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 #ifdef HAVE_LINUX_KEYTYPE_H
52 #include <linux/key-type.h>
54 #include <linux/mutex.h>
55 #include <asm/atomic.h>
57 #include <liblustre.h>
61 #include <obd_class.h>
62 #include <obd_support.h>
63 #include <lustre/lustre_idl.h>
64 #include <lustre_sec.h>
65 #include <lustre_net.h>
66 #include <lustre_import.h>
69 #include "gss_internal.h"
72 static struct ptlrpc_sec_policy gss_policy_keyring;
73 static struct ptlrpc_ctx_ops gss_keyring_ctxops;
74 static struct key_type gss_key_type;
76 static int sec_install_rctx_kr(struct ptlrpc_sec *sec,
77 struct ptlrpc_svc_ctx *svc_ctx);
80 * the timeout is only for the case that upcall child process die abnormally.
81 * in any other cases it should finally update kernel key.
83 * FIXME we'd better to incorporate the client & server side upcall timeouts
84 * into the framework of Adaptive Timeouts, but we need to figure out how to
85 * make sure that kernel knows the upcall processes is in-progress or died
88 #define KEYRING_UPCALL_TIMEOUT (obd_timeout + obd_timeout)
90 /****************************************
92 ****************************************/
94 #define DUMP_PROCESS_KEYRINGS(tsk) \
96 CWARN("DUMP PK: %s[%u,%u/%u](<-%s[%u,%u/%u]): " \
97 "a %d, t %d, p %d, s %d, u %d, us %d, df %d\n", \
98 tsk->comm, tsk->pid, tsk->uid, tsk->fsuid, \
99 tsk->parent->comm, tsk->parent->pid, \
100 tsk->parent->uid, tsk->parent->fsuid, \
101 tsk->request_key_auth ? \
102 tsk->request_key_auth->serial : 0, \
103 key_cred(tsk)->thread_keyring ? \
104 key_cred(tsk)->thread_keyring->serial : 0, \
105 key_tgcred(tsk)->process_keyring ? \
106 key_tgcred(tsk)->process_keyring->serial : 0, \
107 key_tgcred(tsk)->session_keyring ? \
108 key_tgcred(tsk)->session_keyring->serial : 0, \
109 key_cred(tsk)->user->uid_keyring ? \
110 key_cred(tsk)->user->uid_keyring->serial : 0, \
111 key_cred(tsk)->user->session_keyring ? \
112 key_cred(tsk)->user->session_keyring->serial : 0, \
113 key_cred(tsk)->jit_keyring \
117 #define DUMP_KEY(key) \
119 CWARN("DUMP KEY: %p(%d) ref %d u%u/g%u desc %s\n", \
120 key, key->serial, atomic_read(&key->usage), \
121 key->uid, key->gid, \
122 key->description ? key->description : "n/a" \
126 #ifdef HAVE_STRUCT_CRED /* Since 2.6.29 */
127 #define key_cred(tsk) ((tsk)->cred)
128 #define key_tgcred(tsk) ((tsk)->cred->tgcred)
130 #define key_cred(tsk) (tsk)
131 #define key_tgcred(tsk) ((tsk)->signal)
134 static inline void keyring_upcall_lock(struct gss_sec_keyring *gsec_kr)
136 #ifdef HAVE_KEYRING_UPCALL_SERIALIZED
137 mutex_lock(&gsec_kr->gsk_uc_lock);
141 static inline void keyring_upcall_unlock(struct gss_sec_keyring *gsec_kr)
143 #ifdef HAVE_KEYRING_UPCALL_SERIALIZED
144 mutex_unlock(&gsec_kr->gsk_uc_lock);
148 static inline void key_revoke_locked(struct key *key)
150 set_bit(KEY_FLAG_REVOKED, &key->flags);
153 static void ctx_upcall_timeout_kr(unsigned long data)
155 struct ptlrpc_cli_ctx *ctx = (struct ptlrpc_cli_ctx *) data;
156 struct key *key = ctx2gctx_keyring(ctx)->gck_key;
158 CWARN("ctx %p, key %p\n", ctx, key);
163 key_revoke_locked(key);
167 void ctx_start_timer_kr(struct ptlrpc_cli_ctx *ctx, long timeout)
169 struct gss_cli_ctx_keyring *gctx_kr = ctx2gctx_keyring(ctx);
170 struct timer_list *timer = gctx_kr->gck_timer;
174 CDEBUG(D_SEC, "ctx %p: start timer %lds\n", ctx, timeout);
175 timeout = timeout * CFS_HZ + cfs_time_current();
178 timer->expires = timeout;
179 timer->data = (unsigned long ) ctx;
180 timer->function = ctx_upcall_timeout_kr;
186 * caller should make sure no race with other threads
189 void ctx_clear_timer_kr(struct ptlrpc_cli_ctx *ctx)
191 struct gss_cli_ctx_keyring *gctx_kr = ctx2gctx_keyring(ctx);
192 struct timer_list *timer = gctx_kr->gck_timer;
197 CDEBUG(D_SEC, "ctx %p, key %p\n", ctx, gctx_kr->gck_key);
199 gctx_kr->gck_timer = NULL;
201 del_singleshot_timer_sync(timer);
207 struct ptlrpc_cli_ctx *ctx_create_kr(struct ptlrpc_sec *sec,
208 struct vfs_cred *vcred)
210 struct ptlrpc_cli_ctx *ctx;
211 struct gss_cli_ctx_keyring *gctx_kr;
213 OBD_ALLOC_PTR(gctx_kr);
217 OBD_ALLOC_PTR(gctx_kr->gck_timer);
218 if (gctx_kr->gck_timer == NULL) {
219 OBD_FREE_PTR(gctx_kr);
222 init_timer(gctx_kr->gck_timer);
224 ctx = &gctx_kr->gck_base.gc_base;
