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/key-type.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 key_cred(tsk)->thread_keyring ? \
102 key_cred(tsk)->thread_keyring->serial : 0, \
103 key_tgcred(tsk)->process_keyring ? \
104 key_tgcred(tsk)->process_keyring->serial : 0, \
105 key_tgcred(tsk)->session_keyring ? \
106 key_tgcred(tsk)->session_keyring->serial : 0, \
107 key_cred(tsk)->user->uid_keyring ? \
108 key_cred(tsk)->user->uid_keyring->serial : 0, \
109 key_cred(tsk)->user->session_keyring ? \
110 key_cred(tsk)->user->session_keyring->serial : 0, \
111 key_cred(tsk)->jit_keyring \
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" \
124 #define key_cred(tsk) ((tsk)->cred)
125 #define key_tgcred(tsk) ((tsk)->cred->tgcred)
127 static inline void keyring_upcall_lock(struct gss_sec_keyring *gsec_kr)
129 #ifdef HAVE_KEYRING_UPCALL_SERIALIZED
130 mutex_lock(&gsec_kr->gsk_uc_lock);
134 static inline void keyring_upcall_unlock(struct gss_sec_keyring *gsec_kr)
136 #ifdef HAVE_KEYRING_UPCALL_SERIALIZED
137 mutex_unlock(&gsec_kr->gsk_uc_lock);
141 static inline void key_revoke_locked(struct key *key)
143 set_bit(KEY_FLAG_REVOKED, &key->flags);
146 static void ctx_upcall_timeout_kr(unsigned long data)
148 struct ptlrpc_cli_ctx *ctx = (struct ptlrpc_cli_ctx *) data;
149 struct key *key = ctx2gctx_keyring(ctx)->gck_key;
151 CWARN("ctx %p, key %p\n", ctx, key);
156 key_revoke_locked(key);
159 static void ctx_start_timer_kr(struct ptlrpc_cli_ctx *ctx, long timeout)
161 struct gss_cli_ctx_keyring *gctx_kr = ctx2gctx_keyring(ctx);
162 struct timer_list *timer = gctx_kr->gck_timer;
166 CDEBUG(D_SEC, "ctx %p: start timer %lds\n", ctx, timeout);
167 timeout = timeout * HZ + cfs_time_current();
170 timer->expires = timeout;
171 timer->data = (unsigned long ) ctx;
172 timer->function = ctx_upcall_timeout_kr;
178 * caller should make sure no race with other threads
181 void ctx_clear_timer_kr(struct ptlrpc_cli_ctx *ctx)
183 struct gss_cli_ctx_keyring *gctx_kr = ctx2gctx_keyring(ctx);
184 struct timer_list *timer = gctx_kr->gck_timer;
189 CDEBUG(D_SEC, "ctx %p, key %p\n", ctx, gctx_kr->gck_key);
191 gctx_kr->gck_timer = NULL;
193 del_singleshot_timer_sync(timer);
199 struct ptlrpc_cli_ctx *ctx_create_kr(struct ptlrpc_sec *sec,
200 struct vfs_cred *vcred)
202 struct ptlrpc_cli_ctx *ctx;
203 struct gss_cli_ctx_keyring *gctx_kr;
205 OBD_ALLOC_PTR(gctx_kr);
209 OBD_ALLOC_PTR(gctx_kr->gck_timer);
210 if (gctx_kr->gck_timer == NULL) {
211 OBD_FREE_PTR(gctx_kr);
214 init_timer(gctx_kr->gck_timer);
216 ctx = &gctx_kr->gck_base.gc_base;
218 if (gss_cli_ctx_init_common(sec, ctx, &gss_keyring_ctxops, vcred)) {
219 OBD_FREE_PTR(gctx_kr->gck_timer);
220 OBD_FREE_PTR(gctx_kr);
224 ctx->cc_expire = cfs_time_current_sec() + KEYRING_UPCALL_TIMEOUT;
225 clear_bit(PTLRPC_CTX_NEW_BIT, &ctx->cc_flags);
226 cfs_atomic_inc(&ctx->cc_refcount); /* for the caller */
231 static void ctx_destroy_kr(struct ptlrpc_cli_ctx *ctx)
233 struct ptlrpc_sec *sec = ctx->cc_sec;
234 struct gss_cli_ctx_keyring *gctx_kr = ctx2gctx_keyring(ctx);
236 CDEBUG(D_SEC, "destroying ctx %p\n", ctx);
238 /* at this time the association with key has been broken. */
240 LASSERT(cfs_atomic_read(&sec->ps_refcount) > 0);
241 LASSERT(cfs_atomic_read(&sec->ps_nctx) > 0);
242 LASSERT(test_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags) == 0);
243 LASSERT(gctx_kr->gck_key == NULL);
245 ctx_clear_timer_kr(ctx);
246 LASSERT(gctx_kr->gck_timer == NULL);
248 if (gss_cli_ctx_fini_common(sec, ctx))
251 OBD_FREE_PTR(gctx_kr);
253 cfs_atomic_dec(&sec->ps_nctx);
254 sptlrpc_sec_put(sec);
257 static void ctx_release_kr(struct ptlrpc_cli_ctx *ctx, int sync)
262 cfs_atomic_inc(&ctx->cc_refcount);
263 sptlrpc_gc_add_ctx(ctx);
267 static void ctx_put_kr(struct ptlrpc_cli_ctx *ctx, int sync)
269 LASSERT(cfs_atomic_read(&ctx->cc_refcount) > 0);
271 if (cfs_atomic_dec_and_test(&ctx->cc_refcount))
272 ctx_release_kr(ctx, sync);
276 * key <-> ctx association and rules:
277 * - ctx might not bind with any key
278 * - key/ctx binding is protected by key semaphore (if the key present)
279 * - key and ctx each take a reference of the other
280 * - ctx enlist/unlist is protected by ctx spinlock
