring itself to find out
843 * contexts which belong to me.
845 * Note here we suppose only to flush _my_ context, the "uid" will
846 * be ignored in the search.
849 void flush_user_ctx_cache_kr(struct ptlrpc_sec *sec,
851 int grace, int force)
856 /* nothing to do for reverse or rootonly sec */
857 if (sec_is_reverse(sec) || sec_is_rootonly(sec))
860 construct_key_desc(desc, sizeof(desc), sec, uid);
862 /* there should be only one valid key, but we put it in the
863 * loop in case of any weird cases */
865 key = request_key(&gss_key_type, desc, NULL);
867 CDEBUG(D_SEC, "No more key found for current user\n");
871 down_write(&key->sem);
873 kill_key_locked(key);
875 /* kill_key_locked() should usually revoke the key, but we
876 * revoke it again to make sure, e.g. some case the key may
877 * not well coupled with a context. */
878 key_revoke_locked(key);
887 * flush context of root or all, we iterate through the list.
890 void flush_spec_ctx_cache_kr(struct ptlrpc_sec *sec,
892 int grace, int force)
894 struct gss_sec_keyring *gsec_kr;
895 cfs_hlist_head_t freelist = CFS_HLIST_HEAD_INIT;
896 cfs_hlist_node_t *pos, *next;
897 struct ptlrpc_cli_ctx *ctx;
900 gsec_kr = sec2gsec_keyring(sec);
902 cfs_spin_lock(&sec->ps_lock);
903 cfs_hlist_for_each_entry_safe(ctx, pos, next,
904 &gsec_kr->gsk_clist, cc_cache) {
905 LASSERT(cfs_atomic_read(&ctx->cc_refcount) > 0);
907 if (uid != -1 && uid != ctx->cc_vcred.vc_uid)
910 /* at this moment there's at least 2 base reference:
911 * key association and in-list. */
912 if (cfs_atomic_read(&ctx->cc_refcount) > 2) {
915 CWARN("flush busy ctx %p(%u->%s, extra ref %d)\n",
916 ctx, ctx->cc_vcred.vc_uid,
917 sec2target_str(ctx->cc_sec),
918 cfs_atomic_read(&ctx->cc_refcount) - 2);
921 cfs_set_bit(PTLRPC_CTX_DEAD_BIT, &ctx->cc_flags);
923 cfs_clear_bit(PTLRPC_CTX_UPTODATE_BIT, &ctx->cc_flags);
925 cfs_atomic_inc(&ctx->cc_refcount);
927 if (ctx_unlist_kr(ctx, 1)) {
928 cfs_hlist_add_head(&ctx->cc_cache, &freelist);
930 LASSERT(cfs_atomic_read(&ctx->cc_refcount) >= 2);
931 cfs_atomic_dec(&ctx->cc_refcount);
934 cfs_spin_unlock(&sec->ps_lock);
936 dispose_ctx_list_kr(&freelist);
941 int gss_sec_flush_ctx_cache_kr(struct ptlrpc_sec *sec,
942 uid_t uid, int grace, int force)
946 CDEBUG(D_SEC, "sec %p(%d, nctx %d), uid %d, grace %d, force %d\n",
947 sec, cfs_atomic_read(&sec->ps_refcount),
948 cfs_atomic_read(&sec->ps_nctx),
951 if (uid != -1 && uid != 0)
952 flush_user_ctx_cache_kr(sec, uid, grace, force);
954 flush_spec_ctx_cache_kr(sec, uid, grace, force);
960 void gss_sec_gc_ctx_kr(struct ptlrpc_sec *sec)
962 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
963 cfs_hlist_head_t freelist = CFS_HLIST_HEAD_INIT;
