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, 2014, 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
42 #include <linux/init.h>
43 #include <linux/module.h>
44 #include <linux/slab.h>
45 #include <linux/dcache.h>
47 #include <linux/crypto.h>
48 #include <linux/key.h>
49 #include <linux/keyctl.h>
50 #include <linux/key-type.h>
51 #include <linux/mutex.h>
52 #include <asm/atomic.h>
55 #include <obd_class.h>
56 #include <obd_support.h>
57 #include <lustre/lustre_idl.h>
58 #include <lustre_sec.h>
59 #include <lustre_net.h>
60 #include <lustre_import.h>
63 #include "gss_internal.h"
66 static struct ptlrpc_sec_policy gss_policy_keyring;
67 static struct ptlrpc_ctx_ops gss_keyring_ctxops;
68 static struct key_type gss_key_type;
70 static int sec_install_rctx_kr(struct ptlrpc_sec *sec,
71 struct ptlrpc_svc_ctx *svc_ctx);
74 * the timeout is only for the case that upcall child process die abnormally.
75 * in any other cases it should finally update kernel key.
77 * FIXME we'd better to incorporate the client & server side upcall timeouts
78 * into the framework of Adaptive Timeouts, but we need to figure out how to
79 * make sure that kernel knows the upcall processes is in-progress or died
82 #define KEYRING_UPCALL_TIMEOUT (obd_timeout + obd_timeout)
84 /****************************************
86 ****************************************/
88 #define DUMP_PROCESS_KEYRINGS(tsk) \
90 CWARN("DUMP PK: %s[%u,%u/%u](<-%s[%u,%u/%u]): " \
91 "a %d, t %d, p %d, s %d, u %d, us %d, df %d\n", \
92 tsk->comm, tsk->pid, tsk->uid, tsk->fsuid, \
93 tsk->parent->comm, tsk->parent->pid, \
94 tsk->parent->uid, tsk->parent->fsuid, \
95 tsk->request_key_auth ? \
96 tsk->request_key_auth->serial : 0, \
97 key_cred(tsk)->thread_keyring ? \
98 key_cred(tsk)->thread_keyring->serial : 0, \
99 key_tgcred(tsk)->process_keyring ? \
100 key_tgcred(tsk)->process_keyring->serial : 0, \
101 key_tgcred(tsk)->session_keyring ? \
102 key_tgcred(tsk)->session_keyring->serial : 0, \
103 key_cred(tsk)->user->uid_keyring ? \
104 key_cred(tsk)->user->uid_keyring->serial : 0, \
105 key_cred(tsk)->user->session_keyring ? \
106 key_cred(tsk)->user->session_keyring->serial : 0, \
107 key_cred(tsk)->jit_keyring \
111 #define DUMP_KEY(key) \
113 CWARN("DUMP KEY: %p(%d) ref %d u%u/g%u desc %s\n", \
114 key, key->serial, atomic_read(&key->usage), \
115 key->uid, key->gid, \
116 key->description ? key->description : "n/a" \
120 #define key_cred(tsk) ((tsk)->cred)
121 #ifdef HAVE_CRED_TGCRED
122 #define key_tgcred(tsk) ((tsk)->cred->tgcred)
124 #define key_tgcred(tsk) key_cred(tsk)
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 = msecs_to_jiffies(timeout * MSEC_PER_SEC) +
171 timer->expires = timeout;
172 timer->data = (unsigned long ) ctx;
173 timer->function = ctx_upcall_timeout_kr;
179 * caller should make sure no race with other threads
182 void ctx_clear_timer_kr(struct ptlrpc_cli_ctx *ctx)
184 struct gss_cli_ctx_keyring *gctx_kr = ctx2gctx_keyring(ctx);
185 struct timer_list *timer = gctx_kr->gck_timer;
190 CDEBUG(D_SEC, "ctx %p, key %p\n", ctx, gctx_kr->gck_key);
192 gctx_kr->gck_timer = NULL;
194 del_singleshot_timer_sync(timer);
200 struct ptlrpc_cli_ctx *ctx_create_kr(struct ptlrpc_sec *sec,
201 struct vfs_cred *vcred)
203 struct ptlrpc_cli_ctx *ctx;
204 struct gss_cli_ctx_keyring *gctx_kr;
206 OBD_ALLOC_PTR(gctx_kr);
210 OBD_ALLOC_PTR(gctx_kr->gck_timer);
211 if (gctx_kr->gck_timer == NULL) {
212 OBD_FREE_PTR(gctx_kr);
215 init_timer(gctx_kr->gck_timer);
217 ctx = &gctx_kr->gck_base.gc_base;
219 if (gss_cli_ctx_init_common(sec, ctx, &gss_keyring_ctxops, vcred)) {
220 OBD_FREE_PTR(gctx_kr->gck_timer);
221 OBD_FREE_PTR(gctx_kr);
225 ctx->cc_expire = cfs_time_current_sec() + KEYRING_UPCALL_TIMEOUT;
226 clear_bit(PTLRPC_CTX_NEW_BIT, &ctx->cc_flags);
227 atomic_inc(&ctx->cc_refcount); /* for the caller */
232 static void ctx_destroy_kr(struct ptlrpc_cli_ctx *ctx)
234 struct ptlrpc_sec *sec = ctx->cc_sec;
235 struct gss_cli_ctx_keyring *gctx_kr = ctx2gctx_keyring(ctx);
237 CDEBUG(D_SEC, "destroying ctx %p\n", ctx);
239 /* at this time the association with key has been broken. */
241 LASSERT(atomic_read(&sec->ps_refcount) > 0);
242 LASSERT(atomic_read(&sec->ps_nctx) > 0);
243 LASSERT(test_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags) == 0);
244 LASSERT(gctx_kr->gck_key == NULL);
246 ctx_clear_timer_kr(ctx);
247 LASSERT(gctx_kr->gck_timer == NULL);
249 if (gss_cli_ctx_fini_common(sec, ctx))
252 OBD_FREE_PTR(gctx_kr);
254 atomic_dec(&sec->ps_nctx);
255 sptlrpc_sec_put(sec);
