1======================================================= 2Configfs - Userspace-driven Kernel Object Configuration 3======================================================= 4 5Joel Becker <joel.becker@oracle.com> 6 7Updated: 31 March 2005 8 9Copyright (c) 2005 Oracle Corporation, 10 Joel Becker <joel.becker@oracle.com> 11 12 13What is configfs? 14================= 15 16configfs is a ram-based filesystem that provides the converse of 17sysfs's functionality. Where sysfs is a filesystem-based view of 18kernel objects, configfs is a filesystem-based manager of kernel 19objects, or config_items. 20 21With sysfs, an object is created in kernel (for example, when a device 22is discovered) and it is registered with sysfs. Its attributes then 23appear in sysfs, allowing userspace to read the attributes via 24readdir(3)/read(2). It may allow some attributes to be modified via 25write(2). The important point is that the object is created and 26destroyed in kernel, the kernel controls the lifecycle of the sysfs 27representation, and sysfs is merely a window on all this. 28 29A configfs config_item is created via an explicit userspace operation: 30mkdir(2). It is destroyed via rmdir(2). The attributes appear at 31mkdir(2) time, and can be read or modified via read(2) and write(2). 32As with sysfs, readdir(3) queries the list of items and/or attributes. 33symlink(2) can be used to group items together. Unlike sysfs, the 34lifetime of the representation is completely driven by userspace. The 35kernel modules backing the items must respond to this. 36 37Both sysfs and configfs can and should exist together on the same 38system. One is not a replacement for the other. 39 40Using configfs 41============== 42 43configfs can be compiled as a module or into the kernel. You can access 44it by doing:: 45 46 mount -t configfs none /config 47 48The configfs tree will be empty unless client modules are also loaded. 49These are modules that register their item types with configfs as 50subsystems. Once a client subsystem is loaded, it will appear as a 51subdirectory (or more than one) under /config. Like sysfs, the 52configfs tree is always there, whether mounted on /config or not. 53 54An item is created via mkdir(2). The item's attributes will also 55appear at this time. readdir(3) can determine what the attributes are, 56read(2) can query their default values, and write(2) can store new 57values. Don't mix more than one attribute in one attribute file. 58 59There are two types of configfs attributes: 60 61* Normal attributes, which similar to sysfs attributes, are small ASCII text 62 files, with a maximum size of one page (PAGE_SIZE, 4096 on i386). Preferably 63 only one value per file should be used, and the same caveats from sysfs apply. 64 Configfs expects write(2) to store the entire buffer at once. When writing to 65 normal configfs attributes, userspace processes should first read the entire 66 file, modify the portions they wish to change, and then write the entire 67 buffer back. 68 69* Binary attributes, which are somewhat similar to sysfs binary attributes, 70 but with a few slight changes to semantics. The PAGE_SIZE limitation does not 71 apply, but the whole binary item must fit in single kernel vmalloc'ed buffer. 72 The write(2) calls from user space are buffered, and the attributes' 73 write_bin_attribute method will be invoked on the final close, therefore it is 74 imperative for user-space to check the return code of close(2) in order to 75 verify that the operation finished successfully. 76 To avoid a malicious user OOMing the kernel, there's a per-binary attribute 77 maximum buffer value. 78 79When an item needs to be destroyed, remove it with rmdir(2). An 80item cannot be destroyed if any other item has a link to it (via 81symlink(2)). Links can be removed via unlink(2). 82 83Configuring FakeNBD: an Example 84=============================== 85 86Imagine there's a Network Block Device (NBD) driver that allows you to 87access remote block devices. Call it FakeNBD. FakeNBD uses configfs 88for its configuration. Obviously, there will be a nice program that 89sysadmins use to configure FakeNBD, but somehow that program has to tell 90the driver about it. Here's where configfs comes in. 91 92When the FakeNBD driver is loaded, it registers itself with configfs. 93readdir(3) sees this just fine:: 94 95 # ls /config 96 fakenbd 97 98A fakenbd connection can be created with mkdir(2). The name is 99arbitrary, but likely the tool will make some use of the name. Perhaps 100it is a uuid or a disk name:: 101 102 # mkdir /config/fakenbd/disk1 103 # ls /config/fakenbd/disk1 104 target device rw 105 106The target attribute contains the IP address of the server FakeNBD will 107connect to. The device attribute is the device on the server. 