226 if (gss_cli_ctx_init_common(sec, ctx, &gss_keyring_ctxops, vcred)) {
227 OBD_FREE_PTR(gctx_kr->gck_timer);
228 OBD_FREE_PTR(gctx_kr);
232 ctx->cc_expire = cfs_time_current_sec() + KEYRING_UPCALL_TIMEOUT;
233 clear_bit(PTLRPC_CTX_NEW_BIT, &ctx->cc_flags);
234 cfs_atomic_inc(&ctx->cc_refcount); /* for the caller */
239 static void ctx_destroy_kr(struct ptlrpc_cli_ctx *ctx)
241 struct ptlrpc_sec *sec = ctx->cc_sec;
242 struct gss_cli_ctx_keyring *gctx_kr = ctx2gctx_keyring(ctx);
244 CDEBUG(D_SEC, "destroying ctx %p\n", ctx);
246 /* at this time the association with key has been broken. */
248 LASSERT(cfs_atomic_read(&sec->ps_refcount) > 0);
249 LASSERT(cfs_atomic_read(&sec->ps_nctx) > 0);
250 LASSERT(test_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags) == 0);
251 LASSERT(gctx_kr->gck_key == NULL);
253 ctx_clear_timer_kr(ctx);
254 LASSERT(gctx_kr->gck_timer == NULL);
256 if (gss_cli_ctx_fini_common(sec, ctx))
259 OBD_FREE_PTR(gctx_kr);
261 cfs_atomic_dec(&sec->ps_nctx);
262 sptlrpc_sec_put(sec);
265 static void ctx_release_kr(struct ptlrpc_cli_ctx *ctx, int sync)
270 cfs_atomic_inc(&ctx->cc_refcount);
271 sptlrpc_gc_add_ctx(ctx);
275 static void ctx_put_kr(struct ptlrpc_cli_ctx *ctx, int sync)
277 LASSERT(cfs_atomic_read(&ctx->cc_refcount) > 0);
279 if (cfs_atomic_dec_and_test(&ctx->cc_refcount))
280 ctx_release_kr(ctx, sync);
284 * key <-> ctx association and rules:
285 * - ctx might not bind with any key
286 * - key/ctx binding is protected by key semaphore (if the key present)
287 * - key and ctx each take a reference of the other
288 * - ctx enlist/unlist is protected by ctx spinlock
289 * - never enlist a ctx after it's been unlisted
290 * - whoever do enlist should also do bind, lock key before enlist:
291 * - lock key -> lock ctx -> enlist -> unlock ctx -> bind -> unlock key
292 * - whoever do unlist should also do unbind:
293 * - lock key -> lock ctx -> unlist -> unlock ctx -> unbind -> unlock key
294 * - lock ctx -> unlist -> unlock ctx -> lock key -> unbind -> unlock key
297 static inline void spin_lock_if(spinlock_t *lock, int condition)
303 static inline void spin_unlock_if(spinlock_t *lock, int condition)
309 static void ctx_enlist_kr(struct ptlrpc_cli_ctx *ctx, int is_root, int locked)
311 struct ptlrpc_sec *sec = ctx->cc_sec;
312 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
314 LASSERT(!test_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags));
315 LASSERT(cfs_atomic_read(&ctx->cc_refcount) > 0);
317 spin_lock_if(&sec->ps_lock, !locked);
319 cfs_atomic_inc(&ctx->cc_refcount);
320 set_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags);
321 cfs_hlist_add_head(&ctx->cc_cache, &gsec_kr->gsk_clist);
323 gsec_kr->gsk_root_ctx = ctx;
325 spin_unlock_if(&sec->ps_lock, !locked);
329 * Note after this get called, caller should not access ctx again because
330 * it might have been freed, unless caller hold at least one refcount of
333 * return non-zero if we indeed unlist this ctx.
335 static int ctx_unlist_kr(struct ptlrpc_cli_ctx *ctx, int locked)
337 struct ptlrpc_sec *sec = ctx->cc_sec;
338 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
340 /* if hashed bit has gone, leave the job to somebody who is doing it */
341 if (test_and_clear_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags) == 0)
344 /* drop ref inside spin lock to prevent race with other operations */
345 spin_lock_if(&sec->ps_lock, !locked);
347 if (gsec_kr->gsk_root_ctx == ctx)
348 gsec_kr->gsk_root_ctx = NULL;
349 cfs_hlist_del_init(&ctx->cc_cache);
350 cfs_atomic_dec(&ctx->cc_refcount);
352 spin_unlock_if(&sec->ps_lock, !locked);
358 * bind a key with a ctx together.
359 * caller must hold write lock of the key, as well as ref on key & ctx.
361 static void bind_key_ctx(struct key *key, struct ptlrpc_cli_ctx *ctx)
363 LASSERT(cfs_atomic_read(&ctx->cc_refcount) > 0);
364 LASSERT(atomic_read(&key->usage) > 0);
365 LASSERT(ctx2gctx_keyring(ctx)->gck_key == NULL);
366 LASSERT(key->payload.data == NULL);
368 /* at this time context may or may not in list. */
370 cfs_atomic_inc(&ctx->cc_refcount);
371 ctx2gctx_keyring(ctx)->gck_key = key;
372 key->payload.data = ctx;
376 * unbind a key and a ctx.
377 * caller must hold write lock, as well as a ref of the key.