281 * - never enlist a ctx after it's been unlisted
282 * - whoever do enlist should also do bind, lock key before enlist:
283 * - lock key -> lock ctx -> enlist -> unlock ctx -> bind -> unlock key
284 * - whoever do unlist should also do unbind:
285 * - lock key -> lock ctx -> unlist -> unlock ctx -> unbind -> unlock key
286 * - lock ctx -> unlist -> unlock ctx -> lock key -> unbind -> unlock key
289 static inline void spin_lock_if(spinlock_t *lock, int condition)
295 static inline void spin_unlock_if(spinlock_t *lock, int condition)
301 static void ctx_enlist_kr(struct ptlrpc_cli_ctx *ctx, int is_root, int locked)
303 struct ptlrpc_sec *sec = ctx->cc_sec;
304 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
306 LASSERT(!test_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags));
307 LASSERT(cfs_atomic_read(&ctx->cc_refcount) > 0);
309 spin_lock_if(&sec->ps_lock, !locked);
311 cfs_atomic_inc(&ctx->cc_refcount);
312 set_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags);
313 cfs_hlist_add_head(&ctx->cc_cache, &gsec_kr->gsk_clist);
315 gsec_kr->gsk_root_ctx = ctx;
317 spin_unlock_if(&sec->ps_lock, !locked);
321 * Note after this get called, caller should not access ctx again because
322 * it might have been freed, unless caller hold at least one refcount of
325 * return non-zero if we indeed unlist this ctx.
327 static int ctx_unlist_kr(struct ptlrpc_cli_ctx *ctx, int locked)
329 struct ptlrpc_sec *sec = ctx->cc_sec;
330 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
332 /* if hashed bit has gone, leave the job to somebody who is doing it */
333 if (test_and_clear_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags) == 0)
336 /* drop ref inside spin lock to prevent race with other operations */
337 spin_lock_if(&sec->ps_lock, !locked);
339 if (gsec_kr->gsk_root_ctx == ctx)
340 gsec_kr->gsk_root_ctx = NULL;
341 cfs_hlist_del_init(&ctx->cc_cache);
342 cfs_atomic_dec(&ctx->cc_refcount);
344 spin_unlock_if(&sec->ps_lock, !locked);
350 * bind a key with a ctx together.
351 * caller must hold write lock of the key, as well as ref on key & ctx.
353 static void bind_key_ctx(struct key *key, struct ptlrpc_cli_ctx *ctx)
355 LASSERT(cfs_atomic_read(&ctx->cc_refcount) > 0);
356 LASSERT(atomic_read(&key->usage) > 0);
357 LASSERT(ctx2gctx_keyring(ctx)->gck_key == NULL);
358 LASSERT(key->payload.data == NULL);
360 /* at this time context may or may not in list. */
362 cfs_atomic_inc(&ctx->cc_refcount);
363 ctx2gctx_keyring(ctx)->gck_key = key;
364 key->payload.data = ctx;
368 * unbind a key and a ctx.
369 * caller must hold write lock, as well as a ref of the key.
371 static void unbind_key_ctx(struct key *key, struct ptlrpc_cli_ctx *ctx)
373 LASSERT(key->payload.data == ctx);
374 LASSERT(test_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags) == 0);
376 /* must revoke the key, or others may treat it as newly created */
377 key_revoke_locked(key);
379 key->payload.data = NULL;
380 ctx2gctx_keyring(ctx)->gck_key = NULL;
382 /* once ctx get split from key, the timer is meaningless */
383 ctx_clear_timer_kr(ctx);
390 * given a ctx, unbind with its coupled key, if any.
391 * unbind could only be called once, so we don't worry the key be released
394 static void unbind_ctx_kr(struct ptlrpc_cli_ctx *ctx)
396 struct key *key = ctx2gctx_keyring(ctx)->gck_key;
399 LASSERT(key->payload.data == ctx);
402 down_write(&key->sem);
403 unbind_key_ctx(key, ctx);
410 * given a key, unbind with its coupled ctx, if any.
411 * caller must hold write lock, as well as a ref of the key.
413 static void unbind_key_locked(struct key *key)
415 struct ptlrpc_cli_ctx *ctx = key->payload.data;
418 unbind_key_ctx(key, ctx);
422 * unlist a ctx, and unbind from coupled key
424 static void kill_ctx_kr(struct ptlrpc_cli_ctx *ctx)
426 if (ctx_unlist_kr(ctx, 0))
431 * given a key, unlist and unbind with the coupled ctx (if any).
432 * caller must hold write lock, as well as a ref of the key.
434 static void kill_key_locked(struct key *key)
436 struct ptlrpc_cli_ctx *ctx = key->payload.data;
438 if (ctx && ctx_unlist_kr(ctx, 0))
439 unbind_key_locked(key);
443 * caller should hold one ref on contexts in freelist.