258 static void ctx_release_kr(struct ptlrpc_cli_ctx *ctx, int sync)
263 atomic_inc(&ctx->cc_refcount);
264 sptlrpc_gc_add_ctx(ctx);
268 static void ctx_put_kr(struct ptlrpc_cli_ctx *ctx, int sync)
270 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
272 if (atomic_dec_and_test(&ctx->cc_refcount))
273 ctx_release_kr(ctx, sync);
277 * key <-> ctx association and rules:
278 * - ctx might not bind with any key
279 * - key/ctx binding is protected by key semaphore (if the key present)
280 * - key and ctx each take a reference of the other
281 * - ctx enlist/unlist is protected by ctx spinlock
282 * - never enlist a ctx after it's been unlisted
283 * - whoever do enlist should also do bind, lock key before enlist:
284 * - lock key -> lock ctx -> enlist -> unlock ctx -> bind -> unlock key
285 * - whoever do unlist should also do unbind:
286 * - lock key -> lock ctx -> unlist -> unlock ctx -> unbind -> unlock key
287 * - lock ctx -> unlist -> unlock ctx -> lock key -> unbind -> unlock key
290 static inline void spin_lock_if(spinlock_t *lock, int condition)
296 static inline void spin_unlock_if(spinlock_t *lock, int condition)
302 static void ctx_enlist_kr(struct ptlrpc_cli_ctx *ctx, int is_root, int locked)
304 struct ptlrpc_sec *sec = ctx->cc_sec;
305 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
307 LASSERT(!test_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags));
308 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
310 spin_lock_if(&sec->ps_lock, !locked);
312 atomic_inc(&ctx->cc_refcount);
313 set_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags);
314 hlist_add_head(&ctx->cc_cache, &gsec_kr->gsk_clist);
316 gsec_kr->gsk_root_ctx = ctx;
318 spin_unlock_if(&sec->ps_lock, !locked);
322 * Note after this get called, caller should not access ctx again because
323 * it might have been freed, unless caller hold at least one refcount of
326 * return non-zero if we indeed unlist this ctx.
328 static int ctx_unlist_kr(struct ptlrpc_cli_ctx *ctx, int locked)
330 struct ptlrpc_sec *sec = ctx->cc_sec;
331 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
333 /* if hashed bit has gone, leave the job to somebody who is doing it */
334 if (test_and_clear_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags) == 0)
337 /* drop ref inside spin lock to prevent race with other operations */
338 spin_lock_if(&sec->ps_lock, !locked);
340 if (gsec_kr->gsk_root_ctx == ctx)
341 gsec_kr->gsk_root_ctx = NULL;
342 hlist_del_init(&ctx->cc_cache);
343 atomic_dec(&ctx->cc_refcount);
345 spin_unlock_if(&sec->ps_lock, !locked);
351 * bind a key with a ctx together.
352 * caller must hold write lock of the key, as well as ref on key & ctx.
354 static void bind_key_ctx(struct key *key, struct ptlrpc_cli_ctx *ctx)
356 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
357 LASSERT(atomic_read(&key->usage) > 0);
358 LASSERT(ctx2gctx_keyring(ctx)->gck_key == NULL);
359 LASSERT(key->payload.data == NULL);
361 /* at this time context may or may not in list. */
363 atomic_inc(&ctx->cc_refcount);
364 ctx2gctx_keyring(ctx)->gck_key = key;
365 key->payload.data = ctx;
369 * unbind a key and a ctx.
370 * caller must hold write lock, as well as a ref of the key.
372 static void unbind_key_ctx(struct key *key, struct ptlrpc_cli_ctx *ctx)
374 LASSERT(key->payload.data == ctx);
375 LASSERT(test_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags) == 0);
377 /* must revoke the key, or others may treat it as newly created */
378 key_revoke_locked(key);
380 key->payload.data = NULL;
381 ctx2gctx_keyring(ctx)->gck_key = NULL;
383 /* once ctx get split from key, the timer is meaningless */
384 ctx_clear_timer_kr(ctx);
391 * given a ctx, unbind with its coupled key, if any.
392 * unbind could only be called once, so we don't worry the key be released
395 static void unbind_ctx_kr(struct ptlrpc_cli_ctx *ctx)
397 struct key *key = ctx2gctx_keyring(ctx)->gck_key;
400 LASSERT(key->payload.data == ctx);
403 down_write(&key->sem);
404 unbind_key_ctx(key, ctx);
411 * given a key, unbind with its coupled ctx, if any.
412 * caller must hold write lock, as well as a ref of the key.
414 static void unbind_key_locked(struct key *key)
416 struct ptlrpc_cli_ctx *ctx = key->payload.data;
419 unbind_key_ctx(key, ctx);
423 * unlist a ctx, and unbind from coupled key
425 static void kill_ctx_kr(struct ptlrpc_cli_ctx *ctx)
427 if (ctx_unlist_kr(ctx, 0))
432 * given a key, unlist and unbind with the coupled ctx (if any).
433 * caller must hold write lock, as well as a ref of the key.
