1The padata parallel execution mechanism
   2Last updated for 2.6.36
   4Padata is a mechanism by which the kernel can farm work out to be done in
   5parallel on multiple CPUs while retaining the ordering of tasks.  It was
   6developed for use with the IPsec code, which needs to be able to perform
   7encryption and decryption on large numbers of packets without reordering
   8those packets.  The crypto developers made a point of writing padata in a
   9sufficiently general fashion that it could be put to other uses as well.
  11The first step in using padata is to set up a padata_instance structure for
  12overall control of how tasks are to be run:
  14    #include <linux/padata.h>
  16    struct padata_instance *padata_alloc(struct workqueue_struct *wq,
  17                                         const struct cpumask *pcpumask,
  18                                         const struct cpumask *cbcpumask);
  20The pcpumask describes which processors will be used to execute work
  21submitted to this instance in parallel. The cbcpumask defines which
  22processors are allowed to be used as the serialization callback processor.
  23The workqueue wq is where the work will actually be done; it should be
  24a multithreaded queue, naturally.
  26To allocate a padata instance with the cpu_possible_mask for both
  27cpumasks this helper function can be used:
  29    struct padata_instance *padata_alloc_possible(struct workqueue_struct *wq);
  31Note: Padata maintains two kinds of cpumasks internally. The user supplied
  32cpumasks, submitted by padata_alloc/padata_alloc_possible and the 'usable'
  33cpumasks. The usable cpumasks are always a subset of active CPUs in the
  34user supplied cpumasks; these are the cpumasks padata actually uses. So
  35it is legal to supply a cpumask to padata that contains offline CPUs.
  36Once an offline CPU in the user supplied cpumask comes online, padata
  37is going to use it.
  39There are functions for enabling and disabling the instance:
  41    int padata_start(struct padata_instance *pinst);
  42    void padata_stop(struct padata_instance *pinst);
  44These functions are setting or clearing the "PADATA_INIT" flag;
  45if that flag is not set, other functions will refuse to work.
  46padata_start returns zero on success (flag set) or -EINVAL if the
  47padata cpumask contains no active CPU (flag not set).
  48padata_stop clears the flag and blocks until the padata instance
  49is unused.
  51The list of CPUs to be used can be adjusted with these functions:
  53    int padata_set_cpumasks(struct padata_instance *pinst,
  54                            cpumask_var_t pcpumask,
  55                            cpumask_var_t cbcpumask);
  56    int padata_set_cpumask(struct padata_instance *pinst, int cpumask_type,
  57                           cpumask_var_t cpumask);
  58    int padata_add_cpu(struct padata_instance *pinst, int cpu, int mask);
  59    int padata_remove_cpu(struct padata_instance *pinst, int cpu, int mask);
  61Changing the CPU masks are expensive operations, though, so it should not be
  62done with great frequency.
  64It's possible to change both cpumasks of a padata instance with
  65padata_set_cpumasks by specifying the cpumasks for parallel execution (pcpumask)
  66and for the serial callback function (cbcpumask). padata_set_cpumask is used to
  67change just one of the cpumasks. Here cpumask_type is one of PADATA_CPU_SERIAL,
  68PADATA_CPU_PARALLEL and cpumask specifies the new cpumask to use.
  69To simply add or remove one CPU from a certain cpumask the functions
  70padata_add_cpu/padata_remove_cpu are used. cpu specifies the CPU to add or
  71remove and mask is one of PADATA_CPU_SERIAL, PADATA_CPU_PARALLEL.
  73If a user is interested in padata cpumask changes, he can register to
  74the padata cpumask change notifier:
  76    int padata_register_cpumask_notifier(struct padata_instance *pinst,
  77                                         struct notifier_block *nblock);
  79To unregister from that notifier:
  81    int padata_unregister_cpumask_notifier(struct padata_instance *pinst,
  82                                           struct notifier_block *nblock);
  84The padata cpumask change notifier notifies about changes of the usable
  85cpumasks, i.e. the subset of active CPUs in the user supplied cpumask.
  87Padata calls the notifier chain with:
  89    blocking_notifier_call_chain(&pinst->cpumask_change_notifier,
  90                                 notification_mask,
  91                                 &pd_new->cpumask);
  93Here cpumask_change_notifier is registered notifier, notification_mask
  94is one of PADATA_CPU_SERIAL, PADATA_CPU_PARALLEL and cpumask is a pointer
  95to a struct padata_cpumask that contains the new cpumask information.
  97Actually submitting work to the padata instance requires the creation of a
  98padata_priv structure:
 100    struct padata_priv {
 101        /* Other stuff here... */
 102        void                    (*parallel)(struct padata_priv *padata);
 103        void                    (*serial)(struct padata_priv *padata);
 104    };
 106This structure will almost certainly be embedded within some larger
 107structure specific to the work to be done.  Most of its fields are private to
 108padata, but the structure should be zeroed at initialisation time, and the
 109parallel() and serial() functions should be provided.  Those functions will
 110be called in the process of getting the work done as we will see
 113The submission of work is done with:
 115    int padata_do_parallel(struct padata_instance *pinst,
 116                           struct padata_priv *padata, int cb_cpu);
 118The pinst and padata structures must be set up as described above; cb_cpu
 119specifies which CPU will be used for the final callback when the work is
 120done; it must be in the current instance's CPU mask.  The return value from
 121padata_do_parallel() is zero on success, indicating that the work is in
 122progress. -EBUSY means that somebody, somewhere else is messing with the
 123instance's CPU mask, while -EINVAL is a complaint about cb_cpu not being
 124in that CPU mask or about a not running instance.
 126Each task submitted to padata_do_parallel() will, in turn, be passed to
 127exactly one call to the above-mentioned parallel() function, on one CPU, so
 128true parallelism is achieved by submitting multiple tasks.  Despite the
 129fact that the workqueue is used to make these calls, parallel() is run with
 130software interrupts disabled and thus cannot sleep.  The parallel()
 131function gets the padata_priv structure pointer as its lone parameter;
 132information about the actual work to be done is probably obtained by using
 133container_of() to find the enclosing structure.
 135Note that parallel() has no return value; the padata subsystem assumes that
 136parallel() will take responsibility for the task from this point.  The work
 137need not be completed during this call, but, if parallel() leaves work
 138outstanding, it should be prepared to be called again with a new job before
 139the previous one completes.  When a task does complete, parallel() (or
 140whatever function actually finishes the job) should inform padata of the
 141fact with a call to:
 143    void padata_do_serial(struct padata_priv *padata);
 145At some point in the future, padata_do_serial() will trigger a call to the
 146serial() function in the padata_priv structure.  That call will happen on
 147the CPU requested in the initial call to padata_do_parallel(); it, too, is
 148done through the workqueue, but with local software interrupts disabled.
 149Note that this call may be deferred for a while since the padata code takes
 150pains to ensure that tasks are completed in the order in which they were
 153The one remaining function in the padata API should be called to clean up
 154when a padata instance is no longer needed:
 156    void padata_free(struct padata_instance *pinst);
 158This function will busy-wait while any remaining tasks are completed, so it
 159might be best not to call it while there is work outstanding.  Shutting
 160down the workqueue, if necessary, should be done separately.