xref: /linux/Documentation/driver-api/device_link.rst (revision a4eb44a6435d6d8f9e642407a4a06f65eb90ca04)
1.. _device_link:
2
3============
4Device links
5============
6
7By default, the driver core only enforces dependencies between devices
8that are borne out of a parent/child relationship within the device
9hierarchy: When suspending, resuming or shutting down the system, devices
10are ordered based on this relationship, i.e. children are always suspended
11before their parent, and the parent is always resumed before its children.
12
13Sometimes there is a need to represent device dependencies beyond the
14mere parent/child relationship, e.g. between siblings, and have the
15driver core automatically take care of them.
16
17Secondly, the driver core by default does not enforce any driver presence
18dependencies, i.e. that one device must be bound to a driver before
19another one can probe or function correctly.
20
21Often these two dependency types come together, so a device depends on
22another one both with regards to driver presence *and* with regards to
23suspend/resume and shutdown ordering.
24
25Device links allow representation of such dependencies in the driver core.
26
27In its standard or *managed* form, a device link combines *both* dependency
28types:  It guarantees correct suspend/resume and shutdown ordering between a
29"supplier" device and its "consumer" devices, and it guarantees driver
30presence on the supplier.  The consumer devices are not probed before the
31supplier is bound to a driver, and they're unbound before the supplier
32is unbound.
33
34When driver presence on the supplier is irrelevant and only correct
35suspend/resume and shutdown ordering is needed, the device link may
36simply be set up with the ``DL_FLAG_STATELESS`` flag.  In other words,
37enforcing driver presence on the supplier is optional.
38
39Another optional feature is runtime PM integration:  By setting the
40``DL_FLAG_PM_RUNTIME`` flag on addition of the device link, the PM core
41is instructed to runtime resume the supplier and keep it active
42whenever and for as long as the consumer is runtime resumed.
43
44Usage
45=====
46
47The earliest point in time when device links can be added is after
48:c:func:`device_add()` has been called for the supplier and
49:c:func:`device_initialize()` has been called for the consumer.
50
51It is legal to add them later, but care must be taken that the system
52remains in a consistent state:  E.g. a device link cannot be added in
53the midst of a suspend/resume transition, so either commencement of
54such a transition needs to be prevented with :c:func:`lock_system_sleep()`,
55or the device link needs to be added from a function which is guaranteed
56not to run in parallel to a suspend/resume transition, such as from a
57device ``->probe`` callback or a boot-time PCI quirk.
58
59Another example for an inconsistent state would be a device link that
60represents a driver presence dependency, yet is added from the consumer's
61``->probe`` callback while the supplier hasn't started to probe yet:  Had the
62driver core known about the device link earlier, it wouldn't have probed the
63consumer in the first place.  The onus is thus on the consumer to check
64presence of the supplier after adding the link, and defer probing on
65non-presence.  [Note that it is valid to create a link from the consumer's
66``->probe`` callback while the supplier is still probing, but the consumer must
67know that the supplier is functional already at the link creation time (that is
68the case, for instance, if the consumer has just acquired some resources that
69would not have been available had the supplier not been functional then).]
70
71If a device link with ``DL_FLAG_STATELESS`` set (i.e. a stateless device link)
72is added in the ``->probe`` callback of the supplier or consumer driver, it is
73typically deleted in its ``->remove`` callback for symmetry.  That way, if the
74driver is compiled as a module, the device link is added on module load and
75orderly deleted on unload.  The same restrictions that apply to device link
76addition (e.g. exclusion of a parallel suspend/resume transition) apply equally
77to deletion.  Device links managed by the driver core are deleted automatically
78by it.
79
80Several flags may be specified on device link addition, two of which
81have already been mentioned above:  ``DL_FLAG_STATELESS`` to express that no
82driver presence dependency is needed (but only correct suspend/resume and
83shutdown ordering) and ``DL_FLAG_PM_RUNTIME`` to express that runtime PM
84integration is desired.
