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>
54 #include <libcfs/linux/linux-list.h>
56 #include <obd_class.h>
57 #include <obd_support.h>
58 #include <lustre/lustre_idl.h>
59 #include <lustre_sec.h>
60 #include <lustre_net.h>
61 #include <lustre_import.h>
64 #include "gss_internal.h"
67 static struct ptlrpc_sec_policy gss_policy_keyring;
68 static struct ptlrpc_ctx_ops gss_keyring_ctxops;
69 static struct key_type gss_key_type;
71 static int sec_install_rctx_kr(struct ptlrpc_sec *sec,
72 struct ptlrpc_svc_ctx *svc_ctx);
75 * the timeout is only for the case that upcall child process die abnormally.
76 * in any other cases it should finally update kernel key.
78 * FIXME we'd better to incorporate the client & server side upcall timeouts
79 * into the framework of Adaptive Timeouts, but we need to figure out how to
80 * make sure that kernel knows the upcall processes is in-progress or died
83 #define KEYRING_UPCALL_TIMEOUT (obd_timeout + obd_timeout)
85 /****************************************
87 ****************************************/
89 #define DUMP_PROCESS_KEYRINGS(tsk) \
91 CWARN("DUMP PK: %s[%u,%u/%u](<-%s[%u,%u/%u]): " \
92 "a %d, t %d, p %d, s %d, u %d, us %d, df %d\n", \
93 tsk->comm, tsk->pid, tsk->uid, tsk->fsuid, \
94 tsk->parent->comm, tsk->parent->pid, \
95 tsk->parent->uid, tsk->parent->fsuid, \
96 tsk->request_key_auth ? \
97 tsk->request_key_auth->serial : 0, \
98 key_cred(tsk)->thread_keyring ? \
99 key_cred(tsk)->thread_keyring->serial : 0, \
100 key_tgcred(tsk)->process_keyring ? \
101 key_tgcred(tsk)->process_keyring->serial : 0, \
102 key_tgcred(tsk)->session_keyring ? \
103 key_tgcred(tsk)->session_keyring->serial : 0, \
104 key_cred(tsk)->user->uid_keyring ? \
105 key_cred(tsk)->user->uid_keyring->serial : 0, \
106 key_cred(tsk)->user->session_keyring ? \
107 key_cred(tsk)->user->session_keyring->serial : 0, \
108 key_cred(tsk)->jit_keyring \
112 #define DUMP_KEY(key) \
114 CWARN("DUMP KEY: %p(%d) ref %d u%u/g%u desc %s\n", \
115 key, key->serial, atomic_read(&key->usage), \
116 key->uid, key->gid, \
117 key->description ? key->description : "n/a" \
121 #define key_cred(tsk) ((tsk)->cred)
122 #ifdef HAVE_CRED_TGCRED
123 #define key_tgcred(tsk) ((tsk)->cred->tgcred)
125 #define key_tgcred(tsk) key_cred(tsk)
128 static inline void keyring_upcall_lock(struct gss_sec_keyring *gsec_kr)
130 #ifdef HAVE_KEYRING_UPCALL_SERIALIZED
131 mutex_lock(&gsec_kr->gsk_uc_lock);
135 static inline void keyring_upcall_unlock(struct gss_sec_keyring *gsec_kr)
137 #ifdef HAVE_KEYRING_UPCALL_SERIALIZED
138 mutex_unlock(&gsec_kr->gsk_uc_lock);
142 static inline void key_revoke_locked(struct key *key)
144 set_bit(KEY_FLAG_REVOKED, &key->flags);
147 static void ctx_upcall_timeout_kr(unsigned long data)
149 struct ptlrpc_cli_ctx *ctx = (struct ptlrpc_cli_ctx *) data;
150 struct key *key = ctx2gctx_keyring(ctx)->gck_key;
152 CWARN("ctx %p, key %p\n", ctx, key);
157 key_revoke_locked(key);
160 static void ctx_start_timer_kr(struct ptlrpc_cli_ctx *ctx, long timeout)
162 struct gss_cli_ctx_keyring *gctx_kr = ctx2gctx_keyring(ctx);
163 struct timer_list *timer = gctx_kr->gck_timer;
167 CDEBUG(D_SEC, "ctx %p: start timer %lds\n", ctx, timeout);
168 timeout = msecs_to_jiffies(timeout * MSEC_PER_SEC) +
172 timer->expires = timeout;
173 timer->data = (unsigned long ) ctx;
174 timer->function = ctx_upcall_timeout_kr;
180 * caller should make sure no race with other threads
183 void ctx_clear_timer_kr(struct ptlrpc_cli_ctx *ctx)
185 struct gss_cli_ctx_keyring *gctx_kr = ctx2gctx_keyring(ctx);
186 struct timer_list *timer = gctx_kr->gck_timer;
191 CDEBUG(D_SEC, "ctx %p, key %p\n", ctx, gctx_kr->gck_key);
193 gctx_kr->gck_timer = NULL;
195 del_singleshot_timer_sync(timer);
201 struct ptlrpc_cli_ctx *ctx_create_kr(struct ptlrpc_sec *sec,
202 struct vfs_cred *vcred)
204 struct ptlrpc_cli_ctx *ctx;
205 struct gss_cli_ctx_keyring *gctx_kr;
207 OBD_ALLOC_PTR(gctx_kr);
211 OBD_ALLOC_PTR(gctx_kr->gck_timer);
212 if (gctx_kr->gck_timer == NULL) {
213 OBD_FREE_PTR(gctx_kr);
216 init_timer(gctx_kr->gck_timer);
218 ctx = &gctx_kr->gck_base.gc_base;
220 if (gss_cli_ctx_init_common(sec, ctx, &gss_keyring_ctxops, vcred)) {
221 OBD_FREE_PTR(gctx_kr->gck_timer);
222 OBD_FREE_PTR(gctx_kr);
226 ctx->cc_expire = cfs_time_current_sec() + KEYRING_UPCALL_TIMEOUT;
227 clear_bit(PTLRPC_CTX_NEW_BIT, &ctx->cc_flags);
228 atomic_inc(&ctx->cc_refcount); /* for the caller */
233 static void ctx_destroy_kr(struct ptlrpc_cli_ctx *ctx)
235 struct ptlrpc_sec *sec = ctx->cc_sec;
236 struct gss_cli_ctx_keyring *gctx_kr = ctx2gctx_keyring(ctx);
238 CDEBUG(D_SEC, "destroying ctx %p\n", ctx);
240 /* at this time the association with key has been broken. */
