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.
30 * Copyright (c) 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/ptlrpc/gss/gss_keyring.c
38 * Author: Eric Mei <ericm@clusterfs.com>
41 #define DEBUG_SUBSYSTEM S_SEC
43 #include <linux/init.h>
44 #include <linux/module.h>
45 #include <linux/slab.h>
46 #include <linux/dcache.h>
48 #include <linux/crypto.h>
49 #include <linux/key.h>
50 #include <linux/keyctl.h>
51 #include <linux/mutex.h>
52 #include <asm/atomic.h>
54 #include <liblustre.h>
58 #include <obd_class.h>
59 #include <obd_support.h>
60 #include <lustre/lustre_idl.h>
61 #include <lustre_sec.h>
62 #include <lustre_net.h>
63 #include <lustre_import.h>
66 #include "gss_internal.h"
69 static struct ptlrpc_sec_policy gss_policy_keyring;
70 static struct ptlrpc_ctx_ops gss_keyring_ctxops;
71 static struct key_type gss_key_type;
73 static int sec_install_rctx_kr(struct ptlrpc_sec *sec,
74 struct ptlrpc_svc_ctx *svc_ctx);
77 * the timeout is only for the case that upcall child process die abnormally.
78 * in any other cases it should finally update kernel key.
80 * FIXME we'd better to incorporate the client & server side upcall timeouts
81 * into the framework of Adaptive Timeouts, but we need to figure out how to
82 * make sure that kernel knows the upcall processes is in-progress or died
85 #define KEYRING_UPCALL_TIMEOUT (obd_timeout + obd_timeout)
87 /****************************************
89 ****************************************/
91 #define DUMP_PROCESS_KEYRINGS(tsk) \
93 CWARN("DUMP PK: %s[%u,%u/%u](<-%s[%u,%u/%u]): " \
94 "a %d, t %d, p %d, s %d, u %d, us %d, df %d\n", \
95 tsk->comm, tsk->pid, tsk->uid, tsk->fsuid, \
96 tsk->parent->comm, tsk->parent->pid, \
97 tsk->parent->uid, tsk->parent->fsuid, \
98 tsk->request_key_auth ? \
99 tsk->request_key_auth->serial : 0, \
100 tsk->thread_keyring ? \
101 tsk->thread_keyring->serial : 0, \
102 tsk->signal->process_keyring ? \
103 tsk->signal->process_keyring->serial : 0, \
104 tsk->signal->session_keyring ? \
105 tsk->signal->session_keyring->serial : 0, \
106 tsk->user->uid_keyring ? \
107 tsk->user->uid_keyring->serial : 0, \
108 tsk->user->session_keyring ? \
109 tsk->user->session_keyring->serial : 0, \
114 #define DUMP_KEY(key) \
116 CWARN("DUMP KEY: %p(%d) ref %d u%u/g%u desc %s\n", \
117 key, key->serial, atomic_read(&key->usage), \
118 key->uid, key->gid, \
119 key->description ? key->description : "n/a" \
124 static inline void keyring_upcall_lock(struct gss_sec_keyring *gsec_kr)
126 #ifdef HAVE_KEYRING_UPCALL_SERIALIZED
127 mutex_lock(&gsec_kr->gsk_uc_lock);
131 static inline void keyring_upcall_unlock(struct gss_sec_keyring *gsec_kr)
133 #ifdef HAVE_KEYRING_UPCALL_SERIALIZED
134 mutex_unlock(&gsec_kr->gsk_uc_lock);
138 static inline void key_revoke_locked(struct key *key)
140 set_bit(KEY_FLAG_REVOKED, &key->flags);
143 static void ctx_upcall_timeout_kr(unsigned long data)
145 struct ptlrpc_cli_ctx *ctx = (struct ptlrpc_cli_ctx *) data;
146 struct key *key = ctx2gctx_keyring(ctx)->gck_key;
148 CWARN("ctx %p, key %p\n", ctx, key);
153 key_revoke_locked(key);
157 void ctx_start_timer_kr(struct ptlrpc_cli_ctx *ctx, long timeout)
159 struct gss_cli_ctx_keyring *gctx_kr = ctx2gctx_keyring(ctx);
160 struct timer_list *timer = gctx_kr->gck_timer;
164 CDEBUG(D_SEC, "ctx %p: start timer %lds\n", ctx, timeout);
165 timeout = timeout * CFS_HZ + cfs_time_current();
168 timer->expires = timeout;
169 timer->data = (unsigned long ) ctx;
170 timer->function = ctx_upcall_timeout_kr;
176 * caller should make sure no race with other threads
179 void ctx_clear_timer_kr(struct ptlrpc_cli_ctx *ctx)
181 struct gss_cli_ctx_keyring *gctx_kr = ctx2gctx_keyring(ctx);
182 struct timer_list *timer = gctx_kr->gck_timer;
187 CDEBUG(D_SEC, "ctx %p, key %p\n", ctx, gctx_kr->gck_key);
189 gctx_kr->gck_timer = NULL;
191 del_singleshot_timer_sync(timer);
197 struct ptlrpc_cli_ctx *ctx_create_kr(struct ptlrpc_sec *sec,
198 struct vfs_cred *vcred)
200 struct ptlrpc_cli_ctx *ctx;
201 struct gss_cli_ctx_keyring *gctx_kr;
203 OBD_ALLOC_PTR(gctx_kr);
207 OBD_ALLOC_PTR(gctx_kr->gck_timer);
208 if (gctx_kr->gck_timer == NULL) {
209 OBD_FREE_PTR(gctx_kr);
212 init_timer(gctx_kr->gck_timer);
214 ctx = &gctx_kr->gck_base.gc_base;
