xref: /titanic_41/usr/src/uts/common/fs/zfs/zfs_onexit.c (revision be6fd75a69ae679453d9cda5bff3326111e6d1ca)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright (c) 2013 by Delphix. All rights reserved.
24  */
25 
26 #include <sys/types.h>
27 #include <sys/param.h>
28 #include <sys/errno.h>
29 #include <sys/open.h>
30 #include <sys/kmem.h>
31 #include <sys/conf.h>
32 #include <sys/ddi.h>
33 #include <sys/sunddi.h>
34 #include <sys/zfs_ioctl.h>
35 #include <sys/mkdev.h>
36 #include <sys/zfs_onexit.h>
37 #include <sys/zvol.h>
38 
39 /*
40  * ZFS kernel routines may add/delete callback routines to be invoked
41  * upon process exit (triggered via the close operation from the /dev/zfs
42  * driver).
43  *
44  * These cleanup callbacks are intended to allow for the accumulation
45  * of kernel state across multiple ioctls.  User processes participate
46  * by opening ZFS_DEV with O_EXCL. This causes the ZFS driver to do a
47  * clone-open, generating a unique minor number. The process then passes
48  * along that file descriptor to each ioctl that might have a cleanup operation.
49  *
50  * Consumers of the onexit routines should call zfs_onexit_fd_hold() early
51  * on to validate the given fd and add a reference to its file table entry.
52  * This allows the consumer to do its work and then add a callback, knowing
53  * that zfs_onexit_add_cb() won't fail with EBADF.  When finished, consumers
54  * should call zfs_onexit_fd_rele().
55  *
56  * A simple example is zfs_ioc_recv(), where we might create an AVL tree
57  * with dataset/GUID mappings and then reuse that tree on subsequent
58  * zfs_ioc_recv() calls.
59  *
60  * On the first zfs_ioc_recv() call, dmu_recv_stream() will kmem_alloc()
61  * the AVL tree and pass it along with a callback function to
62  * zfs_onexit_add_cb(). The zfs_onexit_add_cb() routine will register the
63  * callback and return an action handle.
64  *
65  * The action handle is then passed from user space to subsequent
66  * zfs_ioc_recv() calls, so that dmu_recv_stream() can fetch its AVL tree
67  * by calling zfs_onexit_cb_data() with the device minor number and
68  * action handle.
69  *
70  * If the user process exits abnormally, the callback is invoked implicitly
71  * as part of the driver close operation.  Once the user space process is
72  * finished with the accumulated kernel state, it can also just call close(2)
73  * on the cleanup fd to trigger the cleanup callback.
74  */
75 
76 void
zfs_onexit_init(zfs_onexit_t ** zop)77 zfs_onexit_init(zfs_onexit_t **zop)
78 {
79 	zfs_onexit_t *zo;
80 
81 	zo = *zop = kmem_zalloc(sizeof (zfs_onexit_t), KM_SLEEP);
82 	mutex_init(&zo->zo_lock, NULL, MUTEX_DEFAULT, NULL);
83 	list_create(&zo->zo_actions, sizeof (zfs_onexit_action_node_t),
84 	    offsetof(zfs_onexit_action_node_t, za_link));
85 }
86 
87 void
zfs_onexit_destroy(zfs_onexit_t * zo)88 zfs_onexit_destroy(zfs_onexit_t *zo)
89 {
90 	zfs_onexit_action_node_t *ap;
91 
92 	mutex_enter(&zo->zo_lock);
93 	while ((ap = list_head(&zo->zo_actions)) != NULL) {
94 		list_remove(&zo->zo_actions, ap);
95 		mutex_exit(&zo->zo_lock);
96 		ap->za_func(ap->za_data);
97 		kmem_free(ap, sizeof (zfs_onexit_action_node_t));
98 		mutex_enter(&zo->zo_lock);
99 	}
100 	mutex_exit(&zo->zo_lock);
101 
102 	list_destroy(&zo->zo_actions);
103 	mutex_destroy(&zo->zo_lock);
104 	kmem_free(zo, sizeof (zfs_onexit_t));
105 }
106 
107 static int
zfs_onexit_minor_to_state(minor_t minor,zfs_onexit_t ** zo)108 zfs_onexit_minor_to_state(minor_t minor, zfs_onexit_t **zo)
109 {
110 	*zo = zfsdev_get_soft_state(minor, ZSST_CTLDEV);
111 	if (*zo == NULL)
112 		return (SET_ERROR(EBADF));
113 
114 	return (0);
115 }
116 
117 /*
118  * Consumers might need to operate by minor number instead of fd, since
119  * they might be running in another thread (e.g. txg_sync_thread). Callers
120  * of this function must call zfs_onexit_fd_rele() when they're finished
121  * using the minor number.
