Hello,
> Takashi,
>
> This patch does not apply cleanly on my perfmon_cell.c.
> Could you check against the one in the GIT tree?
Yes.
> Also, is you mailer somehow formatting to 80-columns?
> Some lines of the patch are malformed due to insertion
> of carriage return.
My mailer formed it to 76-culums. I changed it.
I checked this patch againt the GIT tree
(commit 75c9e70942a28aa9366f969930305f172d2d2e0b) again.
Please use this patch!
And I want to wait for comments againt this patch for a while.
Thanks.
Takashi Yamamoto.
--------------------------------------------------------------
Cell PPU IU(Instruction Unit)/XU(Execution Unit) event support in
per-thread(task) mode.
In per-thread mode, the PPU IU/XU events should be measured
while the target SW thread(task) runs on the target HW thread.
To realize that, the counter enable bit of the pmX_control PMC is
manipulated in pfm_cell_restore_pmcs().
Signed-off-by: Takashi Yamamoto <TakashiA.Yamamoto at jp.sony.com>
---
arch/powerpc/perfmon/perfmon_cell.c | 202 +++++++++++++++++++++++++++++-------
1 file changed, 164 insertions(+), 38 deletions(-)
--- a/arch/powerpc/perfmon/perfmon_cell.c
+++ b/arch/powerpc/perfmon/perfmon_cell.c
@@ -96,7 +96,9 @@ static struct pfm_regmap_desc pfm_cell_p
};
#define PFM_PM_NUM_PMCS ARRAY_SIZE(pfm_cell_pmc_desc)
-#define CELL_PMC_PM_STATUS 20
+#define CELL_PMC_GROUP_CONTROL 16
+#define CELL_PMC_PM_STATUS 20
+
/*
* Mapping from Perfmon logical data counters to Cell hardware counters.
*/
@@ -112,6 +114,19 @@ static struct pfm_regmap_desc pfm_cell_p
};
#define PFM_PM_NUM_PMDS ARRAY_SIZE(pfm_cell_pmd_desc)
+#define PFM_EVENT_PMC_BUS_WORD(x) (((x) >> 48) & 0x00ff)
+#define PFM_EVENT_PMC_FULL_SIGNAL_NUMBER(x) ((x) & 0xffffffff)
+#define PFM_EVENT_PMC_SIGNAL_GROUP(x) (((x) & 0xffffffff) / 100)
+#define PFM_PM_CTR_INPUT_MUX_BIT(pm07_control) (((pm07_control) >> 26) & 0x1f)
+#define PFM_PM_CTR_INPUT_MUX_GROUP_INDEX(pm07_control) ((pm07_control) >> 31)
+#define PFM_GROUP_CONTROL_GROUP0_WORD(grp_ctrl) ((grp_ctrl) >> 30)
+#define PFM_GROUP_CONTROL_GROUP1_WORD(grp_ctrl) (((grp_ctrl) >> 28) & 0x3)
+#define PFM_NUM_OF_GROUPS 2
+#define PFM_PPU_IU1_THREAD1_BASE_BIT 19
+#define PFM_PPU_XU_THREAD1_BASE_BIT 16
+#define PFM_COUNTER_CTRL_PMC_PPU_THREAD0 0x100000000ULL
+#define PFM_COUNTER_CTRL_PMC_PPU_THREAD1 0x200000000ULL
+
/*
* Debug-bus signal handling.
*
@@ -691,6 +706,72 @@ static void pfm_cell_disable_counters(st
reset_signals(smp_processor_id());
}
+/*
+ * Return the thread id of the specified ppu signal.
+ */
+static inline u32 get_target_ppu_thread_id(u32 group, u32 bit)
+{
+ if ((group == SIG_GROUP_PPU_IU1 &&
+ bit < PFM_PPU_IU1_THREAD1_BASE_BIT) ||
+ (group == SIG_GROUP_PPU_XU &&
+ bit < PFM_PPU_XU_THREAD1_BASE_BIT))
+ return 0;
+ else
+ return 1;
+}
+
+/*
+ * Return whether the specified counter is for PPU signal group.
