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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2014, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/ptlrpc/gss/gss_keyring.c
34 * Author: Eric Mei <ericm@clusterfs.com>
37 #define DEBUG_SUBSYSTEM S_SEC
38 #include <linux/init.h>
39 #include <linux/module.h>
40 #include <linux/slab.h>
41 #include <linux/dcache.h>
43 #include <linux/crypto.h>
44 #include <linux/key.h>
45 #include <linux/keyctl.h>
46 #include <linux/key-type.h>
47 #include <linux/mutex.h>
48 #include <asm/atomic.h>
50 #include <libcfs/linux/linux-list.h>
52 #include <obd_class.h>
53 #include <obd_support.h>
54 #include <uapi/linux/lustre/lustre_idl.h>
55 #include <lustre_sec.h>
56 #include <lustre_net.h>
57 #include <lustre_import.h>
60 #include "gss_internal.h"
63 #ifdef HAVE_GET_REQUEST_KEY_AUTH
64 #include <keys/request_key_auth-type.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);
73 static void request_key_unlink(struct key *key);
76 * the timeout is only for the case that upcall child process die abnormally.
77 * in any other cases it should finally update kernel key.
79 * FIXME we'd better to incorporate the client & server side upcall timeouts
80 * into the framework of Adaptive Timeouts, but we need to figure out how to
81 * make sure that kernel knows the upcall processes is in-progress or died
84 #define KEYRING_UPCALL_TIMEOUT (obd_timeout + obd_timeout)
86 /* Check caller's namespace in gss_keyring upcall */
87 unsigned int gss_check_upcall_ns = 1;
89 /****************************************
91 ****************************************/
93 static inline void keyring_upcall_lock(struct gss_sec_keyring *gsec_kr)
95 #ifdef HAVE_KEYRING_UPCALL_SERIALIZED
96 mutex_lock(&gsec_kr->gsk_uc_lock);
100 static inline void keyring_upcall_unlock(struct gss_sec_keyring *gsec_kr)
102 #ifdef HAVE_KEYRING_UPCALL_SERIALIZED
103 mutex_unlock(&gsec_kr->gsk_uc_lock);
107 static inline void key_revoke_locked(struct key *key)
109 set_bit(KEY_FLAG_REVOKED, &key->flags);
112 static void ctx_upcall_timeout_kr(cfs_timer_cb_arg_t data)
114 struct gss_cli_ctx_keyring *gctx_kr = cfs_from_timer(gctx_kr,
116 struct ptlrpc_cli_ctx *ctx = &(gctx_kr->gck_base.gc_base);
117 struct key *key = gctx_kr->gck_key;
119 CWARN("ctx %p, key %p\n", ctx, key);
124 key_revoke_locked(key);
127 static void ctx_start_timer_kr(struct ptlrpc_cli_ctx *ctx, time64_t timeout)
129 struct gss_cli_ctx_keyring *gctx_kr = ctx2gctx_keyring(ctx);
130 struct timer_list *timer = &gctx_kr->gck_timer;
134 CDEBUG(D_SEC, "ctx %p: start timer %llds\n", ctx, timeout);
136 cfs_timer_setup(timer, ctx_upcall_timeout_kr,
137 (unsigned long)gctx_kr, 0);
138 timer->expires = cfs_time_seconds(timeout) + jiffies;
143 * caller should make sure no race with other threads
146 void ctx_clear_timer_kr(struct ptlrpc_cli_ctx *ctx)
148 struct gss_cli_ctx_keyring *gctx_kr = ctx2gctx_keyring(ctx);
149 struct timer_list *timer = &gctx_kr->gck_timer;
151 CDEBUG(D_SEC, "ctx %p, key %p\n", ctx, gctx_kr->gck_key);
153 del_singleshot_timer_sync(timer);
157 struct ptlrpc_cli_ctx *ctx_create_kr(struct ptlrpc_sec *sec,
158 struct vfs_cred *vcred)
160 struct ptlrpc_cli_ctx *ctx;
161 struct gss_cli_ctx_keyring *gctx_kr;
163 OBD_ALLOC_PTR(gctx_kr);
167 cfs_timer_setup(&gctx_kr->gck_timer, NULL, 0, 0);
169 ctx = &gctx_kr->gck_base.gc_base;
171 if (gss_cli_ctx_init_common(sec, ctx, &gss_keyring_ctxops, vcred)) {
172 OBD_FREE_PTR(gctx_kr);
176 ctx->cc_expire = ktime_get_real_seconds() + KEYRING_UPCALL_TIMEOUT;
177 clear_bit(PTLRPC_CTX_NEW_BIT, &ctx->cc_flags);
178 atomic_inc(&ctx->cc_refcount); /* for the caller */
183 static void ctx_destroy_kr(struct ptlrpc_cli_ctx *ctx)
185 struct ptlrpc_sec *sec = ctx->cc_sec;
186 struct gss_cli_ctx_keyring *gctx_kr = ctx2gctx_keyring(ctx);
188 CDEBUG(D_SEC, "destroying ctx %p\n", ctx);
190 /* at this time the association with key has been broken. */
192 LASSERT(atomic_read(&sec->ps_refcount) > 0);
193 LASSERT(atomic_read(&sec->ps_nctx) > 0);
194 LASSERT(test_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags) == 0);
195 LASSERT(gctx_kr->gck_key == NULL);
197 ctx_clear_timer_kr(ctx);
199 if (gss_cli_ctx_fini_common(sec, ctx))
202 OBD_FREE_PTR(gctx_kr);
204 atomic_dec(&sec->ps_nctx);
205 sptlrpc_sec_put(sec);
208 static void ctx_release_kr(struct ptlrpc_cli_ctx *ctx, int sync)
213 atomic_inc(&ctx->cc_refcount);
214 sptlrpc_gc_add_ctx(ctx);
218 static void ctx_put_kr(struct ptlrpc_cli_ctx *ctx, int sync)
220 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
222 if (atomic_dec_and_test(&ctx->cc_refcount))
223 ctx_release_kr(ctx, sync);
227 * key <-> ctx association and rules:
228 * - ctx might not bind with any key
229 * - key/ctx binding is protected by key semaphore (if the key present)
230 * - key and ctx each take a reference of the other
231 * - ctx enlist/unlist is protected by ctx spinlock
232 * - never enlist a ctx after it's been unlisted
233 * - whoever do enlist should also do bind, lock key before enlist:
234 * - lock key -> lock ctx -> enlist -> unlock ctx -> bind -> unlock key
235 * - whoever do unlist should also do unbind:
236 * - lock key -> lock ctx -> unlist -> unlock ctx -> unbind -> unlock key
237 * - lock ctx -> unlist -> unlock ctx -> lock key -> unbind -> unlock key
240 static inline void spin_lock_if(spinlock_t *lock, int condition)
246 static inline void spin_unlock_if(spinlock_t *lock, int condition)
252 static void ctx_enlist_kr(struct ptlrpc_cli_ctx *ctx, int is_root, int locked)
254 struct ptlrpc_sec *sec = ctx->cc_sec;
255 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
257 LASSERT(!test_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags));
258 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
260 spin_lock_if(&sec->ps_lock, !locked);
262 atomic_inc(&ctx->cc_refcount);
263 set_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags);
264 hlist_add_head(&ctx->cc_cache, &gsec_kr->gsk_clist);
266 gsec_kr->gsk_root_ctx = ctx;
268 spin_unlock_if(&sec->ps_lock, !locked);
272 * Note after this get called, caller should not access ctx again because
273 * it might have been freed, unless caller hold at least one refcount of
276 * return non-zero if we indeed unlist this ctx.
