1.. SPDX-License-Identifier: GPL-2.0
2
3=======================================
4The padata parallel execution mechanism
5=======================================
6
7:Date: May 2020
8
9Padata is a mechanism by which the kernel can farm jobs out to be done in
10parallel on multiple CPUs while optionally retaining their ordering.
11
12It was originally developed for IPsec, which needs to perform encryption and
13decryption on large numbers of packets without reordering those packets.  This
14is currently the sole consumer of padata's serialized job support.
15
16Padata also supports multithreaded jobs, splitting up the job evenly while load
17balancing and coordinating between threads.
18
19Running Serialized Jobs
20=======================
21
22Initializing
23------------
24
25The first step in using padata to run serialized jobs is to set up a
26padata_instance structure for overall control of how jobs are to be run::
27
28    #include <linux/padata.h>
29
30    struct padata_instance *padata_alloc_possible(const char *name);
31
32'name' simply identifies the instance.
33
34There are functions for enabling and disabling the instance::
35
36    int padata_start(struct padata_instance *pinst);
37    void padata_stop(struct padata_instance *pinst);
38
39These functions are setting or clearing the "PADATA_INIT" flag; if that flag is
40not set, other functions will refuse to work.  padata_start() returns zero on
41success (flag set) or -EINVAL if the padata cpumask contains no active CPU
42(flag not set).  padata_stop() clears the flag and blocks until the padata
43instance is unused.
44
45Finally, complete padata initialization by allocating a padata_shell::
46
47   struct padata_shell *padata_alloc_shell(struct padata_instance *pinst);
48
49A padata_shell is used to submit a job to padata and allows a series of such
50jobs to be serialized independently.  A padata_instance may have one or more
51padata_shells associated with it, each allowing a separate series of jobs.
52
53Modifying cpumasks
54------------------
55
56The CPUs used to run jobs can be changed in two ways, programatically with
57padata_set_cpumask() or via sysfs.  The former is defined::
58
59    int padata_set_cpumask(struct padata_instance *pinst, int cpumask_type,
60			   cpumask_var_t cpumask);
61
62Here cpumask_type is one of PADATA_CPU_PARALLEL or PADATA_CPU_SERIAL, where a
63parallel cpumask describes which processors will be used to execute jobs
64submitted to this instance in parallel and a serial cpumask defines which
65processors are allowed to be used as the serialization callback processor.
66cpumask specifies the new cpumask to use.
67
68There may be sysfs files for an instance's cpumasks.  For example, pcrypt's
69live in /sys/kernel/pcrypt/<instance-name>.  Within an instance's directory
70there are two files, parallel_cpumask and serial_cpumask, and either cpumask
71may be changed by echoing a bitmask into the file, for example::
72
73    echo f > /sys/kernel/pcrypt/pencrypt/parallel_cpumask
74
75Reading one of these files shows the user-supplied cpumask, which may be
76different from the 'usable' cpumask.
77
78Padata maintains two pairs of cpumasks internally, the user-supplied cpumasks
79and the 'usable' cpumasks.  (Each pair consists of a parallel and a serial
80cpumask.)  The user-supplied cpumasks default to all possible CPUs on instance
81allocation and may be changed as above.  The usable cpumasks are always a
82subset of the user-supplied cpumasks and contain only the online CPUs in the
83user-supplied masks; these are the cpumasks padata actually uses.  So it is
84legal to supply a cpumask to padata that contains offline CPUs.  Once an
85offline CPU in the user-supplied cpumask comes online, padata is going to use
86it.
87
88Changing the CPU masks are expensive operations, so it should not be done with
89great frequency.
90
91Running A Job
92-------------
93
94Actually submitting work to the padata instance requires the creation of a
95padata_priv structure, which represents one job::
96
97    struct padata_priv {
98        /* Other stuff here... */
99	void                    (*parallel)(struct padata_priv *padata);
100	void                    (*serial)(struct padata_priv *padata);
101    };
102
103This structure will almost certainly be embedded within some larger
104structure specific to the work to be done.  Most of its fields are private to
105padata, but the structure should be zeroed at initialisation time, and the
106parallel() and serial() functions should be provided.  Those functions will
107be called in the process of getting the work done as we will see
108momentarily.
109
110The submission of the job is done with::
111
112    int padata_do_parallel(struct padata_shell *ps,
113		           struct padata_priv *padata, int *cb_cpu);
114
115The ps and padata structures must be set up as described above; cb_cpu
116points to the preferred CPU to be used for the final callback when the job is
117done; it must be in the current instance's CPU mask (if not the cb_cpu pointer
118is updated to point to the CPU actually chosen).  The return value from
119padata_do_parallel() is zero on success, indicating that the job is in
120progress. -EBUSY means that somebody, somewhere else is messing with the
121instance's CPU mask, while -EINVAL is a complaint about cb_cpu not being in the
122serial cpumask, no online CPUs in the parallel or serial cpumasks, or a stopped
123instance.
124
125Each job submitted to padata_do_parallel() will, in turn, be passed to
126exactly one call to the above-mentioned parallel() function, on one CPU, so
127true parallelism is achieved by submitting multiple jobs.  parallel() runs with
128software interrupts disabled and thus cannot sleep.  The parallel()
129function gets the padata_priv structure pointer as its lone parameter;
130information about the actual work to be done is probably obtained by using
131container_of() to find the enclosing structure.
132
133Note that parallel() has no return value; the padata subsystem assumes that
134parallel() will take responsibility for the job from this point.  The job
135need not be completed during this call, but, if parallel() leaves work
136outstanding, it should be prepared to be called again with a new job before
137the previous one completes.
138
139Serializing Jobs
140----------------
141
142When a job does complete, parallel() (or whatever function actually finishes
143the work) should inform padata of the fact with a call to::
144
145    void padata_do_serial(struct padata_priv *padata);
146
147At some point in the future, padata_do_serial() will trigger a call to the
148serial() function in the padata_priv structure.  That call will happen on
149the CPU requested in the initial call to padata_do_parallel(); it, too, is
150run with local software interrupts disabled.
151Note that this call may be deferred for a while since the padata code takes
152pains to ensure that jobs are completed in the order in which they were
153submitted.
154
155Destroying
156----------
157
158Cleaning up a padata instance predictably involves calling the three free
159functions that correspond to the allocation in reverse::
160
161    void padata_free_shell(struct padata_shell *ps);
162    void padata_stop(struct padata_instance *pinst);
163    void padata_free(struct padata_instance *pinst);
164
165It is the user's responsibility to ensure all outstanding jobs are complete
166before any of the above are called.
167
168Running Multithreaded Jobs
169==========================
170
171A multithreaded job has a main thread and zero or more helper threads, with the
172main thread participating in the job and then waiting until all helpers have
173finished.  padata splits the job into units called chunks, where a chunk is a
174piece of the job that one thread completes in one call to the thread function.
175
176A user has to do three things to run a multithreaded job.  First, describe the
177job by defining a padata_mt_job structure, which is explained in the Interface
178section.  This includes a pointer to the thread function, which padata will
179call each time it assigns a job chunk to a thread.  Then, define the thread
180function, which accepts three arguments, ``start``, ``end``, and ``arg``, where
181the first two delimit the range that the thread operates on and the last is a
182pointer to the job's shared state, if any.  Prepare the shared state, which is
183typically allocated on the main thread's stack.  Last, call
184padata_do_multithreaded(), which will return once the job is finished.
185
186Interface
187=========
188
189.. kernel-doc:: include/linux/padata.h
190.. kernel-doc:: kernel/padata.c
191