435 static void kill_key_locked(struct key *key)
437 struct ptlrpc_cli_ctx *ctx = key->payload.data;
439 if (ctx && ctx_unlist_kr(ctx, 0))
440 unbind_key_locked(key);
444 * caller should hold one ref on contexts in freelist.
446 static void dispose_ctx_list_kr(struct hlist_head *freelist)
448 struct hlist_node __maybe_unused *pos, *next;
449 struct ptlrpc_cli_ctx *ctx;
450 struct gss_cli_ctx *gctx;
452 cfs_hlist_for_each_entry_safe(ctx, pos, next, freelist, cc_cache) {
453 hlist_del_init(&ctx->cc_cache);
455 /* reverse ctx: update current seq to buddy svcctx if exist.
456 * ideally this should be done at gss_cli_ctx_finalize(), but
457 * the ctx destroy could be delayed by:
458 * 1) ctx still has reference;
459 * 2) ctx destroy is asynchronous;
460 * and reverse import call inval_all_ctx() require this be done
461 * _immediately_ otherwise newly created reverse ctx might copy
462 * the very old sequence number from svcctx. */
463 gctx = ctx2gctx(ctx);
464 if (!rawobj_empty(&gctx->gc_svc_handle) &&
465 sec_is_reverse(gctx->gc_base.cc_sec)) {
466 gss_svc_upcall_update_sequence(&gctx->gc_svc_handle,
467 (__u32) atomic_read(&gctx->gc_seq));
470 /* we need to wakeup waiting reqs here. the context might
471 * be forced released before upcall finished, then the
472 * late-arrived downcall can't find the ctx even. */
473 sptlrpc_cli_ctx_wakeup(ctx);
481 * lookup a root context directly in a sec, return root ctx with a
482 * reference taken or NULL.
485 struct ptlrpc_cli_ctx * sec_lookup_root_ctx_kr(struct ptlrpc_sec *sec)
487 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
488 struct ptlrpc_cli_ctx *ctx = NULL;
490 spin_lock(&sec->ps_lock);
492 ctx = gsec_kr->gsk_root_ctx;
494 if (ctx == NULL && unlikely(sec_is_reverse(sec))) {
495 struct hlist_node __maybe_unused *node;
496 struct ptlrpc_cli_ctx *tmp;
498 /* reverse ctx, search root ctx in list, choose the one
499 * with shortest expire time, which is most possibly have
500 * an established peer ctx at client side. */
501 cfs_hlist_for_each_entry(tmp, node, &gsec_kr->gsk_clist,
503 if (ctx == NULL || ctx->cc_expire == 0 ||
504 ctx->cc_expire > tmp->cc_expire) {
506 /* promote to be root_ctx */
507 gsec_kr->gsk_root_ctx = ctx;
513 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
514 LASSERT(!hlist_empty(&gsec_kr->gsk_clist));
515 atomic_inc(&ctx->cc_refcount);
518 spin_unlock(&sec->ps_lock);
523 #define RVS_CTX_EXPIRE_NICE (10)
526 void rvs_sec_install_root_ctx_kr(struct ptlrpc_sec *sec,
527 struct ptlrpc_cli_ctx *new_ctx,
530 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
531 struct hlist_node __maybe_unused *hnode;
532 struct ptlrpc_cli_ctx *ctx;
536 LASSERT(sec_is_reverse(sec));
538 spin_lock(&sec->ps_lock);
540 now = cfs_time_current_sec();
542 /* set all existing ctxs short expiry */
543 cfs_hlist_for_each_entry(ctx, hnode, &gsec_kr->gsk_clist, cc_cache) {
544 if (ctx->cc_expire > now + RVS_CTX_EXPIRE_NICE) {
545 ctx->cc_early_expire = 1;
546 ctx->cc_expire = now + RVS_CTX_EXPIRE_NICE;
550 /* if there's root_ctx there, instead obsolete the current
551 * immediately, we leave it continue operating for a little while.
552 * hopefully when the first backward rpc with newest ctx send out,
553 * the client side already have the peer ctx well established. */
554 ctx_enlist_kr(new_ctx, gsec_kr->gsk_root_ctx ? 0 : 1, 1);
557 bind_key_ctx(key, new_ctx);
559 spin_unlock(&sec->ps_lock);
562 static void construct_key_desc(void *buf, int bufsize,
563 struct ptlrpc_sec *sec, uid_t uid)
565 snprintf(buf, bufsize, "%d@%x", uid, sec->ps_id);
566 ((char *)buf)[bufsize - 1] = '\0';
569 /****************************************
571 ****************************************/
574 struct ptlrpc_sec * gss_sec_create_kr(struct obd_import *imp,
575 struct ptlrpc_svc_ctx *svcctx,
576 struct sptlrpc_flavor *sf)
578 struct gss_sec_keyring *gsec_kr;
581 OBD_ALLOC(gsec_kr, sizeof(*gsec_kr));
585 INIT_HLIST_HEAD(&gsec_kr->gsk_clist);
586 gsec_kr->gsk_root_ctx = NULL;
587 mutex_init(&gsec_kr->gsk_root_uc_lock);
588 #ifdef HAVE_KEYRING_UPCALL_SERIALIZED
589 mutex_init(&gsec_kr->gsk_uc_lock);
592 if (gss_sec_create_common(&gsec_kr->gsk_base, &gss_policy_keyring,
596 if (svcctx != NULL &&
597 sec_install_rctx_kr(&gsec_kr->gsk_base.gs_base, svcctx)) {
598 gss_sec_destroy_common(&gsec_kr->gsk_base);
602 RETURN(&gsec_kr->gsk_base.gs_base);
605 OBD_FREE(gsec_kr, sizeof(*gsec_kr));
610 void gss_sec_destroy_kr(struct ptlrpc_sec *sec)
612 struct gss_sec *gsec = sec2gsec(sec);
613 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
615 CDEBUG(D_SEC, "destroy %s@%p\n", sec->ps_policy->sp_name, sec);
617 LASSERT(hlist_empty(&gsec_kr->gsk_clist));
618 LASSERT(gsec_kr->gsk_root_ctx == NULL);
620 gss_sec_destroy_common(gsec);
622 OBD_FREE(gsec_kr, sizeof(*gsec_kr));
625 static inline int user_is_root(struct ptlrpc_sec *sec, struct vfs_cred *vcred)
627 /* except the ROOTONLY flag, treat it as root user only if real uid
628 * is 0, euid/fsuid being 0 are handled as setuid scenarios */
629 if (sec_is_rootonly(sec) || (vcred->vc_uid == 0))
636 * unlink request key from it's ring, which is linked during request_key().