216 if (gss_cli_ctx_init_common(sec, ctx, &gss_keyring_ctxops, vcred)) {
217 OBD_FREE_PTR(gctx_kr->gck_timer);
218 OBD_FREE_PTR(gctx_kr);
222 ctx->cc_expire = cfs_time_current_sec() + KEYRING_UPCALL_TIMEOUT;
223 clear_bit(PTLRPC_CTX_NEW_BIT, &ctx->cc_flags);
224 cfs_atomic_inc(&ctx->cc_refcount); /* for the caller */
229 static void ctx_destroy_kr(struct ptlrpc_cli_ctx *ctx)
231 struct ptlrpc_sec *sec = ctx->cc_sec;
232 struct gss_cli_ctx_keyring *gctx_kr = ctx2gctx_keyring(ctx);
234 CDEBUG(D_SEC, "destroying ctx %p\n", ctx);
236 /* at this time the association with key has been broken. */
238 LASSERT(cfs_atomic_read(&sec->ps_refcount) > 0);
239 LASSERT(cfs_atomic_read(&sec->ps_nctx) > 0);
240 LASSERT(test_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags) == 0);
241 LASSERT(gctx_kr->gck_key == NULL);
243 ctx_clear_timer_kr(ctx);
244 LASSERT(gctx_kr->gck_timer == NULL);
246 if (gss_cli_ctx_fini_common(sec, ctx))
249 OBD_FREE_PTR(gctx_kr);
251 cfs_atomic_dec(&sec->ps_nctx);
252 sptlrpc_sec_put(sec);
255 static void ctx_release_kr(struct ptlrpc_cli_ctx *ctx, int sync)
260 cfs_atomic_inc(&ctx->cc_refcount);
261 sptlrpc_gc_add_ctx(ctx);
265 static void ctx_put_kr(struct ptlrpc_cli_ctx *ctx, int sync)
267 LASSERT(cfs_atomic_read(&ctx->cc_refcount) > 0);
269 if (cfs_atomic_dec_and_test(&ctx->cc_refcount))
270 ctx_release_kr(ctx, sync);
274 * key <-> ctx association and rules:
275 * - ctx might not bind with any key
276 * - key/ctx binding is protected by key semaphore (if the key present)
277 * - key and ctx each take a reference of the other
278 * - ctx enlist/unlist is protected by ctx spinlock
279 * - never enlist a ctx after it's been unlisted
280 * - whoever do enlist should also do bind, lock key before enlist:
281 * - lock key -> lock ctx -> enlist -> unlock ctx -> bind -> unlock key
282 * - whoever do unlist should also do unbind:
283 * - lock key -> lock ctx -> unlist -> unlock ctx -> unbind -> unlock key
284 * - lock ctx -> unlist -> unlock ctx -> lock key -> unbind -> unlock key
287 static inline void spin_lock_if(spinlock_t *lock, int condition)
293 static inline void spin_unlock_if(spinlock_t *lock, int condition)
299 static void ctx_enlist_kr(struct ptlrpc_cli_ctx *ctx, int is_root, int locked)
301 struct ptlrpc_sec *sec = ctx->cc_sec;
302 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
304 LASSERT(!test_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags));
305 LASSERT(cfs_atomic_read(&ctx->cc_refcount) > 0);
307 spin_lock_if(&sec->ps_lock, !locked);
309 cfs_atomic_inc(&ctx->cc_refcount);
310 set_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags);
311 cfs_hlist_add_head(&ctx->cc_cache, &gsec_kr->gsk_clist);
313 gsec_kr->gsk_root_ctx = ctx;
315 spin_unlock_if(&sec->ps_lock, !locked);
319 * Note after this get called, caller should not access ctx again because
320 * it might have been freed, unless caller hold at least one refcount of
323 * return non-zero if we indeed unlist this ctx.
325 static int ctx_unlist_kr(struct ptlrpc_cli_ctx *ctx, int locked)
327 struct ptlrpc_sec *sec = ctx->cc_sec;
328 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
330 /* if hashed bit has gone, leave the job to somebody who is doing it */
331 if (test_and_clear_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags) == 0)
334 /* drop ref inside spin lock to prevent race with other operations */
335 spin_lock_if(&sec->ps_lock, !locked);
337 if (gsec_kr->gsk_root_ctx == ctx)
338 gsec_kr->gsk_root_ctx = NULL;
339 cfs_hlist_del_init(&ctx->cc_cache);
340 cfs_atomic_dec(&ctx->cc_refcount);
342 spin_unlock_if(&sec->ps_lock, !locked);
348 * bind a key with a ctx together.
349 * caller must hold write lock of the key, as well as ref on key & ctx.
351 static void bind_key_ctx(struct key *key, struct ptlrpc_cli_ctx *ctx)
353 LASSERT(cfs_atomic_read(&ctx->cc_refcount) > 0);
354 LASSERT(atomic_read(&key->usage) > 0);
355 LASSERT(ctx2gctx_keyring(ctx)->gck_key == NULL);
356 LASSERT(key->payload.data == NULL);
358 /* at this time context may or may not in list. */
360 cfs_atomic_inc(&ctx->cc_refcount);
361 ctx2gctx_keyring(ctx)->gck_key = key;
362 key->payload.data = ctx;
366 * unbind a key and a ctx.
367 * caller must hold write lock, as well as a ref of the key.