278 static int ctx_unlist_kr(struct ptlrpc_cli_ctx *ctx, int locked)
280 struct ptlrpc_sec *sec = ctx->cc_sec;
281 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
283 /* if hashed bit has gone, leave the job to somebody who is doing it */
284 if (test_and_clear_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags) == 0)
287 /* drop ref inside spin lock to prevent race with other operations */
288 spin_lock_if(&sec->ps_lock, !locked);
290 if (gsec_kr->gsk_root_ctx == ctx)
291 gsec_kr->gsk_root_ctx = NULL;
292 hlist_del_init(&ctx->cc_cache);
293 atomic_dec(&ctx->cc_refcount);
295 spin_unlock_if(&sec->ps_lock, !locked);
301 * Get specific payload. Newer kernels support 4 slots.
304 key_get_payload(struct key *key, unsigned int index)
306 void *key_ptr = NULL;
308 #ifdef HAVE_KEY_PAYLOAD_DATA_ARRAY
309 key_ptr = key->payload.data[index];
312 key_ptr = key->payload.data;
318 * Set specific payload. Newer kernels support 4 slots.
320 static int key_set_payload(struct key *key, unsigned int index,
321 struct ptlrpc_cli_ctx *ctx)
325 #ifdef HAVE_KEY_PAYLOAD_DATA_ARRAY
327 key->payload.data[index] = ctx;
330 key->payload.data = ctx;
338 * bind a key with a ctx together.
339 * caller must hold write lock of the key, as well as ref on key & ctx.
341 static void bind_key_ctx(struct key *key, struct ptlrpc_cli_ctx *ctx)
343 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
344 LASSERT(ll_read_key_usage(key) > 0);
345 LASSERT(ctx2gctx_keyring(ctx)->gck_key == NULL);
346 LASSERT(!key_get_payload(key, 0));
348 /* at this time context may or may not in list. */
350 atomic_inc(&ctx->cc_refcount);
351 ctx2gctx_keyring(ctx)->gck_key = key;
352 LASSERT(!key_set_payload(key, 0, ctx));
356 * unbind a key and a ctx.
357 * caller must hold write lock, as well as a ref of the key.
359 static void unbind_key_ctx(struct key *key, struct ptlrpc_cli_ctx *ctx)
361 LASSERT(key_get_payload(key, 0) == ctx);
362 LASSERT(test_bit(PTLRPC_CTX_CACHED_BIT, &ctx->cc_flags) == 0);
364 /* must revoke the key, or others may treat it as newly created */
365 key_revoke_locked(key);
367 key_set_payload(key, 0, NULL);
368 ctx2gctx_keyring(ctx)->gck_key = NULL;
370 /* once ctx get split from key, the timer is meaningless */
371 ctx_clear_timer_kr(ctx);
378 * given a ctx, unbind with its coupled key, if any.
379 * unbind could only be called once, so we don't worry the key be released
382 static void unbind_ctx_kr(struct ptlrpc_cli_ctx *ctx)
384 struct key *key = ctx2gctx_keyring(ctx)->gck_key;
387 LASSERT(key_get_payload(key, 0) == ctx);
390 down_write(&key->sem);
391 unbind_key_ctx(key, ctx);
394 request_key_unlink(key);
399 * given a key, unbind with its coupled ctx, if any.
400 * caller must hold write lock, as well as a ref of the key.
402 static void unbind_key_locked(struct key *key)
404 struct ptlrpc_cli_ctx *ctx = key_get_payload(key, 0);
407 unbind_key_ctx(key, ctx);
411 * unlist a ctx, and unbind from coupled key
413 static void kill_ctx_kr(struct ptlrpc_cli_ctx *ctx)
415 if (ctx_unlist_kr(ctx, 0))
420 * given a key, unlist and unbind with the coupled ctx (if any).
421 * caller must hold write lock, as well as a ref of the key.
423 static void kill_key_locked(struct key *key)
425 struct ptlrpc_cli_ctx *ctx = key_get_payload(key, 0);
427 if (ctx && ctx_unlist_kr(ctx, 0))
428 unbind_key_locked(key);
432 * caller should hold one ref on contexts in freelist.
434 static void dispose_ctx_list_kr(struct hlist_head *freelist)
436 struct hlist_node *next;
437 struct ptlrpc_cli_ctx *ctx;
438 struct gss_cli_ctx *gctx;
440 hlist_for_each_entry_safe(ctx, next, freelist, cc_cache) {
441 hlist_del_init(&ctx->cc_cache);
443 /* reverse ctx: update current seq to buddy svcctx if exist.