964 cfs_hlist_node_t *pos, *next;
965 struct ptlrpc_cli_ctx *ctx;
968 CWARN("running gc\n");
970 cfs_spin_lock(&sec->ps_lock);
971 cfs_hlist_for_each_entry_safe(ctx, pos, next,
972 &gsec_kr->gsk_clist, cc_cache) {
973 LASSERT(cfs_atomic_read(&ctx->cc_refcount) > 0);
975 cfs_atomic_inc(&ctx->cc_refcount);
977 if (cli_ctx_check_death(ctx) && ctx_unlist_kr(ctx, 1)) {
978 cfs_hlist_add_head(&ctx->cc_cache, &freelist);
979 CWARN("unhashed ctx %p\n", ctx);
981 LASSERT(cfs_atomic_read(&ctx->cc_refcount) >= 2);
982 cfs_atomic_dec(&ctx->cc_refcount);
985 cfs_spin_unlock(&sec->ps_lock);
987 dispose_ctx_list_kr(&freelist);
993 int gss_sec_display_kr(struct ptlrpc_sec *sec, struct seq_file *seq)
995 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
996 cfs_hlist_node_t *pos, *next;
997 struct ptlrpc_cli_ctx *ctx;
998 struct gss_cli_ctx *gctx;
999 time_t now = cfs_time_current_sec();
1002 cfs_spin_lock(&sec->ps_lock);
1003 cfs_hlist_for_each_entry_safe(ctx, pos, next,
1004 &gsec_kr->gsk_clist, cc_cache) {
1009 gctx = ctx2gctx(ctx);
1010 key = ctx2gctx_keyring(ctx)->gck_key;
1012 gss_cli_ctx_flags2str(ctx->cc_flags,
1013 flags_str, sizeof(flags_str));
1015 if (gctx->gc_mechctx)
1016 lgss_display(gctx->gc_mechctx, mech, sizeof(mech));
1018 snprintf(mech, sizeof(mech), "N/A");
1019 mech[sizeof(mech) - 1] = '\0';
1021 seq_printf(seq, "%p: uid %u, ref %d, expire %ld(%+ld), fl %s, "
1022 "seq %d, win %u, key %08x(ref %d), "
1023 "hdl "LPX64":"LPX64", mech: %s\n",
1024 ctx, ctx->cc_vcred.vc_uid,
1025 cfs_atomic_read(&ctx->cc_refcount),
1027 ctx->cc_expire ? ctx->cc_expire - now : 0,
1029 cfs_atomic_read(&gctx->gc_seq),
1031 key ? key->serial : 0,
1032 key ? atomic_read(&key->usage) : 0,
1033 gss_handle_to_u64(&gctx->gc_handle),
1034 gss_handle_to_u64(&gctx->gc_svc_handle),
1037 cfs_spin_unlock(&sec->ps_lock);
1042 /****************************************
1044 ****************************************/
1047 int gss_cli_ctx_refresh_kr(struct ptlrpc_cli_ctx *ctx)
1049 /* upcall is already on the way */
1054 int gss_cli_ctx_validate_kr(struct ptlrpc_cli_ctx *ctx)
1056 LASSERT(cfs_atomic_read(&ctx->cc_refcount) > 0);
1057 LASSERT(ctx->cc_sec);
1059 if (cli_ctx_check_death(ctx)) {
1064 if (cli_ctx_is_ready(ctx))
1070 void gss_cli_ctx_die_kr(struct ptlrpc_cli_ctx *ctx, int grace)
1072 LASSERT(cfs_atomic_read(&ctx->cc_refcount) > 0);
1073 LASSERT(ctx->cc_sec);
1075 cli_ctx_expire(ctx);
1079 /****************************************
1080 * (reverse) service *
1081 ****************************************/
1084 * reverse context could have nothing to do with keyrings. here we still keep
1085 * the version which bind to a key, for future reference.