369 static void unbind_key_ctx(struct key *key, struct ptlrpc_cli_ctx *ctx)
371 LASSERT(key->payload.data == ctx);
372 LASSERT(test_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags) == 0);
374 /* must revoke the key, or others may treat it as newly created */
375 key_revoke_locked(key);
377 key->payload.data = NULL;
378 ctx2gctx_keyring(ctx)->gck_key = NULL;
380 /* once ctx get split from key, the timer is meaningless */
381 ctx_clear_timer_kr(ctx);
388 * given a ctx, unbind with its coupled key, if any.
389 * unbind could only be called once, so we don't worry the key be released
392 static void unbind_ctx_kr(struct ptlrpc_cli_ctx *ctx)
394 struct key *key = ctx2gctx_keyring(ctx)->gck_key;
397 LASSERT(key->payload.data == ctx);
400 down_write(&key->sem);
401 unbind_key_ctx(key, ctx);
408 * given a key, unbind with its coupled ctx, if any.
409 * caller must hold write lock, as well as a ref of the key.
411 static void unbind_key_locked(struct key *key)
413 struct ptlrpc_cli_ctx *ctx = key->payload.data;
416 unbind_key_ctx(key, ctx);
420 * unlist a ctx, and unbind from coupled key
422 static void kill_ctx_kr(struct ptlrpc_cli_ctx *ctx)
424 if (ctx_unlist_kr(ctx, 0))
429 * given a key, unlist and unbind with the coupled ctx (if any).
430 * caller must hold write lock, as well as a ref of the key.
432 static void kill_key_locked(struct key *key)
434 struct ptlrpc_cli_ctx *ctx = key->payload.data;
436 if (ctx && ctx_unlist_kr(ctx, 0))
437 unbind_key_locked(key);
441 * caller should hold one ref on contexts in freelist.
443 static void dispose_ctx_list_kr(cfs_hlist_head_t *freelist)
445 cfs_hlist_node_t *pos, *next;
446 struct ptlrpc_cli_ctx *ctx;
447 struct gss_cli_ctx *gctx;
449 cfs_hlist_for_each_entry_safe(ctx, pos, next, freelist, cc_cache) {
450 cfs_hlist_del_init(&ctx->cc_cache);
452 /* reverse ctx: update current seq to buddy svcctx if exist.
453 * ideally this should be done at gss_cli_ctx_finalize(), but
454 * the ctx destroy could be delayed by:
455 * 1) ctx still has reference;
456 * 2) ctx destroy is asynchronous;
457 * and reverse import call inval_all_ctx() require this be done
458 *_immediately_ otherwise newly created reverse ctx might copy
459 * the very old sequence number from svcctx. */
460 gctx = ctx2gctx(ctx);
461 if (!rawobj_empty(&gctx->gc_svc_handle) &&
462 sec_is_reverse(gctx->gc_base.cc_sec)) {
463 gss_svc_upcall_update_sequence(&gctx->gc_svc_handle,
464 (__u32) cfs_atomic_read(&gctx->gc_seq));
467 /* we need to wakeup waiting reqs here. the context might
468 * be forced released before upcall finished, then the
469 * late-arrived downcall can't find the ctx even. */
470 sptlrpc_cli_ctx_wakeup(ctx);
478 * lookup a root context directly in a sec, return root ctx with a
479 * reference taken or NULL.
482 struct ptlrpc_cli_ctx * sec_lookup_root_ctx_kr(struct ptlrpc_sec *sec)
484 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
485 struct ptlrpc_cli_ctx *ctx = NULL;
487 spin_lock(&sec->ps_lock);
489 ctx = gsec_kr->gsk_root_ctx;
491 if (ctx == NULL && unlikely(sec_is_reverse(sec))) {
492 cfs_hlist_node_t *node;
493 struct ptlrpc_cli_ctx *tmp;
495 /* reverse ctx, search root ctx in list, choose the one
496 * with shortest expire time, which is most possibly have
497 * an established peer ctx at client side. */
498 cfs_hlist_for_each_entry(tmp, node, &gsec_kr->gsk_clist,
500 if (ctx == NULL || ctx->cc_expire == 0 ||
501 ctx->cc_expire > tmp->cc_expire) {
503 /* promote to be root_ctx */
504 gsec_kr->gsk_root_ctx = ctx;
510 LASSERT(cfs_atomic_read(&ctx->cc_refcount) > 0);
511 LASSERT(!cfs_hlist_empty(&gsec_kr->gsk_clist));
512 cfs_atomic_inc(&ctx->cc_refcount);
515 spin_unlock(&sec->ps_lock);
520 #define RVS_CTX_EXPIRE_NICE (10)
523 void rvs_sec_install_root_ctx_kr(struct ptlrpc_sec *sec,
524 struct ptlrpc_cli_ctx *new_ctx,
527 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
528 cfs_hlist_node_t *hnode;
529 struct ptlrpc_cli_ctx *ctx;
533 LASSERT(sec_is_reverse(sec));
535 spin_lock(&sec->ps_lock);
537 now = cfs_time_current_sec();
539 /* set all existing ctxs short expiry */
540 cfs_hlist_for_each_entry(ctx, hnode, &gsec_kr->gsk_clist, cc_cache) {
541 if (ctx->cc_expire > now + RVS_CTX_EXPIRE_NICE) {
542 ctx->cc_early_expire = 1;
543 ctx->cc_expire = now + RVS_CTX_EXPIRE_NICE;
547 /* if there's root_ctx there, instead obsolete the current
548 * immediately, we leave it continue operating for a little while.