637 * sadly, we have to 'guess' which keyring it's linked to.
639 * FIXME this code is fragile, depend on how request_key_link() is implemented.
641 static void request_key_unlink(struct key *key)
643 struct task_struct *tsk = current;
646 switch (key_cred(tsk)->jit_keyring) {
647 case KEY_REQKEY_DEFL_DEFAULT:
648 case KEY_REQKEY_DEFL_THREAD_KEYRING:
649 ring = key_get(key_cred(tsk)->thread_keyring);
652 case KEY_REQKEY_DEFL_PROCESS_KEYRING:
653 ring = key_get(key_tgcred(tsk)->process_keyring);
656 case KEY_REQKEY_DEFL_SESSION_KEYRING:
658 ring = key_get(rcu_dereference(key_tgcred(tsk)
663 case KEY_REQKEY_DEFL_USER_SESSION_KEYRING:
664 ring = key_get(key_cred(tsk)->user->session_keyring);
666 case KEY_REQKEY_DEFL_USER_KEYRING:
667 ring = key_get(key_cred(tsk)->user->uid_keyring);
669 case KEY_REQKEY_DEFL_GROUP_KEYRING:
675 key_unlink(ring, key);
680 struct ptlrpc_cli_ctx * gss_sec_lookup_ctx_kr(struct ptlrpc_sec *sec,
681 struct vfs_cred *vcred,
682 int create, int remove_dead)
684 struct obd_import *imp = sec->ps_import;
685 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
686 struct ptlrpc_cli_ctx *ctx = NULL;
687 unsigned int is_root = 0, create_new = 0;
695 LASSERT(imp != NULL);
697 is_root = user_is_root(sec, vcred);
699 /* a little bit optimization for root context */
701 ctx = sec_lookup_root_ctx_kr(sec);
703 * Only lookup directly for REVERSE sec, which should
706 if (ctx || sec_is_reverse(sec))
710 LASSERT(create != 0);
712 /* for root context, obtain lock and check again, this time hold
713 * the root upcall lock, make sure nobody else populated new root
714 * context after last check. */
716 mutex_lock(&gsec_kr->gsk_root_uc_lock);
718 ctx = sec_lookup_root_ctx_kr(sec);
722 /* update reverse handle for root user */
723 sec2gsec(sec)->gs_rvs_hdl = gss_get_next_ctx_index();
725 switch (sec->ps_part) {
744 /* in case of setuid, key will be constructed as owner of fsuid/fsgid,
745 * but we do authentication based on real uid/gid. the key permission
746 * bits will be exactly as POS_ALL, so only processes who subscribed
747 * this key could have the access, although the quota might be counted
748 * on others (fsuid/fsgid).
750 * keyring will use fsuid/fsgid as upcall parameters, so we have to
751 * encode real uid/gid into callout info.