108Predictably, the rw attribute determines whether the connection is 109read-only or read-write:: 110 111 # echo 10.0.0.1 > /config/fakenbd/disk1/target 112 # echo /dev/sda1 > /config/fakenbd/disk1/device 113 # echo 1 > /config/fakenbd/disk1/rw 114 115That's it. That's all there is. Now the device is configured, via the 116shell no less. 117 118Coding With configfs 119==================== 120 121Every object in configfs is a config_item. A config_item reflects an 122object in the subsystem. It has attributes that match values on that 123object. configfs handles the filesystem representation of that object 124and its attributes, allowing the subsystem to ignore all but the 125basic show/store interaction. 126 127Items are created and destroyed inside a config_group. A group is a 128collection of items that share the same attributes and operations. 129Items are created by mkdir(2) and removed by rmdir(2), but configfs 130handles that. The group has a set of operations to perform these tasks 131 132A subsystem is the top level of a client module. During initialization, 133the client module registers the subsystem with configfs, the subsystem 134appears as a directory at the top of the configfs filesystem. A 135subsystem is also a config_group, and can do everything a config_group 136can. 137 138struct config_item 139================== 140 141:: 142 143 struct config_item { 144 char *ci_name; 145 char ci_namebuf[UOBJ_NAME_LEN]; 146 struct kref ci_kref; 147 struct list_head ci_entry; 148 struct config_item *ci_parent; 149 struct config_group *ci_group; 150 struct config_item_type *ci_type; 151 struct dentry *ci_dentry; 152 }; 153 154 void config_item_init(struct config_item *); 155 void config_item_init_type_name(struct config_item *, 156 const char *name, 157 struct config_item_type *type); 158 struct config_item *config_item_get(struct config_item *); 159 void config_item_put(struct config_item *); 160 161Generally, struct config_item is embedded in a container structure, a 162structure that actually represents what the subsystem is doing. The 163config_item portion of that structure is how the object interacts with 164configfs. 165 166Whether statically defined in a source file or created by a parent 167config_group, a config_item must have one of the _init() functions 168called on it. This initializes the reference count and sets up the 169appropriate fields. 170 171All users of a config_item should have a reference on it via 172config_item_get(), and drop the reference when they are done via 173config_item_put(). 174 175By itself, a config_item cannot do much more than appear in configfs. 176Usually a subsystem wants the item to display and/or store attributes, 177among other things. For that, it needs a type. 178 179struct config_item_type 180======================= 181 182:: 183 184 struct configfs_item_operations { 185 void (*release)(struct config_item *); 186 int (*allow_link)(struct config_item *src, 187 struct config_item *target); 188 void (*drop_link)(struct config_item *src, 189 struct config_item *target); 190 }; 191 192 struct config_item_type { 193 struct module *ct_owner; 194 struct configfs_item_operations *ct_item_ops; 195 struct configfs_group_operations *ct_group_ops; 196 struct configfs_attribute **ct_attrs; 197 struct configfs_bin_attribute **ct_bin_attrs; 198 }; 199 200The most basic function of a config_item_type is to define what 201operations can be performed on a config_item. All items that have been 202allocated dynamically will need to provide the ct_item_ops->release() 203method. This method is called when the config_item's reference count 204reaches zero. 205 206struct configfs_attribute 207========================= 208 209:: 210 211 struct configfs_attribute { 212 char *ca_name; 213 struct module *ca_owner; 214 umode_t ca_mode; 215 ssize_t (*show)(struct config_item *, char *); 216 ssize_t (*store)(struct config_item *, const char *, size_t); 217 }; 218 219When a config_item wants an attribute to appear as a file in the item's 220configfs directory, it must define a configfs_attribute describing it. 221It then adds the attribute to the NULL-terminated array 222config_item_type->ct_attrs. When the item appears in configfs, the 223attribute file will appear with the configfs_attribute->ca_name 224filename. configfs_attribute->ca_mode specifies the file permissions. 225 226If an attribute is readable and provides a ->show method, that method will 227be called whenever userspace asks for a read(2) on the attribute. If an 228attribute is writable and provides a ->store method, that method will be 229called whenever userspace asks for a write(2) on the attribute. 