549 * hopefully when the first backward rpc with newest ctx send out,
550 * the client side already have the peer ctx well established. */
551 ctx_enlist_kr(new_ctx, gsec_kr->gsk_root_ctx ? 0 : 1, 1);
554 bind_key_ctx(key, new_ctx);
556 spin_unlock(&sec->ps_lock);
559 static void construct_key_desc(void *buf, int bufsize,
560 struct ptlrpc_sec *sec, uid_t uid)
562 snprintf(buf, bufsize, "%d@%x", uid, sec->ps_id);
563 ((char *)buf)[bufsize - 1] = '\0';
566 /****************************************
568 ****************************************/
571 struct ptlrpc_sec * gss_sec_create_kr(struct obd_import *imp,
572 struct ptlrpc_svc_ctx *svcctx,
573 struct sptlrpc_flavor *sf)
575 struct gss_sec_keyring *gsec_kr;
578 OBD_ALLOC(gsec_kr, sizeof(*gsec_kr));
582 CFS_INIT_HLIST_HEAD(&gsec_kr->gsk_clist);
583 gsec_kr->gsk_root_ctx = NULL;
584 mutex_init(&gsec_kr->gsk_root_uc_lock);
585 #ifdef HAVE_KEYRING_UPCALL_SERIALIZED
586 mutex_init(&gsec_kr->gsk_uc_lock);
589 if (gss_sec_create_common(&gsec_kr->gsk_base, &gss_policy_keyring,
593 if (svcctx != NULL &&
594 sec_install_rctx_kr(&gsec_kr->gsk_base.gs_base, svcctx)) {
595 gss_sec_destroy_common(&gsec_kr->gsk_base);
599 RETURN(&gsec_kr->gsk_base.gs_base);
602 OBD_FREE(gsec_kr, sizeof(*gsec_kr));
607 void gss_sec_destroy_kr(struct ptlrpc_sec *sec)
609 struct gss_sec *gsec = sec2gsec(sec);
610 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
612 CDEBUG(D_SEC, "destroy %s@%p\n", sec->ps_policy->sp_name, sec);
614 LASSERT(cfs_hlist_empty(&gsec_kr->gsk_clist));
615 LASSERT(gsec_kr->gsk_root_ctx == NULL);
617 gss_sec_destroy_common(gsec);
619 OBD_FREE(gsec_kr, sizeof(*gsec_kr));
622 static inline int user_is_root(struct ptlrpc_sec *sec, struct vfs_cred *vcred)
624 /* except the ROOTONLY flag, treat it as root user only if real uid
625 * is 0, euid/fsuid being 0 are handled as setuid scenarios */
626 if (sec_is_rootonly(sec) || (vcred->vc_uid == 0))
633 * unlink request key from it's ring, which is linked during request_key().
634 * sadly, we have to 'guess' which keyring it's linked to.
636 * FIXME this code is fragile, depend on how request_key_link() is implemented.
638 static void request_key_unlink(struct key *key)
640 struct task_struct *tsk = current;
643 switch (tsk->jit_keyring) {
644 case KEY_REQKEY_DEFL_DEFAULT:
645 case KEY_REQKEY_DEFL_THREAD_KEYRING:
646 ring = key_get(tsk->thread_keyring);
649 case KEY_REQKEY_DEFL_PROCESS_KEYRING:
650 ring = key_get(tsk->signal->process_keyring);
653 case KEY_REQKEY_DEFL_SESSION_KEYRING:
655 ring = key_get(rcu_dereference(tsk->signal->session_keyring));
659 case KEY_REQKEY_DEFL_USER_SESSION_KEYRING:
660 ring = key_get(tsk->user->session_keyring);
662 case KEY_REQKEY_DEFL_USER_KEYRING:
663 ring = key_get(tsk->user->uid_keyring);
665 case KEY_REQKEY_DEFL_GROUP_KEYRING:
671 key_unlink(ring, key);
676 struct ptlrpc_cli_ctx * gss_sec_lookup_ctx_kr(struct ptlrpc_sec *sec,
677 struct vfs_cred *vcred,
678 int create, int remove_dead)
680 struct obd_import *imp = sec->ps_import;
681 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
682 struct ptlrpc_cli_ctx *ctx = NULL;
683 unsigned int is_root = 0, create_new = 0;
691 LASSERT(imp != NULL);
693 is_root = user_is_root(sec, vcred);
695 /* a little bit optimization for root context */
697 ctx = sec_lookup_root_ctx_kr(sec);
699 * Only lookup directly for REVERSE sec, which should
702 if (ctx || sec_is_reverse(sec))
706 LASSERT(create != 0);
708 /* for root context, obtain lock and check again, this time hold
709 * the root upcall lock, make sure nobody else populated new root
710 * context after last check. */
712 mutex_lock(&gsec_kr->gsk_root_uc_lock);
714 ctx = sec_lookup_root_ctx_kr(sec);
718 /* update reverse handle for root user */
719 sec2gsec(sec)->gs_rvs_hdl = gss_get_next_ctx_index();
721 switch (sec->ps_part) {
740 /* in case of setuid, key will be constructed as owner of fsuid/fsgid,
741 * but we do authentication based on real uid/gid. the key permission
742 * bits will be exactly as POS_ALL, so only processes who subscribed
743 * this key could have the access, although the quota might be counted
744 * on others (fsuid/fsgid).
746 * keyring will use fsuid/fsgid as upcall parameters, so we have to
747 * encode real uid/gid into callout info.