379 static void unbind_key_ctx(struct key *key, struct ptlrpc_cli_ctx *ctx)
381 LASSERT(key->payload.data == ctx);
382 LASSERT(test_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags) == 0);
384 /* must revoke the key, or others may treat it as newly created */
385 key_revoke_locked(key);
387 key->payload.data = NULL;
388 ctx2gctx_keyring(ctx)->gck_key = NULL;
390 /* once ctx get split from key, the timer is meaningless */
391 ctx_clear_timer_kr(ctx);
398 * given a ctx, unbind with its coupled key, if any.
399 * unbind could only be called once, so we don't worry the key be released
402 static void unbind_ctx_kr(struct ptlrpc_cli_ctx *ctx)
404 struct key *key = ctx2gctx_keyring(ctx)->gck_key;
407 LASSERT(key->payload.data == ctx);
410 down_write(&key->sem);
411 unbind_key_ctx(key, ctx);
418 * given a key, unbind with its coupled ctx, if any.
419 * caller must hold write lock, as well as a ref of the key.
421 static void unbind_key_locked(struct key *key)
423 struct ptlrpc_cli_ctx *ctx = key->payload.data;
426 unbind_key_ctx(key, ctx);
430 * unlist a ctx, and unbind from coupled key
432 static void kill_ctx_kr(struct ptlrpc_cli_ctx *ctx)
434 if (ctx_unlist_kr(ctx, 0))
439 * given a key, unlist and unbind with the coupled ctx (if any).
440 * caller must hold write lock, as well as a ref of the key.
442 static void kill_key_locked(struct key *key)
444 struct ptlrpc_cli_ctx *ctx = key->payload.data;
446 if (ctx && ctx_unlist_kr(ctx, 0))
447 unbind_key_locked(key);
451 * caller should hold one ref on contexts in freelist.
453 static void dispose_ctx_list_kr(cfs_hlist_head_t *freelist)
455 cfs_hlist_node_t *pos, *next;
456 struct ptlrpc_cli_ctx *ctx;
457 struct gss_cli_ctx *gctx;
459 cfs_hlist_for_each_entry_safe(ctx, pos, next, freelist, cc_cache) {
460 cfs_hlist_del_init(&ctx->cc_cache);
462 /* reverse ctx: update current seq to buddy svcctx if exist.
463 * ideally this should be done at gss_cli_ctx_finalize(), but
464 * the ctx destroy could be delayed by:
465 * 1) ctx still has reference;
466 * 2) ctx destroy is asynchronous;
467 * and reverse import call inval_all_ctx() require this be done
468 *_immediately_ otherwise newly created reverse ctx might copy
469 * the very old sequence number from svcctx. */
470 gctx = ctx2gctx(ctx);
471 if (!rawobj_empty(&gctx->gc_svc_handle) &&
472 sec_is_reverse(gctx->gc_base.cc_sec)) {
473 gss_svc_upcall_update_sequence(&gctx->gc_svc_handle,
474 (__u32) cfs_atomic_read(&gctx->gc_seq));
477 /* we need to wakeup waiting reqs here. the context might
478 * be forced released before upcall finished, then the
479 * late-arrived downcall can't find the ctx even. */
480 sptlrpc_cli_ctx_wakeup(ctx);
488 * lookup a root context directly in a sec, return root ctx with a
489 * reference taken or NULL.
492 struct ptlrpc_cli_ctx * sec_lookup_root_ctx_kr(struct ptlrpc_sec *sec)
494 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
495 struct ptlrpc_cli_ctx *ctx = NULL;
497 spin_lock(&sec->ps_lock);
499 ctx = gsec_kr->gsk_root_ctx;
501 if (ctx == NULL && unlikely(sec_is_reverse(sec))) {
502 cfs_hlist_node_t *node;
503 struct ptlrpc_cli_ctx *tmp;
505 /* reverse ctx, search root ctx in list, choose the one
506 * with shortest expire time, which is most possibly have
507 * an established peer ctx at client side. */
508 cfs_hlist_for_each_entry(tmp, node, &gsec_kr->gsk_clist,
510 if (ctx == NULL || ctx->cc_expire == 0 ||
511 ctx->cc_expire > tmp->cc_expire) {
513 /* promote to be root_ctx */
514 gsec_kr->gsk_root_ctx = ctx;
520 LASSERT(cfs_atomic_read(&ctx->cc_refcount) > 0);
521 LASSERT(!cfs_hlist_empty(&gsec_kr->gsk_clist));
522 cfs_atomic_inc(&ctx->cc_refcount);
525 spin_unlock(&sec->ps_lock);
530 #define RVS_CTX_EXPIRE_NICE (10)
533 void rvs_sec_install_root_ctx_kr(struct ptlrpc_sec *sec,
534 struct ptlrpc_cli_ctx *new_ctx,
537 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
538 cfs_hlist_node_t *hnode;
539 struct ptlrpc_cli_ctx *ctx;
543 LASSERT(sec_is_reverse(sec));
545 spin_lock(&sec->ps_lock);
547 now = cfs_time_current_sec();
549 /* set all existing ctxs short expiry */
550 cfs_hlist_for_each_entry(ctx, hnode, &gsec_kr->gsk_clist, cc_cache) {
551 if (ctx->cc_expire > now + RVS_CTX_EXPIRE_NICE) {
552 ctx->cc_early_expire = 1;
553 ctx->cc_expire = now + RVS_CTX_EXPIRE_NICE;
557 /* if there's root_ctx there, instead obsolete the current
558 * immediately, we leave it continue operating for a little while.