1087 #define HAVE_REVERSE_CTX_NOKEY
1089 #ifdef HAVE_REVERSE_CTX_NOKEY
1092 int sec_install_rctx_kr(struct ptlrpc_sec *sec,
1093 struct ptlrpc_svc_ctx *svc_ctx)
1095 struct ptlrpc_cli_ctx *cli_ctx;
1096 struct vfs_cred vcred = { 0, 0 };
1102 cli_ctx = ctx_create_kr(sec, &vcred);
1103 if (cli_ctx == NULL)
1106 rc = gss_copy_rvc_cli_ctx(cli_ctx, svc_ctx);
1108 CERROR("failed copy reverse cli ctx: %d\n", rc);
1110 ctx_put_kr(cli_ctx, 1);
1114 rvs_sec_install_root_ctx_kr(sec, cli_ctx, NULL);
1116 ctx_put_kr(cli_ctx, 1);
1121 #else /* ! HAVE_REVERSE_CTX_NOKEY */
1124 int sec_install_rctx_kr(struct ptlrpc_sec *sec,
1125 struct ptlrpc_svc_ctx *svc_ctx)
1127 struct ptlrpc_cli_ctx *cli_ctx = NULL;
1129 struct vfs_cred vcred = { 0, 0 };
1137 construct_key_desc(desc, sizeof(desc), sec, 0);
1139 key = key_alloc(&gss_key_type, desc, 0, 0,
1140 KEY_POS_ALL | KEY_USR_ALL, 1);
1142 CERROR("failed to alloc key: %ld\n", PTR_ERR(key));
1143 return PTR_ERR(key);
1146 rc = key_instantiate_and_link(key, NULL, 0, NULL, NULL);
1148 CERROR("failed to instantiate key: %d\n", rc);
1152 down_write(&key->sem);
1154 LASSERT(key->payload.data == NULL);
1156 cli_ctx = ctx_create_kr(sec, &vcred);
1157 if (cli_ctx == NULL) {
1162 rc = gss_copy_rvc_cli_ctx(cli_ctx, svc_ctx);
1164 CERROR("failed copy reverse cli ctx: %d\n", rc);
1168 rvs_sec_install_root_ctx_kr(sec, cli_ctx, key);
1170 ctx_put_kr(cli_ctx, 1);
1171 up_write(&key->sem);
1180 ctx_put_kr(cli_ctx, 1);
1182 up_write(&key->sem);
1188 #endif /* HAVE_REVERSE_CTX_NOKEY */
1190 /****************************************
1192 ****************************************/
1195 int gss_svc_accept_kr(struct ptlrpc_request *req)
1197 return gss_svc_accept(&gss_policy_keyring, req);
1201 int gss_svc_install_rctx_kr(struct obd_import *imp,
1202 struct ptlrpc_svc_ctx *svc_ctx)
1204 struct ptlrpc_sec *sec;
1207 sec = sptlrpc_import_sec_ref(imp);
1210 rc = sec_install_rctx_kr(sec, svc_ctx);
1211 sptlrpc_sec_put(sec);
1216 /****************************************
1218 ****************************************/
1221 int gss_kt_instantiate(struct key *key, const void *data, size_t datalen)
1226 if (data != NULL || datalen != 0) {
1227 CERROR("invalid: data %p, len %lu\n", data, (long)datalen);
1231 if (key->payload.data != 0) {
1232 CERROR("key already have payload\n");
1236 /* link the key to session keyring, so following context negotiation
1237 * rpc fired from user space could find this key. This will be unlinked
1238 * automatically when upcall processes die.
1240 * we can't do this through keyctl from userspace, because the upcall
1241 * might be neither possessor nor owner of the key (setuid).
1243 * the session keyring is created upon upcall, and don't change all
1244 * the way until upcall finished, so rcu lock is not needed here.
1246 LASSERT(cfs_current()->signal->session_keyring);
1249 rc = key_link(cfs_current()->signal->session_keyring, key);
1252 CERROR("failed to link key %08x to keyring %08x: %d\n",
1254 cfs_current()->signal->session_keyring->serial, rc);
1258 CDEBUG(D_SEC, "key %p instantiated, ctx %p\n", key, key->payload.data);
1263 * called with key semaphore write locked. it means we can operate
1264 * on the context without fear of loosing refcount.