754 /* But first we need to make sure the obd type is supported */
755 if (strcmp(imp->imp_obd->obd_type->typ_name, LUSTRE_MDC_NAME) &&
756 strcmp(imp->imp_obd->obd_type->typ_name, LUSTRE_OSC_NAME) &&
757 strcmp(imp->imp_obd->obd_type->typ_name, LUSTRE_MGC_NAME) &&
758 strcmp(imp->imp_obd->obd_type->typ_name, LUSTRE_LWP_NAME) &&
759 strcmp(imp->imp_obd->obd_type->typ_name, LUSTRE_OSP_NAME)) {
760 CERROR("obd %s is not a supported device\n",
761 imp->imp_obd->obd_name);
762 GOTO(out, ctx = NULL);
765 construct_key_desc(desc, sizeof(desc), sec, vcred->vc_uid);
767 /* callout info format:
768 * secid:mech:uid:gid:flags:svc_type:peer_nid:target_uuid
770 coinfo_size = sizeof(struct obd_uuid) + MAX_OBD_NAME + 64;
771 OBD_ALLOC(coinfo, coinfo_size);
775 snprintf(coinfo, coinfo_size, "%d:%s:%u:%u:%s:%d:"LPX64":%s:"LPX64,
776 sec->ps_id, sec2gsec(sec)->gs_mech->gm_name,
777 vcred->vc_uid, vcred->vc_gid,
778 co_flags, import_to_gss_svc(imp),
779 imp->imp_connection->c_peer.nid, imp->imp_obd->obd_name,
780 imp->imp_connection->c_self);
782 CDEBUG(D_SEC, "requesting key for %s\n", desc);
784 keyring_upcall_lock(gsec_kr);
785 key = request_key(&gss_key_type, desc, coinfo);
786 keyring_upcall_unlock(gsec_kr);
788 OBD_FREE(coinfo, coinfo_size);
791 CERROR("failed request key: %ld\n", PTR_ERR(key));
794 CDEBUG(D_SEC, "obtained key %08x for %s\n", key->serial, desc);
796 /* once payload.data was pointed to a ctx, it never changes until
797 * we de-associate them; but parallel request_key() may return
798 * a key with payload.data == NULL at the same time. so we still
799 * need wirtelock of key->sem to serialize them. */
800 down_write(&key->sem);
802 if (likely(key->payload.data != NULL)) {
803 ctx = key->payload.data;
805 LASSERT(atomic_read(&ctx->cc_refcount) >= 1);
806 LASSERT(ctx2gctx_keyring(ctx)->gck_key == key);
807 LASSERT(atomic_read(&key->usage) >= 2);
809 /* simply take a ref and return. it's upper layer's
810 * responsibility to detect & replace dead ctx. */
811 atomic_inc(&ctx->cc_refcount);
813 /* pre initialization with a cli_ctx. this can't be done in
814 * key_instantiate() because we'v no enough information
816 ctx = ctx_create_kr(sec, vcred);
818 ctx_enlist_kr(ctx, is_root, 0);
819 bind_key_ctx(key, ctx);
821 ctx_start_timer_kr(ctx, KEYRING_UPCALL_TIMEOUT);
823 CDEBUG(D_SEC, "installed key %p <-> ctx %p (sec %p)\n",
826 /* we'd prefer to call key_revoke(), but we more like
827 * to revoke it within this key->sem locked period. */
828 key_revoke_locked(key);
836 if (is_root && create_new)
837 request_key_unlink(key);
842 mutex_unlock(&gsec_kr->gsk_root_uc_lock);
847 void gss_sec_release_ctx_kr(struct ptlrpc_sec *sec,
848 struct ptlrpc_cli_ctx *ctx,
851 LASSERT(atomic_read(&sec->ps_refcount) > 0);
852 LASSERT(atomic_read(&ctx->cc_refcount) == 0);
853 ctx_release_kr(ctx, sync);
857 * flush context of normal user, we must resort to keyring itself to find out
858 * contexts which belong to me.
860 * Note here we suppose only to flush _my_ context, the "uid" will
861 * be ignored in the search.
864 void flush_user_ctx_cache_kr(struct ptlrpc_sec *sec,
866 int grace, int force)
871 /* nothing to do for reverse or rootonly sec */
872 if (sec_is_reverse(sec) || sec_is_rootonly(sec))
875 construct_key_desc(desc, sizeof(desc), sec, uid);
877 /* there should be only one valid key, but we put it in the
878 * loop in case of any weird cases */
880 key = request_key(&gss_key_type, desc, NULL);
882 CDEBUG(D_SEC, "No more key found for current user\n");
886 down_write(&key->sem);
888 kill_key_locked(key);
890 /* kill_key_locked() should usually revoke the key, but we
891 * revoke it again to make sure, e.g. some case the key may
892 * not well coupled with a context. */
893 key_revoke_locked(key);
902 * flush context of root or all, we iterate through the list.
905 void flush_spec_ctx_cache_kr(struct ptlrpc_sec *sec, uid_t uid, int grace,
908 struct gss_sec_keyring *gsec_kr;
909 struct hlist_head freelist = HLIST_HEAD_INIT;
910 struct hlist_node __maybe_unused *pos, *next;
911 struct ptlrpc_cli_ctx *ctx;
914 gsec_kr = sec2gsec_keyring(sec);
916 spin_lock(&sec->ps_lock);
917 cfs_hlist_for_each_entry_safe(ctx, pos, next,
918 &gsec_kr->gsk_clist, cc_cache) {
919 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
921 if (uid != -1 && uid != ctx->cc_vcred.vc_uid)
924 /* at this moment there's at least 2 base reference:
925 * key association and in-list. */
926 if (atomic_read(&ctx->cc_refcount) > 2) {
929 CWARN("flush busy ctx %p(%u->%s, extra ref %d)\n",
930 ctx, ctx->cc_vcred.vc_uid,
931 sec2target_str(ctx->cc_sec),
932 atomic_read(&ctx->cc_refcount) - 2);
935 set_bit(PTLRPC_CTX_DEAD_BIT, &ctx->cc_flags);
937 clear_bit(PTLRPC_CTX_UPTODATE_BIT, &ctx->cc_flags);
939 atomic_inc(&ctx->cc_refcount);
941 if (ctx_unlist_kr(ctx, 1)) {
942 hlist_add_head(&ctx->cc_cache, &freelist);
944 LASSERT(atomic_read(&ctx->cc_refcount) >= 2);
945 atomic_dec(&ctx->cc_refcount);
948 spin_unlock(&sec->ps_lock);
