1INFINIBAND MIDLAYER LOCKING
2
3  This guide is an attempt to make explicit the locking assumptions
4  made by the InfiniBand midlayer.  It describes the requirements on
5  both low-level drivers that sit below the midlayer and upper level
6  protocols that use the midlayer.
7
8Sleeping and interrupt context
9
10  With the following exceptions, a low-level driver implementation of
11  all of the methods in struct ib_device may sleep.  The exceptions
12  are any methods from the list:
13
14    create_ah
15    modify_ah
16    query_ah
17    destroy_ah
18    bind_mw
19    post_send
20    post_recv
21    poll_cq
22    req_notify_cq
23    map_phys_fmr
24
25  which may not sleep and must be callable from any context.
26
27  The corresponding functions exported to upper level protocol
28  consumers:
29
30    ib_create_ah
31    ib_modify_ah
32    ib_query_ah
33    ib_destroy_ah
34    ib_bind_mw
35    ib_post_send
36    ib_post_recv
37    ib_req_notify_cq
38    ib_map_phys_fmr
39
40  are therefore safe to call from any context.
41
42  In addition, the function
43
44    ib_dispatch_event
45
46  used by low-level drivers to dispatch asynchronous events through
47  the midlayer is also safe to call from any context.
48
49Reentrancy
50
51  All of the methods in struct ib_device exported by a low-level
52  driver must be fully reentrant.  The low-level driver is required to
53  perform all synchronization necessary to maintain consistency, even
54  if multiple function calls using the same object are run
55  simultaneously.
56
57  The IB midlayer does not perform any serialization of function calls.
58
59  Because low-level drivers are reentrant, upper level protocol
60  consumers are not required to perform any serialization.  However,
61  some serialization may be required to get sensible results.  For
62  example, a consumer may safely call ib_poll_cq() on multiple CPUs
63  simultaneously.  However, the ordering of the work completion
64  information between different calls of ib_poll_cq() is not defined.
65
66Callbacks
67
68  A low-level driver must not perform a callback directly from the
69  same callchain as an ib_device method call.  For example, it is not
70  allowed for a low-level driver to call a consumer's completion event
71  handler directly from its post_send method.  Instead, the low-level
72  driver should defer this callback by, for example, scheduling a
73  tasklet to perform the callback.
74
75  The low-level driver is responsible for ensuring that multiple
76  completion event handlers for the same CQ are not called
77  simultaneously.  The driver must guarantee that only one CQ event
78  handler for a given CQ is running at a time.  In other words, the
79  following situation is not allowed:
80
81        CPU1                                    CPU2
82
83  low-level driver ->
84    consumer CQ event callback:
85      /* ... */
86      ib_req_notify_cq(cq, ...);
87                                        low-level driver ->
88      /* ... */                           consumer CQ event callback:
89                                            /* ... */
90      return from CQ event handler
91
92  The context in which completion event and asynchronous event
93  callbacks run is not defined.  Depending on the low-level driver, it
94  may be process context, softirq context, or interrupt context.
95  Upper level protocol consumers may not sleep in a callback.
96
97Hot-plug
98
99  A low-level driver announces that a device is ready for use by
100  consumers when it calls ib_register_device(), all initialization
101  must be complete before this call.  The device must remain usable
102  until the driver's call to ib_unregister_device() has returned.
103
104  A low-level driver must call ib_register_device() and
105  ib_unregister_device() from process context.  It must not hold any
106  semaphores that could cause deadlock if a consumer calls back into
107  the driver across these calls.
108
109  An upper level protocol consumer may begin using an IB device as
110  soon as the add method of its struct ib_client is called for that
111  device.  A consumer must finish all cleanup and free all resources
112  relating to a device before returning from the remove method.
113
114  A consumer is permitted to sleep in its add and remove methods.
115