750 construct_key_desc(desc, sizeof(desc), sec, vcred->vc_uid);
752 /* callout info format:
753 * secid:mech:uid:gid:flags:svc_type:peer_nid:target_uuid
755 coinfo_size = sizeof(struct obd_uuid) + MAX_OBD_NAME + 64;
756 OBD_ALLOC(coinfo, coinfo_size);
760 snprintf(coinfo, coinfo_size, "%d:%s:%u:%u:%s:%d:"LPX64":%s",
761 sec->ps_id, sec2gsec(sec)->gs_mech->gm_name,
762 vcred->vc_uid, vcred->vc_gid,
763 co_flags, import_to_gss_svc(imp),
764 imp->imp_connection->c_peer.nid, imp->imp_obd->obd_name);
766 CDEBUG(D_SEC, "requesting key for %s\n", desc);
768 keyring_upcall_lock(gsec_kr);
769 key = request_key(&gss_key_type, desc, coinfo);
770 keyring_upcall_unlock(gsec_kr);
772 OBD_FREE(coinfo, coinfo_size);
775 CERROR("failed request key: %ld\n", PTR_ERR(key));
778 CDEBUG(D_SEC, "obtained key %08x for %s\n", key->serial, desc);
780 /* once payload.data was pointed to a ctx, it never changes until
781 * we de-associate them; but parallel request_key() may return
782 * a key with payload.data == NULL at the same time. so we still
783 * need wirtelock of key->sem to serialize them. */
784 down_write(&key->sem);
786 if (likely(key->payload.data != NULL)) {
787 ctx = key->payload.data;
789 LASSERT(cfs_atomic_read(&ctx->cc_refcount) >= 1);
790 LASSERT(ctx2gctx_keyring(ctx)->gck_key == key);
791 LASSERT(atomic_read(&key->usage) >= 2);
793 /* simply take a ref and return. it's upper layer's
794 * responsibility to detect & replace dead ctx. */
795 cfs_atomic_inc(&ctx->cc_refcount);
797 /* pre initialization with a cli_ctx. this can't be done in
798 * key_instantiate() because we'v no enough information
800 ctx = ctx_create_kr(sec, vcred);
802 ctx_enlist_kr(ctx, is_root, 0);
803 bind_key_ctx(key, ctx);
805 ctx_start_timer_kr(ctx, KEYRING_UPCALL_TIMEOUT);
807 CDEBUG(D_SEC, "installed key %p <-> ctx %p (sec %p)\n",
810 /* we'd prefer to call key_revoke(), but we more like
811 * to revoke it within this key->sem locked period. */
812 key_revoke_locked(key);
820 if (is_root && create_new)
821 request_key_unlink(key);
826 mutex_unlock(&gsec_kr->gsk_root_uc_lock);
831 void gss_sec_release_ctx_kr(struct ptlrpc_sec *sec,
832 struct ptlrpc_cli_ctx *ctx,
835 LASSERT(cfs_atomic_read(&sec->ps_refcount) > 0);
836 LASSERT(cfs_atomic_read(&ctx->cc_refcount) == 0);
837 ctx_release_kr(ctx, sync);
841 * flush context of normal user, we must resort to keyring itself to find out
842 * contexts which belong to me.
844 * Note here we suppose only to flush _my_ context, the "uid" will
845 * be ignored in the search.
848 void flush_user_ctx_cache_kr(struct ptlrpc_sec *sec,
850 int grace, int force)
855 /* nothing to do for reverse or rootonly sec */
856 if (sec_is_reverse(sec) || sec_is_rootonly(sec))
859 construct_key_desc(desc, sizeof(desc), sec, uid);
861 /* there should be only one valid key, but we put it in the
862 * loop in case of any weird cases */
864 key = request_key(&gss_key_type, desc, NULL);
866 CDEBUG(D_SEC, "No more key found for current user\n");
870 down_write(&key->sem);
872 kill_key_locked(key);
874 /* kill_key_locked() should usually revoke the key, but we
875 * revoke it again to make sure, e.g. some case the key may
876 * not well coupled with a context. */
877 key_revoke_locked(key);
886 * flush context of root or all, we iterate through the list.