445 static void dispose_ctx_list_kr(cfs_hlist_head_t *freelist)
447 cfs_hlist_node_t *pos, *next;
448 struct ptlrpc_cli_ctx *ctx;
449 struct gss_cli_ctx *gctx;
451 cfs_hlist_for_each_entry_safe(ctx, pos, next, freelist, cc_cache) {
452 cfs_hlist_del_init(&ctx->cc_cache);
454 /* reverse ctx: update current seq to buddy svcctx if exist.
455 * ideally this should be done at gss_cli_ctx_finalize(), but
456 * the ctx destroy could be delayed by:
457 * 1) ctx still has reference;
458 * 2) ctx destroy is asynchronous;
459 * and reverse import call inval_all_ctx() require this be done
460 *_immediately_ otherwise newly created reverse ctx might copy
461 * the very old sequence number from svcctx. */
462 gctx = ctx2gctx(ctx);
463 if (!rawobj_empty(&gctx->gc_svc_handle) &&
464 sec_is_reverse(gctx->gc_base.cc_sec)) {
465 gss_svc_upcall_update_sequence(&gctx->gc_svc_handle,
466 (__u32) cfs_atomic_read(&gctx->gc_seq));
469 /* we need to wakeup waiting reqs here. the context might
470 * be forced released before upcall finished, then the
471 * late-arrived downcall can't find the ctx even. */
472 sptlrpc_cli_ctx_wakeup(ctx);
480 * lookup a root context directly in a sec, return root ctx with a
481 * reference taken or NULL.
484 struct ptlrpc_cli_ctx * sec_lookup_root_ctx_kr(struct ptlrpc_sec *sec)
486 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
487 struct ptlrpc_cli_ctx *ctx = NULL;
489 spin_lock(&sec->ps_lock);
491 ctx = gsec_kr->gsk_root_ctx;
493 if (ctx == NULL && unlikely(sec_is_reverse(sec))) {
494 cfs_hlist_node_t *node;
495 struct ptlrpc_cli_ctx *tmp;
497 /* reverse ctx, search root ctx in list, choose the one
498 * with shortest expire time, which is most possibly have
499 * an established peer ctx at client side. */
500 cfs_hlist_for_each_entry(tmp, node, &gsec_kr->gsk_clist,
502 if (ctx == NULL || ctx->cc_expire == 0 ||
503 ctx->cc_expire > tmp->cc_expire) {
505 /* promote to be root_ctx */
506 gsec_kr->gsk_root_ctx = ctx;
512 LASSERT(cfs_atomic_read(&ctx->cc_refcount) > 0);
513 LASSERT(!cfs_hlist_empty(&gsec_kr->gsk_clist));
514 cfs_atomic_inc(&ctx->cc_refcount);
517 spin_unlock(&sec->ps_lock);
522 #define RVS_CTX_EXPIRE_NICE (10)
525 void rvs_sec_install_root_ctx_kr(struct ptlrpc_sec *sec,
526 struct ptlrpc_cli_ctx *new_ctx,
529 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
530 cfs_hlist_node_t *hnode;
531 struct ptlrpc_cli_ctx *ctx;
535 LASSERT(sec_is_reverse(sec));
537 spin_lock(&sec->ps_lock);
539 now = cfs_time_current_sec();
541 /* set all existing ctxs short expiry */
542 cfs_hlist_for_each_entry(ctx, hnode, &gsec_kr->gsk_clist, cc_cache) {
543 if (ctx->cc_expire > now + RVS_CTX_EXPIRE_NICE) {
544 ctx->cc_early_expire = 1;
545 ctx->cc_expire = now + RVS_CTX_EXPIRE_NICE;
549 /* if there's root_ctx there, instead obsolete the current
550 * immediately, we leave it continue operating for a little while.
551 * hopefully when the first backward rpc with newest ctx send out,
552 * the client side already have the peer ctx well established. */
553 ctx_enlist_kr(new_ctx, gsec_kr->gsk_root_ctx ? 0 : 1, 1);
556 bind_key_ctx(key, new_ctx);
558 spin_unlock(&sec->ps_lock);
561 static void construct_key_desc(void *buf, int bufsize,
562 struct ptlrpc_sec *sec, uid_t uid)
564 snprintf(buf, bufsize, "%d@%x", uid, sec->ps_id);
565 ((char *)buf)[bufsize - 1] = '\0';
568 /****************************************
570 ****************************************/
573 struct ptlrpc_sec * gss_sec_create_kr(struct obd_import *imp,
574 struct ptlrpc_svc_ctx *svcctx,
575 struct sptlrpc_flavor *sf)
577 struct gss_sec_keyring *gsec_kr;
580 OBD_ALLOC(gsec_kr, sizeof(*gsec_kr));
584 CFS_INIT_HLIST_HEAD(&gsec_kr->gsk_clist);
585 gsec_kr->gsk_root_ctx = NULL;
586 mutex_init(&gsec_kr->gsk_root_uc_lock);
587 #ifdef HAVE_KEYRING_UPCALL_SERIALIZED
588 mutex_init(&gsec_kr->gsk_uc_lock);
591 if (gss_sec_create_common(&gsec_kr->gsk_base, &gss_policy_keyring,
595 if (svcctx != NULL &&
596 sec_install_rctx_kr(&gsec_kr->gsk_base.gs_base, svcctx)) {
597 gss_sec_destroy_common(&gsec_kr->gsk_base);
601 RETURN(&gsec_kr->gsk_base.gs_base);
604 OBD_FREE(gsec_kr, sizeof(*gsec_kr));
609 void gss_sec_destroy_kr(struct ptlrpc_sec *sec)
611 struct gss_sec *gsec = sec2gsec(sec);
612 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
614 CDEBUG(D_SEC, "destroy %s@%p\n", sec->ps_policy->sp_name, sec);
616 LASSERT(cfs_hlist_empty(&gsec_kr->gsk_clist));
617 LASSERT(gsec_kr->gsk_root_ctx == NULL);
619 gss_sec_destroy_common(gsec);
621 OBD_FREE(gsec_kr, sizeof(*gsec_kr));
624 static inline int user_is_root(struct ptlrpc_sec *sec, struct vfs_cred *vcred)
626 /* except the ROOTONLY flag, treat it as root user only if real uid
627 * is 0, euid/fsuid being 0 are handled as setuid scenarios */
628 if (sec_is_rootonly(sec) || (vcred->vc_uid == 0))
635 * unlink request key from it's ring, which is linked during request_key().