444 * ideally this should be done at gss_cli_ctx_finalize(), but
445 * the ctx destroy could be delayed by:
446 * 1) ctx still has reference;
447 * 2) ctx destroy is asynchronous;
448 * and reverse import call inval_all_ctx() require this be done
449 * _immediately_ otherwise newly created reverse ctx might copy
450 * the very old sequence number from svcctx. */
451 gctx = ctx2gctx(ctx);
452 if (!rawobj_empty(&gctx->gc_svc_handle) &&
453 sec_is_reverse(gctx->gc_base.cc_sec)) {
454 gss_svc_upcall_update_sequence(&gctx->gc_svc_handle,
455 (__u32) atomic_read(&gctx->gc_seq));
458 /* we need to wakeup waiting reqs here. the context might
459 * be forced released before upcall finished, then the
460 * late-arrived downcall can't find the ctx even. */
461 sptlrpc_cli_ctx_wakeup(ctx);
469 * lookup a root context directly in a sec, return root ctx with a
470 * reference taken or NULL.
473 struct ptlrpc_cli_ctx * sec_lookup_root_ctx_kr(struct ptlrpc_sec *sec)
475 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
476 struct ptlrpc_cli_ctx *ctx = NULL;
478 spin_lock(&sec->ps_lock);
480 ctx = gsec_kr->gsk_root_ctx;
482 if (ctx == NULL && unlikely(sec_is_reverse(sec))) {
483 struct ptlrpc_cli_ctx *tmp;
485 /* reverse ctx, search root ctx in list, choose the one
486 * with shortest expire time, which is most possibly have
487 * an established peer ctx at client side. */
488 hlist_for_each_entry(tmp, &gsec_kr->gsk_clist, cc_cache) {
489 if (ctx == NULL || ctx->cc_expire == 0 ||
490 ctx->cc_expire > tmp->cc_expire) {
492 /* promote to be root_ctx */
493 gsec_kr->gsk_root_ctx = ctx;
499 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
500 LASSERT(!hlist_empty(&gsec_kr->gsk_clist));
501 atomic_inc(&ctx->cc_refcount);
504 spin_unlock(&sec->ps_lock);
509 #define RVS_CTX_EXPIRE_NICE (10)
512 void rvs_sec_install_root_ctx_kr(struct ptlrpc_sec *sec,
513 struct ptlrpc_cli_ctx *new_ctx,
516 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
517 struct ptlrpc_cli_ctx *ctx;
521 LASSERT(sec_is_reverse(sec));
523 spin_lock(&sec->ps_lock);
525 now = ktime_get_real_seconds();
527 /* set all existing ctxs short expiry */
528 hlist_for_each_entry(ctx, &gsec_kr->gsk_clist, cc_cache) {
529 if (ctx->cc_expire > now + RVS_CTX_EXPIRE_NICE) {
530 ctx->cc_early_expire = 1;
531 ctx->cc_expire = now + RVS_CTX_EXPIRE_NICE;
535 /* if there's root_ctx there, instead obsolete the current
536 * immediately, we leave it continue operating for a little while.
537 * hopefully when the first backward rpc with newest ctx send out,
538 * the client side already have the peer ctx well established. */
539 ctx_enlist_kr(new_ctx, gsec_kr->gsk_root_ctx ? 0 : 1, 1);
542 bind_key_ctx(key, new_ctx);
544 spin_unlock(&sec->ps_lock);
547 static void construct_key_desc(void *buf, int bufsize,
548 struct ptlrpc_sec *sec, uid_t uid)
550 snprintf(buf, bufsize, "%d@%x", uid, sec->ps_id);
551 ((char *)buf)[bufsize - 1] = '\0';
554 /****************************************
556 ****************************************/
559 struct ptlrpc_sec * gss_sec_create_kr(struct obd_import *imp,
560 struct ptlrpc_svc_ctx *svcctx,
561 struct sptlrpc_flavor *sf)
563 struct gss_sec_keyring *gsec_kr;
566 OBD_ALLOC(gsec_kr, sizeof(*gsec_kr));
570 INIT_HLIST_HEAD(&gsec_kr->gsk_clist);
571 gsec_kr->gsk_root_ctx = NULL;
572 mutex_init(&gsec_kr->gsk_root_uc_lock);
573 #ifdef HAVE_KEYRING_UPCALL_SERIALIZED
574 mutex_init(&gsec_kr->gsk_uc_lock);
577 if (gss_sec_create_common(&gsec_kr->gsk_base, &gss_policy_keyring,
581 if (svcctx != NULL &&
582 sec_install_rctx_kr(&gsec_kr->gsk_base.gs_base, svcctx)) {
583 gss_sec_destroy_common(&gsec_kr->gsk_base);
587 RETURN(&gsec_kr->gsk_base.gs_base);
590 OBD_FREE(gsec_kr, sizeof(*gsec_kr));
595 void gss_sec_destroy_kr(struct ptlrpc_sec *sec)
597 struct gss_sec *gsec = sec2gsec(sec);
598 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
600 CDEBUG(D_SEC, "destroy %s@%p\n", sec->ps_policy->sp_name, sec);
602 LASSERT(hlist_empty(&gsec_kr->gsk_clist));
603 LASSERT(gsec_kr->gsk_root_ctx == NULL);
605 gss_sec_destroy_common(gsec);
607 OBD_FREE(gsec_kr, sizeof(*gsec_kr));
610 static inline int user_is_root(struct ptlrpc_sec *sec, struct vfs_cred *vcred)
612 /* except the ROOTONLY flag, treat it as root user only if real uid
613 * is 0, euid/fsuid being 0 are handled as setuid scenarios */
614 if (sec_is_rootonly(sec) || (vcred->vc_uid == 0))
621 * kernel 5.3: commit 0f44e4d976f96c6439da0d6717238efa4b91196e
622 * keys: Move the user and user-session keyrings to the user_namespace
624 * When lookup_user_key is available use the kernel API rather than directly
625 * accessing the uid_keyring and session_keyring via the current process
628 #ifdef HAVE_LOOKUP_USER_KEY
630 /* from Linux security/keys/internal.h: */
631 #ifndef KEY_LOOKUP_FOR_UNLINK
632 #define KEY_LOOKUP_FOR_UNLINK 0x04
635 static struct key *_user_key(key_serial_t id)
640 ref = lookup_user_key(id, KEY_LOOKUP_FOR_UNLINK, 0);
643 return key_ref_to_ptr(ref);
646 static inline struct key *get_user_session_keyring(const struct cred *cred)
648 return _user_key(KEY_SPEC_USER_SESSION_KEYRING);
651 static inline struct key *get_user_keyring(const struct cred *cred)
653 return _user_key(KEY_SPEC_USER_KEYRING);
656 static inline struct key *get_user_session_keyring(const struct cred *cred)
658 return key_get(cred->user->session_keyring);
661 static inline struct key *get_user_keyring(const struct cred *cred)
663 return key_get(cred->user->uid_keyring);
668 * unlink request key from it's ring, which is linked during request_key().