950 dispose_ctx_list_kr(&freelist);
955 int gss_sec_flush_ctx_cache_kr(struct ptlrpc_sec *sec,
956 uid_t uid, int grace, int force)
960 CDEBUG(D_SEC, "sec %p(%d, nctx %d), uid %d, grace %d, force %d\n",
961 sec, atomic_read(&sec->ps_refcount),
962 atomic_read(&sec->ps_nctx),
965 if (uid != -1 && uid != 0)
966 flush_user_ctx_cache_kr(sec, uid, grace, force);
968 flush_spec_ctx_cache_kr(sec, uid, grace, force);
974 void gss_sec_gc_ctx_kr(struct ptlrpc_sec *sec)
976 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
977 struct hlist_head freelist = HLIST_HEAD_INIT;
978 struct hlist_node __maybe_unused *pos, *next;
979 struct ptlrpc_cli_ctx *ctx;
982 CWARN("running gc\n");
984 spin_lock(&sec->ps_lock);
985 cfs_hlist_for_each_entry_safe(ctx, pos, next,
986 &gsec_kr->gsk_clist, cc_cache) {
987 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
989 atomic_inc(&ctx->cc_refcount);
991 if (cli_ctx_check_death(ctx) && ctx_unlist_kr(ctx, 1)) {
992 hlist_add_head(&ctx->cc_cache, &freelist);
993 CWARN("unhashed ctx %p\n", ctx);
995 LASSERT(atomic_read(&ctx->cc_refcount) >= 2);
996 atomic_dec(&ctx->cc_refcount);
999 spin_unlock(&sec->ps_lock);
1001 dispose_ctx_list_kr(&freelist);
1007 int gss_sec_display_kr(struct ptlrpc_sec *sec, struct seq_file *seq)
1009 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
1010 struct hlist_node __maybe_unused *pos, *next;
1011 struct ptlrpc_cli_ctx *ctx;
1012 struct gss_cli_ctx *gctx;
1013 time_t now = cfs_time_current_sec();
1016 spin_lock(&sec->ps_lock);
1017 cfs_hlist_for_each_entry_safe(ctx, pos, next,
1018 &gsec_kr->gsk_clist, cc_cache) {
1023 gctx = ctx2gctx(ctx);
1024 key = ctx2gctx_keyring(ctx)->gck_key;
1026 gss_cli_ctx_flags2str(ctx->cc_flags,
1027 flags_str, sizeof(flags_str));
1029 if (gctx->gc_mechctx)
1030 lgss_display(gctx->gc_mechctx, mech, sizeof(mech));
1032 snprintf(mech, sizeof(mech), "N/A");
1033 mech[sizeof(mech) - 1] = '\0';
1035 seq_printf(seq, "%p: uid %u, ref %d, expire %ld(%+ld), fl %s, "
1036 "seq %d, win %u, key %08x(ref %d), "
1037 "hdl "LPX64":"LPX64", mech: %s\n",
1038 ctx, ctx->cc_vcred.vc_uid,
1039 atomic_read(&ctx->cc_refcount),
1041 ctx->cc_expire ? ctx->cc_expire - now : 0,
1043 atomic_read(&gctx->gc_seq),
1045 key ? key->serial : 0,
1046 key ? atomic_read(&key->usage) : 0,
1047 gss_handle_to_u64(&gctx->gc_handle),
1048 gss_handle_to_u64(&gctx->gc_svc_handle),
1051 spin_unlock(&sec->ps_lock);
1056 /****************************************
1058 ****************************************/
1061 int gss_cli_ctx_refresh_kr(struct ptlrpc_cli_ctx *ctx)
1063 /* upcall is already on the way */
1068 int gss_cli_ctx_validate_kr(struct ptlrpc_cli_ctx *ctx)
1070 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
1071 LASSERT(ctx->cc_sec);
1073 if (cli_ctx_check_death(ctx)) {
1078 if (cli_ctx_is_ready(ctx))
1084 void gss_cli_ctx_die_kr(struct ptlrpc_cli_ctx *ctx, int grace)
1086 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
1087 LASSERT(ctx->cc_sec);
1089 cli_ctx_expire(ctx);
1093 /****************************************
1094 * (reverse) service *
1095 ****************************************/
1098 * reverse context could have nothing to do with keyrings. here we still keep
1099 * the version which bind to a key, for future reference.
1101 #define HAVE_REVERSE_CTX_NOKEY
1103 #ifdef HAVE_REVERSE_CTX_NOKEY
1106 int sec_install_rctx_kr(struct ptlrpc_sec *sec,
1107 struct ptlrpc_svc_ctx *svc_ctx)
1109 struct ptlrpc_cli_ctx *cli_ctx;
1110 struct vfs_cred vcred = { 0, 0 };
1116 cli_ctx = ctx_create_kr(sec, &vcred);
1117 if (cli_ctx == NULL)
1120 rc = gss_copy_rvc_cli_ctx(cli_ctx, svc_ctx);
1122 CERROR("failed copy reverse cli ctx: %d\n", rc);
1124 ctx_put_kr(cli_ctx, 1);
1128 rvs_sec_install_root_ctx_kr(sec, cli_ctx, NULL);
1130 ctx_put_kr(cli_ctx, 1);
1135 #else /* ! HAVE_REVERSE_CTX_NOKEY */
1138 int sec_install_rctx_kr(struct ptlrpc_sec *sec,
1139 struct ptlrpc_svc_ctx *svc_ctx)
1141 struct ptlrpc_cli_ctx *cli_ctx = NULL;
1143 struct vfs_cred vcred = { 0, 0 };
1151 construct_key_desc(desc, sizeof(desc), sec, 0);
1153 key = key_alloc(&gss_key_type, desc, 0, 0,
1154 KEY_POS_ALL | KEY_USR_ALL, 1);
1156 CERROR("failed to alloc key: %ld\n", PTR_ERR(key));
1157 return PTR_ERR(key);
1160 rc = key_instantiate_and_link(key, NULL, 0, NULL, NULL);
1162 CERROR("failed to instantiate key: %d\n", rc);
1166 down_write(&key->sem);
1168 LASSERT(key->payload.data == NULL);
1170 cli_ctx = ctx_create_kr(sec, &vcred);
1171 if (cli_ctx == NULL) {
1176 rc = gss_copy_rvc_cli_ctx(cli_ctx, svc_ctx);
1178 CERROR("failed copy reverse cli ctx: %d\n", rc);
1182 rvs_sec_install_root_ctx_kr(sec, cli_ctx, key);
1184 ctx_put_kr(cli_ctx, 1);
1185 up_write(&key->sem);
1194 ctx_put_kr(cli_ctx, 1);
1196 up_write(&key->sem);
1202 #endif /* HAVE_REVERSE_CTX_NOKEY */
1204 /****************************************
1206 ****************************************/
1209 int gss_svc_accept_kr(struct ptlrpc_request *req)
1211 return gss_svc_accept(&gss_policy_keyring, req);