242 LASSERT(atomic_read(&sec->ps_refcount) > 0);
243 LASSERT(atomic_read(&sec->ps_nctx) > 0);
244 LASSERT(test_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags) == 0);
245 LASSERT(gctx_kr->gck_key == NULL);
247 ctx_clear_timer_kr(ctx);
248 LASSERT(gctx_kr->gck_timer == NULL);
250 if (gss_cli_ctx_fini_common(sec, ctx))
253 OBD_FREE_PTR(gctx_kr);
255 atomic_dec(&sec->ps_nctx);
256 sptlrpc_sec_put(sec);
259 static void ctx_release_kr(struct ptlrpc_cli_ctx *ctx, int sync)
264 atomic_inc(&ctx->cc_refcount);
265 sptlrpc_gc_add_ctx(ctx);
269 static void ctx_put_kr(struct ptlrpc_cli_ctx *ctx, int sync)
271 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
273 if (atomic_dec_and_test(&ctx->cc_refcount))
274 ctx_release_kr(ctx, sync);
278 * key <-> ctx association and rules:
279 * - ctx might not bind with any key
280 * - key/ctx binding is protected by key semaphore (if the key present)
281 * - key and ctx each take a reference of the other
282 * - ctx enlist/unlist is protected by ctx spinlock
283 * - never enlist a ctx after it's been unlisted
284 * - whoever do enlist should also do bind, lock key before enlist:
285 * - lock key -> lock ctx -> enlist -> unlock ctx -> bind -> unlock key
286 * - whoever do unlist should also do unbind:
287 * - lock key -> lock ctx -> unlist -> unlock ctx -> unbind -> unlock key
288 * - lock ctx -> unlist -> unlock ctx -> lock key -> unbind -> unlock key
291 static inline void spin_lock_if(spinlock_t *lock, int condition)
297 static inline void spin_unlock_if(spinlock_t *lock, int condition)
303 static void ctx_enlist_kr(struct ptlrpc_cli_ctx *ctx, int is_root, int locked)
305 struct ptlrpc_sec *sec = ctx->cc_sec;
306 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
308 LASSERT(!test_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags));
309 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
311 spin_lock_if(&sec->ps_lock, !locked);
313 atomic_inc(&ctx->cc_refcount);
314 set_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags);
315 hlist_add_head(&ctx->cc_cache, &gsec_kr->gsk_clist);
317 gsec_kr->gsk_root_ctx = ctx;
319 spin_unlock_if(&sec->ps_lock, !locked);
323 * Note after this get called, caller should not access ctx again because
324 * it might have been freed, unless caller hold at least one refcount of
327 * return non-zero if we indeed unlist this ctx.
329 static int ctx_unlist_kr(struct ptlrpc_cli_ctx *ctx, int locked)
331 struct ptlrpc_sec *sec = ctx->cc_sec;
332 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
334 /* if hashed bit has gone, leave the job to somebody who is doing it */
335 if (test_and_clear_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags) == 0)
338 /* drop ref inside spin lock to prevent race with other operations */
339 spin_lock_if(&sec->ps_lock, !locked);
341 if (gsec_kr->gsk_root_ctx == ctx)
342 gsec_kr->gsk_root_ctx = NULL;
343 hlist_del_init(&ctx->cc_cache);
344 atomic_dec(&ctx->cc_refcount);
346 spin_unlock_if(&sec->ps_lock, !locked);
352 * Get specific payload. Newer kernels support 4 slots.
355 key_get_payload(struct key *key, unsigned int index)
357 void *key_ptr = NULL;
359 #ifdef HAVE_KEY_PAYLOAD_DATA_ARRAY
360 key_ptr = key->payload.data[index];
363 key_ptr = key->payload.data;
369 * Set specific payload. Newer kernels support 4 slots.
371 static int key_set_payload(struct key *key, unsigned int index,
372 struct ptlrpc_cli_ctx *ctx)
376 #ifdef HAVE_KEY_PAYLOAD_DATA_ARRAY
378 key->payload.data[index] = ctx;
381 key->payload.data = ctx;
389 * bind a key with a ctx together.
390 * caller must hold write lock of the key, as well as ref on key & ctx.
392 static void bind_key_ctx(struct key *key, struct ptlrpc_cli_ctx *ctx)
394 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
395 LASSERT(atomic_read(&key->usage) > 0);
396 LASSERT(ctx2gctx_keyring(ctx)->gck_key == NULL);
397 LASSERT(!key_get_payload(key, 0));
399 /* at this time context may or may not in list. */
401 atomic_inc(&ctx->cc_refcount);
402 ctx2gctx_keyring(ctx)->gck_key = key;
403 LASSERT(!key_set_payload(key, 0, ctx));
407 * unbind a key and a ctx.
408 * caller must hold write lock, as well as a ref of the key.
410 static void unbind_key_ctx(struct key *key, struct ptlrpc_cli_ctx *ctx)
412 LASSERT(key_get_payload(key, 0) == ctx);
413 LASSERT(test_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags) == 0);
415 /* must revoke the key, or others may treat it as newly created */
416 key_revoke_locked(key);
418 key_set_payload(key, 0, NULL);
419 ctx2gctx_keyring(ctx)->gck_key = NULL;
421 /* once ctx get split from key, the timer is meaningless */
422 ctx_clear_timer_kr(ctx);
429 * given a ctx, unbind with its coupled key, if any.