669 * sadly, we have to 'guess' which keyring it's linked to.
671 * FIXME this code is fragile, it depends on how request_key() is implemented.
673 static void request_key_unlink(struct key *key)
675 const struct cred *cred = current_cred();
676 struct key *ring = NULL;
678 switch (cred->jit_keyring) {
679 case KEY_REQKEY_DEFL_DEFAULT:
680 case KEY_REQKEY_DEFL_REQUESTOR_KEYRING:
681 #ifdef HAVE_GET_REQUEST_KEY_AUTH
682 if (cred->request_key_auth) {
683 struct request_key_auth *rka;
684 struct key *authkey = cred->request_key_auth;
686 down_read(&authkey->sem);
687 rka = get_request_key_auth(authkey);
688 if (!test_bit(KEY_FLAG_REVOKED, &authkey->flags))
689 ring = key_get(rka->dest_keyring);
690 up_read(&authkey->sem);
696 case KEY_REQKEY_DEFL_THREAD_KEYRING:
697 ring = key_get(cred->thread_keyring);
701 case KEY_REQKEY_DEFL_PROCESS_KEYRING:
702 ring = key_get(cred->process_keyring);
706 case KEY_REQKEY_DEFL_SESSION_KEYRING:
708 ring = key_get(rcu_dereference(cred->session_keyring));
713 case KEY_REQKEY_DEFL_USER_SESSION_KEYRING:
714 ring = get_user_session_keyring(cred);
716 case KEY_REQKEY_DEFL_USER_KEYRING:
717 ring = get_user_keyring(cred);
719 case KEY_REQKEY_DEFL_GROUP_KEYRING:
725 key_unlink(ring, key);
730 struct ptlrpc_cli_ctx * gss_sec_lookup_ctx_kr(struct ptlrpc_sec *sec,
731 struct vfs_cred *vcred,
732 int create, int remove_dead)
734 struct obd_import *imp = sec->ps_import;
735 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
736 struct ptlrpc_cli_ctx *ctx = NULL;
737 unsigned int is_root = 0, create_new = 0;
742 const char *sec_part_flags = "";
747 LASSERT(imp != NULL);
749 is_root = user_is_root(sec, vcred);
751 /* a little bit optimization for root context */
753 ctx = sec_lookup_root_ctx_kr(sec);
755 * Only lookup directly for REVERSE sec, which should
758 if (ctx || sec_is_reverse(sec))
762 LASSERT(create != 0);
764 /* for root context, obtain lock and check again, this time hold
765 * the root upcall lock, make sure nobody else populated new root
766 * context after last check.
769 mutex_lock(&gsec_kr->gsk_root_uc_lock);
771 ctx = sec_lookup_root_ctx_kr(sec);
775 /* update reverse handle for root user */
776 sec2gsec(sec)->gs_rvs_hdl = gss_get_next_ctx_index();
778 switch (sec->ps_part) {
780 sec_part_flags = "m";
783 sec_part_flags = "o";
786 sec_part_flags = "rmo";
789 sec_part_flags = "r";
796 switch (SPTLRPC_FLVR_SVC(sec->ps_flvr.sf_rpc)) {
797 case SPTLRPC_SVC_NULL:
800 case SPTLRPC_SVC_AUTH:
803 case SPTLRPC_SVC_INTG:
806 case SPTLRPC_SVC_PRIV:
814 /* in case of setuid, key will be constructed as owner of fsuid/fsgid,
815 * but we do authentication based on real uid/gid. the key permission
816 * bits will be exactly as POS_ALL, so only processes who subscribed
817 * this key could have the access, although the quota might be counted
818 * on others (fsuid/fsgid).
820 * keyring will use fsuid/fsgid as upcall parameters, so we have to
821 * encode real uid/gid into callout info.
824 /* But first we need to make sure the obd type is supported */
825 if (strcmp(imp->imp_obd->obd_type->typ_name, LUSTRE_MDC_NAME) &&
826 strcmp(imp->imp_obd->obd_type->typ_name, LUSTRE_OSC_NAME) &&
827 strcmp(imp->imp_obd->obd_type->typ_name, LUSTRE_MGC_NAME) &&
828 strcmp(imp->imp_obd->obd_type->typ_name, LUSTRE_LWP_NAME) &&
829 strcmp(imp->imp_obd->obd_type->typ_name, LUSTRE_OSP_NAME)) {
830 CERROR("obd %s is not a supported device\n",
831 imp->imp_obd->obd_name);
832 GOTO(out, ctx = NULL);
835 construct_key_desc(desc, sizeof(desc), sec, vcred->vc_uid);
837 /* callout info format:
838 * secid:mech:uid:gid:sec_flags:svc_flag:svc_type:peer_nid:target_uuid:
841 coinfo_size = sizeof(struct obd_uuid) + MAX_OBD_NAME + 64;
842 OBD_ALLOC(coinfo, coinfo_size);
846 /* Last callout parameter is pid of process whose namespace will be used
847 * for credentials' retrieval.