1215 int gss_svc_install_rctx_kr(struct obd_import *imp,
1216 struct ptlrpc_svc_ctx *svc_ctx)
1218 struct ptlrpc_sec *sec;
1221 sec = sptlrpc_import_sec_ref(imp);
1224 rc = sec_install_rctx_kr(sec, svc_ctx);
1225 sptlrpc_sec_put(sec);
1230 /****************************************
1232 ****************************************/
1235 #ifdef HAVE_KEY_TYPE_INSTANTIATE_2ARGS
1236 int gss_kt_instantiate(struct key *key, struct key_preparsed_payload *prep)
1238 const void *data = prep->data;
1239 size_t datalen = prep->datalen;
1241 int gss_kt_instantiate(struct key *key, const void *data, size_t datalen)
1247 if (data != NULL || datalen != 0) {
1248 CERROR("invalid: data %p, len %lu\n", data, (long)datalen);
1252 if (key->payload.data != NULL) {
1253 CERROR("key already have payload\n");
1257 /* link the key to session keyring, so following context negotiation
1258 * rpc fired from user space could find this key. This will be unlinked
1259 * automatically when upcall processes die.
1261 * we can't do this through keyctl from userspace, because the upcall
1262 * might be neither possessor nor owner of the key (setuid).
1264 * the session keyring is created upon upcall, and don't change all
1265 * the way until upcall finished, so rcu lock is not needed here.
1267 LASSERT(key_tgcred(current)->session_keyring);
1270 rc = key_link(key_tgcred(current)->session_keyring, key);
1273 CERROR("failed to link key %08x to keyring %08x: %d\n",
1275 key_tgcred(current)->session_keyring->serial, rc);
1279 CDEBUG(D_SEC, "key %p instantiated, ctx %p\n", key, key->payload.data);
1284 * called with key semaphore write locked. it means we can operate
1285 * on the context without fear of loosing refcount.
1288 #ifdef HAVE_KEY_TYPE_INSTANTIATE_2ARGS
1289 int gss_kt_update(struct key *key, struct key_preparsed_payload *prep)
1291 const void *data = prep->data;
1292 __u32 datalen32 = (__u32) prep->datalen;
1294 int gss_kt_update(struct key *key, const void *data, size_t datalen)
1296 __u32 datalen32 = (__u32) datalen;
1298 struct ptlrpc_cli_ctx *ctx = key->payload.data;
1299 struct gss_cli_ctx *gctx;
1300 rawobj_t tmpobj = RAWOBJ_EMPTY;
1304 if (data == NULL || datalen32 == 0) {
1305 CWARN("invalid: data %p, len %lu\n", data, (long)datalen32);
1309 /* if upcall finished negotiation too fast (mostly likely because
1310 * of local error happened) and call kt_update(), the ctx
1311 * might be still NULL. but the key will finally be associate
1312 * with a context, or be revoked. if key status is fine, return
1313 * -EAGAIN to allow userspace sleep a while and call again. */
1315 CDEBUG(D_SEC, "update too soon: key %p(%x) flags %lx\n",
1316 key, key->serial, key->flags);
1318 rc = key_validate(key);
1325 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
1326 LASSERT(ctx->cc_sec);
1328 ctx_clear_timer_kr(ctx);
1330 /* don't proceed if already refreshed */
1331 if (cli_ctx_is_refreshed(ctx)) {
1332 CWARN("ctx already done refresh\n");
1336 sptlrpc_cli_ctx_get(ctx);
1337 gctx = ctx2gctx(ctx);
1339 rc = buffer_extract_bytes(&data, &datalen32, &gctx->gc_win,
1340 sizeof(gctx->gc_win));
1342 CERROR("failed extract seq_win\n");
1346 if (gctx->gc_win == 0) {
1347 __u32 nego_rpc_err, nego_gss_err;
1349 rc = buffer_extract_bytes(&data, &datalen32, &nego_rpc_err,
1350 sizeof(nego_rpc_err));
1352 CERROR("failed to extrace rpc rc\n");
1356 rc = buffer_extract_bytes(&data, &datalen32, &nego_gss_err,
1357 sizeof(nego_gss_err));
1359 CERROR("failed to extrace gss rc\n");
1363 CERROR("negotiation: rpc err %d, gss err %x\n",
1364 nego_rpc_err, nego_gss_err);
1366 rc = nego_rpc_err ? nego_rpc_err : -EACCES;
1368 rc = rawobj_extract_local_alloc(&gctx->gc_handle,
1369 (__u32 **) &data, &datalen32);
1371 CERROR("failed extract handle\n");
1375 rc = rawobj_extract_local(&tmpobj, (__u32 **) &data,&datalen32);
1377 CERROR("failed extract mech\n");
1381 rc = lgss_import_sec_context(&tmpobj,
1382 sec2gsec(ctx->cc_sec)->gs_mech,
1384 if (rc != GSS_S_COMPLETE)
1385 CERROR("failed import context\n");
1390 /* we don't care what current status of this ctx, even someone else
1391 * is operating on the ctx at the same time. we just add up our own
1394 gss_cli_ctx_uptodate(gctx);
1396 /* this will also revoke the key. has to be done before
1397 * wakeup waiters otherwise they can find the stale key */
1398 kill_key_locked(key);
1400 cli_ctx_expire(ctx);
1402 if (rc != -ERESTART)
1403 set_bit(PTLRPC_CTX_ERROR_BIT, &ctx->cc_flags);
1406 /* let user space think it's a success */
1407 sptlrpc_cli_ctx_put(ctx, 1);
1411 #ifndef HAVE_KEY_MATCH_DATA
1413 gss_kt_match(const struct key *key, const void *desc)
1415 return (strcmp(key->description, (const char *) desc) == 0);
1417 #else /* ! HAVE_KEY_MATCH_DATA */
1419 gss_kt_match(const struct key *key, const struct key_match_data *match_data)
1421 const char *desc = match_data->raw_data;
1423 return (strcmp(key->description, desc) == 0);
1427 * Preparse the match criterion.