430 * unbind could only be called once, so we don't worry the key be released
433 static void unbind_ctx_kr(struct ptlrpc_cli_ctx *ctx)
435 struct key *key = ctx2gctx_keyring(ctx)->gck_key;
438 LASSERT(key_get_payload(key, 0) == ctx);
441 down_write(&key->sem);
442 unbind_key_ctx(key, ctx);
449 * given a key, unbind with its coupled ctx, if any.
450 * caller must hold write lock, as well as a ref of the key.
452 static void unbind_key_locked(struct key *key)
454 struct ptlrpc_cli_ctx *ctx = key_get_payload(key, 0);
457 unbind_key_ctx(key, ctx);
461 * unlist a ctx, and unbind from coupled key
463 static void kill_ctx_kr(struct ptlrpc_cli_ctx *ctx)
465 if (ctx_unlist_kr(ctx, 0))
470 * given a key, unlist and unbind with the coupled ctx (if any).
471 * caller must hold write lock, as well as a ref of the key.
473 static void kill_key_locked(struct key *key)
475 struct ptlrpc_cli_ctx *ctx = key_get_payload(key, 0);
477 if (ctx && ctx_unlist_kr(ctx, 0))
478 unbind_key_locked(key);
482 * caller should hold one ref on contexts in freelist.
484 static void dispose_ctx_list_kr(struct hlist_head *freelist)
486 struct hlist_node __maybe_unused *pos, *next;
487 struct ptlrpc_cli_ctx *ctx;
488 struct gss_cli_ctx *gctx;
490 cfs_hlist_for_each_entry_safe(ctx, pos, next, freelist, cc_cache) {
491 hlist_del_init(&ctx->cc_cache);
493 /* reverse ctx: update current seq to buddy svcctx if exist.
494 * ideally this should be done at gss_cli_ctx_finalize(), but
495 * the ctx destroy could be delayed by:
496 * 1) ctx still has reference;
497 * 2) ctx destroy is asynchronous;
498 * and reverse import call inval_all_ctx() require this be done
499 * _immediately_ otherwise newly created reverse ctx might copy
500 * the very old sequence number from svcctx. */
501 gctx = ctx2gctx(ctx);
502 if (!rawobj_empty(&gctx->gc_svc_handle) &&
503 sec_is_reverse(gctx->gc_base.cc_sec)) {
504 gss_svc_upcall_update_sequence(&gctx->gc_svc_handle,
505 (__u32) atomic_read(&gctx->gc_seq));
508 /* we need to wakeup waiting reqs here. the context might
509 * be forced released before upcall finished, then the
510 * late-arrived downcall can't find the ctx even. */
511 sptlrpc_cli_ctx_wakeup(ctx);
519 * lookup a root context directly in a sec, return root ctx with a
520 * reference taken or NULL.
523 struct ptlrpc_cli_ctx * sec_lookup_root_ctx_kr(struct ptlrpc_sec *sec)
525 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
526 struct ptlrpc_cli_ctx *ctx = NULL;
528 spin_lock(&sec->ps_lock);
530 ctx = gsec_kr->gsk_root_ctx;
532 if (ctx == NULL && unlikely(sec_is_reverse(sec))) {
533 struct hlist_node __maybe_unused *node;
534 struct ptlrpc_cli_ctx *tmp;
536 /* reverse ctx, search root ctx in list, choose the one
537 * with shortest expire time, which is most possibly have
538 * an established peer ctx at client side. */
539 cfs_hlist_for_each_entry(tmp, node, &gsec_kr->gsk_clist,
541 if (ctx == NULL || ctx->cc_expire == 0 ||
542 ctx->cc_expire > tmp->cc_expire) {
544 /* promote to be root_ctx */
545 gsec_kr->gsk_root_ctx = ctx;
551 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
552 LASSERT(!hlist_empty(&gsec_kr->gsk_clist));
553 atomic_inc(&ctx->cc_refcount);
556 spin_unlock(&sec->ps_lock);
561 #define RVS_CTX_EXPIRE_NICE (10)
564 void rvs_sec_install_root_ctx_kr(struct ptlrpc_sec *sec,
565 struct ptlrpc_cli_ctx *new_ctx,
568 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
569 struct hlist_node __maybe_unused *hnode;
570 struct ptlrpc_cli_ctx *ctx;
574 LASSERT(sec_is_reverse(sec));
576 spin_lock(&sec->ps_lock);
578 now = cfs_time_current_sec();
580 /* set all existing ctxs short expiry */
581 cfs_hlist_for_each_entry(ctx, hnode, &gsec_kr->gsk_clist, cc_cache) {
582 if (ctx->cc_expire > now + RVS_CTX_EXPIRE_NICE) {
583 ctx->cc_early_expire = 1;
584 ctx->cc_expire = now + RVS_CTX_EXPIRE_NICE;
588 /* if there's root_ctx there, instead obsolete the current
589 * immediately, we leave it continue operating for a little while.
