1 =======================================
2 The padata parallel execution mechanism
3 =======================================
5 :Last updated: for 2.6.36
7 Padata is a mechanism by which the kernel can farm work out to be done in
8 parallel on multiple CPUs while retaining the ordering of tasks. It was
9 developed for use with the IPsec code, which needs to be able to perform
10 encryption and decryption on large numbers of packets without reordering
11 those packets. The crypto developers made a point of writing padata in a
12 sufficiently general fashion that it could be put to other uses as well.
14 The first step in using padata is to set up a padata_instance structure for
15 overall control of how tasks are to be run::
17 #include <linux/padata.h>
19 struct padata_instance *padata_alloc(struct workqueue_struct *wq,
20 const struct cpumask *pcpumask,
21 const struct cpumask *cbcpumask);
23 The pcpumask describes which processors will be used to execute work
24 submitted to this instance in parallel. The cbcpumask defines which
25 processors are allowed to be used as the serialization callback processor.
26 The workqueue wq is where the work will actually be done; it should be
27 a multithreaded queue, naturally.
29 To allocate a padata instance with the cpu_possible_mask for both
30 cpumasks this helper function can be used::
32 struct padata_instance *padata_alloc_possible(struct workqueue_struct *wq);
34 Note: Padata maintains two kinds of cpumasks internally. The user supplied
35 cpumasks, submitted by padata_alloc/padata_alloc_possible and the 'usable'
36 cpumasks. The usable cpumasks are always a subset of active CPUs in the
37 user supplied cpumasks; these are the cpumasks padata actually uses. So
38 it is legal to supply a cpumask to padata that contains offline CPUs.
39 Once an offline CPU in the user supplied cpumask comes online, padata
42 There are functions for enabling and disabling the instance::
44 int padata_start(struct padata_instance *pinst);
45 void padata_stop(struct padata_instance *pinst);
47 These functions are setting or clearing the "PADATA_INIT" flag;
48 if that flag is not set, other functions will refuse to work.
49 padata_start returns zero on success (flag set) or -EINVAL if the
50 padata cpumask contains no active CPU (flag not set).
51 padata_stop clears the flag and blocks until the padata instance
54 The list of CPUs to be used can be adjusted with these functions::
56 int padata_set_cpumasks(struct padata_instance *pinst,
57 cpumask_var_t pcpumask,
58 cpumask_var_t cbcpumask);
59 int padata_set_cpumask(struct padata_instance *pinst, int cpumask_type,
60 cpumask_var_t cpumask);
61 int padata_add_cpu(struct padata_instance *pinst, int cpu, int mask);
62 int padata_remove_cpu(struct padata_instance *pinst, int cpu, int mask);
64 Changing the CPU masks are expensive operations, though, so it should not be
65 done with great frequency.
67 It's possible to change both cpumasks of a padata instance with
68 padata_set_cpumasks by specifying the cpumasks for parallel execution (pcpumask)
69 and for the serial callback function (cbcpumask). padata_set_cpumask is used to
70 change just one of the cpumasks. Here cpumask_type is one of PADATA_CPU_SERIAL,
71 PADATA_CPU_PARALLEL and cpumask specifies the new cpumask to use.
72 To simply add or remove one CPU from a certain cpumask the functions
73 padata_add_cpu/padata_remove_cpu are used. cpu specifies the CPU to add or
74 remove and mask is one of PADATA_CPU_SERIAL, PADATA_CPU_PARALLEL.
76 If a user is interested in padata cpumask changes, he can register to
77 the padata cpumask change notifier::
79 int padata_register_cpumask_notifier(struct padata_instance *pinst,
80 struct notifier_block *nblock);
82 To unregister from that notifier::
84 int padata_unregister_cpumask_notifier(struct padata_instance *pinst,
85 struct notifier_block *nblock);
87 The padata cpumask change notifier notifies about changes of the usable
88 cpumasks, i.e. the subset of active CPUs in the user supplied cpumask.
90 Padata calls the notifier chain with::
92 blocking_notifier_call_chain(&pinst->cpumask_change_notifier,
96 Here cpumask_change_notifier is registered notifier, notification_mask
97 is one of PADATA_CPU_SERIAL, PADATA_CPU_PARALLEL and cpumask is a pointer
98 to a struct padata_cpumask that contains the new cpumask information.
100 Actually submitting work to the padata instance requires the creation of a
101 padata_priv structure::
104 /* Other stuff here... */
105 void (*parallel)(struct padata_priv *padata);
106 void (*serial)(struct padata_priv *padata);
109 This structure will almost certainly be embedded within some larger
110 structure specific to the work to be done. Most of its fields are private to
111 padata, but the structure should be zeroed at initialisation time, and the
112 parallel() and serial() functions should be provided. Those functions will
113 be called in the process of getting the work done as we will see
116 The submission of work is done with::
118 int padata_do_parallel(struct padata_instance *pinst,
119 struct padata_priv *padata, int cb_cpu);
121 The pinst and padata structures must be set up as described above; cb_cpu
122 specifies which CPU will be used for the final callback when the work is
123 done; it must be in the current instance's CPU mask. The return value from
124 padata_do_parallel() is zero on success, indicating that the work is in
125 progress. -EBUSY means that somebody, somewhere else is messing with the
126 instance's CPU mask, while -EINVAL is a complaint about cb_cpu not being
127 in that CPU mask or about a not running instance.
129 Each task submitted to padata_do_parallel() will, in turn, be passed to
130 exactly one call to the above-mentioned parallel() function, on one CPU, so
131 true parallelism is achieved by submitting multiple tasks. Despite the
132 fact that the workqueue is used to make these calls, parallel() is run with
133 software interrupts disabled and thus cannot sleep. The parallel()
134 function gets the padata_priv structure pointer as its lone parameter;
135 information about the actual work to be done is probably obtained by using
136 container_of() to find the enclosing structure.
138 Note that parallel() has no return value; the padata subsystem assumes that
139 parallel() will take responsibility for the task from this point. The work
140 need not be completed during this call, but, if parallel() leaves work
141 outstanding, it should be prepared to be called again with a new job before
142 the previous one completes. When a task does complete, parallel() (or
143 whatever function actually finishes the job) should inform padata of the
144 fact with a call to::
146 void padata_do_serial(struct padata_priv *padata);
148 At some point in the future, padata_do_serial() will trigger a call to the
149 serial() function in the padata_priv structure. That call will happen on
150 the CPU requested in the initial call to padata_do_parallel(); it, too, is
151 done through the workqueue, but with local software interrupts disabled.
152 Note that this call may be deferred for a while since the padata code takes
153 pains to ensure that tasks are completed in the order in which they were
156 The one remaining function in the padata API should be called to clean up
157 when a padata instance is no longer needed::
159 void padata_free(struct padata_instance *pinst);
161 This function will busy-wait while any remaining tasks are completed, so it
162 might be best not to call it while there is work outstanding. Shutting
163 down the workqueue, if necessary, should be done separately.