889 void flush_spec_ctx_cache_kr(struct ptlrpc_sec *sec,
891 int grace, int force)
893 struct gss_sec_keyring *gsec_kr;
894 cfs_hlist_head_t freelist = CFS_HLIST_HEAD_INIT;
895 cfs_hlist_node_t *pos, *next;
896 struct ptlrpc_cli_ctx *ctx;
899 gsec_kr = sec2gsec_keyring(sec);
901 spin_lock(&sec->ps_lock);
902 cfs_hlist_for_each_entry_safe(ctx, pos, next,
903 &gsec_kr->gsk_clist, cc_cache) {
904 LASSERT(cfs_atomic_read(&ctx->cc_refcount) > 0);
906 if (uid != -1 && uid != ctx->cc_vcred.vc_uid)
909 /* at this moment there's at least 2 base reference:
910 * key association and in-list. */
911 if (cfs_atomic_read(&ctx->cc_refcount) > 2) {
914 CWARN("flush busy ctx %p(%u->%s, extra ref %d)\n",
915 ctx, ctx->cc_vcred.vc_uid,
916 sec2target_str(ctx->cc_sec),
917 cfs_atomic_read(&ctx->cc_refcount) - 2);
920 set_bit(PTLRPC_CTX_DEAD_BIT, &ctx->cc_flags);
922 clear_bit(PTLRPC_CTX_UPTODATE_BIT, &ctx->cc_flags);
924 cfs_atomic_inc(&ctx->cc_refcount);
926 if (ctx_unlist_kr(ctx, 1)) {
927 cfs_hlist_add_head(&ctx->cc_cache, &freelist);
929 LASSERT(cfs_atomic_read(&ctx->cc_refcount) >= 2);
930 cfs_atomic_dec(&ctx->cc_refcount);
933 spin_unlock(&sec->ps_lock);
935 dispose_ctx_list_kr(&freelist);
940 int gss_sec_flush_ctx_cache_kr(struct ptlrpc_sec *sec,
941 uid_t uid, int grace, int force)
945 CDEBUG(D_SEC, "sec %p(%d, nctx %d), uid %d, grace %d, force %d\n",
946 sec, cfs_atomic_read(&sec->ps_refcount),
947 cfs_atomic_read(&sec->ps_nctx),
950 if (uid != -1 && uid != 0)
951 flush_user_ctx_cache_kr(sec, uid, grace, force);
953 flush_spec_ctx_cache_kr(sec, uid, grace, force);
959 void gss_sec_gc_ctx_kr(struct ptlrpc_sec *sec)
961 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
962 cfs_hlist_head_t freelist = CFS_HLIST_HEAD_INIT;
963 cfs_hlist_node_t *pos, *next;
964 struct ptlrpc_cli_ctx *ctx;
967 CWARN("running gc\n");
969 spin_lock(&sec->ps_lock);
970 cfs_hlist_for_each_entry_safe(ctx, pos, next,
971 &gsec_kr->gsk_clist, cc_cache) {
972 LASSERT(cfs_atomic_read(&ctx->cc_refcount) > 0);
974 cfs_atomic_inc(&ctx->cc_refcount);
976 if (cli_ctx_check_death(ctx) && ctx_unlist_kr(ctx, 1)) {
977 cfs_hlist_add_head(&ctx->cc_cache, &freelist);
978 CWARN("unhashed ctx %p\n", ctx);
980 LASSERT(cfs_atomic_read(&ctx->cc_refcount) >= 2);
981 cfs_atomic_dec(&ctx->cc_refcount);
984 spin_unlock(&sec->ps_lock);
986 dispose_ctx_list_kr(&freelist);
992 int gss_sec_display_kr(struct ptlrpc_sec *sec, struct seq_file *seq)
994 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
995 cfs_hlist_node_t *pos, *next;
996 struct ptlrpc_cli_ctx *ctx;
997 struct gss_cli_ctx *gctx;
998 time_t now = cfs_time_current_sec();
1001 spin_lock(&sec->ps_lock);
1002 cfs_hlist_for_each_entry_safe(ctx, pos, next,
1003 &gsec_kr->gsk_clist, cc_cache) {
1008 gctx = ctx2gctx(ctx);
1009 key = ctx2gctx_keyring(ctx)->gck_key;
1011 gss_cli_ctx_flags2str(ctx->cc_flags,
1012 flags_str, sizeof(flags_str));
1014 if (gctx->gc_mechctx)
1015 lgss_display(gctx->gc_mechctx, mech, sizeof(mech));
1017 snprintf(mech, sizeof(mech), "N/A");
1018 mech[sizeof(mech) - 1] = '\0';
1020 seq_printf(seq, "%p: uid %u, ref %d, expire %ld(%+ld), fl %s, "
1021 "seq %d, win %u, key %08x(ref %d), "
1022 "hdl "LPX64":"LPX64", mech: %s\n",
1023 ctx, ctx->cc_vcred.vc_uid,
1024 cfs_atomic_read(&ctx->cc_refcount),
1026 ctx->cc_expire ? ctx->cc_expire - now : 0,
1028 cfs_atomic_read(&gctx->gc_seq),
1030 key ? key->serial : 0,
1031 key ? atomic_read(&key->usage) : 0,
1032 gss_handle_to_u64(&gctx->gc_handle),
1033 gss_handle_to_u64(&gctx->gc_svc_handle),
1036 spin_unlock(&sec->ps_lock);
1041 /****************************************
1043 ****************************************/
1046 int gss_cli_ctx_refresh_kr(struct ptlrpc_cli_ctx *ctx)
1048 /* upcall is already on the way */
1053 int gss_cli_ctx_validate_kr(struct ptlrpc_cli_ctx *ctx)
1055 LASSERT(cfs_atomic_read(&ctx->cc_refcount) > 0);
1056 LASSERT(ctx->cc_sec);
1058 if (cli_ctx_check_death(ctx)) {
1063 if (cli_ctx_is_ready(ctx))
1069 void gss_cli_ctx_die_kr(struct ptlrpc_cli_ctx *ctx, int grace)
1071 LASSERT(cfs_atomic_read(&ctx->cc_refcount) > 0);
1072 LASSERT(ctx->cc_sec);
1074 cli_ctx_expire(ctx);
1078 /****************************************
1079 * (reverse) service *
1080 ****************************************/
1083 * reverse context could have nothing to do with keyrings. here we still keep
1084 * the version which bind to a key, for future reference.
