/** * @file op_alloc_counter.c * hardware counter allocation * * You can have silliness here. * * @remark Copyright 2002 OProfile authors * @remark Read the file COPYING * * @author John Levon * @author Philippe Elie */ #include #include #include #include "op_events.h" #include "op_libiberty.h" typedef struct counter_arc_head { /** the head of allowed counter for this event */ struct list_head next; } counter_arc_head; typedef struct counter_arc { /** counter nr */ int counter; /** the next counter allowed for this event */ struct list_head next; } counter_arc; /** * @param pev an array of event * @param nr_events number of entry in pev * * build an array of counter list allowed for each events * counter_arc_head[i] is the list of allowed counter for pev[i] events * The returned pointer is an array of nr_events entry */ static counter_arc_head * build_counter_arc(struct op_event const * pev[], int nr_events) { counter_arc_head * ctr_arc; int i; ctr_arc = xmalloc(nr_events * sizeof(*ctr_arc)); for (i = 0; i < nr_events; ++i) { int j; u32 mask = pev[i]->counter_mask; list_init(&ctr_arc[i].next); for (j = 0; mask; ++j) { if (mask & (1 << j)) { counter_arc * arc = xmalloc(sizeof(counter_arc)); arc->counter = j; /* we are looping by increasing counter number, * allocation use a left to right tree walking * so we add at end to ensure counter will * be allocated by increasing number: it's not * required but a bit less surprising when * debugging code */ list_add_tail(&arc->next, &ctr_arc[i].next); mask &= ~(1 << j); } } } return ctr_arc; } /** * @param ctr_arc the array to deallocate * @param nr_events number of entry in array * * deallocate all previously allocated resource by build_counter_arc() */ static void delete_counter_arc(counter_arc_head * ctr_arc, int nr_events) { int i; for (i = 0; i < nr_events; ++i) { struct list_head * pos, * pos2; list_for_each_safe(pos, pos2, &ctr_arc[i].next) { counter_arc * arc = list_entry(pos, counter_arc, next); list_del(&arc->next); free(arc); } } free(ctr_arc); } /** * @param ctr_arc tree description, ctr_arc[i] is the i-th level of tree. * @param max_depth number of entry in array ctr_arc == depth of tree * @param depth current level we are exploring * @param allocated_mask current counter already allocated mask * @param counter_map array of counter number mapping, returned results go * here * * return non zero on succees, in this case counter_map is set to the counter * mapping number. * * Solution is searched through a simple backtracking exploring recursively all * possible solution until one is found, prunning is done in O(1) by tracking * a bitmask of already allocated counter. Walking through node is done in * preorder left to right. * * In case of extended events (required no phisical counters), the associated * counter_map entry will be -1. * * Possible improvment if neccessary: partition counters in class of counter, * two counter belong to the same class if they allow exactly the same set of * event. Now using a variant of the backtrack algo can works on class of * counter rather on counter (this is not an improvment if each counter goes * in it's own class) */ static int allocate_counter(counter_arc_head const * ctr_arc, int max_depth, int depth, u32 allocated_mask, size_t * counter_map) { struct list_head * pos; if (depth == max_depth) return 1; /* If ctr_arc is not available, counter_map is -1 */ if((&ctr_arc[depth].next)->next == &ctr_arc[depth].next) { counter_map[depth] = -1; if (allocate_counter(ctr_arc, max_depth, depth + 1, allocated_mask, counter_map)) return 1; } else { list_for_each(pos, &ctr_arc[depth].next) { counter_arc const * arc = list_entry(pos, counter_arc, next); if (allocated_mask & (1 << arc->counter)) continue; counter_map[depth] = arc->counter; if (allocate_counter(ctr_arc, max_depth, depth + 1, allocated_mask | (1 << arc->counter), counter_map)) return 1; } } return 0; } /* determine which directories are counter directories */ static int perfcounterdir(const struct dirent * entry) { return (isdigit(entry->d_name[0])); } /** * @param mask pointer where to place bit mask of unavailable counters * * return >= 0 number of counters that are available * < 0 could not determine number of counters * */ static int op_get_counter_mask(u32 * mask) { struct dirent **counterlist; int count, i; /* assume nothing is available */ u32 available=0; count = scandir("/dev/oprofile", &counterlist, perfcounterdir, alphasort); if (count < 0) /* unable to determine bit mask */ return -1; /* convert to bit map (0 where counter exists) */ for (i=0; id_name); free(counterlist[i]); } *mask=~available; free(counterlist); return count; } size_t * map_event_to_counter(struct op_event const * pev[], int nr_events, op_cpu cpu_type) { counter_arc_head * ctr_arc; size_t * counter_map; int i, nr_counters, nr_pmc_events; op_cpu curr_cpu_type; u32 unavailable_counters = 0; /* Either ophelp or one of the libop tests may invoke this * function with a non-native cpu_type. If so, we should not * call op_get_counter_mask because that will look for real counter * information in oprofilefs. */ curr_cpu_type = op_get_cpu_type(); if (cpu_type != curr_cpu_type) nr_counters = op_get_nr_counters(cpu_type); else nr_counters = op_get_counter_mask(&unavailable_counters); /* no counters then probably perfmon managing perfmon hw */ if (nr_counters <= 0) { nr_counters = op_get_nr_counters(cpu_type); unavailable_counters = (~0) << nr_counters; } /* Check to see if we have enough physical counters to map events*/ for (i = 0, nr_pmc_events = 0; i < nr_events; i++) if(pev[i]->ext == NULL) if (++nr_pmc_events > nr_counters) return 0; ctr_arc = build_counter_arc(pev, nr_events); counter_map = xmalloc(nr_events * sizeof(size_t)); if (!allocate_counter(ctr_arc, nr_events, 0, unavailable_counters, counter_map)) { free(counter_map); counter_map = 0; } delete_counter_arc(ctr_arc, nr_events); return counter_map; }