tjh.dev
Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
1Rules on how to access information in the Linux kernel sysfs
2
3The kernel-exported sysfs exports internal kernel implementation details
4and depends on internal kernel structures and layout. It is agreed upon
5by the kernel developers that the Linux kernel does not provide a stable
6internal API. As sysfs is a direct export of kernel internal
7structures, the sysfs interface cannot provide a stable interface either;
8it may always change along with internal kernel changes.
9
10To minimize the risk of breaking users of sysfs, which are in most cases
11low-level userspace applications, with a new kernel release, the users
12of sysfs must follow some rules to use an as-abstract-as-possible way to
13access this filesystem. The current udev and HAL programs already
14implement this and users are encouraged to plug, if possible, into the
15abstractions these programs provide instead of accessing sysfs directly.
16
17But if you really do want or need to access sysfs directly, please follow
18the following rules and then your programs should work with future
19versions of the sysfs interface.
20
21- Do not use libsysfs
22 It makes assumptions about sysfs which are not true. Its API does not
23 offer any abstraction, it exposes all the kernel driver-core
24 implementation details in its own API. Therefore it is not better than
25 reading directories and opening the files yourself.
26 Also, it is not actively maintained, in the sense of reflecting the
27 current kernel development. The goal of providing a stable interface
28 to sysfs has failed; it causes more problems than it solves. It
29 violates many of the rules in this document.
30
31- sysfs is always at /sys
32 Parsing /proc/mounts is a waste of time. Other mount points are a
33 system configuration bug you should not try to solve. For test cases,
34 possibly support a SYSFS_PATH environment variable to overwrite the
35 application's behavior, but never try to search for sysfs. Never try
36 to mount it, if you are not an early boot script.
37
38- devices are only "devices"
39 There is no such thing like class-, bus-, physical devices,
40 interfaces, and such that you can rely on in userspace. Everything is
41 just simply a "device". Class-, bus-, physical, ... types are just
42 kernel implementation details which should not be expected by
43 applications that look for devices in sysfs.
44
45 The properties of a device are:
46 o devpath (/devices/pci0000:00/0000:00:1d.1/usb2/2-2/2-2:1.0)
47 - identical to the DEVPATH value in the event sent from the kernel
48 at device creation and removal
49 - the unique key to the device at that point in time
50 - the kernel's path to the device directory without the leading
51 /sys, and always starting with with a slash
52 - all elements of a devpath must be real directories. Symlinks
53 pointing to /sys/devices must always be resolved to their real
54 target and the target path must be used to access the device.
55 That way the devpath to the device matches the devpath of the
56 kernel used at event time.
57 - using or exposing symlink values as elements in a devpath string
58 is a bug in the application
59
60 o kernel name (sda, tty, 0000:00:1f.2, ...)
61 - a directory name, identical to the last element of the devpath
62 - applications need to handle spaces and characters like '!' in
63 the name
64
65 o subsystem (block, tty, pci, ...)
66 - simple string, never a path or a link
67 - retrieved by reading the "subsystem"-link and using only the
68 last element of the target path
69
70 o driver (tg3, ata_piix, uhci_hcd)
71 - a simple string, which may contain spaces, never a path or a
72 link
73 - it is retrieved by reading the "driver"-link and using only the
74 last element of the target path
75 - devices which do not have "driver"-link just do not have a
76 driver; copying the driver value in a child device context is a
77 bug in the application
78
79 o attributes
80 - the files in the device directory or files below subdirectories
81 of the same device directory
82 - accessing attributes reached by a symlink pointing to another device,
83 like the "device"-link, is a bug in the application
84
85 Everything else is just a kernel driver-core implementation detail
86 that should not be assumed to be stable across kernel releases.
87
88- Properties of parent devices never belong into a child device.
89 Always look at the parent devices themselves for determining device
90 context properties. If the device 'eth0' or 'sda' does not have a
91 "driver"-link, then this device does not have a driver. Its value is empty.
92 Never copy any property of the parent-device into a child-device. Parent
93 device properties may change dynamically without any notice to the
94 child device.
95
96- Hierarchy in a single device tree
97 There is only one valid place in sysfs where hierarchy can be examined
98 and this is below: /sys/devices.
99 It is planned that all device directories will end up in the tree
100 below this directory.
101
102- Classification by subsystem
103 There are currently three places for classification of devices:
104 /sys/block, /sys/class and /sys/bus. It is planned that these will
105 not contain any device directories themselves, but only flat lists of
106 symlinks pointing to the unified /sys/devices tree.
107 All three places have completely different rules on how to access
108 device information. It is planned to merge all three
109 classification directories into one place at /sys/subsystem,
110 following the layout of the bus directories. All buses and
111 classes, including the converted block subsystem, will show up
112 there.
113 The devices belonging to a subsystem will create a symlink in the
114 "devices" directory at /sys/subsystem/<name>/devices.
115
116 If /sys/subsystem exists, /sys/bus, /sys/class and /sys/block can be
117 ignored. If it does not exist, you have always to scan all three
118 places, as the kernel is free to move a subsystem from one place to
119 the other, as long as the devices are still reachable by the same
120 subsystem name.
121
122 Assuming /sys/class/<subsystem> and /sys/bus/<subsystem>, or
123 /sys/block and /sys/class/block are not interchangeable is a bug in
124 the application.
125
126- Block
127 The converted block subsystem at /sys/class/block or
128 /sys/subsystem/block will contain the links for disks and partitions
129 at the same level, never in a hierarchy. Assuming the block subsytem to
130 contain only disks and not partition devices in the same flat list is
131 a bug in the application.
132
133- "device"-link and <subsystem>:<kernel name>-links
134 Never depend on the "device"-link. The "device"-link is a workaround
135 for the old layout, where class devices are not created in
136 /sys/devices/ like the bus devices. If the link-resolving of a
137 device directory does not end in /sys/devices/, you can use the
138 "device"-link to find the parent devices in /sys/devices/. That is the
139 single valid use of the "device"-link; it must never appear in any
140 path as an element. Assuming the existence of the "device"-link for
141 a device in /sys/devices/ is a bug in the application.
142 Accessing /sys/class/net/eth0/device is a bug in the application.
143
144 Never depend on the class-specific links back to the /sys/class
145 directory. These links are also a workaround for the design mistake
146 that class devices are not created in /sys/devices. If a device
147 directory does not contain directories for child devices, these links
148 may be used to find the child devices in /sys/class. That is the single
149 valid use of these links; they must never appear in any path as an
150 element. Assuming the existence of these links for devices which are
151 real child device directories in the /sys/devices tree is a bug in
152 the application.
153
154 It is planned to remove all these links when all class device
155 directories live in /sys/devices.
156
157- Position of devices along device chain can change.
158 Never depend on a specific parent device position in the devpath,
159 or the chain of parent devices. The kernel is free to insert devices into
160 the chain. You must always request the parent device you are looking for
161 by its subsystem value. You need to walk up the chain until you find
162 the device that matches the expected subsystem. Depending on a specific
163 position of a parent device or exposing relative paths using "../" to
164 access the chain of parents is a bug in the application.