590 * hopefully when the first backward rpc with newest ctx send out,
591 * the client side already have the peer ctx well established. */
592 ctx_enlist_kr(new_ctx, gsec_kr->gsk_root_ctx ? 0 : 1, 1);
595 bind_key_ctx(key, new_ctx);
597 spin_unlock(&sec->ps_lock);
600 static void construct_key_desc(void *buf, int bufsize,
601 struct ptlrpc_sec *sec, uid_t uid)
603 snprintf(buf, bufsize, "%d@%x", uid, sec->ps_id);
604 ((char *)buf)[bufsize - 1] = '\0';
607 /****************************************
609 ****************************************/
612 struct ptlrpc_sec * gss_sec_create_kr(struct obd_import *imp,
613 struct ptlrpc_svc_ctx *svcctx,
614 struct sptlrpc_flavor *sf)
616 struct gss_sec_keyring *gsec_kr;
619 OBD_ALLOC(gsec_kr, sizeof(*gsec_kr));
623 INIT_HLIST_HEAD(&gsec_kr->gsk_clist);
624 gsec_kr->gsk_root_ctx = NULL;
625 mutex_init(&gsec_kr->gsk_root_uc_lock);
626 #ifdef HAVE_KEYRING_UPCALL_SERIALIZED
627 mutex_init(&gsec_kr->gsk_uc_lock);
630 if (gss_sec_create_common(&gsec_kr->gsk_base, &gss_policy_keyring,
634 if (svcctx != NULL &&
635 sec_install_rctx_kr(&gsec_kr->gsk_base.gs_base, svcctx)) {
636 gss_sec_destroy_common(&gsec_kr->gsk_base);
640 RETURN(&gsec_kr->gsk_base.gs_base);
643 OBD_FREE(gsec_kr, sizeof(*gsec_kr));
648 void gss_sec_destroy_kr(struct ptlrpc_sec *sec)
650 struct gss_sec *gsec = sec2gsec(sec);
651 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
653 CDEBUG(D_SEC, "destroy %s@%p\n", sec->ps_policy->sp_name, sec);
655 LASSERT(hlist_empty(&gsec_kr->gsk_clist));
656 LASSERT(gsec_kr->gsk_root_ctx == NULL);
658 gss_sec_destroy_common(gsec);
660 OBD_FREE(gsec_kr, sizeof(*gsec_kr));
663 static inline int user_is_root(struct ptlrpc_sec *sec, struct vfs_cred *vcred)
665 /* except the ROOTONLY flag, treat it as root user only if real uid
666 * is 0, euid/fsuid being 0 are handled as setuid scenarios */
667 if (sec_is_rootonly(sec) || (vcred->vc_uid == 0))
674 * unlink request key from it's ring, which is linked during request_key().
675 * sadly, we have to 'guess' which keyring it's linked to.
677 * FIXME this code is fragile, depend on how request_key_link() is implemented.
679 static void request_key_unlink(struct key *key)
681 struct task_struct *tsk = current;
684 switch (key_cred(tsk)->jit_keyring) {
685 case KEY_REQKEY_DEFL_DEFAULT:
686 case KEY_REQKEY_DEFL_THREAD_KEYRING:
687 ring = key_get(key_cred(tsk)->thread_keyring);
690 case KEY_REQKEY_DEFL_PROCESS_KEYRING:
691 ring = key_get(key_tgcred(tsk)->process_keyring);
694 case KEY_REQKEY_DEFL_SESSION_KEYRING:
696 ring = key_get(rcu_dereference(key_tgcred(tsk)
701 case KEY_REQKEY_DEFL_USER_SESSION_KEYRING:
702 ring = key_get(key_cred(tsk)->user->session_keyring);
704 case KEY_REQKEY_DEFL_USER_KEYRING:
705 ring = key_get(key_cred(tsk)->user->uid_keyring);
707 case KEY_REQKEY_DEFL_GROUP_KEYRING:
713 key_unlink(ring, key);
718 struct ptlrpc_cli_ctx * gss_sec_lookup_ctx_kr(struct ptlrpc_sec *sec,
719 struct vfs_cred *vcred,
720 int create, int remove_dead)
722 struct obd_import *imp = sec->ps_import;
723 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
724 struct ptlrpc_cli_ctx *ctx = NULL;
725 unsigned int is_root = 0, create_new = 0;
730 const char *sec_part_flags = "";
731 char svc_flag = '\0';
734 LASSERT(imp != NULL);
736 is_root = user_is_root(sec, vcred);
738 /* a little bit optimization for root context */
740 ctx = sec_lookup_root_ctx_kr(sec);
742 * Only lookup directly for REVERSE sec, which should
745 if (ctx || sec_is_reverse(sec))
749 LASSERT(create != 0);
751 /* for root context, obtain lock and check again, this time hold
752 * the root upcall lock, make sure nobody else populated new root
753 * context after last check. */
755 mutex_lock(&gsec_kr->gsk_root_uc_lock);
757 ctx = sec_lookup_root_ctx_kr(sec);
761 /* update reverse handle for root user */
762 sec2gsec(sec)->gs_rvs_hdl = gss_get_next_ctx_index();
764 switch (sec->ps_part) {
766 sec_part_flags = "m";
769 sec_part_flags = "o";
772 sec_part_flags = "rmo";
775 sec_part_flags = "r";
782 switch (SPTLRPC_FLVR_SVC(sec->ps_flvr.sf_rpc)) {
783 case SPTLRPC_SVC_NULL:
786 case SPTLRPC_SVC_AUTH:
789 case SPTLRPC_SVC_INTG:
792 case SPTLRPC_SVC_PRIV:
800 /* in case of setuid, key will be constructed as owner of fsuid/fsgid,
801 * but we do authentication based on real uid/gid. the key permission
802 * bits will be exactly as POS_ALL, so only processes who subscribed
803 * this key could have the access, although the quota might be counted
804 * on others (fsuid/fsgid).
806 * keyring will use fsuid/fsgid as upcall parameters, so we have to
807 * encode real uid/gid into callout info.