230 231struct configfs_bin_attribute 232============================= 233 234:: 235 236 struct configfs_bin_attribute { 237 struct configfs_attribute cb_attr; 238 void *cb_private; 239 size_t cb_max_size; 240 }; 241 242The binary attribute is used when the one needs to use binary blob to 243appear as the contents of a file in the item's configfs directory. 244To do so add the binary attribute to the NULL-terminated array 245config_item_type->ct_bin_attrs, and the item appears in configfs, the 246attribute file will appear with the configfs_bin_attribute->cb_attr.ca_name 247filename. configfs_bin_attribute->cb_attr.ca_mode specifies the file 248permissions. 249The cb_private member is provided for use by the driver, while the 250cb_max_size member specifies the maximum amount of vmalloc buffer 251to be used. 252 253If binary attribute is readable and the config_item provides a 254ct_item_ops->read_bin_attribute() method, that method will be called 255whenever userspace asks for a read(2) on the attribute. The converse 256will happen for write(2). The reads/writes are bufferred so only a 257single read/write will occur; the attributes' need not concern itself 258with it. 259 260struct config_group 261=================== 262 263A config_item cannot live in a vacuum. The only way one can be created 264is via mkdir(2) on a config_group. This will trigger creation of a 265child item:: 266 267 struct config_group { 268 struct config_item cg_item; 269 struct list_head cg_children; 270 struct configfs_subsystem *cg_subsys; 271 struct list_head default_groups; 272 struct list_head group_entry; 273 }; 274 275 void config_group_init(struct config_group *group); 276 void config_group_init_type_name(struct config_group *group, 277 const char *name, 278 struct config_item_type *type); 279 280 281The config_group structure contains a config_item. Properly configuring 282that item means that a group can behave as an item in its own right. 283However, it can do more: it can create child items or groups. This is 284accomplished via the group operations specified on the group's 285config_item_type:: 286 287 struct configfs_group_operations { 288 struct config_item *(*make_item)(struct config_group *group, 289 const char *name); 290 struct config_group *(*make_group)(struct config_group *group, 291 const char *name); 292 int (*commit_item)(struct config_item *item); 293 void (*disconnect_notify)(struct config_group *group, 294 struct config_item *item); 295 void (*drop_item)(struct config_group *group, 296 struct config_item *item); 297 }; 298 299A group creates child items by providing the 300ct_group_ops->make_item() method. If provided, this method is called from 301mkdir(2) in the group's directory. The subsystem allocates a new 302config_item (or more likely, its container structure), initializes it, 303and returns it to configfs. Configfs will then populate the filesystem 304tree to reflect the new item. 305 306If the subsystem wants the child to be a group itself, the subsystem 307provides ct_group_ops->make_group(). Everything else behaves the same, 308using the group _init() functions on the group. 309 310Finally, when userspace calls rmdir(2) on the item or group, 311ct_group_ops->drop_item() is called. As a config_group is also a 312config_item, it is not necessary for a separate drop_group() method. 313The subsystem must config_item_put() the reference that was initialized 314upon item allocation. If a subsystem has no work to do, it may omit 315the ct_group_ops->drop_item() method, and configfs will call 316config_item_put() on the item on behalf of the subsystem. 317 318Important: 319 drop_item() is void, and as such cannot fail. When rmdir(2) 320 is called, configfs WILL remove the item from the filesystem tree 321 (assuming that it has no children to keep it busy). The subsystem is 322 responsible for responding to this. If the subsystem has references to 323 the item in other threads, the memory is safe. It may take some time 324 for the item to actually disappear from the subsystem's usage. But it 325 is gone from configfs. 326 327When drop_item() is called, the item's linkage has already been torn 328down. It no longer has a reference on its parent and has no place in 329the item hierarchy. If a client needs to do some cleanup before this 330teardown happens, the subsystem can implement the 331ct_group_ops->disconnect_notify() method. The method is called after 332configfs has removed the item from the filesystem view but before the 333item is removed from its parent group. Like drop_item(), 334disconnect_notify() is void and cannot fail. Client subsystems should 335not drop any references here, as they still must do it in drop_item(). 336 337A config_group cannot be removed while it still has child items. This 338is implemented in the configfs rmdir(2) code. ->drop_item() will not be 339called, as the item has not been dropped. rmdir(2) will fail, as the 340directory is not empty. 