+ */
+static inline int is_counter_for_ppu_sig_grp(u32 counter_control, u32 sig_grp)
+{
+ if (!(counter_control & CBE_PM_CTR_INPUT_CONTROL) &&
+ (counter_control & CBE_PM_CTR_ENABLE) &&
+ ((sig_grp == SIG_GROUP_PPU_IU1) || (sig_grp == SIG_GROUP_PPU_XU)))
+ return 1;
+ else
+ return 0;
+}
+
+/*
+ * Search ppu signal groups.
+ */
+static int get_ppu_signal_groups(struct pfm_event_set *set,
+ u32 *ppu_sig_grp0, u32 *ppu_sig_grp1)
+{
+ u64 pm_event, *used_pmcs = set->used_pmcs;
+ int i, j;
+ u32 grp0_wd, grp1_wd, wd, sig_grp;
+
+ *ppu_sig_grp0 = 0;
+ *ppu_sig_grp1 = 0;
+ grp0_wd = PFM_GROUP_CONTROL_GROUP0_WORD(
+ set->pmcs[CELL_PMC_GROUP_CONTROL]);
+ grp1_wd = PFM_GROUP_CONTROL_GROUP1_WORD(
+ set->pmcs[CELL_PMC_GROUP_CONTROL]);
+
+ for (i = 0, j = 0; (i < NR_CTRS) && (j < PFM_NUM_OF_GROUPS); i++) {
+ if (test_bit(i + NR_CTRS, used_pmcs)) {
+ pm_event = set->pmcs[i + NR_CTRS];
+ wd = PFM_EVENT_PMC_BUS_WORD(pm_event);
+ sig_grp = PFM_EVENT_PMC_SIGNAL_GROUP(pm_event);
+ if ((sig_grp == SIG_GROUP_PPU_IU1) ||
+ (sig_grp == SIG_GROUP_PPU_XU)) {
+
+ if (wd == grp0_wd && *ppu_sig_grp0 == 0) {
+ *ppu_sig_grp0 = sig_grp;
+ j++;
+ } else if (wd == grp1_wd &&
+ *ppu_sig_grp1 == 0) {
+ *ppu_sig_grp1 = sig_grp;
+ j++;
+ }
+ }
+ }
+ }
+ return j;
+}
+
/**
* pfm_cell_restore_pmcs
*
@@ -699,56 +780,54 @@ static void pfm_cell_disable_counters(st
* individually (as is done in other architectures), but that results in
* multiple RTAS calls. As an optimization, we will setup the RTAS argument
* array so we can do all event-control registers in one RTAS call.
+ *
+ * In per-thread mode,
+ * The counter enable bit of the pmX_control PMC is enabled while the target
+ * task runs on the target HW thread.
**/
void pfm_cell_restore_pmcs(struct pfm_event_set *set)
{
- struct cell_rtas_arg signals[NR_CTRS];
- u64 value, *used_pmcs = set->used_pmcs;
- int i, rc, num_used = 0, cpu = smp_processor_id();
- s32 signal_number;
+ u64 ctr_ctrl;
+ u64 *used_pmcs = set->used_pmcs;
+ int i;
+ int cpu = smp_processor_id();
+ u32 current_th_id;
struct pfm_cell_platform_pmu_info *info =
((struct pfm_arch_pmu_info *)
(pfm_pmu_conf->arch_info))->platform_info;
- memset(signals, 0, sizeof(signals));
-
for (i = 0; i < NR_CTRS; i++) {
- /* Write the per-counter control register. If the PMC is not
- * in use, then it will simply clear the register, which will
- * disable the associated counter.
- */
- info->write_pm07_control(cpu, i, set->pmcs[i]);
+ ctr_ctrl = set->pmcs[i];
- /* Set up the next RTAS array entry for this counter. Only
- * include pm07_event registers that are in use by this set
- * so the RTAS call doesn't have to process blank array entries.