849 if (gss_check_upcall_ns) {
850 /* For user's credentials (in which case sec_part_flags is
851 * empty), use current PID instead of import's reference
852 * PID to get reference namespace.
854 if (sec_part_flags[0] == '\0')
855 caller_pid = current->pid;
857 caller_pid = imp->imp_sec_refpid;
859 /* Do not switch namespace in gss keyring upcall. */
862 snprintf(coinfo, coinfo_size, "%d:%s:%u:%u:%s:%c:%d:%#llx:%s:%#llx:%d",
863 sec->ps_id, sec2gsec(sec)->gs_mech->gm_name,
864 vcred->vc_uid, vcred->vc_gid,
865 sec_part_flags, svc_flag, import_to_gss_svc(imp),
866 imp->imp_connection->c_peer.nid, imp->imp_obd->obd_name,
867 imp->imp_connection->c_self, caller_pid);
869 CDEBUG(D_SEC, "requesting key for %s\n", desc);
871 keyring_upcall_lock(gsec_kr);
872 key = request_key(&gss_key_type, desc, coinfo);
873 keyring_upcall_unlock(gsec_kr);
875 OBD_FREE(coinfo, coinfo_size);
878 CERROR("failed request key: %ld\n", PTR_ERR(key));
881 CDEBUG(D_SEC, "obtained key %08x for %s\n", key->serial, desc);
883 /* once payload.data was pointed to a ctx, it never changes until
884 * we de-associate them; but parallel request_key() may return
885 * a key with payload.data == NULL at the same time. so we still
886 * need wirtelock of key->sem to serialize them.
888 down_write(&key->sem);
890 ctx = key_get_payload(key, 0);
892 LASSERT(atomic_read(&ctx->cc_refcount) >= 1);
893 LASSERT(ctx2gctx_keyring(ctx)->gck_key == key);
894 LASSERT(ll_read_key_usage(key) >= 2);
896 /* simply take a ref and return. it's upper layer's
897 * responsibility to detect & replace dead ctx.
899 atomic_inc(&ctx->cc_refcount);
901 /* pre initialization with a cli_ctx. this can't be done in
902 * key_instantiate() because we'v no enough information
905 ctx = ctx_create_kr(sec, vcred);
907 ctx_enlist_kr(ctx, is_root, 0);
908 bind_key_ctx(key, ctx);
910 ctx_start_timer_kr(ctx, KEYRING_UPCALL_TIMEOUT);
912 CDEBUG(D_SEC, "installed key %p <-> ctx %p (sec %p)\n",
915 /* we'd prefer to call key_revoke(), but we more like
916 * to revoke it within this key->sem locked period.
918 key_revoke_locked(key);
926 if (is_root && create_new)
927 request_key_unlink(key);
932 mutex_unlock(&gsec_kr->gsk_root_uc_lock);
937 void gss_sec_release_ctx_kr(struct ptlrpc_sec *sec,
938 struct ptlrpc_cli_ctx *ctx,
941 LASSERT(atomic_read(&sec->ps_refcount) > 0);
942 LASSERT(atomic_read(&ctx->cc_refcount) == 0);
943 ctx_release_kr(ctx, sync);
947 * flush context of normal user, we must resort to keyring itself to find out
948 * contexts which belong to me.
950 * Note here we suppose only to flush _my_ context, the "uid" will
951 * be ignored in the search.
954 void flush_user_ctx_cache_kr(struct ptlrpc_sec *sec,
956 int grace, int force)
961 /* nothing to do for reverse or rootonly sec */
962 if (sec_is_reverse(sec) || sec_is_rootonly(sec))
965 construct_key_desc(desc, sizeof(desc), sec, uid);
967 /* there should be only one valid key, but we put it in the
968 * loop in case of any weird cases */
970 key = request_key(&gss_key_type, desc, NULL);
972 CDEBUG(D_SEC, "No more key found for current user\n");
976 down_write(&key->sem);
978 kill_key_locked(key);
980 /* kill_key_locked() should usually revoke the key, but we
981 * revoke it again to make sure, e.g. some case the key may
982 * not well coupled with a context. */
983 key_revoke_locked(key);
987 request_key_unlink(key);
994 * flush context of root or all, we iterate through the list.