1086 #define HAVE_REVERSE_CTX_NOKEY
1088 #ifdef HAVE_REVERSE_CTX_NOKEY
1091 int sec_install_rctx_kr(struct ptlrpc_sec *sec,
1092 struct ptlrpc_svc_ctx *svc_ctx)
1094 struct ptlrpc_cli_ctx *cli_ctx;
1095 struct vfs_cred vcred = { 0, 0 };
1101 cli_ctx = ctx_create_kr(sec, &vcred);
1102 if (cli_ctx == NULL)
1105 rc = gss_copy_rvc_cli_ctx(cli_ctx, svc_ctx);
1107 CERROR("failed copy reverse cli ctx: %d\n", rc);
1109 ctx_put_kr(cli_ctx, 1);
1113 rvs_sec_install_root_ctx_kr(sec, cli_ctx, NULL);
1115 ctx_put_kr(cli_ctx, 1);
1120 #else /* ! HAVE_REVERSE_CTX_NOKEY */
1123 int sec_install_rctx_kr(struct ptlrpc_sec *sec,
1124 struct ptlrpc_svc_ctx *svc_ctx)
1126 struct ptlrpc_cli_ctx *cli_ctx = NULL;
1128 struct vfs_cred vcred = { 0, 0 };
1136 construct_key_desc(desc, sizeof(desc), sec, 0);
1138 key = key_alloc(&gss_key_type, desc, 0, 0,
1139 KEY_POS_ALL | KEY_USR_ALL, 1);
1141 CERROR("failed to alloc key: %ld\n", PTR_ERR(key));
1142 return PTR_ERR(key);
1145 rc = key_instantiate_and_link(key, NULL, 0, NULL, NULL);
1147 CERROR("failed to instantiate key: %d\n", rc);
1151 down_write(&key->sem);
1153 LASSERT(key->payload.data == NULL);
1155 cli_ctx = ctx_create_kr(sec, &vcred);
1156 if (cli_ctx == NULL) {
1161 rc = gss_copy_rvc_cli_ctx(cli_ctx, svc_ctx);
1163 CERROR("failed copy reverse cli ctx: %d\n", rc);
1167 rvs_sec_install_root_ctx_kr(sec, cli_ctx, key);
1169 ctx_put_kr(cli_ctx, 1);
1170 up_write(&key->sem);
1179 ctx_put_kr(cli_ctx, 1);
1181 up_write(&key->sem);
1187 #endif /* HAVE_REVERSE_CTX_NOKEY */
1189 /****************************************
1191 ****************************************/
1194 int gss_svc_accept_kr(struct ptlrpc_request *req)
1196 return gss_svc_accept(&gss_policy_keyring, req);
1200 int gss_svc_install_rctx_kr(struct obd_import *imp,
1201 struct ptlrpc_svc_ctx *svc_ctx)
1203 struct ptlrpc_sec *sec;
1206 sec = sptlrpc_import_sec_ref(imp);
1209 rc = sec_install_rctx_kr(sec, svc_ctx);
1210 sptlrpc_sec_put(sec);
1215 /****************************************
1217 ****************************************/
1220 int gss_kt_instantiate(struct key *key, const void *data, size_t datalen)
1225 if (data != NULL || datalen != 0) {
1226 CERROR("invalid: data %p, len %lu\n", data, (long)datalen);
1230 if (key->payload.data != 0) {
1231 CERROR("key already have payload\n");
1235 /* link the key to session keyring, so following context negotiation
1236 * rpc fired from user space could find this key. This will be unlinked
1237 * automatically when upcall processes die.
1239 * we can't do this through keyctl from userspace, because the upcall
1240 * might be neither possessor nor owner of the key (setuid).
1242 * the session keyring is created upon upcall, and don't change all
1243 * the way until upcall finished, so rcu lock is not needed here.
1245 LASSERT(cfs_current()->signal->session_keyring);
1248 rc = key_link(cfs_current()->signal->session_keyring, key);
1251 CERROR("failed to link key %08x to keyring %08x: %d\n",
1253 cfs_current()->signal->session_keyring->serial, rc);
1257 CDEBUG(D_SEC, "key %p instantiated, ctx %p\n", key, key->payload.data);
1262 * called with key semaphore write locked. it means we can operate
1263 * on the context without fear of loosing refcount.
1266 int gss_kt_update(struct key *key, const void *data, size_t datalen)
1268 struct ptlrpc_cli_ctx *ctx = key->payload.data;
1269 struct gss_cli_ctx *gctx;
1270 rawobj_t tmpobj = RAWOBJ_EMPTY;
1271 __u32 datalen32 = (__u32) datalen;
1275 if (data == NULL || datalen == 0) {
1276 CWARN("invalid: data %p, len %lu\n", data, (long)datalen);
1280 /* if upcall finished negotiation too fast (mostly likely because
1281 * of local error happened) and call kt_update(), the ctx
1282 * might be still NULL. but the key will finally be associate
1283 * with a context, or be revoked. if key status is fine, return
1284 * -EAGAIN to allow userspace sleep a while and call again. */
1286 CDEBUG(D_SEC, "update too soon: key %p(%x) flags %lx\n",
1287 key, key->serial, key->flags);
1289 rc = key_validate(key);
1296 LASSERT(cfs_atomic_read(&ctx->cc_refcount) > 0);
1297 LASSERT(ctx->cc_sec);
1299 ctx_clear_timer_kr(ctx);
1301 /* don't proceed if already refreshed */
1302 if (cli_ctx_is_refreshed(ctx)) {
1303 CWARN("ctx already done refresh\n");
1307 sptlrpc_cli_ctx_get(ctx);
1308 gctx = ctx2gctx(ctx);