636 * sadly, we have to 'guess' which keyring it's linked to.
638 * FIXME this code is fragile, depend on how request_key_link() is implemented.
640 static void request_key_unlink(struct key *key)
642 struct task_struct *tsk = current;
645 switch (key_cred(tsk)->jit_keyring) {
646 case KEY_REQKEY_DEFL_DEFAULT:
647 case KEY_REQKEY_DEFL_THREAD_KEYRING:
648 ring = key_get(key_cred(tsk)->thread_keyring);
651 case KEY_REQKEY_DEFL_PROCESS_KEYRING:
652 ring = key_get(key_tgcred(tsk)->process_keyring);
655 case KEY_REQKEY_DEFL_SESSION_KEYRING:
657 ring = key_get(rcu_dereference(key_tgcred(tsk)
662 case KEY_REQKEY_DEFL_USER_SESSION_KEYRING:
663 ring = key_get(key_cred(tsk)->user->session_keyring);
665 case KEY_REQKEY_DEFL_USER_KEYRING:
666 ring = key_get(key_cred(tsk)->user->uid_keyring);
668 case KEY_REQKEY_DEFL_GROUP_KEYRING:
674 key_unlink(ring, key);
679 struct ptlrpc_cli_ctx * gss_sec_lookup_ctx_kr(struct ptlrpc_sec *sec,
680 struct vfs_cred *vcred,
681 int create, int remove_dead)
683 struct obd_import *imp = sec->ps_import;
684 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
685 struct ptlrpc_cli_ctx *ctx = NULL;
686 unsigned int is_root = 0, create_new = 0;
694 LASSERT(imp != NULL);
696 is_root = user_is_root(sec, vcred);
698 /* a little bit optimization for root context */
700 ctx = sec_lookup_root_ctx_kr(sec);
702 * Only lookup directly for REVERSE sec, which should
705 if (ctx || sec_is_reverse(sec))
709 LASSERT(create != 0);
711 /* for root context, obtain lock and check again, this time hold
712 * the root upcall lock, make sure nobody else populated new root
713 * context after last check. */
715 mutex_lock(&gsec_kr->gsk_root_uc_lock);
717 ctx = sec_lookup_root_ctx_kr(sec);
721 /* update reverse handle for root user */
722 sec2gsec(sec)->gs_rvs_hdl = gss_get_next_ctx_index();
724 switch (sec->ps_part) {
743 /* in case of setuid, key will be constructed as owner of fsuid/fsgid,
744 * but we do authentication based on real uid/gid. the key permission
745 * bits will be exactly as POS_ALL, so only processes who subscribed
746 * this key could have the access, although the quota might be counted
747 * on others (fsuid/fsgid).
749 * keyring will use fsuid/fsgid as upcall parameters, so we have to
750 * encode real uid/gid into callout info.