997 void flush_spec_ctx_cache_kr(struct ptlrpc_sec *sec, uid_t uid, int grace,
1000 struct gss_sec_keyring *gsec_kr;
1001 struct hlist_head freelist = HLIST_HEAD_INIT;
1002 struct hlist_node *next;
1003 struct ptlrpc_cli_ctx *ctx;
1006 gsec_kr = sec2gsec_keyring(sec);
1008 spin_lock(&sec->ps_lock);
1009 hlist_for_each_entry_safe(ctx, next, &gsec_kr->gsk_clist,
1011 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
1013 if (uid != -1 && uid != ctx->cc_vcred.vc_uid)
1016 /* at this moment there's at least 2 base reference:
1017 * key association and in-list. */
1018 if (atomic_read(&ctx->cc_refcount) > 2) {
1021 CWARN("flush busy ctx %p(%u->%s, extra ref %d)\n",
1022 ctx, ctx->cc_vcred.vc_uid,
1023 sec2target_str(ctx->cc_sec),
1024 atomic_read(&ctx->cc_refcount) - 2);
1027 set_bit(PTLRPC_CTX_DEAD_BIT, &ctx->cc_flags);
1029 clear_bit(PTLRPC_CTX_UPTODATE_BIT, &ctx->cc_flags);
1031 atomic_inc(&ctx->cc_refcount);
1033 if (ctx_unlist_kr(ctx, 1)) {
1034 hlist_add_head(&ctx->cc_cache, &freelist);
1036 LASSERT(atomic_read(&ctx->cc_refcount) >= 2);
1037 atomic_dec(&ctx->cc_refcount);
1040 spin_unlock(&sec->ps_lock);
1042 dispose_ctx_list_kr(&freelist);
1047 int gss_sec_flush_ctx_cache_kr(struct ptlrpc_sec *sec,
1048 uid_t uid, int grace, int force)
1052 CDEBUG(D_SEC, "sec %p(%d, nctx %d), uid %d, grace %d, force %d\n",
1053 sec, atomic_read(&sec->ps_refcount),
1054 atomic_read(&sec->ps_nctx),
1057 if (uid != -1 && uid != 0)
1058 flush_user_ctx_cache_kr(sec, uid, grace, force);
1060 flush_spec_ctx_cache_kr(sec, uid, grace, force);
1066 void gss_sec_gc_ctx_kr(struct ptlrpc_sec *sec)
1068 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
1069 struct hlist_head freelist = HLIST_HEAD_INIT;
1070 struct hlist_node *next;
1071 struct ptlrpc_cli_ctx *ctx;
1074 CWARN("running gc\n");
1076 spin_lock(&sec->ps_lock);
1077 hlist_for_each_entry_safe(ctx, next, &gsec_kr->gsk_clist,
1079 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
1081 atomic_inc(&ctx->cc_refcount);
1083 if (cli_ctx_check_death(ctx) && ctx_unlist_kr(ctx, 1)) {
1084 hlist_add_head(&ctx->cc_cache, &freelist);
1085 CWARN("unhashed ctx %p\n", ctx);
1087 LASSERT(atomic_read(&ctx->cc_refcount) >= 2);
1088 atomic_dec(&ctx->cc_refcount);
1091 spin_unlock(&sec->ps_lock);
1093 dispose_ctx_list_kr(&freelist);
1098 int gss_sec_display_kr(struct ptlrpc_sec *sec, struct seq_file *seq)
1100 struct gss_sec_keyring *gsec_kr = sec2gsec_keyring(sec);
1101 struct hlist_node *next;
1102 struct ptlrpc_cli_ctx *ctx;
1103 struct gss_cli_ctx *gctx;
1104 time64_t now = ktime_get_real_seconds();
1107 spin_lock(&sec->ps_lock);
1108 hlist_for_each_entry_safe(ctx, next, &gsec_kr->gsk_clist,
1114 gctx = ctx2gctx(ctx);
1115 key = ctx2gctx_keyring(ctx)->gck_key;
1117 gss_cli_ctx_flags2str(ctx->cc_flags,
1118 flags_str, sizeof(flags_str));
1120 if (gctx->gc_mechctx)
1121 lgss_display(gctx->gc_mechctx, mech, sizeof(mech));
1123 snprintf(mech, sizeof(mech), "N/A");
1124 mech[sizeof(mech) - 1] = '\0';
1127 "%p: uid %u, ref %d, expire %lld(%+lld), fl %s, seq %d, win %u, key %08x(ref %d), hdl %#llx:%#llx, mech: %s\n",
1128 ctx, ctx->cc_vcred.vc_uid,
1129 atomic_read(&ctx->cc_refcount),
1131 ctx->cc_expire ? ctx->cc_expire - now : 0,
1133 atomic_read(&gctx->gc_seq),
1135 key ? key->serial : 0,
1136 key ? ll_read_key_usage(key) : 0,
1137 gss_handle_to_u64(&gctx->gc_handle),
1138 gss_handle_to_u64(&gctx->gc_svc_handle),
1141 spin_unlock(&sec->ps_lock);
1146 /****************************************
1148 ****************************************/
1151 int gss_cli_ctx_refresh_kr(struct ptlrpc_cli_ctx *ctx)
1153 /* upcall is already on the way */
1154 struct gss_cli_ctx *gctx = ctx ? ctx2gctx(ctx) : NULL;
1156 /* record latest sequence number in buddy svcctx */
1157 if (gctx && !rawobj_empty(&gctx->gc_svc_handle) &&
1158 sec_is_reverse(gctx->gc_base.cc_sec)) {
1159 return gss_svc_upcall_update_sequence(&gctx->gc_svc_handle,
1160 (__u32)atomic_read(&gctx->gc_seq));
1166 int gss_cli_ctx_validate_kr(struct ptlrpc_cli_ctx *ctx)
1168 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
1169 LASSERT(ctx->cc_sec);
1171 if (cli_ctx_check_death(ctx)) {
1176 if (cli_ctx_is_ready(ctx))
1182 void gss_cli_ctx_die_kr(struct ptlrpc_cli_ctx *ctx, int grace)
1184 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
1185 LASSERT(ctx->cc_sec);
1187 cli_ctx_expire(ctx);
1191 /****************************************
1192 * (reverse) service *
1193 ****************************************/
1196 * reverse context could have nothing to do with keyrings. here we still keep
1197 * the version which bind to a key, for future reference.