810 /* But first we need to make sure the obd type is supported */
811 if (strcmp(imp->imp_obd->obd_type->typ_name, LUSTRE_MDC_NAME) &&
812 strcmp(imp->imp_obd->obd_type->typ_name, LUSTRE_OSC_NAME) &&
813 strcmp(imp->imp_obd->obd_type->typ_name, LUSTRE_MGC_NAME) &&
814 strcmp(imp->imp_obd->obd_type->typ_name, LUSTRE_LWP_NAME) &&
815 strcmp(imp->imp_obd->obd_type->typ_name, LUSTRE_OSP_NAME)) {
816 CERROR("obd %s is not a supported device\n",
817 imp->imp_obd->obd_name);
818 GOTO(out, ctx = NULL);
821 construct_key_desc(desc, sizeof(desc), sec, vcred->vc_uid);
823 /* callout info format:
824 * secid:mech:uid:gid:sec_flags:svc_flag:svc_type:peer_nid:target_uuid
826 coinfo_size = sizeof(struct obd_uuid) + MAX_OBD_NAME + 64;
827 OBD_ALLOC(coinfo, coinfo_size);
831 snprintf(coinfo, coinfo_size, "%d:%s:%u:%u:%s:%c:%d:%#llx:%s:%#llx",
832 sec->ps_id, sec2gsec(sec)->gs_mech->gm_name,
833 vcred->vc_uid, vcred->vc_gid,
834 sec_part_flags, svc_flag, import_to_gss_svc(imp),
835 imp->imp_connection->c_peer.nid, imp->imp_obd->obd_name,
836 imp->imp_connection->c_self);
838 CDEBUG(D_SEC, "requesting key for %s\n", desc);
840 keyring_upcall_lock(gsec_kr);
841 key = request_key(&gss_key_type, desc, coinfo);
842 keyring_upcall_unlock(gsec_kr);
844 OBD_FREE(coinfo, coinfo_size);
847 CERROR("failed request key: %ld\n", PTR_ERR(key));
850 CDEBUG(D_SEC, "obtained key %08x for %s\n", key->serial, desc);
852 /* once payload.data was pointed to a ctx, it never changes until
853 * we de-associate them; but parallel request_key() may return
854 * a key with payload.data == NULL at the same time. so we still
855 * need wirtelock of key->sem to serialize them. */
856 down_write(&key->sem);
858 ctx = key_get_payload(key, 0);
860 LASSERT(atomic_read(&ctx->cc_refcount) >= 1);
861 LASSERT(ctx2gctx_keyring(ctx)->gck_key == key);
862 LASSERT(atomic_read(&key->usage) >= 2);
864 /* simply take a ref and return. it's upper layer's
865 * responsibility to detect & replace dead ctx. */
866 atomic_inc(&ctx->cc_refcount);
868 /* pre initialization with a cli_ctx. this can't be done in
869 * key_instantiate() because we'v no enough information
871 ctx = ctx_create_kr(sec, vcred);
873 ctx_enlist_kr(ctx, is_root, 0);
874 bind_key_ctx(key, ctx);
876 ctx_start_timer_kr(ctx, KEYRING_UPCALL_TIMEOUT);
878 CDEBUG(D_SEC, "installed key %p <-> ctx %p (sec %p)\n",
881 /* we'd prefer to call key_revoke(), but we more like
882 * to revoke it within this key->sem locked period. */
883 key_revoke_locked(key);
891 if (is_root && create_new)
892 request_key_unlink(key);
897 mutex_unlock(&gsec_kr->gsk_root_uc_lock);
902 void gss_sec_release_ctx_kr(struct ptlrpc_sec *sec,
903 struct ptlrpc_cli_ctx *ctx,
906 LASSERT(atomic_read(&sec->ps_refcount) > 0);
907 LASSERT(atomic_read(&ctx->cc_refcount) == 0);
908 ctx_release_kr(ctx, sync);
912 * flush context of normal user, we must resort to keyring itself to find out
913 * contexts which belong to me.
915 * Note here we suppose only to flush _my_ context, the "uid" will
916 * be ignored in the search.
919 void flush_user_ctx_cache_kr(struct ptlrpc_sec *sec,
921 int grace, int force)
926 /* nothing to do for reverse or rootonly sec */
927 if (sec_is_reverse(sec) || sec_is_rootonly(sec))
930 construct_key_desc(desc, sizeof(desc), sec, uid);
932 /* there should be only one valid key, but we put it in the
933 * loop in case of any weird cases */
935 key = request_key(&gss_key_type, desc, NULL);
937 CDEBUG(D_SEC, "No more key found for current user\n");
941 down_write(&key->sem);
943 kill_key_locked(key);
945 /* kill_key_locked() should usually revoke the key, but we
946 * revoke it again to make sure, e.g. some case the key may
947 * not well coupled with a context. */
948 key_revoke_locked(key);
952 request_key_unlink(key);
959 * flush context of root or all, we iterate through the list.