341 342struct configfs_subsystem 343========================= 344 345A subsystem must register itself, usually at module_init time. This 346tells configfs to make the subsystem appear in the file tree:: 347 348 struct configfs_subsystem { 349 struct config_group su_group; 350 struct mutex su_mutex; 351 }; 352 353 int configfs_register_subsystem(struct configfs_subsystem *subsys); 354 void configfs_unregister_subsystem(struct configfs_subsystem *subsys); 355 356A subsystem consists of a toplevel config_group and a mutex. 357The group is where child config_items are created. For a subsystem, 358this group is usually defined statically. Before calling 359configfs_register_subsystem(), the subsystem must have initialized the 360group via the usual group _init() functions, and it must also have 361initialized the mutex. 362 363When the register call returns, the subsystem is live, and it 364will be visible via configfs. At that point, mkdir(2) can be called and 365the subsystem must be ready for it. 366 367An Example 368========== 369 370The best example of these basic concepts is the simple_children 371subsystem/group and the simple_child item in 372samples/configfs/configfs_sample.c. It shows a trivial object displaying 373and storing an attribute, and a simple group creating and destroying 374these children. 375 376Hierarchy Navigation and the Subsystem Mutex 377============================================ 378 379There is an extra bonus that configfs provides. The config_groups and 380config_items are arranged in a hierarchy due to the fact that they 381appear in a filesystem. A subsystem is NEVER to touch the filesystem 382parts, but the subsystem might be interested in this hierarchy. For 383this reason, the hierarchy is mirrored via the config_group->cg_children 384and config_item->ci_parent structure members. 385 386A subsystem can navigate the cg_children list and the ci_parent pointer 387to see the tree created by the subsystem. This can race with configfs' 388management of the hierarchy, so configfs uses the subsystem mutex to 389protect modifications. Whenever a subsystem wants to navigate the 390hierarchy, it must do so under the protection of the subsystem 391mutex. 392 393A subsystem will be prevented from acquiring the mutex while a newly 394allocated item has not been linked into this hierarchy. Similarly, it 395will not be able to acquire the mutex while a dropping item has not 396yet been unlinked. This means that an item's ci_parent pointer will 397never be NULL while the item is in configfs, and that an item will only 398be in its parent's cg_children list for the same duration. This allows 399a subsystem to trust ci_parent and cg_children while they hold the 400mutex. 401 402Item Aggregation Via symlink(2) 403=============================== 404 405configfs provides a simple group via the group->item parent/child 406relationship. Often, however, a larger environment requires aggregation 407outside of the parent/child connection. This is implemented via 408symlink(2). 409 410A config_item may provide the ct_item_ops->allow_link() and 411ct_item_ops->drop_link() methods. If the ->allow_link() method exists, 412symlink(2) may be called with the config_item as the source of the link. 413These links are only allowed between configfs config_items. Any 414symlink(2) attempt outside the configfs filesystem will be denied. 415 416When symlink(2) is called, the source config_item's ->allow_link() 417method is called with itself and a target item. If the source item 418allows linking to target item, it returns 0. A source item may wish to 419reject a link if it only wants links to a certain type of object (say, 420in its own subsystem). 421 422When unlink(2) is called on the symbolic link, the source item is 423notified via the ->drop_link() method. Like the ->drop_item() method, 424this is a void function and cannot return failure. The subsystem is 425responsible for responding to the change. 426 427A config_item cannot be removed while it links to any other item, nor 428can it be removed while an item links to it. Dangling symlinks are not 429allowed in configfs. 430 431Automatically Created Subgroups 432=============================== 433 434A new config_group may want to have two types of child config_items. 435While this could be codified by magic names in ->make_item(), it is much 436more explicit to have a method whereby userspace sees this divergence. 437 438Rather than have a group where some items behave differently than 439others, configfs provides a method whereby one or many subgroups are 440automatically created inside the parent at its creation. Thus, 441mkdir("parent") results in "parent", "parent/subgroup1", up through 442"parent/subgroupN". Items of type 1 can now be created in 443"parent/subgroup1", and items of type N can be created in 444"parent/subgroupN". 