753 construct_key_desc(desc, sizeof(desc), sec, vcred->vc_uid);
755 /* callout info format:
756 * secid:mech:uid:gid:flags:svc_type:peer_nid:target_uuid
758 coinfo_size = sizeof(struct obd_uuid) + MAX_OBD_NAME + 64;
759 OBD_ALLOC(coinfo, coinfo_size);
763 snprintf(coinfo, coinfo_size, "%d:%s:%u:%u:%s:%d:"LPX64":%s",
764 sec->ps_id, sec2gsec(sec)->gs_mech->gm_name,
765 vcred->vc_uid, vcred->vc_gid,
766 co_flags, import_to_gss_svc(imp),
767 imp->imp_connection->c_peer.nid, imp->imp_obd->obd_name);
769 CDEBUG(D_SEC, "requesting key for %s\n", desc);
771 keyring_upcall_lock(gsec_kr);
772 key = request_key(&gss_key_type, desc, coinfo);
773 keyring_upcall_unlock(gsec_kr);
775 OBD_FREE(coinfo, coinfo_size);
778 CERROR("failed request key: %ld\n", PTR_ERR(key));
781 CDEBUG(D_SEC, "obtained key %08x for %s\n", key->serial, desc);
783 /* once payload.data was pointed to a ctx, it never changes until
784 * we de-associate them; but parallel request_key() may return
785 * a key with payload.data == NULL at the same time. so we still
786 * need wirtelock of key->sem to serialize them. */
787 down_write(&key->sem);
789 if (likely(key->payload.data != NULL)) {
790 ctx = key->payload.data;
792 LASSERT(cfs_atomic_read(&ctx->cc_refcount) >= 1);
793 LASSERT(ctx2gctx_keyring(ctx)->gck_key == key);
794 LASSERT(atomic_read(&key->usage) >= 2);
796 /* simply take a ref and return. it's upper layer's
797 * responsibility to detect & replace dead ctx. */
798 cfs_atomic_inc(&ctx->cc_refcount);
800 /* pre initialization with a cli_ctx. this can't be done in
801 * key_instantiate() because we'v no enough information
803 ctx = ctx_create_kr(sec, vcred);
805 ctx_enlist_kr(ctx, is_root, 0);
806 bind_key_ctx(key, ctx);
808 ctx_start_timer_kr(ctx, KEYRING_UPCALL_TIMEOUT);
810 CDEBUG(D_SEC, "installed key %p <-> ctx %p (sec %p)\n",
813 /* we'd prefer to call key_revoke(), but we more like
814 * to revoke it within this key->sem locked period. */
815 key_revoke_locked(key);
823 if (is_root && create_new)
824 request_key_unlink(key);
829 mutex_unlock(&gsec_kr->gsk_root_uc_lock);
834 void gss_sec_release_ctx_kr(struct ptlrpc_sec *sec,
835 struct ptlrpc_cli_ctx *ctx,
838 LASSERT(cfs_atomic_read(&sec->ps_refcount) > 0);
839 LASSERT(cfs_atomic_read(&ctx->cc_refcount) == 0);
840 ctx_release_kr(ctx, sync);
844 * flush context of normal user, we must resort to keyring itself to find out
845 * contexts which belong to me.
847 * Note here we suppose only to flush _my_ context, the "uid" will
848 * be ignored in the search.
851 void flush_user_ctx_cache_kr(struct ptlrpc_sec *sec,
853 int grace, int force)
858 /* nothing to do for reverse or rootonly sec */
859 if (sec_is_reverse(sec) || sec_is_rootonly(sec))
862 construct_key_desc(desc, sizeof(desc), sec, uid);
864 /* there should be only one valid key, but we put it in the
865 * loop in case of any weird cases */
867 key = request_key(&gss_key_type, desc, NULL);
869 CDEBUG(D_SEC, "No more key found for current user\n");
873 down_write(&key->sem);
875 kill_key_locked(key);
877 /* kill_key_locked() should usually revoke the key, but we
878 * revoke it again to make sure, e.g. some case the key may
879 * not well coupled with a context. */
880 key_revoke_locked(key);
889 * flush context of root or all, we iterate through the list.
892 void flush_spec_ctx_cache_kr(struct ptlrpc_sec *sec,
894 int grace, int force)
896 struct gss_sec_keyring *gsec_kr;
897 cfs_hlist_head_t freelist = CFS_HLIST_HEAD_INIT;
898 cfs_hlist_node_t *pos, *next;
899 struct ptlrpc_cli_ctx *ctx;
902 gsec_kr = sec2gsec_keyring(sec);
904 spin_lock(&sec->ps_lock);
905 cfs_hlist_for_each_entry_safe(ctx, pos, next,
906 &gsec_kr->gsk_clist, cc_cache) {
907 LASSERT(cfs_atomic_read(&ctx->cc_refcount) > 0);
909 if (uid != -1 && uid != ctx->cc_vcred.vc_uid)
912 /* at this moment there's at least 2 base reference:
913 * key association and in-list. */
914 if (cfs_atomic_read(&ctx->cc_refcount) > 2) {
917 CWARN("flush busy ctx %p(%u->%s, extra ref %d)\n",
918 ctx, ctx->cc_vcred.vc_uid,
919 sec2target_str(ctx->cc_sec),
920 cfs_atomic_read(&ctx->cc_refcount) - 2);
923 set_bit(PTLRPC_CTX_DEAD_BIT, &ctx->cc_flags);
925 clear_bit(PTLRPC_CTX_UPTODATE_BIT, &ctx->cc_flags);
927 cfs_atomic_inc(&ctx->cc_refcount);
929 if (ctx_unlist_kr(ctx, 1)) {
930 cfs_hlist_add_head(&ctx->cc_cache, &freelist);
932 LASSERT(cfs_atomic_read(&ctx->cc_refcount) >= 2);
933 cfs_atomic_dec(&ctx->cc_refcount);
936 spin_unlock(&sec->ps_lock);
938 dispose_ctx_list_kr(&freelist);
943 int gss_sec_flush_ctx_cache_kr(struct ptlrpc_sec *sec,
944 uid_t uid, int grace, int force)
948 CDEBUG(D_SEC, "sec %p(%d, nctx %d), uid %d, grace %d, force %d\n",
949 sec, cfs_atomic_read(&sec->ps_refcount),
950 cfs_atomic_read(&sec->ps_nctx),
953 if (uid != -1 && uid != 0)
954 flush_user_ctx_cache_kr(sec, uid, grace, force);
956 flush_spec_ctx_cache_kr(sec, uid, grace, force);
962 void gss_sec_gc_ctx_kr(struct ptlrpc_sec *sec)
964 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
965 cfs_hlist_head_t freelist = CFS_HLIST_HEAD_INIT;
966 cfs_hlist_node_t *pos, *next;
967 struct ptlrpc_cli_ctx *ctx;
970 CWARN("running gc\n");
972 spin_lock(&sec->ps_lock);
973 cfs_hlist_for_each_entry_safe(ctx, pos, next,
974 &gsec_kr->gsk_clist, cc_cache) {
975 LASSERT(cfs_atomic_read(&ctx->cc_refcount) > 0);
977 cfs_atomic_inc(&ctx->cc_refcount);
979 if (cli_ctx_check_death(ctx) && ctx_unlist_kr(ctx, 1)) {
980 cfs_hlist_add_head(&ctx->cc_cache, &freelist);
981 CWARN("unhashed ctx %p\n", ctx);
983 LASSERT(cfs_atomic_read(&ctx->cc_refcount) >= 2);
984 cfs_atomic_dec(&ctx->cc_refcount);
987 spin_unlock(&sec->ps_lock);
989 dispose_ctx_list_kr(&freelist);
995 int gss_sec_display_kr(struct ptlrpc_sec *sec, struct seq_file *seq)
997 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
998 cfs_hlist_node_t *pos, *next;
999 struct ptlrpc_cli_ctx *ctx;
1000 struct gss_cli_ctx *gctx;
1001 time_t now = cfs_time_current_sec();
1004 spin_lock(&sec->ps_lock);
1005 cfs_hlist_for_each_entry_safe(ctx, pos, next,
1006 &gsec_kr->gsk_clist, cc_cache) {
1011 gctx = ctx2gctx(ctx);
1012 key = ctx2gctx_keyring(ctx)->gck_key;
1014 gss_cli_ctx_flags2str(ctx->cc_flags,
1015 flags_str, sizeof(flags_str));
1017 if (gctx->gc_mechctx)
1018 lgss_display(gctx->gc_mechctx, mech, sizeof(mech));
1020 snprintf(mech, sizeof(mech), "N/A");
1021 mech[sizeof(mech) - 1] = '\0';
1023 seq_printf(seq, "%p: uid %u, ref %d, expire %ld(%+ld), fl %s, "
1024 "seq %d, win %u, key %08x(ref %d), "
1025 "hdl "LPX64":"LPX64", mech: %s\n",
1026 ctx, ctx->cc_vcred.vc_uid,
1027 cfs_atomic_read(&ctx->cc_refcount),
1029 ctx->cc_expire ? ctx->cc_expire - now : 0,
1031 cfs_atomic_read(&gctx->gc_seq),
1033 key ? key->serial : 0,
1034 key ? atomic_read(&key->usage) : 0,
1035 gss_handle_to_u64(&gctx->gc_handle),
1036 gss_handle_to_u64(&gctx->gc_svc_handle),
1039 spin_unlock(&sec->ps_lock);
1044 /****************************************
1046 ****************************************/
1049 int gss_cli_ctx_refresh_kr(struct ptlrpc_cli_ctx *ctx)
1051 /* upcall is already on the way */
1056 int gss_cli_ctx_validate_kr(struct ptlrpc_cli_ctx *ctx)
1058 LASSERT(cfs_atomic_read(&ctx->cc_refcount) > 0);
1059 LASSERT(ctx->cc_sec);
1061 if (cli_ctx_check_death(ctx)) {
1066 if (cli_ctx_is_ready(ctx))
1072 void gss_cli_ctx_die_kr(struct ptlrpc_cli_ctx *ctx, int grace)
1074 LASSERT(cfs_atomic_read(&ctx->cc_refcount) > 0);
1075 LASSERT(ctx->cc_sec);
1077 cli_ctx_expire(ctx);
1081 /****************************************
1082 * (reverse) service *
1083 ****************************************/
1086 * reverse context could have nothing to do with keyrings. here we still keep
1087 * the version which bind to a key, for future reference.