962 void flush_spec_ctx_cache_kr(struct ptlrpc_sec *sec, uid_t uid, int grace,
965 struct gss_sec_keyring *gsec_kr;
966 struct hlist_head freelist = HLIST_HEAD_INIT;
967 struct hlist_node __maybe_unused *pos, *next;
968 struct ptlrpc_cli_ctx *ctx;
971 gsec_kr = sec2gsec_keyring(sec);
973 spin_lock(&sec->ps_lock);
974 cfs_hlist_for_each_entry_safe(ctx, pos, next,
975 &gsec_kr->gsk_clist, cc_cache) {
976 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
978 if (uid != -1 && uid != ctx->cc_vcred.vc_uid)
981 /* at this moment there's at least 2 base reference:
982 * key association and in-list. */
983 if (atomic_read(&ctx->cc_refcount) > 2) {
986 CWARN("flush busy ctx %p(%u->%s, extra ref %d)\n",
987 ctx, ctx->cc_vcred.vc_uid,
988 sec2target_str(ctx->cc_sec),
989 atomic_read(&ctx->cc_refcount) - 2);
992 set_bit(PTLRPC_CTX_DEAD_BIT, &ctx->cc_flags);
994 clear_bit(PTLRPC_CTX_UPTODATE_BIT, &ctx->cc_flags);
996 atomic_inc(&ctx->cc_refcount);
998 if (ctx_unlist_kr(ctx, 1)) {
999 hlist_add_head(&ctx->cc_cache, &freelist);
1001 LASSERT(atomic_read(&ctx->cc_refcount) >= 2);
1002 atomic_dec(&ctx->cc_refcount);
1005 spin_unlock(&sec->ps_lock);
1007 dispose_ctx_list_kr(&freelist);
1012 int gss_sec_flush_ctx_cache_kr(struct ptlrpc_sec *sec,
1013 uid_t uid, int grace, int force)
1017 CDEBUG(D_SEC, "sec %p(%d, nctx %d), uid %d, grace %d, force %d\n",
1018 sec, atomic_read(&sec->ps_refcount),
1019 atomic_read(&sec->ps_nctx),
1022 if (uid != -1 && uid != 0)
1023 flush_user_ctx_cache_kr(sec, uid, grace, force);
1025 flush_spec_ctx_cache_kr(sec, uid, grace, force);
1031 void gss_sec_gc_ctx_kr(struct ptlrpc_sec *sec)
1033 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
1034 struct hlist_head freelist = HLIST_HEAD_INIT;
1035 struct hlist_node __maybe_unused *pos, *next;
1036 struct ptlrpc_cli_ctx *ctx;
1039 CWARN("running gc\n");
1041 spin_lock(&sec->ps_lock);
1042 cfs_hlist_for_each_entry_safe(ctx, pos, next,
1043 &gsec_kr->gsk_clist, cc_cache) {
1044 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
1046 atomic_inc(&ctx->cc_refcount);
1048 if (cli_ctx_check_death(ctx) && ctx_unlist_kr(ctx, 1)) {
1049 hlist_add_head(&ctx->cc_cache, &freelist);
1050 CWARN("unhashed ctx %p\n", ctx);
1052 LASSERT(atomic_read(&ctx->cc_refcount) >= 2);
1053 atomic_dec(&ctx->cc_refcount);
1056 spin_unlock(&sec->ps_lock);
1058 dispose_ctx_list_kr(&freelist);
1064 int gss_sec_display_kr(struct ptlrpc_sec *sec, struct seq_file *seq)
1066 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
1067 struct hlist_node __maybe_unused *pos, *next;
1068 struct ptlrpc_cli_ctx *ctx;
1069 struct gss_cli_ctx *gctx;
1070 time_t now = cfs_time_current_sec();
1073 spin_lock(&sec->ps_lock);
1074 cfs_hlist_for_each_entry_safe(ctx, pos, next,
1075 &gsec_kr->gsk_clist, cc_cache) {
1080 gctx = ctx2gctx(ctx);
1081 key = ctx2gctx_keyring(ctx)->gck_key;
1083 gss_cli_ctx_flags2str(ctx->cc_flags,
1084 flags_str, sizeof(flags_str));
1086 if (gctx->gc_mechctx)
1087 lgss_display(gctx->gc_mechctx, mech, sizeof(mech));
1089 snprintf(mech, sizeof(mech), "N/A");
1090 mech[sizeof(mech) - 1] = '\0';
1092 seq_printf(seq, "%p: uid %u, ref %d, expire %lu(%+ld), fl %s, "
1093 "seq %d, win %u, key %08x(ref %d), "
1094 "hdl %#llx:%#llx, mech: %s\n",
1095 ctx, ctx->cc_vcred.vc_uid,
1096 atomic_read(&ctx->cc_refcount),
1098 ctx->cc_expire ? ctx->cc_expire - now : 0,
1100 atomic_read(&gctx->gc_seq),
1102 key ? key->serial : 0,
1103 key ? atomic_read(&key->usage) : 0,
1104 gss_handle_to_u64(&gctx->gc_handle),
1105 gss_handle_to_u64(&gctx->gc_svc_handle),
1108 spin_unlock(&sec->ps_lock);
1113 /****************************************
1115 ****************************************/
1118 int gss_cli_ctx_refresh_kr(struct ptlrpc_cli_ctx *ctx)
1120 /* upcall is already on the way */
1125 int gss_cli_ctx_validate_kr(struct ptlrpc_cli_ctx *ctx)
1127 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
1128 LASSERT(ctx->cc_sec);
1130 if (cli_ctx_check_death(ctx)) {
1135 if (cli_ctx_is_ready(ctx))
1141 void gss_cli_ctx_die_kr(struct ptlrpc_cli_ctx *ctx, int grace)
1143 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
1144 LASSERT(ctx->cc_sec);
1146 cli_ctx_expire(ctx);
1150 /****************************************
1151 * (reverse) service *
1152 ****************************************/
1155 * reverse context could have nothing to do with keyrings. here we still keep
1156 * the version which bind to a key, for future reference.