559 * hopefully when the first backward rpc with newest ctx send out,
560 * the client side already have the peer ctx well established. */
561 ctx_enlist_kr(new_ctx, gsec_kr->gsk_root_ctx ? 0 : 1, 1);
564 bind_key_ctx(key, new_ctx);
566 spin_unlock(&sec->ps_lock);
569 static void construct_key_desc(void *buf, int bufsize,
570 struct ptlrpc_sec *sec, uid_t uid)
572 snprintf(buf, bufsize, "%d@%x", uid, sec->ps_id);
573 ((char *)buf)[bufsize - 1] = '\0';
576 /****************************************
578 ****************************************/
581 struct ptlrpc_sec * gss_sec_create_kr(struct obd_import *imp,
582 struct ptlrpc_svc_ctx *svcctx,
583 struct sptlrpc_flavor *sf)
585 struct gss_sec_keyring *gsec_kr;
588 OBD_ALLOC(gsec_kr, sizeof(*gsec_kr));
592 CFS_INIT_HLIST_HEAD(&gsec_kr->gsk_clist);
593 gsec_kr->gsk_root_ctx = NULL;
594 mutex_init(&gsec_kr->gsk_root_uc_lock);
595 #ifdef HAVE_KEYRING_UPCALL_SERIALIZED
596 mutex_init(&gsec_kr->gsk_uc_lock);
599 if (gss_sec_create_common(&gsec_kr->gsk_base, &gss_policy_keyring,
603 if (svcctx != NULL &&
604 sec_install_rctx_kr(&gsec_kr->gsk_base.gs_base, svcctx)) {
605 gss_sec_destroy_common(&gsec_kr->gsk_base);
609 RETURN(&gsec_kr->gsk_base.gs_base);
612 OBD_FREE(gsec_kr, sizeof(*gsec_kr));
617 void gss_sec_destroy_kr(struct ptlrpc_sec *sec)
619 struct gss_sec *gsec = sec2gsec(sec);
620 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
622 CDEBUG(D_SEC, "destroy %s@%p\n", sec->ps_policy->sp_name, sec);
624 LASSERT(cfs_hlist_empty(&gsec_kr->gsk_clist));
625 LASSERT(gsec_kr->gsk_root_ctx == NULL);
627 gss_sec_destroy_common(gsec);
629 OBD_FREE(gsec_kr, sizeof(*gsec_kr));
632 static inline int user_is_root(struct ptlrpc_sec *sec, struct vfs_cred *vcred)
634 /* except the ROOTONLY flag, treat it as root user only if real uid
635 * is 0, euid/fsuid being 0 are handled as setuid scenarios */
636 if (sec_is_rootonly(sec) || (vcred->vc_uid == 0))
643 * unlink request key from it's ring, which is linked during request_key().
644 * sadly, we have to 'guess' which keyring it's linked to.
646 * FIXME this code is fragile, depend on how request_key_link() is implemented.
648 static void request_key_unlink(struct key *key)
650 struct task_struct *tsk = current;
653 switch (key_cred(tsk)->jit_keyring) {
654 case KEY_REQKEY_DEFL_DEFAULT:
655 case KEY_REQKEY_DEFL_THREAD_KEYRING:
656 ring = key_get(key_cred(tsk)->thread_keyring);
659 case KEY_REQKEY_DEFL_PROCESS_KEYRING:
660 ring = key_get(key_tgcred(tsk)->process_keyring);
663 case KEY_REQKEY_DEFL_SESSION_KEYRING:
665 ring = key_get(rcu_dereference(key_tgcred(tsk)
670 case KEY_REQKEY_DEFL_USER_SESSION_KEYRING:
671 ring = key_get(key_cred(tsk)->user->session_keyring);
673 case KEY_REQKEY_DEFL_USER_KEYRING:
674 ring = key_get(key_cred(tsk)->user->uid_keyring);
676 case KEY_REQKEY_DEFL_GROUP_KEYRING:
682 key_unlink(ring, key);
687 struct ptlrpc_cli_ctx * gss_sec_lookup_ctx_kr(struct ptlrpc_sec *sec,
688 struct vfs_cred *vcred,
689 int create, int remove_dead)
691 struct obd_import *imp = sec->ps_import;
692 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
693 struct ptlrpc_cli_ctx *ctx = NULL;
694 unsigned int is_root = 0, create_new = 0;
702 LASSERT(imp != NULL);
704 is_root = user_is_root(sec, vcred);
706 /* a little bit optimization for root context */
708 ctx = sec_lookup_root_ctx_kr(sec);
710 * Only lookup directly for REVERSE sec, which should
713 if (ctx || sec_is_reverse(sec))
717 LASSERT(create != 0);
719 /* for root context, obtain lock and check again, this time hold
720 * the root upcall lock, make sure nobody else populated new root
721 * context after last check. */
723 mutex_lock(&gsec_kr->gsk_root_uc_lock);
725 ctx = sec_lookup_root_ctx_kr(sec);
729 /* update reverse handle for root user */
730 sec2gsec(sec)->gs_rvs_hdl = gss_get_next_ctx_index();
732 switch (sec->ps_part) {
751 /* in case of setuid, key will be constructed as owner of fsuid/fsgid,
752 * but we do authentication based on real uid/gid. the key permission
753 * bits will be exactly as POS_ALL, so only processes who subscribed
754 * this key could have the access, although the quota might be counted
755 * on others (fsuid/fsgid).
757 * keyring will use fsuid/fsgid as upcall parameters, so we have to
758 * encode real uid/gid into callout info.