1089 #define HAVE_REVERSE_CTX_NOKEY
1091 #ifdef HAVE_REVERSE_CTX_NOKEY
1094 int sec_install_rctx_kr(struct ptlrpc_sec *sec,
1095 struct ptlrpc_svc_ctx *svc_ctx)
1097 struct ptlrpc_cli_ctx *cli_ctx;
1098 struct vfs_cred vcred = { 0, 0 };
1104 cli_ctx = ctx_create_kr(sec, &vcred);
1105 if (cli_ctx == NULL)
1108 rc = gss_copy_rvc_cli_ctx(cli_ctx, svc_ctx);
1110 CERROR("failed copy reverse cli ctx: %d\n", rc);
1112 ctx_put_kr(cli_ctx, 1);
1116 rvs_sec_install_root_ctx_kr(sec, cli_ctx, NULL);
1118 ctx_put_kr(cli_ctx, 1);
1123 #else /* ! HAVE_REVERSE_CTX_NOKEY */
1126 int sec_install_rctx_kr(struct ptlrpc_sec *sec,
1127 struct ptlrpc_svc_ctx *svc_ctx)
1129 struct ptlrpc_cli_ctx *cli_ctx = NULL;
1131 struct vfs_cred vcred = { 0, 0 };
1139 construct_key_desc(desc, sizeof(desc), sec, 0);
1141 key = key_alloc(&gss_key_type, desc, 0, 0,
1142 KEY_POS_ALL | KEY_USR_ALL, 1);
1144 CERROR("failed to alloc key: %ld\n", PTR_ERR(key));
1145 return PTR_ERR(key);
1148 rc = key_instantiate_and_link(key, NULL, 0, NULL, NULL);
1150 CERROR("failed to instantiate key: %d\n", rc);
1154 down_write(&key->sem);
1156 LASSERT(key->payload.data == NULL);
1158 cli_ctx = ctx_create_kr(sec, &vcred);
1159 if (cli_ctx == NULL) {
1164 rc = gss_copy_rvc_cli_ctx(cli_ctx, svc_ctx);
1166 CERROR("failed copy reverse cli ctx: %d\n", rc);
1170 rvs_sec_install_root_ctx_kr(sec, cli_ctx, key);
1172 ctx_put_kr(cli_ctx, 1);
1173 up_write(&key->sem);
1182 ctx_put_kr(cli_ctx, 1);
1184 up_write(&key->sem);
1190 #endif /* HAVE_REVERSE_CTX_NOKEY */
1192 /****************************************
1194 ****************************************/
1197 int gss_svc_accept_kr(struct ptlrpc_request *req)
1199 return gss_svc_accept(&gss_policy_keyring, req);
1203 int gss_svc_install_rctx_kr(struct obd_import *imp,
1204 struct ptlrpc_svc_ctx *svc_ctx)
1206 struct ptlrpc_sec *sec;
1209 sec = sptlrpc_import_sec_ref(imp);
1212 rc = sec_install_rctx_kr(sec, svc_ctx);
1213 sptlrpc_sec_put(sec);
1218 /****************************************
1220 ****************************************/
1223 int gss_kt_instantiate(struct key *key, const void *data, size_t datalen)
1228 if (data != NULL || datalen != 0) {
1229 CERROR("invalid: data %p, len %lu\n", data, (long)datalen);
1233 if (key->payload.data != 0) {
1234 CERROR("key already have payload\n");
1238 /* link the key to session keyring, so following context negotiation
1239 * rpc fired from user space could find this key. This will be unlinked
1240 * automatically when upcall processes die.
1242 * we can't do this through keyctl from userspace, because the upcall
1243 * might be neither possessor nor owner of the key (setuid).
1245 * the session keyring is created upon upcall, and don't change all
1246 * the way until upcall finished, so rcu lock is not needed here.
1248 LASSERT(key_tgcred(current)->session_keyring);
1251 rc = key_link(key_tgcred(current)->session_keyring, key);
1254 CERROR("failed to link key %08x to keyring %08x: %d\n",
1256 key_tgcred(current)->session_keyring->serial, rc);
1260 CDEBUG(D_SEC, "key %p instantiated, ctx %p\n", key, key->payload.data);
1265 * called with key semaphore write locked. it means we can operate
1266 * on the context without fear of loosing refcount.
1269 int gss_kt_update(struct key *key, const void *data, size_t datalen)
1271 struct ptlrpc_cli_ctx *ctx = key->payload.data;
1272 struct gss_cli_ctx *gctx;
1273 rawobj_t tmpobj = RAWOBJ_EMPTY;
1274 __u32 datalen32 = (__u32) datalen;
1278 if (data == NULL || datalen == 0) {
1279 CWARN("invalid: data %p, len %lu\n", data, (long)datalen);
1283 /* if upcall finished negotiation too fast (mostly likely because
1284 * of local error happened) and call kt_update(), the ctx
1285 * might be still NULL. but the key will finally be associate
1286 * with a context, or be revoked. if key status is fine, return
1287 * -EAGAIN to allow userspace sleep a while and call again. */
1289 CDEBUG(D_SEC, "update too soon: key %p(%x) flags %lx\n",
1290 key, key->serial, key->flags);
1292 rc = key_validate(key);
1299 LASSERT(cfs_atomic_read(&ctx->cc_refcount) > 0);
1300 LASSERT(ctx->cc_sec);
1302 ctx_clear_timer_kr(ctx);
1304 /* don't proceed if already refreshed */
1305 if (cli_ctx_is_refreshed(ctx)) {
1306 CWARN("ctx already done refresh\n");
1310 sptlrpc_cli_ctx_get(ctx);
1311 gctx = ctx2gctx(ctx);