1158 #define HAVE_REVERSE_CTX_NOKEY
1160 #ifdef HAVE_REVERSE_CTX_NOKEY
1163 int sec_install_rctx_kr(struct ptlrpc_sec *sec,
1164 struct ptlrpc_svc_ctx *svc_ctx)
1166 struct ptlrpc_cli_ctx *cli_ctx;
1167 struct vfs_cred vcred = { 0, 0 };
1173 cli_ctx = ctx_create_kr(sec, &vcred);
1174 if (cli_ctx == NULL)
1177 rc = gss_copy_rvc_cli_ctx(cli_ctx, svc_ctx);
1179 CERROR("failed copy reverse cli ctx: %d\n", rc);
1181 ctx_put_kr(cli_ctx, 1);
1185 rvs_sec_install_root_ctx_kr(sec, cli_ctx, NULL);
1187 ctx_put_kr(cli_ctx, 1);
1192 #else /* ! HAVE_REVERSE_CTX_NOKEY */
1195 int sec_install_rctx_kr(struct ptlrpc_sec *sec,
1196 struct ptlrpc_svc_ctx *svc_ctx)
1198 struct ptlrpc_cli_ctx *cli_ctx = NULL;
1200 struct vfs_cred vcred = { 0, 0 };
1208 construct_key_desc(desc, sizeof(desc), sec, 0);
1210 key = key_alloc(&gss_key_type, desc, 0, 0,
1211 KEY_POS_ALL | KEY_USR_ALL, 1);
1213 CERROR("failed to alloc key: %ld\n", PTR_ERR(key));
1214 return PTR_ERR(key);
1217 rc = key_instantiate_and_link(key, NULL, 0, NULL, NULL);
1219 CERROR("failed to instantiate key: %d\n", rc);
1223 down_write(&key->sem);
1225 LASSERT(!key_get_payload(key, 0));
1227 cli_ctx = ctx_create_kr(sec, &vcred);
1228 if (cli_ctx == NULL) {
1233 rc = gss_copy_rvc_cli_ctx(cli_ctx, svc_ctx);
1235 CERROR("failed copy reverse cli ctx: %d\n", rc);
1239 rvs_sec_install_root_ctx_kr(sec, cli_ctx, key);
1241 ctx_put_kr(cli_ctx, 1);
1242 up_write(&key->sem);
1251 ctx_put_kr(cli_ctx, 1);
1253 up_write(&key->sem);
1259 #endif /* HAVE_REVERSE_CTX_NOKEY */
1261 /****************************************
1263 ****************************************/
1266 int gss_svc_accept_kr(struct ptlrpc_request *req)
1268 return gss_svc_accept(&gss_policy_keyring, req);
1272 int gss_svc_install_rctx_kr(struct obd_import *imp,
1273 struct ptlrpc_svc_ctx *svc_ctx)
1275 struct ptlrpc_sec *sec;
1278 sec = sptlrpc_import_sec_ref(imp);
1281 rc = sec_install_rctx_kr(sec, svc_ctx);
1282 sptlrpc_sec_put(sec);
1287 /****************************************
1289 ****************************************/
1292 #ifdef HAVE_KEY_TYPE_INSTANTIATE_2ARGS
1293 int gss_kt_instantiate(struct key *key, struct key_preparsed_payload *prep)
1295 const void *data = prep->data;
1296 size_t datalen = prep->datalen;
1298 int gss_kt_instantiate(struct key *key, const void *data, size_t datalen)
1304 if (data != NULL || datalen != 0) {
1305 CERROR("invalid: data %p, len %lu\n", data, (long)datalen);
1309 if (key_get_payload(key, 0)) {
1310 CERROR("key already have payload\n");
1314 /* link the key to session keyring, so following context negotiation
1315 * rpc fired from user space could find this key. This will be unlinked
1316 * automatically when upcall processes die.
1318 * we can't do this through keyctl from userspace, because the upcall
1319 * might be neither possessor nor owner of the key (setuid).
1321 * the session keyring is created upon upcall, and don't change all
1322 * the way until upcall finished, so rcu lock is not needed here.
1324 LASSERT(key_tgcred(current)->session_keyring);
1327 rc = key_link(key_tgcred(current)->session_keyring, key);
1330 CERROR("failed to link key %08x to keyring %08x: %d\n",
1332 key_tgcred(current)->session_keyring->serial, rc);
1336 CDEBUG(D_SEC, "key %p instantiated, ctx %p\n", key,
1337 key_get_payload(key, 0));
1342 * called with key semaphore write locked. it means we can operate
1343 * on the context without fear of loosing refcount.
1346 #ifdef HAVE_KEY_TYPE_INSTANTIATE_2ARGS
1347 int gss_kt_update(struct key *key, struct key_preparsed_payload *prep)
1349 const void *data = prep->data;
1350 __u32 datalen32 = (__u32) prep->datalen;
1352 int gss_kt_update(struct key *key, const void *data, size_t datalen)
1354 __u32 datalen32 = (__u32) datalen;
1356 struct ptlrpc_cli_ctx *ctx = key_get_payload(key, 0);
1357 struct gss_cli_ctx *gctx;
1358 rawobj_t tmpobj = RAWOBJ_EMPTY;
1362 if (data == NULL || datalen32 == 0) {
1363 CWARN("invalid: data %p, len %lu\n", data, (long)datalen32);
1367 /* if upcall finished negotiation too fast (mostly likely because
1368 * of local error happened) and call kt_update(), the ctx
1369 * might be still NULL. but the key will finally be associate
1370 * with a context, or be revoked. if key status is fine, return
1371 * -EAGAIN to allow userspace sleep a while and call again. */
1373 CDEBUG(D_SEC, "update too soon: key %p(%x) flags %lx\n",
1374 key, key->serial, key->flags);
1376 rc = key_validate(key);
1383 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
1384 LASSERT(ctx->cc_sec);
1386 ctx_clear_timer_kr(ctx);
1388 /* don't proceed if already refreshed */
1389 if (cli_ctx_is_refreshed(ctx)) {
1390 CWARN("ctx already done refresh\n");
1394 sptlrpc_cli_ctx_get(ctx);
1395 gctx = ctx2gctx(ctx);
1397 rc = buffer_extract_bytes(&data, &datalen32, &gctx->gc_win,
1398 sizeof(gctx->gc_win));
1400 CERROR("failed extract seq_win\n");
1404 if (gctx->gc_win == 0) {
1405 __u32 nego_rpc_err, nego_gss_err;
1407 rc = buffer_extract_bytes(&data, &datalen32, &nego_rpc_err,
1408 sizeof(nego_rpc_err));
1410 CERROR("cannot extract RPC: rc = %d\n", rc);
1414 rc = buffer_extract_bytes(&data, &datalen32, &nego_gss_err,
1415 sizeof(nego_gss_err));
1417 CERROR("failed to extract gss rc = %d\n", rc);
1421 CERROR("negotiation: rpc err %d, gss err %x\n",
1422 nego_rpc_err, nego_gss_err);
1424 rc = nego_rpc_err ? nego_rpc_err : -EACCES;
1426 rc = rawobj_extract_local_alloc(&gctx->gc_handle,
1427 (__u32 **) &data, &datalen32);
1429 CERROR("failed extract handle\n");
1433 rc = rawobj_extract_local(&tmpobj,
1434 (__u32 **) &data, &datalen32);
1436 CERROR("failed extract mech\n");
1440 rc = lgss_import_sec_context(&tmpobj,
1441 sec2gsec(ctx->cc_sec)->gs_mech,
1443 if (rc != GSS_S_COMPLETE)
1444 CERROR("failed import context\n");
1449 /* we don't care what current status of this ctx, even someone else
1450 * is operating on the ctx at the same time. we just add up our own
1453 gss_cli_ctx_uptodate(gctx);
1455 /* this will also revoke the key. has to be done before
1456 * wakeup waiters otherwise they can find the stale key */
1457 kill_key_locked(key);
1459 cli_ctx_expire(ctx);
1461 if (rc != -ERESTART)
1462 set_bit(PTLRPC_CTX_ERROR_BIT, &ctx->cc_flags);
1465 /* let user space think it's a success */
1466 sptlrpc_cli_ctx_put(ctx, 1);
1470 #ifndef HAVE_KEY_MATCH_DATA
1472 gss_kt_match(const struct key *key, const void *desc)
1474 return strcmp(key->description, (const char *) desc) == 0 &&
1475 !test_bit(KEY_FLAG_REVOKED, &key->flags);
1477 #else /* ! HAVE_KEY_MATCH_DATA */
1479 gss_kt_match(const struct key *key, const struct key_match_data *match_data)
1481 const char *desc = match_data->raw_data;
1483 return strcmp(key->description, desc) == 0 &&
1484 !test_bit(KEY_FLAG_REVOKED, &key->flags);
1488 * Preparse the match criterion.
