perfmon2 minimal v3: attach/detach support

Index: o/perfmon/perfmon_attach.c
===================================================================
@@ -42,6 +42,243 @@
 #include "perfmon_priv.h"
 
 /**
+ * __pfm_load_ctx_thread - attach context to a thread
+ * @ctx: context to operate on
+ * @task: thread to attach to
+ *
+ * The function must be called with the context locked and interrupts disabled.
+ */
+static int pfm_load_ctx_thread(struct pfm_context *ctx,
+			       struct task_struct *task)
+{
+	struct pfm_event_set *set;
+	struct pfm_context *old;
+	int ret;
+	u16 max;
+
+	PFM_DBG("pid=%d",  task->pid);
+
+	/*
+	 * per-thread:
+	 *   - task to attach to is checked in sys_pfm_load_context() to avoid
+	 *     locking issues. if found, and not self,  task refcount was
+	 *     incremented.
+	 */
+	old = cmpxchg(&task->pfm_context, NULL, ctx);
+	if (old) {
+		PFM_DBG("load_pid=%d has a context "
+			"old=%p new=%p cur=%p",
+			task->pid,
+			old,
+			ctx,
+			task->pfm_context);
+		return -EEXIST;
+	}
+
+	/*
+	 * initialize sets
+	 */
+	set = ctx->active_set;
+
+	max = pfm_pmu_conf->regs.max_intr_pmd;
+
+	/*
+	 * cleanup bitvectors
+	 */
+	bitmap_zero(cast_ulp(set->povfl_pmds), max);
+
+	set->npend_ovfls = 0;
+
+	/*
+	 * we cannot just use plain clear because of arch-specific flags
+	 */
+	set->priv_flags &= ~PFM_SETFL_PRIV_MOD_BOTH;
+
+	ctx->task = task;
+
+	/*
+	 * perform any architecture specific actions
+	 */
+	ret = pfm_arch_load_context(ctx);
+	if (ret)
+		goto error_noload;
+
+	/*
+	 * now reserve the session, before we can proceed with
+	 * actually accessing the PMU hardware
+	 */
+	ret = pfm_session_acquire();
+	if (ret)
+		goto error;
+
+
+	if (ctx->task != current) {
+
+		ctx->flags.is_self = 0;
+
+		/* force a full reload */
+		ctx->last_act = PFM_INVALID_ACTIVATION;
+		ctx->last_cpu = -1;
+		set->priv_flags |= PFM_SETFL_PRIV_MOD_BOTH;
+
+	} else {
+		pfm_check_save_prev_ctx();
+
+		ctx->last_cpu = smp_processor_id();
+		__get_cpu_var(pmu_activation_number)++;
+		ctx->last_act = __get_cpu_var(pmu_activation_number);
+
+		ctx->flags.is_self = 1;
+
+		/*
+		 * load PMD from set
+		 * load PMC from set
+		 */
+		pfm_arch_restore_pmds(ctx, set);
+		pfm_arch_restore_pmcs(ctx, set);
+
+		/*
+		 * set new ownership
+		 */
+		pfm_set_pmu_owner(ctx->task, ctx);
+	}
+	set_tsk_thread_flag(task, TIF_PERFMON_CTXSW);
+
+	ctx->state = PFM_CTX_LOADED;
+
+	return 0;
+
+error:
+	pfm_arch_unload_context(ctx);
+	ctx->task = NULL;
+error_noload:
+	/*
+	 * detach context
+	 */
+	task->pfm_context = NULL;
+	return ret;
+}
+
+/**
+ * __pfm_load_context - attach context to a thread
+ * @ctx: context to operate on
+ * @task: thread to attach to
+ */
+int __pfm_load_context(struct pfm_context *ctx, struct task_struct *task)
+{
+	return pfm_load_ctx_thread(ctx, task);
+}
+
+/**
+ * pfm_update_ovfl_pmds - account for pending ovfls on PMDs