445 446These automatic subgroups, or default groups, do not preclude other 447children of the parent group. If ct_group_ops->make_group() exists, 448other child groups can be created on the parent group directly. 449 450A configfs subsystem specifies default groups by adding them using the 451configfs_add_default_group() function to the parent config_group 452structure. Each added group is populated in the configfs tree at the same 453time as the parent group. Similarly, they are removed at the same time 454as the parent. No extra notification is provided. When a ->drop_item() 455method call notifies the subsystem the parent group is going away, it 456also means every default group child associated with that parent group. 457 458As a consequence of this, default groups cannot be removed directly via 459rmdir(2). They also are not considered when rmdir(2) on the parent 460group is checking for children. 461 462Dependent Subsystems 463==================== 464 465Sometimes other drivers depend on particular configfs items. For 466example, ocfs2 mounts depend on a heartbeat region item. If that 467region item is removed with rmdir(2), the ocfs2 mount must BUG or go 468readonly. Not happy. 469 470configfs provides two additional API calls: configfs_depend_item() and 471configfs_undepend_item(). A client driver can call 472configfs_depend_item() on an existing item to tell configfs that it is 473depended on. configfs will then return -EBUSY from rmdir(2) for that 474item. When the item is no longer depended on, the client driver calls 475configfs_undepend_item() on it. 476 477These API cannot be called underneath any configfs callbacks, as 478they will conflict. They can block and allocate. A client driver 479probably shouldn't calling them of its own gumption. Rather it should 480be providing an API that external subsystems call. 481 482How does this work? Imagine the ocfs2 mount process. When it mounts, 483it asks for a heartbeat region item. This is done via a call into the 484heartbeat code. Inside the heartbeat code, the region item is looked 485up. Here, the heartbeat code calls configfs_depend_item(). If it 486succeeds, then heartbeat knows the region is safe to give to ocfs2. 487If it fails, it was being torn down anyway, and heartbeat can gracefully 488pass up an error. 489 490Committable Items 491================= 492 493Note: 494 Committable items are currently unimplemented. 495 496Some config_items cannot have a valid initial state. That is, no 497default values can be specified for the item's attributes such that the 498item can do its work. Userspace must configure one or more attributes, 499after which the subsystem can start whatever entity this item 500represents. 501 502Consider the FakeNBD device from above. Without a target address *and* 503a target device, the subsystem has no idea what block device to import. 504The simple example assumes that the subsystem merely waits until all the 505appropriate attributes are configured, and then connects. This will, 506indeed, work, but now every attribute store must check if the attributes 507are initialized. Every attribute store must fire off the connection if 508that condition is met. 509 510Far better would be an explicit action notifying the subsystem that the 511config_item is ready to go. More importantly, an explicit action allows 512the subsystem to provide feedback as to whether the attributes are 513initialized in a way that makes sense. configfs provides this as 514committable items. 515 516configfs still uses only normal filesystem operations. An item is 517committed via rename(2). The item is moved from a directory where it 518can be modified to a directory where it cannot. 519 520Any group that provides the ct_group_ops->commit_item() method has 521committable items. When this group appears in configfs, mkdir(2) will 522not work directly in the group. Instead, the group will have two 523subdirectories: "live" and "pending". The "live" directory does not 524support mkdir(2) or rmdir(2) either. It only allows rename(2). The 525"pending" directory does allow mkdir(2) and rmdir(2). An item is 526created in the "pending" directory. Its attributes can be modified at 527will. Userspace commits the item by renaming it into the "live" 528directory. At this point, the subsystem receives the ->commit_item() 529callback. If all required attributes are filled to satisfaction, the 530method returns zero and the item is moved to the "live" directory. 531 532As rmdir(2) does not work in the "live" directory, an item must be 533shutdown, or "uncommitted". Again, this is done via rename(2), this 534time from the "live" directory back to the "pending" one. The subsystem 535is notified by the ct_group_ops->uncommit_object() method. 536