1310 rc = buffer_extract_bytes(&data, &datalen32, &gctx->gc_win,
1311 sizeof(gctx->gc_win));
1313 CERROR("failed extract seq_win\n");
1317 if (gctx->gc_win == 0) {
1318 __u32 nego_rpc_err, nego_gss_err;
1320 rc = buffer_extract_bytes(&data, &datalen32, &nego_rpc_err,
1321 sizeof(nego_rpc_err));
1323 CERROR("failed to extrace rpc rc\n");
1327 rc = buffer_extract_bytes(&data, &datalen32, &nego_gss_err,
1328 sizeof(nego_gss_err));
1330 CERROR("failed to extrace gss rc\n");
1334 CERROR("negotiation: rpc err %d, gss err %x\n",
1335 nego_rpc_err, nego_gss_err);
1337 rc = nego_rpc_err ? nego_rpc_err : -EACCES;
1339 rc = rawobj_extract_local_alloc(&gctx->gc_handle,
1340 (__u32 **) &data, &datalen32);
1342 CERROR("failed extract handle\n");
1346 rc = rawobj_extract_local(&tmpobj, (__u32 **) &data,&datalen32);
1348 CERROR("failed extract mech\n");
1352 rc = lgss_import_sec_context(&tmpobj,
1353 sec2gsec(ctx->cc_sec)->gs_mech,
1355 if (rc != GSS_S_COMPLETE)
1356 CERROR("failed import context\n");
1361 /* we don't care what current status of this ctx, even someone else
1362 * is operating on the ctx at the same time. we just add up our own
1365 gss_cli_ctx_uptodate(gctx);
1367 /* this will also revoke the key. has to be done before
1368 * wakeup waiters otherwise they can find the stale key */
1369 kill_key_locked(key);
1371 cli_ctx_expire(ctx);
1373 if (rc != -ERESTART)
1374 set_bit(PTLRPC_CTX_ERROR_BIT, &ctx->cc_flags);
1377 /* let user space think it's a success */
1378 sptlrpc_cli_ctx_put(ctx, 1);
1383 int gss_kt_match(const struct key *key, const void *desc)
1385 return (strcmp(key->description, (const char *) desc) == 0);
1389 void gss_kt_destroy(struct key *key)
1392 LASSERT(key->payload.data == NULL);
1393 CDEBUG(D_SEC, "destroy key %p\n", key);
1398 void gss_kt_describe(const struct key *key, struct seq_file *s)
1400 if (key->description == NULL)
1401 seq_puts(s, "[null]");
1403 seq_puts(s, key->description);
1406 static struct key_type gss_key_type =
1410 .instantiate = gss_kt_instantiate,
1411 .update = gss_kt_update,
1412 .match = gss_kt_match,
1413 .destroy = gss_kt_destroy,
1414 .describe = gss_kt_describe,
1417 /****************************************
1418 * lustre gss keyring policy *
1419 ****************************************/
1421 static struct ptlrpc_ctx_ops gss_keyring_ctxops = {
1422 .match = gss_cli_ctx_match,
1423 .refresh = gss_cli_ctx_refresh_kr,
1424 .validate = gss_cli_ctx_validate_kr,
1425 .die = gss_cli_ctx_die_kr,
1426 .sign = gss_cli_ctx_sign,
1427 .verify = gss_cli_ctx_verify,
1428 .seal = gss_cli_ctx_seal,
1429 .unseal = gss_cli_ctx_unseal,
1430 .wrap_bulk = gss_cli_ctx_wrap_bulk,
1431 .unwrap_bulk = gss_cli_ctx_unwrap_bulk,
1434 static struct ptlrpc_sec_cops gss_sec_keyring_cops = {
1435 .create_sec = gss_sec_create_kr,
1436 .destroy_sec = gss_sec_destroy_kr,
1437 .kill_sec = gss_sec_kill,
1438 .lookup_ctx = gss_sec_lookup_ctx_kr,
1439 .release_ctx = gss_sec_release_ctx_kr,
1440 .flush_ctx_cache = gss_sec_flush_ctx_cache_kr,
1441 .gc_ctx = gss_sec_gc_ctx_kr,
1442 .install_rctx = gss_sec_install_rctx,
1443 .alloc_reqbuf = gss_alloc_reqbuf,
1444 .free_reqbuf = gss_free_reqbuf,
1445 .alloc_repbuf = gss_alloc_repbuf,
1446 .free_repbuf = gss_free_repbuf,
1447 .enlarge_reqbuf = gss_enlarge_reqbuf,
1448 .display = gss_sec_display_kr,
1451 static struct ptlrpc_sec_sops gss_sec_keyring_sops = {
1452 .accept = gss_svc_accept_kr,
1453 .invalidate_ctx = gss_svc_invalidate_ctx,
1454 .alloc_rs = gss_svc_alloc_rs,
1455 .authorize = gss_svc_authorize,
1456 .free_rs = gss_svc_free_rs,
1457 .free_ctx = gss_svc_free_ctx,
1458 .prep_bulk = gss_svc_prep_bulk,
1459 .unwrap_bulk = gss_svc_unwrap_bulk,
1460 .wrap_bulk = gss_svc_wrap_bulk,
1461 .install_rctx = gss_svc_install_rctx_kr,
1464 static struct ptlrpc_sec_policy gss_policy_keyring = {
1465 .sp_owner = THIS_MODULE,
1466 .sp_name = "gss.keyring",
1467 .sp_policy = SPTLRPC_POLICY_GSS,
1468 .sp_cops = &gss_sec_keyring_cops,
1469 .sp_sops = &gss_sec_keyring_sops,
1473 int __init gss_init_keyring(void)
1477 rc = register_key_type(&gss_key_type);
1479 CERROR("failed to register keyring type: %d\n", rc);
1483 rc = sptlrpc_register_policy(&gss_policy_keyring);
1485 unregister_key_type(&gss_key_type);
1492 void __exit gss_exit_keyring(void)
1494 unregister_key_type(&gss_key_type);
1495 sptlrpc_unregister_policy(&gss_policy_keyring);