85
86Two other flags are specifically targeted at use cases where the device
87link is added from the consumer's ``->probe`` callback:  ``DL_FLAG_RPM_ACTIVE``
88can be specified to runtime resume the supplier and prevent it from suspending
89before the consumer is runtime suspended.  ``DL_FLAG_AUTOREMOVE_CONSUMER``
90causes the device link to be automatically purged when the consumer fails to
91probe or later unbinds.
92
93Similarly, when the device link is added from supplier's ``->probe`` callback,
94``DL_FLAG_AUTOREMOVE_SUPPLIER`` causes the device link to be automatically
95purged when the supplier fails to probe or later unbinds.
96
97If neither ``DL_FLAG_AUTOREMOVE_CONSUMER`` nor ``DL_FLAG_AUTOREMOVE_SUPPLIER``
98is set, ``DL_FLAG_AUTOPROBE_CONSUMER`` can be used to request the driver core
99to probe for a driver for the consumer driver on the link automatically after
100a driver has been bound to the supplier device.
101
102Note, however, that any combinations of ``DL_FLAG_AUTOREMOVE_CONSUMER``,
103``DL_FLAG_AUTOREMOVE_SUPPLIER`` or ``DL_FLAG_AUTOPROBE_CONSUMER`` with
104``DL_FLAG_STATELESS`` are invalid and cannot be used.
105
106Limitations
107===========
108
109Driver authors should be aware that a driver presence dependency for managed
110device links (i.e. when ``DL_FLAG_STATELESS`` is not specified on link addition)
111may cause probing of the consumer to be deferred indefinitely.  This can become
112a problem if the consumer is required to probe before a certain initcall level
113is reached.  Worse, if the supplier driver is blacklisted or missing, the
114consumer will never be probed.
115
116Moreover, managed device links cannot be deleted directly.  They are deleted
117by the driver core when they are not necessary any more in accordance with the
118``DL_FLAG_AUTOREMOVE_CONSUMER`` and ``DL_FLAG_AUTOREMOVE_SUPPLIER`` flags.
119However, stateless device links (i.e. device links with ``DL_FLAG_STATELESS``
120set) are expected to be removed by whoever called :c:func:`device_link_add()`
121to add them with the help of either :c:func:`device_link_del()` or
122:c:func:`device_link_remove()`.
123
124Passing ``DL_FLAG_RPM_ACTIVE`` along with ``DL_FLAG_STATELESS`` to
125:c:func:`device_link_add()` may cause the PM-runtime usage counter of the
126supplier device to remain nonzero after a subsequent invocation of either
127:c:func:`device_link_del()` or :c:func:`device_link_remove()` to remove the
128device link returned by it.  This happens if :c:func:`device_link_add()` is
129called twice in a row for the same consumer-supplier pair without removing the
130link between these calls, in which case allowing the PM-runtime usage counter
131of the supplier to drop on an attempt to remove the link may cause it to be
132suspended while the consumer is still PM-runtime-active and that has to be
133avoided.  [To work around this limitation it is sufficient to let the consumer
134runtime suspend at least once, or call :c:func:`pm_runtime_set_suspended()` for
135it with PM-runtime disabled, between the :c:func:`device_link_add()` and
136:c:func:`device_link_del()` or :c:func:`device_link_remove()` calls.]
137
138Sometimes drivers depend on optional resources.  They are able to operate
139in a degraded mode (reduced feature set or performance) when those resources
140are not present.  An example is an SPI controller that can use a DMA engine
141or work in PIO mode.  The controller can determine presence of the optional
142resources at probe time but on non-presence there is no way to know whether
143they will become available in the near future (due to a supplier driver
144probing) or never.  Consequently it cannot be determined whether to defer
145probing or not.  It would be possible to notify drivers when optional
146resources become available after probing, but it would come at a high cost
147for drivers as switching between modes of operation at runtime based on the
148availability of such resources would be much more complex than a mechanism
149based on probe deferral.  In any case optional resources are beyond the
150scope of device links.