+ * @ctx: context to operate on
+ *
+ * This function is always called after pfm_stop has been issued
+ */
+static void pfm_update_ovfl_pmds(struct pfm_context *ctx)
+{
+	struct pfm_event_set *set;
+	u64 *cnt_pmds;
+	u64 ovfl_mask;
+	u16 num_ovfls, i;
+
+	ovfl_mask = pfm_pmu_conf->ovfl_mask;
+	cnt_pmds = pfm_pmu_conf->regs.cnt_pmds;
+	set = ctx->active_set;
+
+	if (!set->npend_ovfls)
+		return;
+
+	num_ovfls = set->npend_ovfls;
+	PFM_DBG("novfls=%u", num_ovfls);
+
+	for (i = 0; num_ovfls; i++) {
+		if (test_bit(i, cast_ulp(set->povfl_pmds))) {
+			/* only correct value for counters */
+			if (test_bit(i, cast_ulp(cnt_pmds)))
+				set->pmds[i].value += 1 + ovfl_mask;
+			num_ovfls--;
+		}
+		PFM_DBG("pmd%u val=0x%llx",
+			i,
+			(unsigned long long)set->pmds[i].value);
+	}
+	/*
+	 * we need to clear to prevent a pfm_getinfo_evtsets() from
+	 * returning stale data even after the context is unloaded
+	 */
+	set->npend_ovfls = 0;
+	bitmap_zero(cast_ulp(set->povfl_pmds),
+		    pfm_pmu_conf->regs.max_intr_pmd);
+}
+
+/**
+ * __pfm_unload_context - detach context from CPU or thread
+ * @ctx: context to operate on
+ *
+ * The function must be called with the context locked and interrupts disabled.
+ */
+int __pfm_unload_context(struct pfm_context *ctx)
+{
+	int ret;
+
+	PFM_DBG("ctx_state=%d task [%d]",
+		ctx->state,
+		ctx->task ? ctx->task->pid : -1);
+
+	/*
+	 * check unload-able state
+	 */
+	if (ctx->state == PFM_CTX_UNLOADED)
+		return -EINVAL;
+
+	/*
+	 * stop monitoring
+	 */
+	ret = __pfm_stop(ctx);
+	if (ret)
+		return ret;
+
+	ctx->state = PFM_CTX_UNLOADED;
+
+	/*
+	 * save active set
+	 * UP:
+	 * 	if not current task and due to lazy, state may
+	 * 	still be live
+	 * for system-wide, guaranteed to run on correct CPU
+	 */
+	if (__get_cpu_var(pmu_ctx) == ctx) {
+		/*
+		 * pending overflows have been saved by pfm_stop()
+		 */
+		pfm_save_pmds(ctx);
+		pfm_set_pmu_owner(NULL, NULL);
+		PFM_DBG("released ownership");
+	}
+
+	/*
+	 * account for pending overflows
+	 */
+	pfm_update_ovfl_pmds(ctx);
+
+	/*
+	 * arch-specific unload operations
+	 */
+	pfm_arch_unload_context(ctx);
+
+	/*
+	 * per-thread: disconnect from monitored task
+	 */
+	if (ctx->task) {
+		ctx->task->pfm_context = NULL;
+		clear_tsk_thread_flag(ctx->task, TIF_PERFMON_CTXSW);
+		ctx->task = NULL;
+	}
+	return 0;
+}
+
+/**
  * __pfm_exit_thread - detach and free context on thread exit
  */
 void __pfm_exit_thread(void)

Index: o/perfmon/perfmon_priv.h
===================================================================
@@ -60,6 +60,9 @@
 int __pfm_stop(struct pfm_context *ctx);
 int __pfm_start(struct pfm_context *ctx);
 
+int __pfm_load_context(struct pfm_context *ctx, struct task_struct *task);
+int __pfm_unload_context(struct pfm_context *ctx);
+
 ssize_t pfm_sysfs_res_show(char *buf, size_t sz, int what);
 
 int pfm_pmu_acquire(void);