[PATCH 0/1] (Revision3) Oprofile Extented Feature Interface

[Suravee S. <sur...@am...>]

Revision 3 Note:
- Design changes in based on Maynard's recommendation from revision2.
    * Replacing "ext_features" to "ext_feature"
    * Changing the usage message for "--ext-feature" option.
- Adding the function "is_ext_sfile_enabled()" function in daemon/opd_extended.c to enhance the sanity check in each opd_ext_sfile_XXX wrapper. 

----------------------------------

Design Summary (Revision3):

    *  New oprofiled command line interface
          -x|--ext-feature=<extended-feature-name>:[args] This option enable a specific extended feature name, and arguments.

    * "opd_ext_feature" structure contains information for each extended feature including feature name, and a structure containing handlers function.

        struct opd_ext_feature {
                const char* feature;               // Feature name
                struct opd_ext_handlers *handlers; // Feature handlers
        };

    * "opd_ext_handlers" structure contains handlers for each extended feature.
          o ext_init: handle the initialization of the feature at the beginning when oprofiled parsing the command line options (oprofiled.c).
          o ext_print_stats: will print out the statistics summary in the oprofiled.log file at the end of the run (daemon/opd_stat.c).
          o ext_sfle: a pointer to structure containing handlers for the operations necessary to extend opd_sfile (daemon/opd_sfile.c).

        struct opd_ext_handlers {
                int (*ext_init)(char const * );           // Extended init
                int (*ext_print_stats)();                 // Extended statistics
                struct opd_ext_sfile_handlers *ext_sfile; // Extended sfile handlers (optional)
        };

        struct opd_ext_sfile_handlers {
                void (*create)(struct sfile *, struct transient const *);
                void (*dup)(struct sfile *, struct sfile * );
                void (*close)(struct sfile *);
                void (*sync)(struct sfile *);
                odb_t * (*get)(struct transient const * , int );
                struct opd_event * (*find_counter_event)(unsigned long);
        };

    * Extended Feature Table: This is a global table of "struct opd_ext_feature" described above. Each extended feature must be listed here.

        struct opd_ext_feature ext_feature_table[] = {
                {"feature1", &feature1_handlers},
                {"feature2", &feature2_handlers},
                {"feature3", &feature3_handlers},
                {....},
                {NULL, NULL}
        };

    * Extend the "struct sfile" in opd_sfile.h by adding "odb_t * ext_file".

    * Typically, opcontrol uses ophelp to determine how to program a set of performance counters with the specified events. To handle the new extended events:
          1. A new tag "ext" will be used by extended events, which does not use the actual performance counter.

             - Traditional events
               event:0x76 counters:0,1,2,3 um:zero minimum:5000 name:CPU_CLK_UNHALTED : Cpu unhalted clock cycle */
               ophelp --check-events=CPU_CLK_UNHALTED:100000 ---> 0

             - Extended events
               event:0xf00 ext:ibs_fetch um:zero minimum:100000 name:IBS_FETCH_ALL : All IBS fetch samples */
               ophelp --check-events=IBS_FETCH_ALL:100000 ---> ibs_fetch

          2. Extend "ophelp --check-events" to support extended events. The output will be the appropriate counters assignment based on the provided events.  For instance,

             ophelp --check-events CPU_CLK_UNHALTED:100000 \
                                   IBS_FETCH_ALL:100000 \  --> (extended event)
                                   IBS_OP_ALL:100000 \     --> (extended event)
                                   RETIRED_INSTRUCTIONS:100000 \
                                   DATA_CACHE_MISSES:100000 \
                                   DATA_CACHE_ACCESSES:100000

             will generate output "0 ibs_fetch ibs_op 1 2 3".

          3. Modify the function "map_event_to_counter" so that the entries of "counter_map" will contain "-1" for the extended events.  For instance, the above input will generate counter_map with contents "0 -1 -1 1 2 3". "ophelp" will then map the entry with counter_map value "-1" to the appropriate value, which is specified with tag "ext" in the events file.