1429 static int gss_kt_match_preparse(struct key_match_data *match_data)
1431 match_data->lookup_type = KEYRING_SEARCH_LOOKUP_DIRECT;
1432 match_data->cmp = gss_kt_match;
1435 #endif /* HAVE_KEY_MATCH_DATA */
1438 void gss_kt_destroy(struct key *key)
1441 LASSERT(key->payload.data == NULL);
1442 CDEBUG(D_SEC, "destroy key %p\n", key);
1447 void gss_kt_describe(const struct key *key, struct seq_file *s)
1449 if (key->description == NULL)
1450 seq_puts(s, "[null]");
1452 seq_puts(s, key->description);
1455 static struct key_type gss_key_type =
1459 .instantiate = gss_kt_instantiate,
1460 .update = gss_kt_update,
1461 #ifdef HAVE_KEY_MATCH_DATA
1462 .match_preparse = gss_kt_match_preparse,
1464 .match = gss_kt_match,
1466 .destroy = gss_kt_destroy,
1467 .describe = gss_kt_describe,
1470 /****************************************
1471 * lustre gss keyring policy *
1472 ****************************************/
1474 static struct ptlrpc_ctx_ops gss_keyring_ctxops = {
1475 .match = gss_cli_ctx_match,
1476 .refresh = gss_cli_ctx_refresh_kr,
1477 .validate = gss_cli_ctx_validate_kr,
1478 .die = gss_cli_ctx_die_kr,
1479 .sign = gss_cli_ctx_sign,
1480 .verify = gss_cli_ctx_verify,
1481 .seal = gss_cli_ctx_seal,
1482 .unseal = gss_cli_ctx_unseal,
1483 .wrap_bulk = gss_cli_ctx_wrap_bulk,
1484 .unwrap_bulk = gss_cli_ctx_unwrap_bulk,
1487 static struct ptlrpc_sec_cops gss_sec_keyring_cops = {
1488 .create_sec = gss_sec_create_kr,
1489 .destroy_sec = gss_sec_destroy_kr,
1490 .kill_sec = gss_sec_kill,
1491 .lookup_ctx = gss_sec_lookup_ctx_kr,
1492 .release_ctx = gss_sec_release_ctx_kr,
1493 .flush_ctx_cache = gss_sec_flush_ctx_cache_kr,
1494 .gc_ctx = gss_sec_gc_ctx_kr,
1495 .install_rctx = gss_sec_install_rctx,
1496 .alloc_reqbuf = gss_alloc_reqbuf,
1497 .free_reqbuf = gss_free_reqbuf,
1498 .alloc_repbuf = gss_alloc_repbuf,
1499 .free_repbuf = gss_free_repbuf,
1500 .enlarge_reqbuf = gss_enlarge_reqbuf,
1501 .display = gss_sec_display_kr,
1504 static struct ptlrpc_sec_sops gss_sec_keyring_sops = {
1505 .accept = gss_svc_accept_kr,
1506 .invalidate_ctx = gss_svc_invalidate_ctx,
1507 .alloc_rs = gss_svc_alloc_rs,
1508 .authorize = gss_svc_authorize,
1509 .free_rs = gss_svc_free_rs,
1510 .free_ctx = gss_svc_free_ctx,
1511 .prep_bulk = gss_svc_prep_bulk,
1512 .unwrap_bulk = gss_svc_unwrap_bulk,
1513 .wrap_bulk = gss_svc_wrap_bulk,
1514 .install_rctx = gss_svc_install_rctx_kr,
1517 static struct ptlrpc_sec_policy gss_policy_keyring = {
1518 .sp_owner = THIS_MODULE,
1519 .sp_name = "gss.keyring",
1520 .sp_policy = SPTLRPC_POLICY_GSS,
1521 .sp_cops = &gss_sec_keyring_cops,
1522 .sp_sops = &gss_sec_keyring_sops,
1526 int __init gss_init_keyring(void)
1530 rc = register_key_type(&gss_key_type);
1532 CERROR("failed to register keyring type: %d\n", rc);
1536 rc = sptlrpc_register_policy(&gss_policy_keyring);
1538 unregister_key_type(&gss_key_type);
1545 void __exit gss_exit_keyring(void)
1547 unregister_key_type(&gss_key_type);
1548 sptlrpc_unregister_policy(&gss_policy_keyring);