1199 #define HAVE_REVERSE_CTX_NOKEY
1201 #ifdef HAVE_REVERSE_CTX_NOKEY
1204 int sec_install_rctx_kr(struct ptlrpc_sec *sec,
1205 struct ptlrpc_svc_ctx *svc_ctx)
1207 struct ptlrpc_cli_ctx *cli_ctx;
1208 struct vfs_cred vcred = { .vc_uid = 0 };
1214 cli_ctx = ctx_create_kr(sec, &vcred);
1215 if (cli_ctx == NULL)
1218 rc = gss_copy_rvc_cli_ctx(cli_ctx, svc_ctx);
1220 CERROR("failed copy reverse cli ctx: %d\n", rc);
1222 ctx_put_kr(cli_ctx, 1);
1226 rvs_sec_install_root_ctx_kr(sec, cli_ctx, NULL);
1228 ctx_put_kr(cli_ctx, 1);
1233 #else /* ! HAVE_REVERSE_CTX_NOKEY */
1236 int sec_install_rctx_kr(struct ptlrpc_sec *sec,
1237 struct ptlrpc_svc_ctx *svc_ctx)
1239 struct ptlrpc_cli_ctx *cli_ctx = NULL;
1241 struct vfs_cred vcred = { .vc_uid = 0 };
1249 construct_key_desc(desc, sizeof(desc), sec, 0);
1251 key = key_alloc(&gss_key_type, desc, 0, 0,
1252 KEY_POS_ALL | KEY_USR_ALL, 1);
1254 CERROR("failed to alloc key: %ld\n", PTR_ERR(key));
1255 return PTR_ERR(key);
1258 rc = key_instantiate_and_link(key, NULL, 0, NULL, NULL);
1260 CERROR("failed to instantiate key: %d\n", rc);
1264 down_write(&key->sem);
1266 LASSERT(!key_get_payload(key, 0));
1268 cli_ctx = ctx_create_kr(sec, &vcred);
1269 if (cli_ctx == NULL) {
1274 rc = gss_copy_rvc_cli_ctx(cli_ctx, svc_ctx);
1276 CERROR("failed copy reverse cli ctx: %d\n", rc);
1280 rvs_sec_install_root_ctx_kr(sec, cli_ctx, key);
1282 ctx_put_kr(cli_ctx, 1);
1283 up_write(&key->sem);
1292 ctx_put_kr(cli_ctx, 1);
1294 up_write(&key->sem);
1300 #endif /* HAVE_REVERSE_CTX_NOKEY */
1302 /****************************************
1304 ****************************************/
1307 int gss_svc_accept_kr(struct ptlrpc_request *req)
1309 return gss_svc_accept(&gss_policy_keyring, req);
1313 int gss_svc_install_rctx_kr(struct obd_import *imp,
1314 struct ptlrpc_svc_ctx *svc_ctx)
1316 struct ptlrpc_sec *sec;
1319 sec = sptlrpc_import_sec_ref(imp);
1322 rc = sec_install_rctx_kr(sec, svc_ctx);
1323 sptlrpc_sec_put(sec);
1328 /****************************************
1330 ****************************************/
1333 #ifdef HAVE_KEY_TYPE_INSTANTIATE_2ARGS
1334 int gss_kt_instantiate(struct key *key, struct key_preparsed_payload *prep)
1336 const void *data = prep->data;
1337 size_t datalen = prep->datalen;
1339 int gss_kt_instantiate(struct key *key, const void *data, size_t datalen)
1345 if (data != NULL || datalen != 0) {
1346 CERROR("invalid: data %p, len %lu\n", data, (long)datalen);
1350 if (key_get_payload(key, 0)) {
1351 CERROR("key already have payload\n");
1355 /* link the key to session keyring, so following context negotiation
1356 * rpc fired from user space could find this key. This will be unlinked
1357 * automatically when upcall processes die.
1359 * we can't do this through keyctl from userspace, because the upcall
1360 * might be neither possessor nor owner of the key (setuid).
1362 * the session keyring is created upon upcall, and don't change all
1363 * the way until upcall finished, so rcu lock is not needed here.
1365 LASSERT(current_cred()->session_keyring);
1368 rc = key_link(current_cred()->session_keyring, key);
1371 CERROR("failed to link key %08x to keyring %08x: %d\n",
1373 current_cred()->session_keyring->serial, rc);
1377 CDEBUG(D_SEC, "key %p instantiated, ctx %p\n", key,
1378 key_get_payload(key, 0));
1383 * called with key semaphore write locked. it means we can operate
1384 * on the context without fear of loosing refcount.
1387 #ifdef HAVE_KEY_TYPE_INSTANTIATE_2ARGS
1388 int gss_kt_update(struct key *key, struct key_preparsed_payload *prep)
1390 const void *data = prep->data;
1391 __u32 datalen32 = (__u32) prep->datalen;
1393 int gss_kt_update(struct key *key, const void *data, size_t datalen)
1395 __u32 datalen32 = (__u32) datalen;
1397 struct ptlrpc_cli_ctx *ctx = key_get_payload(key, 0);
1398 struct gss_cli_ctx *gctx;
1399 rawobj_t tmpobj = RAWOBJ_EMPTY;
1403 if (data == NULL || datalen32 == 0) {
1404 CWARN("invalid: data %p, len %lu\n", data, (long)datalen32);
1408 /* if upcall finished negotiation too fast (mostly likely because
1409 * of local error happened) and call kt_update(), the ctx
1410 * might be still NULL. but the key will finally be associate
1411 * with a context, or be revoked. if key status is fine, return
1412 * -EAGAIN to allow userspace sleep a while and call again. */
1414 CDEBUG(D_SEC, "update too soon: key %p(%x) flags %lx\n",
1415 key, key->serial, key->flags);
1417 rc = key_validate(key);
1424 LASSERT(atomic_read(&ctx->cc_refcount) > 0);
1425 LASSERT(ctx->cc_sec);
1427 ctx_clear_timer_kr(ctx);
1429 /* don't proceed if already refreshed */
1430 if (cli_ctx_is_refreshed(ctx)) {
1431 CWARN("ctx already done refresh\n");
1435 sptlrpc_cli_ctx_get(ctx);
1436 gctx = ctx2gctx(ctx);
1438 rc = buffer_extract_bytes(&data, &datalen32, &gctx->gc_win,
1439 sizeof(gctx->gc_win));
1441 CERROR("failed extract seq_win\n");
1445 if (gctx->gc_win == 0) {
1446 __u32 nego_rpc_err, nego_gss_err;
1448 rc = buffer_extract_bytes(&data, &datalen32, &nego_rpc_err,
1449 sizeof(nego_rpc_err));
1451 CERROR("cannot extract RPC: rc = %d\n", rc);
1455 rc = buffer_extract_bytes(&data, &datalen32, &nego_gss_err,
1456 sizeof(nego_gss_err));
1458 CERROR("failed to extract gss rc = %d\n", rc);
1462 CERROR("negotiation: rpc err %d, gss err %x\n",
1463 nego_rpc_err, nego_gss_err);
1465 rc = nego_rpc_err ? nego_rpc_err : -EACCES;
1467 rc = rawobj_extract_local_alloc(&gctx->gc_handle,
1468 (__u32 **) &data, &datalen32);
1470 CERROR("failed extract handle\n");
1474 rc = rawobj_extract_local(&tmpobj,
1475 (__u32 **) &data, &datalen32);
1477 CERROR("failed extract mech\n");
1481 rc = lgss_import_sec_context(&tmpobj,
1482 sec2gsec(ctx->cc_sec)->gs_mech,
1484 if (rc != GSS_S_COMPLETE)
1485 CERROR("failed import context\n");
1490 /* we don't care what current status of this ctx, even someone else
1491 * is operating on the ctx at the same time. we just add up our own
1494 gss_cli_ctx_uptodate(gctx);
1496 /* this will also revoke the key. has to be done before
1497 * wakeup waiters otherwise they can find the stale key */
1498 kill_key_locked(key);
1500 cli_ctx_expire(ctx);
1502 if (rc != -ERESTART)
1503 set_bit(PTLRPC_CTX_ERROR_BIT, &ctx->cc_flags);
1506 /* let user space think it's a success */
1507 sptlrpc_cli_ctx_put(ctx, 1);
1511 #ifndef HAVE_KEY_MATCH_DATA
1513 gss_kt_match(const struct key *key, const void *desc)
1515 return strcmp(key->description, (const char *) desc) == 0 &&
1516 !test_bit(KEY_FLAG_REVOKED, &key->flags);
1518 #else /* ! HAVE_KEY_MATCH_DATA */
1520 gss_kt_match(const struct key *key, const struct key_match_data *match_data)
1522 const char *desc = match_data->raw_data;
1524 return strcmp(key->description, desc) == 0 &&
1525 !test_bit(KEY_FLAG_REVOKED, &key->flags);
1529 * Preparse the match criterion.