761 construct_key_desc(desc, sizeof(desc), sec, vcred->vc_uid);
763 /* callout info format:
764 * secid:mech:uid:gid:flags:svc_type:peer_nid:target_uuid
766 coinfo_size = sizeof(struct obd_uuid) + MAX_OBD_NAME + 64;
767 OBD_ALLOC(coinfo, coinfo_size);
771 snprintf(coinfo, coinfo_size, "%d:%s:%u:%u:%s:%d:"LPX64":%s",
772 sec->ps_id, sec2gsec(sec)->gs_mech->gm_name,
773 vcred->vc_uid, vcred->vc_gid,
774 co_flags, import_to_gss_svc(imp),
775 imp->imp_connection->c_peer.nid, imp->imp_obd->obd_name);
777 CDEBUG(D_SEC, "requesting key for %s\n", desc);
779 keyring_upcall_lock(gsec_kr);
780 key = request_key(&gss_key_type, desc, coinfo);
781 keyring_upcall_unlock(gsec_kr);
783 OBD_FREE(coinfo, coinfo_size);
786 CERROR("failed request key: %ld\n", PTR_ERR(key));
789 CDEBUG(D_SEC, "obtained key %08x for %s\n", key->serial, desc);
791 /* once payload.data was pointed to a ctx, it never changes until
792 * we de-associate them; but parallel request_key() may return
793 * a key with payload.data == NULL at the same time. so we still
794 * need wirtelock of key->sem to serialize them. */
795 down_write(&key->sem);
797 if (likely(key->payload.data != NULL)) {
798 ctx = key->payload.data;
800 LASSERT(cfs_atomic_read(&ctx->cc_refcount) >= 1);
801 LASSERT(ctx2gctx_keyring(ctx)->gck_key == key);
802 LASSERT(atomic_read(&key->usage) >= 2);
804 /* simply take a ref and return. it's upper layer's
805 * responsibility to detect & replace dead ctx. */
806 cfs_atomic_inc(&ctx->cc_refcount);
808 /* pre initialization with a cli_ctx. this can't be done in
809 * key_instantiate() because we'v no enough information
811 ctx = ctx_create_kr(sec, vcred);
813 ctx_enlist_kr(ctx, is_root, 0);
814 bind_key_ctx(key, ctx);
816 ctx_start_timer_kr(ctx, KEYRING_UPCALL_TIMEOUT);
818 CDEBUG(D_SEC, "installed key %p <-> ctx %p (sec %p)\n",
821 /* we'd prefer to call key_revoke(), but we more like
822 * to revoke it within this key->sem locked period. */
823 key_revoke_locked(key);
831 if (is_root && create_new)
832 request_key_unlink(key);
837 mutex_unlock(&gsec_kr->gsk_root_uc_lock);
842 void gss_sec_release_ctx_kr(struct ptlrpc_sec *sec,
843 struct ptlrpc_cli_ctx *ctx,
846 LASSERT(cfs_atomic_read(&sec->ps_refcount) > 0);
847 LASSERT(cfs_atomic_read(&ctx->cc_refcount) == 0);
848 ctx_release_kr(ctx, sync);
852 * flush context of normal user, we must resort to keyring itself to find out
853 * contexts which belong to me.
855 * Note here we suppose only to flush _my_ context, the "uid" will
856 * be ignored in the search.
859 void flush_user_ctx_cache_kr(struct ptlrpc_sec *sec,
861 int grace, int force)
866 /* nothing to do for reverse or rootonly sec */
867 if (sec_is_reverse(sec) || sec_is_rootonly(sec))
870 construct_key_desc(desc, sizeof(desc), sec, uid);
872 /* there should be only one valid key, but we put it in the
873 * loop in case of any weird cases */
875 key = request_key(&gss_key_type, desc, NULL);
877 CDEBUG(D_SEC, "No more key found for current user\n");
881 down_write(&key->sem);
883 kill_key_locked(key);
885 /* kill_key_locked() should usually revoke the key, but we
886 * revoke it again to make sure, e.g. some case the key may
887 * not well coupled with a context. */
888 key_revoke_locked(key);
897 * flush context of root or all, we iterate through the list.