1490 static int gss_kt_match_preparse(struct key_match_data *match_data)
1492 match_data->lookup_type = KEYRING_SEARCH_LOOKUP_DIRECT;
1493 match_data->cmp = gss_kt_match;
1496 #endif /* HAVE_KEY_MATCH_DATA */
1499 void gss_kt_destroy(struct key *key)
1502 LASSERT(!key_get_payload(key, 0));
1503 CDEBUG(D_SEC, "destroy key %p\n", key);
1508 void gss_kt_describe(const struct key *key, struct seq_file *s)
1510 if (key->description == NULL)
1511 seq_puts(s, "[null]");
1513 seq_puts(s, key->description);
1516 static struct key_type gss_key_type =
1520 .instantiate = gss_kt_instantiate,
1521 .update = gss_kt_update,
1522 #ifdef HAVE_KEY_MATCH_DATA
1523 .match_preparse = gss_kt_match_preparse,
1525 .match = gss_kt_match,
1527 .destroy = gss_kt_destroy,
1528 .describe = gss_kt_describe,
1531 /****************************************
1532 * lustre gss keyring policy *
1533 ****************************************/
1535 static struct ptlrpc_ctx_ops gss_keyring_ctxops = {
1536 .match = gss_cli_ctx_match,
1537 .refresh = gss_cli_ctx_refresh_kr,
1538 .validate = gss_cli_ctx_validate_kr,
1539 .die = gss_cli_ctx_die_kr,
1540 .sign = gss_cli_ctx_sign,
1541 .verify = gss_cli_ctx_verify,
1542 .seal = gss_cli_ctx_seal,
1543 .unseal = gss_cli_ctx_unseal,
1544 .wrap_bulk = gss_cli_ctx_wrap_bulk,
1545 .unwrap_bulk = gss_cli_ctx_unwrap_bulk,
1548 static struct ptlrpc_sec_cops gss_sec_keyring_cops = {
1549 .create_sec = gss_sec_create_kr,
1550 .destroy_sec = gss_sec_destroy_kr,
1551 .kill_sec = gss_sec_kill,
1552 .lookup_ctx = gss_sec_lookup_ctx_kr,
1553 .release_ctx = gss_sec_release_ctx_kr,
1554 .flush_ctx_cache = gss_sec_flush_ctx_cache_kr,
1555 .gc_ctx = gss_sec_gc_ctx_kr,
1556 .install_rctx = gss_sec_install_rctx,
1557 .alloc_reqbuf = gss_alloc_reqbuf,
1558 .free_reqbuf = gss_free_reqbuf,
1559 .alloc_repbuf = gss_alloc_repbuf,
1560 .free_repbuf = gss_free_repbuf,
1561 .enlarge_reqbuf = gss_enlarge_reqbuf,
1562 .display = gss_sec_display_kr,
1565 static struct ptlrpc_sec_sops gss_sec_keyring_sops = {
1566 .accept = gss_svc_accept_kr,
1567 .invalidate_ctx = gss_svc_invalidate_ctx,
1568 .alloc_rs = gss_svc_alloc_rs,
1569 .authorize = gss_svc_authorize,
1570 .free_rs = gss_svc_free_rs,
1571 .free_ctx = gss_svc_free_ctx,
1572 .prep_bulk = gss_svc_prep_bulk,
1573 .unwrap_bulk = gss_svc_unwrap_bulk,
1574 .wrap_bulk = gss_svc_wrap_bulk,
1575 .install_rctx = gss_svc_install_rctx_kr,
1578 static struct ptlrpc_sec_policy gss_policy_keyring = {
1579 .sp_owner = THIS_MODULE,
1580 .sp_name = "gss.keyring",
1581 .sp_policy = SPTLRPC_POLICY_GSS,
1582 .sp_cops = &gss_sec_keyring_cops,
1583 .sp_sops = &gss_sec_keyring_sops,
1587 int __init gss_init_keyring(void)
1591 rc = register_key_type(&gss_key_type);
1593 CERROR("failed to register keyring type: %d\n", rc);
1597 rc = sptlrpc_register_policy(&gss_policy_keyring);
1599 unregister_key_type(&gss_key_type);
1606 void __exit gss_exit_keyring(void)
1608 unregister_key_type(&gss_key_type);
1609 sptlrpc_unregister_policy(&gss_policy_keyring);