151
152Examples
153========
154
155* An MMU device exists alongside a busmaster device, both are in the same
156  power domain.  The MMU implements DMA address translation for the busmaster
157  device and shall be runtime resumed and kept active whenever and as long
158  as the busmaster device is active.  The busmaster device's driver shall
159  not bind before the MMU is bound.  To achieve this, a device link with
160  runtime PM integration is added from the busmaster device (consumer)
161  to the MMU device (supplier).  The effect with regards to runtime PM
162  is the same as if the MMU was the parent of the master device.
163
164  The fact that both devices share the same power domain would normally
165  suggest usage of a struct dev_pm_domain or struct generic_pm_domain,
166  however these are not independent devices that happen to share a power
167  switch, but rather the MMU device serves the busmaster device and is
168  useless without it.  A device link creates a synthetic hierarchical
169  relationship between the devices and is thus more apt.
170
171* A Thunderbolt host controller comprises a number of PCIe hotplug ports
172  and an NHI device to manage the PCIe switch.  On resume from system sleep,
173  the NHI device needs to re-establish PCI tunnels to attached devices
174  before the hotplug ports can resume.  If the hotplug ports were children
175  of the NHI, this resume order would automatically be enforced by the
176  PM core, but unfortunately they're aunts.  The solution is to add
177  device links from the hotplug ports (consumers) to the NHI device
178  (supplier).  A driver presence dependency is not necessary for this
179  use case.
180
181* Discrete GPUs in hybrid graphics laptops often feature an HDA controller
182  for HDMI/DP audio.  In the device hierarchy the HDA controller is a sibling
183  of the VGA device, yet both share the same power domain and the HDA
184  controller is only ever needed when an HDMI/DP display is attached to the
185  VGA device.  A device link from the HDA controller (consumer) to the
186  VGA device (supplier) aptly represents this relationship.
187
188* ACPI allows definition of a device start order by way of _DEP objects.
189  A classical example is when ACPI power management methods on one device
190  are implemented in terms of I\ :sup:`2`\ C accesses and require a specific
191  I\ :sup:`2`\ C controller to be present and functional for the power
192  management of the device in question to work.
193
194* In some SoCs a functional dependency exists from display, video codec and
195  video processing IP cores on transparent memory access IP cores that handle
196  burst access and compression/decompression.
197
198Alternatives
199============
200
201* A struct dev_pm_domain can be used to override the bus,
202  class or device type callbacks.  It is intended for devices sharing
203  a single on/off switch, however it does not guarantee a specific
204  suspend/resume ordering, this needs to be implemented separately.
205  It also does not by itself track the runtime PM status of the involved
206  devices and turn off the power switch only when all of them are runtime
207  suspended.  Furthermore it cannot be used to enforce a specific shutdown
208  ordering or a driver presence dependency.
209
210* A struct generic_pm_domain is a lot more heavyweight than a
211  device link and does not allow for shutdown ordering or driver presence
212  dependencies.  It also cannot be used on ACPI systems.
213
214Implementation
215==============
216
217The device hierarchy, which -- as the name implies -- is a tree,
218becomes a directed acyclic graph once device links are added.
219
220Ordering of these devices during suspend/resume is determined by the
221dpm_list.  During shutdown it is determined by the devices_kset.  With
222no device links present, the two lists are a flattened, one-dimensional
223representations of the device tree such that a device is placed behind
224all its ancestors.  That is achieved by traversing the ACPI namespace
225or OpenFirmware device tree top-down and appending devices to the lists
226as they are discovered.
227
228Once device links are added, the lists need to satisfy the additional
229constraint that a device is placed behind all its suppliers, recursively.
230To ensure this, upon addition of the device link the consumer and the
231entire sub-graph below it (all children and consumers of the consumer)
232are moved to the end of the list.  (Call to :c:func:`device_reorder_to_tail()`
233from :c:func:`device_link_add()`.)
234
235To prevent introduction of dependency loops into the graph, it is
236verified upon device link addition that the supplier is not dependent
237on the consumer or any children or consumers of the consumer.