900 void flush_spec_ctx_cache_kr(struct ptlrpc_sec *sec,
902 int grace, int force)
904 struct gss_sec_keyring *gsec_kr;
905 cfs_hlist_head_t freelist = CFS_HLIST_HEAD_INIT;
906 cfs_hlist_node_t *pos, *next;
907 struct ptlrpc_cli_ctx *ctx;
910 gsec_kr = sec2gsec_keyring(sec);
912 spin_lock(&sec->ps_lock);
913 cfs_hlist_for_each_entry_safe(ctx, pos, next,
914 &gsec_kr->gsk_clist, cc_cache) {
915 LASSERT(cfs_atomic_read(&ctx->cc_refcount) > 0);
917 if (uid != -1 && uid != ctx->cc_vcred.vc_uid)
920 /* at this moment there's at least 2 base reference:
921 * key association and in-list. */
922 if (cfs_atomic_read(&ctx->cc_refcount) > 2) {
925 CWARN("flush busy ctx %p(%u->%s, extra ref %d)\n",
926 ctx, ctx->cc_vcred.vc_uid,
927 sec2target_str(ctx->cc_sec),
928 cfs_atomic_read(&ctx->cc_refcount) - 2);
931 set_bit(PTLRPC_CTX_DEAD_BIT, &ctx->cc_flags);
933 clear_bit(PTLRPC_CTX_UPTODATE_BIT, &ctx->cc_flags);
935 cfs_atomic_inc(&ctx->cc_refcount);
937 if (ctx_unlist_kr(ctx, 1)) {
938 cfs_hlist_add_head(&ctx->cc_cache, &freelist);
940 LASSERT(cfs_atomic_read(&ctx->cc_refcount) >= 2);
941 cfs_atomic_dec(&ctx->cc_refcount);
944 spin_unlock(&sec->ps_lock);
946 dispose_ctx_list_kr(&freelist);
951 int gss_sec_flush_ctx_cache_kr(struct ptlrpc_sec *sec,
952 uid_t uid, int grace, int force)
956 CDEBUG(D_SEC, "sec %p(%d, nctx %d), uid %d, grace %d, force %d\n",
957 sec, cfs_atomic_read(&sec->ps_refcount),
958 cfs_atomic_read(&sec->ps_nctx),
961 if (uid != -1 && uid != 0)
962 flush_user_ctx_cache_kr(sec, uid, grace, force);
964 flush_spec_ctx_cache_kr(sec, uid, grace, force);
970 void gss_sec_gc_ctx_kr(struct ptlrpc_sec *sec)
972 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
973 cfs_hlist_head_t freelist = CFS_HLIST_HEAD_INIT;
974 cfs_hlist_node_t *pos, *next;
975 struct ptlrpc_cli_ctx *ctx;
978 CWARN("running gc\n");
980 spin_lock(&sec->ps_lock);
981 cfs_hlist_for_each_entry_safe(ctx, pos, next,
982 &gsec_kr->gsk_clist, cc_cache) {
983 LASSERT(cfs_atomic_read(&ctx->cc_refcount) > 0);
985 cfs_atomic_inc(&ctx->cc_refcount);
987 if (cli_ctx_check_death(ctx) && ctx_unlist_kr(ctx, 1)) {
988 cfs_hlist_add_head(&ctx->cc_cache, &freelist);
989 CWARN("unhashed ctx %p\n", ctx);
991 LASSERT(cfs_atomic_read(&ctx->cc_refcount) >= 2);
992 cfs_atomic_dec(&ctx->cc_refcount);
995 spin_unlock(&sec->ps_lock);
997 dispose_ctx_list_kr(&freelist);
1003 int gss_sec_display_kr(struct ptlrpc_sec *sec, struct seq_file *seq)
1005 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
1006 cfs_hlist_node_t *pos, *next;
1007 struct ptlrpc_cli_ctx *ctx;
1008 struct gss_cli_ctx *gctx;
1009 time_t now = cfs_time_current_sec();
1012 spin_lock(&sec->ps_lock);
1013 cfs_hlist_for_each_entry_safe(ctx, pos, next,
1014 &gsec_kr->gsk_clist, cc_cache) {
1019 gctx = ctx2gctx(ctx);
1020 key = ctx2gctx_keyring(ctx)->gck_key;
1022 gss_cli_ctx_flags2str(ctx->cc_flags,
1023 flags_str, sizeof(flags_str));
1025 if (gctx->gc_mechctx)
1026 lgss_display(gctx->gc_mechctx, mech, sizeof(mech));
1028 snprintf(mech, sizeof(mech), "N/A");
1029 mech[sizeof(mech) - 1] = '\0';
1031 seq_printf(seq, "%p: uid %u, ref %d, expire %ld(%+ld), fl %s, "
1032 "seq %d, win %u, key %08x(ref %d), "
1033 "hdl "LPX64":"LPX64", mech: %s\n",
1034 ctx, ctx->cc_vcred.vc_uid,
1035 cfs_atomic_read(&ctx->cc_refcount),
1037 ctx->cc_expire ? ctx->cc_expire - now : 0,
1039 cfs_atomic_read(&gctx->gc_seq),
1041 key ? key->serial : 0,
1042 key ? atomic_read(&key->usage) : 0,
1043 gss_handle_to_u64(&gctx->gc_handle),
1044 gss_handle_to_u64(&gctx->gc_svc_handle),
1047 spin_unlock(&sec->ps_lock);
1052 /****************************************
1054 ****************************************/
1057 int gss_cli_ctx_refresh_kr(struct ptlrpc_cli_ctx *ctx)
1059 /* upcall is already on the way */
1064 int gss_cli_ctx_validate_kr(struct ptlrpc_cli_ctx *ctx)
1066 LASSERT(cfs_atomic_read(&ctx->cc_refcount) > 0);
1067 LASSERT(ctx->cc_sec);
1069 if (cli_ctx_check_death(ctx)) {
1074 if (cli_ctx_is_ready(ctx))
1080 void gss_cli_ctx_die_kr(struct ptlrpc_cli_ctx *ctx, int grace)
1082 LASSERT(cfs_atomic_read(&ctx->cc_refcount) > 0);
1083 LASSERT(ctx->cc_sec);
1085 cli_ctx_expire(ctx);
1089 /****************************************
1090 * (reverse) service *
1091 ****************************************/
1094 * reverse context could have nothing to do with keyrings. here we still keep
1095 * the version which bind to a key, for future reference.