1267 int gss_kt_update(struct key *key, const void *data, size_t datalen)
1269 struct ptlrpc_cli_ctx *ctx = key->payload.data;
1270 struct gss_cli_ctx *gctx;
1271 rawobj_t tmpobj = RAWOBJ_EMPTY;
1272 __u32 datalen32 = (__u32) datalen;
1276 if (data == NULL || datalen == 0) {
1277 CWARN("invalid: data %p, len %lu\n", data, (long)datalen);
1281 /* if upcall finished negotiation too fast (mostly likely because
1282 * of local error happened) and call kt_update(), the ctx
1283 * might be still NULL. but the key will finally be associate
1284 * with a context, or be revoked. if key status is fine, return
1285 * -EAGAIN to allow userspace sleep a while and call again. */
1287 CDEBUG(D_SEC, "update too soon: key %p(%x) flags %lx\n",
1288 key, key->serial, key->flags);
1290 rc = key_validate(key);
1297 LASSERT(cfs_atomic_read(&ctx->cc_refcount) > 0);
1298 LASSERT(ctx->cc_sec);
1300 ctx_clear_timer_kr(ctx);
1302 /* don't proceed if already refreshed */
1303 if (cli_ctx_is_refreshed(ctx)) {
1304 CWARN("ctx already done refresh\n");
1308 sptlrpc_cli_ctx_get(ctx);
1309 gctx = ctx2gctx(ctx);
1311 rc = buffer_extract_bytes(&data, &datalen32, &gctx->gc_win,
1312 sizeof(gctx->gc_win));
1314 CERROR("failed extract seq_win\n");
1318 if (gctx->gc_win == 0) {
1319 __u32 nego_rpc_err, nego_gss_err;
1321 rc = buffer_extract_bytes(&data, &datalen32, &nego_rpc_err,
1322 sizeof(nego_rpc_err));
1324 CERROR("failed to extrace rpc rc\n");
1328 rc = buffer_extract_bytes(&data, &datalen32, &nego_gss_err,
1329 sizeof(nego_gss_err));
1331 CERROR("failed to extrace gss rc\n");
1335 CERROR("negotiation: rpc err %d, gss err %x\n",
1336 nego_rpc_err, nego_gss_err);
1338 rc = nego_rpc_err ? nego_rpc_err : -EACCES;
1340 rc = rawobj_extract_local_alloc(&gctx->gc_handle,
1341 (__u32 **) &data, &datalen32);
1343 CERROR("failed extract handle\n");
1347 rc = rawobj_extract_local(&tmpobj, (__u32 **) &data,&datalen32);
1349 CERROR("failed extract mech\n");
1353 rc = lgss_import_sec_context(&tmpobj,
1354 sec2gsec(ctx->cc_sec)->gs_mech,
1356 if (rc != GSS_S_COMPLETE)
1357 CERROR("failed import context\n");
1362 /* we don't care what current status of this ctx, even someone else
1363 * is operating on the ctx at the same time. we just add up our own
1366 gss_cli_ctx_uptodate(gctx);
1368 /* this will also revoke the key. has to be done before
1369 * wakeup waiters otherwise they can find the stale key */
1370 kill_key_locked(key);
1372 cli_ctx_expire(ctx);
1374 if (rc != -ERESTART)
1375 cfs_set_bit(PTLRPC_CTX_ERROR_BIT, &ctx->cc_flags);
1378 /* let user space think it's a success */
1379 sptlrpc_cli_ctx_put(ctx, 1);
1384 int gss_kt_match(const struct key *key, const void *desc)
1386 return (strcmp(key->description, (const char *) desc) == 0);