122  */
123 int
zfs_onexit_fd_hold(int fd,minor_t * minorp)124 zfs_onexit_fd_hold(int fd, minor_t *minorp)
125 {
126 	file_t *fp;
127 	zfs_onexit_t *zo;
128 
129 	fp = getf(fd);
130 	if (fp == NULL)
131 		return (SET_ERROR(EBADF));
132 
133 	*minorp = getminor(fp->f_vnode->v_rdev);
134 	return (zfs_onexit_minor_to_state(*minorp, &zo));
135 }
136 
137 void
zfs_onexit_fd_rele(int fd)138 zfs_onexit_fd_rele(int fd)
139 {
140 	releasef(fd);
141 }
142 
143 /*
144  * Add a callback to be invoked when the calling process exits.
145  */
146 int
zfs_onexit_add_cb(minor_t minor,void (* func)(void *),void * data,uint64_t * action_handle)147 zfs_onexit_add_cb(minor_t minor, void (*func)(void *), void *data,
148     uint64_t *action_handle)
149 {
150 	zfs_onexit_t *zo;
151 	zfs_onexit_action_node_t *ap;
152 	int error;
153 
154 	error = zfs_onexit_minor_to_state(minor, &zo);
155 	if (error)
156 		return (error);
157 
158 	ap = kmem_alloc(sizeof (zfs_onexit_action_node_t), KM_SLEEP);
159 	list_link_init(&ap->za_link);
160 	ap->za_func = func;
161 	ap->za_data = data;
162 
163 	mutex_enter(&zo->zo_lock);
164 	list_insert_tail(&zo->zo_actions, ap);
165 	mutex_exit(&zo->zo_lock);
166 	if (action_handle)
167 		*action_handle = (uint64_t)(uintptr_t)ap;
168 
169 	return (0);
170 }
171 
172 static zfs_onexit_action_node_t *
zfs_onexit_find_cb(zfs_onexit_t * zo,uint64_t action_handle)173 zfs_onexit_find_cb(zfs_onexit_t *zo, uint64_t action_handle)
174 {
175 	zfs_onexit_action_node_t *match;
176 	zfs_onexit_action_node_t *ap;
177 	list_t *l;
178 
179 	ASSERT(MUTEX_HELD(&zo->zo_lock));
180 
181 	match = (zfs_onexit_action_node_t *)(uintptr_t)action_handle;
182 	l = &zo->zo_actions;
183 	for (ap = list_head(l); ap != NULL; ap = list_next(l, ap)) {
184 		if (match == ap)
185 			break;
186 	}
187 	return (ap);
188 }
189 
190 /*
191  * Delete the callback, triggering it first if 'fire' is set.
192  */
193 int
zfs_onexit_del_cb(minor_t minor,uint64_t action_handle,boolean_t fire)194 zfs_onexit_del_cb(minor_t minor, uint64_t action_handle, boolean_t fire)
195 {
196 	zfs_onexit_t *zo;
197 	zfs_onexit_action_node_t *ap;
198 	int error;
199 
200 	error = zfs_onexit_minor_to_state(minor, &zo);
201 	if (error)
202 		return (error);
203 
204 	mutex_enter(&zo->zo_lock);
205 	ap = zfs_onexit_find_cb(zo, action_handle);
206 	if (ap != NULL) {
207 		list_remove(&zo->zo_actions, ap);
208 		mutex_exit(&zo->zo_lock);
209 		if (fire)
210 			ap->za_func(ap->za_data);
211 		kmem_free(ap, sizeof (zfs_onexit_action_node_t));
212 	} else {
213 		mutex_exit(&zo->zo_lock);
214 		error = SET_ERROR(ENOENT);
215 	}
216 
217 	return (error);
218 }
219 
220 /*
221  * Return the data associated with this callback.  This allows consumers
222  * of the cleanup-on-exit interfaces to stash kernel data across system
223  * calls, knowing that it will be cleaned up if the calling process exits.
224  */
225 int
zfs_onexit_cb_data(minor_t minor,uint64_t action_handle,void ** data)226 zfs_onexit_cb_data(minor_t minor, uint64_t action_handle, void **data)
227 {
228 	zfs_onexit_t *zo;
229 	zfs_onexit_action_node_t *ap;
230 	int error;
231 
232 	*data = NULL;
233 
234 	error = zfs_onexit_minor_to_state(minor, &zo);
235 	if (error)
236 		return (error);
237 
238 	mutex_enter(&zo->zo_lock);
239 	ap = zfs_onexit_find_cb(zo, action_handle);
240 	if (ap != NULL)
241 		*data = ap->za_data;
242 	else
243 		error = SET_ERROR(ENOENT);
244 	mutex_exit(&zo->zo_lock);
245 
246 	return (error);
247 }
248