1097 #define HAVE_REVERSE_CTX_NOKEY
1099 #ifdef HAVE_REVERSE_CTX_NOKEY
1102 int sec_install_rctx_kr(struct ptlrpc_sec *sec,
1103 struct ptlrpc_svc_ctx *svc_ctx)
1105 struct ptlrpc_cli_ctx *cli_ctx;
1106 struct vfs_cred vcred = { 0, 0 };
1112 cli_ctx = ctx_create_kr(sec, &vcred);
1113 if (cli_ctx == NULL)
1116 rc = gss_copy_rvc_cli_ctx(cli_ctx, svc_ctx);
1118 CERROR("failed copy reverse cli ctx: %d\n", rc);
1120 ctx_put_kr(cli_ctx, 1);
1124 rvs_sec_install_root_ctx_kr(sec, cli_ctx, NULL);
1126 ctx_put_kr(cli_ctx, 1);
1131 #else /* ! HAVE_REVERSE_CTX_NOKEY */
1134 int sec_install_rctx_kr(struct ptlrpc_sec *sec,
1135 struct ptlrpc_svc_ctx *svc_ctx)
1137 struct ptlrpc_cli_ctx *cli_ctx = NULL;
1139 struct vfs_cred vcred = { 0, 0 };
1147 construct_key_desc(desc, sizeof(desc), sec, 0);
1149 key = key_alloc(&gss_key_type, desc, 0, 0,
1150 KEY_POS_ALL | KEY_USR_ALL, 1);
1152 CERROR("failed to alloc key: %ld\n", PTR_ERR(key));
1153 return PTR_ERR(key);
1156 rc = key_instantiate_and_link(key, NULL, 0, NULL, NULL);
1158 CERROR("failed to instantiate key: %d\n", rc);
1162 down_write(&key->sem);
1164 LASSERT(key->payload.data == NULL);
1166 cli_ctx = ctx_create_kr(sec, &vcred);
1167 if (cli_ctx == NULL) {
1172 rc = gss_copy_rvc_cli_ctx(cli_ctx, svc_ctx);
1174 CERROR("failed copy reverse cli ctx: %d\n", rc);
1178 rvs_sec_install_root_ctx_kr(sec, cli_ctx, key);
1180 ctx_put_kr(cli_ctx, 1);
1181 up_write(&key->sem);
1190 ctx_put_kr(cli_ctx, 1);
1192 up_write(&key->sem);
1198 #endif /* HAVE_REVERSE_CTX_NOKEY */
1200 /****************************************
1202 ****************************************/
1205 int gss_svc_accept_kr(struct ptlrpc_request *req)
1207 return gss_svc_accept(&gss_policy_keyring, req);
1211 int gss_svc_install_rctx_kr(struct obd_import *imp,
1212 struct ptlrpc_svc_ctx *svc_ctx)
1214 struct ptlrpc_sec *sec;
1217 sec = sptlrpc_import_sec_ref(imp);
1220 rc = sec_install_rctx_kr(sec, svc_ctx);
1221 sptlrpc_sec_put(sec);
1226 /****************************************
1228 ****************************************/
1231 int gss_kt_instantiate(struct key *key, const void *data, size_t datalen)
1236 if (data != NULL || datalen != 0) {
1237 CERROR("invalid: data %p, len %lu\n", data, (long)datalen);
1241 if (key->payload.data != 0) {
1242 CERROR("key already have payload\n");
1246 /* link the key to session keyring, so following context negotiation
1247 * rpc fired from user space could find this key. This will be unlinked
1248 * automatically when upcall processes die.
1250 * we can't do this through keyctl from userspace, because the upcall
1251 * might be neither possessor nor owner of the key (setuid).
1253 * the session keyring is created upon upcall, and don't change all
1254 * the way until upcall finished, so rcu lock is not needed here.
1256 LASSERT(key_tgcred(cfs_current())->session_keyring);
1259 rc = key_link(key_tgcred(cfs_current())->session_keyring, key);
1262 CERROR("failed to link key %08x to keyring %08x: %d\n",
1264 key_tgcred(cfs_current())->session_keyring->serial, rc);
1268 CDEBUG(D_SEC, "key %p instantiated, ctx %p\n", key, key->payload.data);
1273 * called with key semaphore write locked. it means we can operate
1274 * on the context without fear of loosing refcount.
1277 int gss_kt_update(struct key *key, const void *data, size_t datalen)
1279 struct ptlrpc_cli_ctx *ctx = key->payload.data;
1280 struct gss_cli_ctx *gctx;
1281 rawobj_t tmpobj = RAWOBJ_EMPTY;
1282 __u32 datalen32 = (__u32) datalen;
1286 if (data == NULL || datalen == 0) {
1287 CWARN("invalid: data %p, len %lu\n", data, (long)datalen);
1291 /* if upcall finished negotiation too fast (mostly likely because
1292 * of local error happened) and call kt_update(), the ctx
1293 * might be still NULL. but the key will finally be associate
1294 * with a context, or be revoked. if key status is fine, return
1295 * -EAGAIN to allow userspace sleep a while and call again. */
1297 CDEBUG(D_SEC, "update too soon: key %p(%x) flags %lx\n",
1298 key, key->serial, key->flags);
1300 rc = key_validate(key);
1307 LASSERT(cfs_atomic_read(&ctx->cc_refcount) > 0);
1308 LASSERT(ctx->cc_sec);
1310 ctx_clear_timer_kr(ctx);
1312 /* don't proceed if already refreshed */
1313 if (cli_ctx_is_refreshed(ctx)) {
1314 CWARN("ctx already done refresh\n");
1318 sptlrpc_cli_ctx_get(ctx);
1319 gctx = ctx2gctx(ctx);