1390 void gss_kt_destroy(struct key *key)
1393 LASSERT(key->payload.data == NULL);
1394 CDEBUG(D_SEC, "destroy key %p\n", key);
1399 void gss_kt_describe(const struct key *key, struct seq_file *s)
1401 if (key->description == NULL)
1402 seq_puts(s, "[null]");
1404 seq_puts(s, key->description);
1407 static struct key_type gss_key_type =
1411 .instantiate = gss_kt_instantiate,
1412 .update = gss_kt_update,
1413 .match = gss_kt_match,
1414 .destroy = gss_kt_destroy,
1415 .describe = gss_kt_describe,
1418 /****************************************
1419 * lustre gss keyring policy *
1420 ****************************************/
1422 static struct ptlrpc_ctx_ops gss_keyring_ctxops = {
1423 .match = gss_cli_ctx_match,
1424 .refresh = gss_cli_ctx_refresh_kr,
1425 .validate = gss_cli_ctx_validate_kr,
1426 .die = gss_cli_ctx_die_kr,
1427 .sign = gss_cli_ctx_sign,
1428 .verify = gss_cli_ctx_verify,
1429 .seal = gss_cli_ctx_seal,
1430 .unseal = gss_cli_ctx_unseal,
1431 .wrap_bulk = gss_cli_ctx_wrap_bulk,
1432 .unwrap_bulk = gss_cli_ctx_unwrap_bulk,
1435 static struct ptlrpc_sec_cops gss_sec_keyring_cops = {
1436 .create_sec = gss_sec_create_kr,
1437 .destroy_sec = gss_sec_destroy_kr,
1438 .kill_sec = gss_sec_kill,
1439 .lookup_ctx = gss_sec_lookup_ctx_kr,
1440 .release_ctx = gss_sec_release_ctx_kr,
1441 .flush_ctx_cache = gss_sec_flush_ctx_cache_kr,
1442 .gc_ctx = gss_sec_gc_ctx_kr,
1443 .install_rctx = gss_sec_install_rctx,
1444 .alloc_reqbuf = gss_alloc_reqbuf,
1445 .free_reqbuf = gss_free_reqbuf,
1446 .alloc_repbuf = gss_alloc_repbuf,
1447 .free_repbuf = gss_free_repbuf,
1448 .enlarge_reqbuf = gss_enlarge_reqbuf,
1449 .display = gss_sec_display_kr,
1452 static struct ptlrpc_sec_sops gss_sec_keyring_sops = {
1453 .accept = gss_svc_accept_kr,
1454 .invalidate_ctx = gss_svc_invalidate_ctx,
1455 .alloc_rs = gss_svc_alloc_rs,
1456 .authorize = gss_svc_authorize,
1457 .free_rs = gss_svc_free_rs,
1458 .free_ctx = gss_svc_free_ctx,
1459 .prep_bulk = gss_svc_prep_bulk,
1460 .unwrap_bulk = gss_svc_unwrap_bulk,
1461 .wrap_bulk = gss_svc_wrap_bulk,
1462 .install_rctx = gss_svc_install_rctx_kr,
1465 static struct ptlrpc_sec_policy gss_policy_keyring = {
1466 .sp_owner = THIS_MODULE,
1467 .sp_name = "gss.keyring",
1468 .sp_policy = SPTLRPC_POLICY_GSS,
1469 .sp_cops = &gss_sec_keyring_cops,
1470 .sp_sops = &gss_sec_keyring_sops,
1474 int __init gss_init_keyring(void)
1478 rc = register_key_type(&gss_key_type);
1480 CERROR("failed to register keyring type: %d\n", rc);
1484 rc = sptlrpc_register_policy(&gss_policy_keyring);
1486 unregister_key_type(&gss_key_type);
1493 void __exit gss_exit_keyring(void)
1495 unregister_key_type(&gss_key_type);
1496 sptlrpc_unregister_policy(&gss_policy_keyring);