238(Call to :c:func:`device_is_dependent()` from :c:func:`device_link_add()`.)
239If that constraint is violated, :c:func:`device_link_add()` will return
240``NULL`` and a ``WARNING`` will be logged.
241
242Notably this also prevents the addition of a device link from a parent
243device to a child.  However the converse is allowed, i.e. a device link
244from a child to a parent.  Since the driver core already guarantees
245correct suspend/resume and shutdown ordering between parent and child,
246such a device link only makes sense if a driver presence dependency is
247needed on top of that.  In this case driver authors should weigh
248carefully if a device link is at all the right tool for the purpose.
249A more suitable approach might be to simply use deferred probing or
250add a device flag causing the parent driver to be probed before the
251child one.
252
253State machine
254=============
255
256.. kernel-doc:: include/linux/device.h
257   :functions: device_link_state
258
259::
260
261                 .=============================.
262                 |                             |
263                 v                             |
264 DORMANT <=> AVAILABLE <=> CONSUMER_PROBE => ACTIVE
265    ^                                          |
266    |                                          |
267    '============ SUPPLIER_UNBIND <============'
268
269* The initial state of a device link is automatically determined by
270  :c:func:`device_link_add()` based on the driver presence on the supplier
271  and consumer.  If the link is created before any devices are probed, it
272  is set to ``DL_STATE_DORMANT``.
273
274* When a supplier device is bound to a driver, links to its consumers
275  progress to ``DL_STATE_AVAILABLE``.
276  (Call to :c:func:`device_links_driver_bound()` from
277  :c:func:`driver_bound()`.)
278
279* Before a consumer device is probed, presence of supplier drivers is
280  verified by checking the consumer device is not in the wait_for_suppliers
281  list and by checking that links to suppliers are in ``DL_STATE_AVAILABLE``
282  state.  The state of the links is updated to ``DL_STATE_CONSUMER_PROBE``.
283  (Call to :c:func:`device_links_check_suppliers()` from
284  :c:func:`really_probe()`.)
285  This prevents the supplier from unbinding.
286  (Call to :c:func:`wait_for_device_probe()` from
287  :c:func:`device_links_unbind_consumers()`.)
288
289* If the probe fails, links to suppliers revert back to ``DL_STATE_AVAILABLE``.
290  (Call to :c:func:`device_links_no_driver()` from :c:func:`really_probe()`.)
291
292* If the probe succeeds, links to suppliers progress to ``DL_STATE_ACTIVE``.
293  (Call to :c:func:`device_links_driver_bound()` from :c:func:`driver_bound()`.)
294
295* When the consumer's driver is later on removed, links to suppliers revert
296  back to ``DL_STATE_AVAILABLE``.
297  (Call to :c:func:`__device_links_no_driver()` from
298  :c:func:`device_links_driver_cleanup()`, which in turn is called from
299  :c:func:`__device_release_driver()`.)
300
301* Before a supplier's driver is removed, links to consumers that are not
302  bound to a driver are updated to ``DL_STATE_SUPPLIER_UNBIND``.
303  (Call to :c:func:`device_links_busy()` from
304  :c:func:`__device_release_driver()`.)
305  This prevents the consumers from binding.
306  (Call to :c:func:`device_links_check_suppliers()` from
307  :c:func:`really_probe()`.)
308  Consumers that are bound are freed from their driver; consumers that are
309  probing are waited for until they are done.
310  (Call to :c:func:`device_links_unbind_consumers()` from
311  :c:func:`__device_release_driver()`.)
312  Once all links to consumers are in ``DL_STATE_SUPPLIER_UNBIND`` state,
313  the supplier driver is released and the links revert to ``DL_STATE_DORMANT``.
314  (Call to :c:func:`device_links_driver_cleanup()` from
315  :c:func:`__device_release_driver()`.)
316
317API
318===
319
320See device_link_add(), device_link_del() and device_link_remove().
321