1313 rc = buffer_extract_bytes(&data, &datalen32, &gctx->gc_win,
1314 sizeof(gctx->gc_win));
1316 CERROR("failed extract seq_win\n");
1320 if (gctx->gc_win == 0) {
1321 __u32 nego_rpc_err, nego_gss_err;
1323 rc = buffer_extract_bytes(&data, &datalen32, &nego_rpc_err,
1324 sizeof(nego_rpc_err));
1326 CERROR("failed to extrace rpc rc\n");
1330 rc = buffer_extract_bytes(&data, &datalen32, &nego_gss_err,
1331 sizeof(nego_gss_err));
1333 CERROR("failed to extrace gss rc\n");
1337 CERROR("negotiation: rpc err %d, gss err %x\n",
1338 nego_rpc_err, nego_gss_err);
1340 rc = nego_rpc_err ? nego_rpc_err : -EACCES;
1342 rc = rawobj_extract_local_alloc(&gctx->gc_handle,
1343 (__u32 **) &data, &datalen32);
1345 CERROR("failed extract handle\n");
1349 rc = rawobj_extract_local(&tmpobj, (__u32 **) &data,&datalen32);
1351 CERROR("failed extract mech\n");
1355 rc = lgss_import_sec_context(&tmpobj,
1356 sec2gsec(ctx->cc_sec)->gs_mech,
1358 if (rc != GSS_S_COMPLETE)
1359 CERROR("failed import context\n");
1364 /* we don't care what current status of this ctx, even someone else
1365 * is operating on the ctx at the same time. we just add up our own
1368 gss_cli_ctx_uptodate(gctx);
1370 /* this will also revoke the key. has to be done before
1371 * wakeup waiters otherwise they can find the stale key */
1372 kill_key_locked(key);
1374 cli_ctx_expire(ctx);
1376 if (rc != -ERESTART)
1377 set_bit(PTLRPC_CTX_ERROR_BIT, &ctx->cc_flags);
1380 /* let user space think it's a success */
1381 sptlrpc_cli_ctx_put(ctx, 1);
1386 int gss_kt_match(const struct key *key, const void *desc)
1388 return (strcmp(key->description, (const char *) desc) == 0);
1392 void gss_kt_destroy(struct key *key)
1395 LASSERT(key->payload.data == NULL);
1396 CDEBUG(D_SEC, "destroy key %p\n", key);
1401 void gss_kt_describe(const struct key *key, struct seq_file *s)
1403 if (key->description == NULL)
1404 seq_puts(s, "[null]");
1406 seq_puts(s, key->description);
1409 static struct key_type gss_key_type =
1413 .instantiate = gss_kt_instantiate,
1414 .update = gss_kt_update,
1415 .match = gss_kt_match,
1416 .destroy = gss_kt_destroy,
1417 .describe = gss_kt_describe,
1420 /****************************************
1421 * lustre gss keyring policy *
1422 ****************************************/
1424 static struct ptlrpc_ctx_ops gss_keyring_ctxops = {
1425 .match = gss_cli_ctx_match,
1426 .refresh = gss_cli_ctx_refresh_kr,
1427 .validate = gss_cli_ctx_validate_kr,
1428 .die = gss_cli_ctx_die_kr,
1429 .sign = gss_cli_ctx_sign,
1430 .verify = gss_cli_ctx_verify,
1431 .seal = gss_cli_ctx_seal,
1432 .unseal = gss_cli_ctx_unseal,
1433 .wrap_bulk = gss_cli_ctx_wrap_bulk,
1434 .unwrap_bulk = gss_cli_ctx_unwrap_bulk,
1437 static struct ptlrpc_sec_cops gss_sec_keyring_cops = {
1438 .create_sec = gss_sec_create_kr,
1439 .destroy_sec = gss_sec_destroy_kr,
1440 .kill_sec = gss_sec_kill,
1441 .lookup_ctx = gss_sec_lookup_ctx_kr,
1442 .release_ctx = gss_sec_release_ctx_kr,
1443 .flush_ctx_cache = gss_sec_flush_ctx_cache_kr,
1444 .gc_ctx = gss_sec_gc_ctx_kr,
1445 .install_rctx = gss_sec_install_rctx,
1446 .alloc_reqbuf = gss_alloc_reqbuf,
1447 .free_reqbuf = gss_free_reqbuf,
1448 .alloc_repbuf = gss_alloc_repbuf,
1449 .free_repbuf = gss_free_repbuf,
1450 .enlarge_reqbuf = gss_enlarge_reqbuf,
1451 .display = gss_sec_display_kr,
1454 static struct ptlrpc_sec_sops gss_sec_keyring_sops = {
1455 .accept = gss_svc_accept_kr,
1456 .invalidate_ctx = gss_svc_invalidate_ctx,
1457 .alloc_rs = gss_svc_alloc_rs,
1458 .authorize = gss_svc_authorize,
1459 .free_rs = gss_svc_free_rs,
1460 .free_ctx = gss_svc_free_ctx,
1461 .prep_bulk = gss_svc_prep_bulk,
1462 .unwrap_bulk = gss_svc_unwrap_bulk,
1463 .wrap_bulk = gss_svc_wrap_bulk,
1464 .install_rctx = gss_svc_install_rctx_kr,
1467 static struct ptlrpc_sec_policy gss_policy_keyring = {
1468 .sp_owner = THIS_MODULE,
1469 .sp_name = "gss.keyring",
1470 .sp_policy = SPTLRPC_POLICY_GSS,
1471 .sp_cops = &gss_sec_keyring_cops,
1472 .sp_sops = &gss_sec_keyring_sops,
1476 int __init gss_init_keyring(void)
1480 rc = register_key_type(&gss_key_type);
1482 CERROR("failed to register keyring type: %d\n", rc);
1486 rc = sptlrpc_register_policy(&gss_policy_keyring);
1488 unregister_key_type(&gss_key_type);
1495 void __exit gss_exit_keyring(void)
1497 unregister_key_type(&gss_key_type);
1498 sptlrpc_unregister_policy(&gss_policy_keyring);