1531 static int gss_kt_match_preparse(struct key_match_data *match_data)
1533 match_data->lookup_type = KEYRING_SEARCH_LOOKUP_DIRECT;
1534 match_data->cmp = gss_kt_match;
1537 #endif /* HAVE_KEY_MATCH_DATA */
1540 void gss_kt_destroy(struct key *key)
1543 LASSERT(!key_get_payload(key, 0));
1544 CDEBUG(D_SEC, "destroy key %p\n", key);
1549 void gss_kt_describe(const struct key *key, struct seq_file *s)
1551 if (key->description == NULL)
1552 seq_puts(s, "[null]");
1554 seq_puts(s, key->description);
1557 static struct key_type gss_key_type =
1561 .instantiate = gss_kt_instantiate,
1562 .update = gss_kt_update,
1563 #ifdef HAVE_KEY_MATCH_DATA
1564 .match_preparse = gss_kt_match_preparse,
1566 .match = gss_kt_match,
1568 .destroy = gss_kt_destroy,
1569 .describe = gss_kt_describe,
1572 /****************************************
1573 * lustre gss keyring policy *
1574 ****************************************/
1576 static struct ptlrpc_ctx_ops gss_keyring_ctxops = {
1577 .match = gss_cli_ctx_match,
1578 .refresh = gss_cli_ctx_refresh_kr,
1579 .validate = gss_cli_ctx_validate_kr,
1580 .die = gss_cli_ctx_die_kr,
1581 .sign = gss_cli_ctx_sign,
1582 .verify = gss_cli_ctx_verify,
1583 .seal = gss_cli_ctx_seal,
1584 .unseal = gss_cli_ctx_unseal,
1585 .wrap_bulk = gss_cli_ctx_wrap_bulk,
1586 .unwrap_bulk = gss_cli_ctx_unwrap_bulk,
1589 static struct ptlrpc_sec_cops gss_sec_keyring_cops = {
1590 .create_sec = gss_sec_create_kr,
1591 .destroy_sec = gss_sec_destroy_kr,
1592 .kill_sec = gss_sec_kill,
1593 .lookup_ctx = gss_sec_lookup_ctx_kr,
1594 .release_ctx = gss_sec_release_ctx_kr,
1595 .flush_ctx_cache = gss_sec_flush_ctx_cache_kr,
1596 .gc_ctx = gss_sec_gc_ctx_kr,
1597 .install_rctx = gss_sec_install_rctx,
1598 .alloc_reqbuf = gss_alloc_reqbuf,
1599 .free_reqbuf = gss_free_reqbuf,
1600 .alloc_repbuf = gss_alloc_repbuf,
1601 .free_repbuf = gss_free_repbuf,
1602 .enlarge_reqbuf = gss_enlarge_reqbuf,
1603 .display = gss_sec_display_kr,
1606 static struct ptlrpc_sec_sops gss_sec_keyring_sops = {
1607 .accept = gss_svc_accept_kr,
1608 .invalidate_ctx = gss_svc_invalidate_ctx,
1609 .alloc_rs = gss_svc_alloc_rs,
1610 .authorize = gss_svc_authorize,
1611 .free_rs = gss_svc_free_rs,
1612 .free_ctx = gss_svc_free_ctx,
1613 .prep_bulk = gss_svc_prep_bulk,
1614 .unwrap_bulk = gss_svc_unwrap_bulk,
1615 .wrap_bulk = gss_svc_wrap_bulk,
1616 .install_rctx = gss_svc_install_rctx_kr,
1619 static struct ptlrpc_sec_policy gss_policy_keyring = {
1620 .sp_owner = THIS_MODULE,
1621 .sp_name = "gss.keyring",
1622 .sp_policy = SPTLRPC_POLICY_GSS,
1623 .sp_cops = &gss_sec_keyring_cops,
1624 .sp_sops = &gss_sec_keyring_sops,
1628 int __init gss_init_keyring(void)
1632 rc = register_key_type(&gss_key_type);
1634 CERROR("failed to register keyring type: %d\n", rc);
1638 rc = sptlrpc_register_policy(&gss_policy_keyring);
1640 unregister_key_type(&gss_key_type);
1647 void __exit gss_exit_keyring(void)
1649 unregister_key_type(&gss_key_type);
1650 sptlrpc_unregister_policy(&gss_policy_keyring);