Hello,

> >
> > I think we need to hide this into an arch_specific function. Which can
> > then have a indirect function pointer for
> > cell just like what you do today on PPC.
> >
> > But I don't understand why you need this extra check JUST for
> > pfm_read_pmds() and not for all 3 accesses.
> > Could you elaborate on that?
> 
> I think that pfm_write_pmcs() and pfm_write_pmds() need this extra check too.
> I did not notice that because I didn't get the problem by pfmon.
> But If the user program calls them on its way, the problem happen.
> 
> I will update the patch tomorrow.

I updated it.
Please review again!

Thanks.
Takashi Yamamoto.

------------------------------
This patch changes the check of PMU access in pfm_read_pmds(),
pfm_write_pmds(), pfm_write_pmcs().

In the cell-spe-follow mode, 
we need to check whether the target SPU program is running
before pmc/pmd access.

Signed-off-by: Takashi Yamamoto <[EMAIL PROTECTED]>
---
 arch/powerpc/perfmon/perfmon_cell.c |   23 ++++++++++++++++++++
 include/asm-powerpc/perfmon_kern.h  |   17 ++++++++++++++
 perfmon/perfmon_rw.c                |   41 +++++++++++++++++++++++++++++++++---
 3 files changed, 78 insertions(+), 3 deletions(-)

--- a/arch/powerpc/perfmon/perfmon_cell.c
+++ b/arch/powerpc/perfmon/perfmon_cell.c
@@ -1415,6 +1415,28 @@ void pfm_cell_release_pmu(void)
 }
 
 /**
+ * pfm_cell_can_access_pmu
+ *
+ * In the per-thread & spe-follow mode,
+ * if the target spu program is running, the pmc/pmd register
+ * access is allowed.
+ **/
+int pfm_cell_can_access_pmu(struct pfm_context *ctx, int can_access)
+{
+       struct pfm_arch_context *ctx_arch = pfm_ctx_arch(ctx);
+
+       if (!ctx->flags.system && ctx->flags.not_dflt_ctxsw)
+               if (ctx_arch->arg)
+                       /* the target spu program is running */
+                       return 1;
+               else
+                       /* the traget spu program is NOT running */
+                       return 0;
+       else
+               return can_access;
+}
+
+/**
  * handle_trace_buffer_interrupts
  *
  * This routine is for processing just the interval timer and trace buffer
@@ -1548,6 +1570,7 @@ static struct pfm_cell_platform_pmu_info
 
 static struct pfm_arch_pmu_info pfm_cell_pmu_info = {
        .pmu_style        = PFM_POWERPC_PMU_CELL,
+       .can_access_pmu   = pfm_cell_can_access_pmu,
        .acquire_pmu      = pfm_cell_acquire_pmu,
        .release_pmu      = pfm_cell_release_pmu,
        .create_context   = pfm_cell_create_context,
--- a/include/asm-powerpc/perfmon_kern.h
+++ b/include/asm-powerpc/perfmon_kern.h
@@ -120,6 +120,7 @@ struct pfm_arch_pmu_info {
        void (*free_context)(struct pfm_context *ctx);
        int  (*acquire_pmu)(void);
        void (*release_pmu)(void);
+       int  (*can_access_pmu)(struct pfm_context *ctx, int can_access);
        void *platform_info;
 };
 
@@ -427,6 +428,22 @@ struct pfm_arch_context {
        void                    *arg;
 };
 
+/*
+ * This function returns whether the pmc/pmd register access is allowed.
+ */
+static inline int pfm_arch_can_access_pmu(
+       struct pfm_context *ctx, int can_access)
+{
+       struct pfm_arch_pmu_info *arch_info;
+
+       arch_info =  pfm_pmu_info();
+
+       if (arch_info->can_access_pmu)
+               return arch_info->can_access_pmu(ctx, can_access);
+       else
+               return can_access;
+}
+
 #define PFM_ARCH_CTX_SIZE sizeof(struct pfm_arch_context)
 /*
  * PowerPC does not need extra alignment requirements for the sampling buffer
--- a/perfmon/perfmon_rw.c
+++ b/perfmon/perfmon_rw.c
@@ -83,10 +83,22 @@ int __pfm_write_pmds(struct pfm_context 
        /*
         * we cannot access the actual PMD registers when monitoring is masked
         */
-       if (unlikely(ctx->state == PFM_CTX_LOADED))
+       if (unlikely(ctx->state == PFM_CTX_LOADED)) {
                can_access_pmu = __get_cpu_var(pmu_owner) == ctx->task
                              || ctx->flags.system;
 
+#ifdef CONFIG_PPC
+               /*
+                * Architecture depended access check.
+                *
+                * cell_spe_follow mode needs this check.
+                * Because the target SPU program may run on
+                * the other CPU's SPU.
+                */
+               can_access_pmu = pfm_arch_can_access_pmu(ctx, can_access_pmu);
+#endif
+       }
+
        ret = -EINVAL;
 
        for (i = 0; i < count; i++, req++) {
@@ -355,10 +367,22 @@ int __pfm_write_pmcs(struct pfm_context 
        /*
         * we cannot access the actual PMC registers when monitoring is masked
         */
-       if (unlikely(ctx->state == PFM_CTX_LOADED))
+       if (unlikely(ctx->state == PFM_CTX_LOADED)) {
                can_access_pmu = __get_cpu_var(pmu_owner) == ctx->task
                                || ctx->flags.system;
 
+#ifdef CONFIG_PPC
+               /*
+                * Architecture depended access check.
+                *
+                * cell_spe_follow mode needs this check.
+                * Because the target SPU program may run on
+                * the other CPU's SPU.
+                */
+               can_access_pmu = pfm_arch_can_access_pmu(ctx, can_access_pmu);
+#endif
+       }
+
        ret = -EINVAL;
 
        for (i = 0; i < count; i++, req++) {
@@ -506,7 +530,18 @@ int __pfm_read_pmds(struct pfm_context *
 
        if (likely(ctx->state == PFM_CTX_LOADED)) {
                can_access_pmu = __get_cpu_var(pmu_owner) == ctx->task
-                              || ctx->flags.system;
+                       || ctx->flags.system;
+
+#ifdef CONFIG_PPC
+               /*
+                * Architecture depended access check.
+                *
+                * cell_spe_follow mode needs this check.
+                * Because the target SPU program may run on
+                * the other CPU's SPU.
+                */
+               can_access_pmu = pfm_arch_can_access_pmu(ctx, can_access_pmu);
+#endif
 
                if (can_access_pmu)
                        pfm_arch_serialize();



Attachment: cell-change_read_pmds.patch
Description: Binary data

-------------------------------------------------------------------------
This SF.net email is sponsored by the 2008 JavaOne(SM) Conference 
Don't miss this year's exciting event. There's still time to save $100. 
Use priority code J8TL2D2. 
http://ad.doubleclick.net/clk;198757673;13503038;p?http://java.sun.com/javaone
_______________________________________________
perfmon2-devel mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/perfmon2-devel

Reply via email to