1321 rc = buffer_extract_bytes(&data, &datalen32, &gctx->gc_win,
1322 sizeof(gctx->gc_win));
1324 CERROR("failed extract seq_win\n");
1328 if (gctx->gc_win == 0) {
1329 __u32 nego_rpc_err, nego_gss_err;
1331 rc = buffer_extract_bytes(&data, &datalen32, &nego_rpc_err,
1332 sizeof(nego_rpc_err));
1334 CERROR("failed to extrace rpc rc\n");
1338 rc = buffer_extract_bytes(&data, &datalen32, &nego_gss_err,
1339 sizeof(nego_gss_err));
1341 CERROR("failed to extrace gss rc\n");
1345 CERROR("negotiation: rpc err %d, gss err %x\n",
1346 nego_rpc_err, nego_gss_err);
1348 rc = nego_rpc_err ? nego_rpc_err : -EACCES;
1350 rc = rawobj_extract_local_alloc(&gctx->gc_handle,
1351 (__u32 **) &data, &datalen32);
1353 CERROR("failed extract handle\n");
1357 rc = rawobj_extract_local(&tmpobj, (__u32 **) &data,&datalen32);
1359 CERROR("failed extract mech\n");
1363 rc = lgss_import_sec_context(&tmpobj,
1364 sec2gsec(ctx->cc_sec)->gs_mech,
1366 if (rc != GSS_S_COMPLETE)
1367 CERROR("failed import context\n");
1372 /* we don't care what current status of this ctx, even someone else
1373 * is operating on the ctx at the same time. we just add up our own
1376 gss_cli_ctx_uptodate(gctx);
1378 /* this will also revoke the key. has to be done before
1379 * wakeup waiters otherwise they can find the stale key */
1380 kill_key_locked(key);
1382 cli_ctx_expire(ctx);
1384 if (rc != -ERESTART)
1385 set_bit(PTLRPC_CTX_ERROR_BIT, &ctx->cc_flags);
1388 /* let user space think it's a success */
1389 sptlrpc_cli_ctx_put(ctx, 1);
1394 int gss_kt_match(const struct key *key, const void *desc)
1396 return (strcmp(key->description, (const char *) desc) == 0);
1400 void gss_kt_destroy(struct key *key)
1403 LASSERT(key->payload.data == NULL);
1404 CDEBUG(D_SEC, "destroy key %p\n", key);
1409 void gss_kt_describe(const struct key *key, struct seq_file *s)
1411 if (key->description == NULL)
1412 seq_puts(s, "[null]");
1414 seq_puts(s, key->description);
1417 static struct key_type gss_key_type =
1421 .instantiate = gss_kt_instantiate,
1422 .update = gss_kt_update,
1423 .match = gss_kt_match,
1424 .destroy = gss_kt_destroy,
1425 .describe = gss_kt_describe,
1428 /****************************************
1429 * lustre gss keyring policy *
1430 ****************************************/
1432 static struct ptlrpc_ctx_ops gss_keyring_ctxops = {
1433 .match = gss_cli_ctx_match,
1434 .refresh = gss_cli_ctx_refresh_kr,
1435 .validate = gss_cli_ctx_validate_kr,
1436 .die = gss_cli_ctx_die_kr,
1437 .sign = gss_cli_ctx_sign,
1438 .verify = gss_cli_ctx_verify,
1439 .seal = gss_cli_ctx_seal,
1440 .unseal = gss_cli_ctx_unseal,
1441 .wrap_bulk = gss_cli_ctx_wrap_bulk,
1442 .unwrap_bulk = gss_cli_ctx_unwrap_bulk,
1445 static struct ptlrpc_sec_cops gss_sec_keyring_cops = {
1446 .create_sec = gss_sec_create_kr,
1447 .destroy_sec = gss_sec_destroy_kr,
1448 .kill_sec = gss_sec_kill,
1449 .lookup_ctx = gss_sec_lookup_ctx_kr,
1450 .release_ctx = gss_sec_release_ctx_kr,
1451 .flush_ctx_cache = gss_sec_flush_ctx_cache_kr,
1452 .gc_ctx = gss_sec_gc_ctx_kr,
1453 .install_rctx = gss_sec_install_rctx,
1454 .alloc_reqbuf = gss_alloc_reqbuf,
1455 .free_reqbuf = gss_free_reqbuf,
1456 .alloc_repbuf = gss_alloc_repbuf,
1457 .free_repbuf = gss_free_repbuf,
1458 .enlarge_reqbuf = gss_enlarge_reqbuf,
1459 .display = gss_sec_display_kr,
1462 static struct ptlrpc_sec_sops gss_sec_keyring_sops = {
1463 .accept = gss_svc_accept_kr,
1464 .invalidate_ctx = gss_svc_invalidate_ctx,
1465 .alloc_rs = gss_svc_alloc_rs,
1466 .authorize = gss_svc_authorize,
1467 .free_rs = gss_svc_free_rs,
1468 .free_ctx = gss_svc_free_ctx,
1469 .prep_bulk = gss_svc_prep_bulk,
1470 .unwrap_bulk = gss_svc_unwrap_bulk,
1471 .wrap_bulk = gss_svc_wrap_bulk,
1472 .install_rctx = gss_svc_install_rctx_kr,
1475 static struct ptlrpc_sec_policy gss_policy_keyring = {
1476 .sp_owner = THIS_MODULE,
1477 .sp_name = "gss.keyring",
1478 .sp_policy = SPTLRPC_POLICY_GSS,
1479 .sp_cops = &gss_sec_keyring_cops,
1480 .sp_sops = &gss_sec_keyring_sops,
1484 int __init gss_init_keyring(void)
1488 rc = register_key_type(&gss_key_type);
1490 CERROR("failed to register keyring type: %d\n", rc);
1494 rc = sptlrpc_register_policy(&gss_policy_keyring);
1496 unregister_key_type(&gss_key_type);
1503 void __exit gss_exit_keyring(void)
1505 unregister_key_type(&gss_key_type);
1506 sptlrpc_unregister_policy(&gss_policy_keyring);