- */
- if (!test_bit(i + NR_CTRS, used_pmcs)) {
- continue;
- }
+ if (ctr_ctrl & PFM_COUNTER_CTRL_PMC_PPU_THREAD0) {
+ current_th_id = info->get_hw_thread_id(cpu);
- value = set->pmcs[i + NR_CTRS];
- signal_number = RTAS_SIGNAL_NUMBER(value);
- if (!signal_number) {
- /* Don't include counters that are counting cycles. */
- continue;
+ /*
+ * Set the counter enable bit down if the current
+ * HW thread is NOT 0
+ **/
+ if (current_th_id)
+ ctr_ctrl = ctr_ctrl & ~CBE_PM_CTR_ENABLE;
+
+ } else if (ctr_ctrl & PFM_COUNTER_CTRL_PMC_PPU_THREAD1) {
+ current_th_id = info->get_hw_thread_id(cpu);
+
+ /*
+ * Set the counter enable bit down if the current
+ * HW thread is 0
+ **/
+ if (!current_th_id)
+ ctr_ctrl = ctr_ctrl & ~CBE_PM_CTR_ENABLE;
}
- signals[num_used].cpu = RTAS_CPU(cpu);
- signals[num_used].sub_unit = RTAS_SUB_UNIT(value);
- signals[num_used].bus_word = 1 << RTAS_BUS_WORD(value);
- signals[num_used].signal_group = signal_number / 100;
- signals[num_used].bit = signal_number % 100;
- num_used++;
- }
-
- rc = activate_signals(signals, num_used);
- if (rc) {
- PFM_WARN("Error calling activate_signal(): %d\n", rc);
- /* FIX: We will also need this routine to be able to return
- * an error if Stephane agrees to change pfm_arch_write_pmc
- * to return an error.
+ /* Write the per-counter control register. If the PMC is not
+ * in use, then it will simply clear the register, which will
+ * disable the associated counter.
*/
+ info->write_pm07_control(cpu, i, ctr_ctrl);
+
+ if (test_bit(i + NR_CTRS, used_pmcs))
+ write_pm07_event(cpu, 0, set->pmcs[i + NR_CTRS]);
}
/* Write all the global PMCs. Need to call pfm_cell_write_pmc()
@@ -760,6 +839,52 @@ void pfm_cell_restore_pmcs(struct pfm_ev
}
/**
+ * pfm_cell_load_context
+ *
+ * In per-thread mode,
+ * The pmX_control PMCs which are used for PPU IU/XU event are marked with
+ * the thread id(PFM_COUNTER_CTRL_PMC_PPU_THREAD0/THREAD1).
+ **/
+static int pfm_cell_load_context(struct pfm_context *ctx,
+ struct pfm_event_set *set,
+ struct task_struct *task)
+{
+ int i;
+ u32 ppu_sig_grp[PFM_NUM_OF_GROUPS] = {SIG_GROUP_NONE, SIG_GROUP_NONE};
+ u32 bit;
+ int grp_index;
+ u32 target_th_id;
+ int ppu_sig_num = 0;
+
+ if (!ctx->flags.system)
+ ppu_sig_num = get_ppu_signal_groups(set, &ppu_sig_grp[0],
+ &ppu_sig_grp[1]);
+
+ for (i = 0; i < NR_CTRS; i++) {
+ grp_index = PFM_PM_CTR_INPUT_MUX_GROUP_INDEX(set->pmcs[i]);
+ if (ppu_sig_num &&
+ (ppu_sig_grp[grp_index] != SIG_GROUP_NONE) &&
+ is_counter_for_ppu_sig_grp(set->pmcs[i],
+ ppu_sig_grp[grp_index])) {
+
+ bit = PFM_PM_CTR_INPUT_MUX_BIT(set->pmcs[i]);
+ target_th_id = get_target_ppu_thread_id(
+ ppu_sig_grp[grp_index], bit);
+ if (!target_th_id)
+ set->pmcs[i] |=
+ PFM_COUNTER_CTRL_PMC_PPU_THREAD0;
+ else
+ set->pmcs[i] |=
+ PFM_COUNTER_CTRL_PMC_PPU_THREAD1;
+ PFM_DBG("mark ctr:%d target_thread:%d",
+ i, target_th_id);
+ }
+ }
+
+ return 0;
+}
+
+/**
* pfm_cell_unload_context
*
* For system-wide contexts and self-monitored contexts, make the RTAS call
@@ -1026,6 +1151,7 @@ static struct pfm_arch_pmu_info pfm_cell
.get_ovfl_pmds = pfm_cell_get_ovfl_pmds,
.restore_pmcs = pfm_cell_restore_pmcs,
.ctxswout_thread = pfm_cell_ctxswout_thread,
+ .load_context = pfm_cell_load_context,
.unload_context = pfm_cell_unload_context,
};
_______________________________________________
perfmon mailing list
[email protected]
http://www.hpl.hp.com/hosted/linux/mail-archives/perfmon/
