]> git.kernelconcepts.de Git - karo-tx-linux.git/blobdiff - kernel/events/core.c
bpf, perf: delay release of BPF prog after grace period
[karo-tx-linux.git] / kernel / events / core.c
index ef2d6ea10736e4805e758cae83524dc13c2f06e9..12ecd4f0329fcb6532eb6391ac8984a9a2e901b3 100644 (file)
@@ -946,6 +946,7 @@ static void put_ctx(struct perf_event_context *ctx)
  * function.
  *
  * Lock order:
+ *    cred_guard_mutex
  *     task_struct::perf_event_mutex
  *       perf_event_context::mutex
  *         perf_event_context::lock
@@ -1580,14 +1581,14 @@ event_sched_out(struct perf_event *event,
 
        perf_pmu_disable(event->pmu);
 
+       event->tstamp_stopped = tstamp;
+       event->pmu->del(event, 0);
+       event->oncpu = -1;
        event->state = PERF_EVENT_STATE_INACTIVE;
        if (event->pending_disable) {
                event->pending_disable = 0;
                event->state = PERF_EVENT_STATE_OFF;
        }
-       event->tstamp_stopped = tstamp;
-       event->pmu->del(event, 0);
-       event->oncpu = -1;
 
        if (!is_software_event(event))
                cpuctx->active_oncpu--;
@@ -3154,15 +3155,16 @@ static int event_enable_on_exec(struct perf_event *event,
  * Enable all of a task's events that have been marked enable-on-exec.
  * This expects task == current.
  */
-static void perf_event_enable_on_exec(struct perf_event_context *ctx)
+static void perf_event_enable_on_exec(int ctxn)
 {
-       struct perf_event_context *clone_ctx = NULL;
+       struct perf_event_context *ctx, *clone_ctx = NULL;
        struct perf_event *event;
        unsigned long flags;
        int enabled = 0;
        int ret;
 
        local_irq_save(flags);
+       ctx = current->perf_event_ctxp[ctxn];
        if (!ctx || !ctx->nr_events)
                goto out;
 
@@ -3205,17 +3207,11 @@ out:
 
 void perf_event_exec(void)
 {
-       struct perf_event_context *ctx;
        int ctxn;
 
        rcu_read_lock();
-       for_each_task_context_nr(ctxn) {
-               ctx = current->perf_event_ctxp[ctxn];
-               if (!ctx)
-                       continue;
-
-               perf_event_enable_on_exec(ctx);
-       }
+       for_each_task_context_nr(ctxn)
+               perf_event_enable_on_exec(ctxn);
        rcu_read_unlock();
 }
 
@@ -3423,7 +3419,6 @@ static struct task_struct *
 find_lively_task_by_vpid(pid_t vpid)
 {
        struct task_struct *task;
-       int err;
 
        rcu_read_lock();
        if (!vpid)
@@ -3437,16 +3432,7 @@ find_lively_task_by_vpid(pid_t vpid)
        if (!task)
                return ERR_PTR(-ESRCH);
 
-       /* Reuse ptrace permission checks for now. */
-       err = -EACCES;
-       if (!ptrace_may_access(task, PTRACE_MODE_READ))
-               goto errout;
-
        return task;
-errout:
-       put_task_struct(task);
-       return ERR_PTR(err);
-
 }
 
 /*
@@ -6493,9 +6479,6 @@ struct swevent_htable {
 
        /* Recursion avoidance in each contexts */
        int                             recursion[PERF_NR_CONTEXTS];
-
-       /* Keeps track of cpu being initialized/exited */
-       bool                            online;
 };
 
 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
@@ -6753,14 +6736,8 @@ static int perf_swevent_add(struct perf_event *event, int flags)
        hwc->state = !(flags & PERF_EF_START);
 
        head = find_swevent_head(swhash, event);
-       if (!head) {
-               /*
-                * We can race with cpu hotplug code. Do not
-                * WARN if the cpu just got unplugged.
-                */
-               WARN_ON_ONCE(swhash->online);
+       if (WARN_ON_ONCE(!head))
                return -EINVAL;
-       }
 
        hlist_add_head_rcu(&event->hlist_entry, head);
        perf_event_update_userpage(event);
@@ -6828,7 +6805,6 @@ static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
        int err = 0;
 
        mutex_lock(&swhash->hlist_mutex);
-
        if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
                struct swevent_hlist *hlist;
 
@@ -7125,7 +7101,7 @@ static void perf_event_free_bpf_prog(struct perf_event *event)
        prog = event->tp_event->prog;
        if (prog) {
                event->tp_event->prog = NULL;
-               bpf_prog_put(prog);
+               bpf_prog_put_rcu(prog);
        }
 }
 
@@ -7994,6 +7970,9 @@ perf_event_alloc(struct perf_event_attr *attr, int cpu,
                }
        }
 
+       /* symmetric to unaccount_event() in _free_event() */
+       account_event(event);
+
        return event;
 
 err_per_task:
@@ -8340,6 +8319,24 @@ SYSCALL_DEFINE5(perf_event_open,
 
        get_online_cpus();
 
+       if (task) {
+               err = mutex_lock_interruptible(&task->signal->cred_guard_mutex);
+               if (err)
+                       goto err_cpus;
+
+               /*
+                * Reuse ptrace permission checks for now.
+                *
+                * We must hold cred_guard_mutex across this and any potential
+                * perf_install_in_context() call for this new event to
+                * serialize against exec() altering our credentials (and the
+                * perf_event_exit_task() that could imply).
+                */
+               err = -EACCES;
+               if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS))
+                       goto err_cred;
+       }
+
        if (flags & PERF_FLAG_PID_CGROUP)
                cgroup_fd = pid;
 
@@ -8347,7 +8344,7 @@ SYSCALL_DEFINE5(perf_event_open,
                                 NULL, NULL, cgroup_fd);
        if (IS_ERR(event)) {
                err = PTR_ERR(event);
-               goto err_cpus;
+               goto err_cred;
        }
 
        if (is_sampling_event(event)) {
@@ -8357,8 +8354,6 @@ SYSCALL_DEFINE5(perf_event_open,
                }
        }
 
-       account_event(event);
-
        /*
         * Special case software events and allow them to be part of
         * any hardware group.
@@ -8408,11 +8403,6 @@ SYSCALL_DEFINE5(perf_event_open,
                goto err_context;
        }
 
-       if (task) {
-               put_task_struct(task);
-               task = NULL;
-       }
-
        /*
         * Look up the group leader (we will attach this event to it):
         */
@@ -8500,6 +8490,11 @@ SYSCALL_DEFINE5(perf_event_open,
 
        WARN_ON_ONCE(ctx->parent_ctx);
 
+       /*
+        * This is the point on no return; we cannot fail hereafter. This is
+        * where we start modifying current state.
+        */
+
        if (move_group) {
                /*
                 * See perf_event_ctx_lock() for comments on the details
@@ -8569,6 +8564,11 @@ SYSCALL_DEFINE5(perf_event_open,
                mutex_unlock(&gctx->mutex);
        mutex_unlock(&ctx->mutex);
 
+       if (task) {
+               mutex_unlock(&task->signal->cred_guard_mutex);
+               put_task_struct(task);
+       }
+
        put_online_cpus();
 
        event->owner = current;
@@ -8597,7 +8597,15 @@ err_context:
        perf_unpin_context(ctx);
        put_ctx(ctx);
 err_alloc:
-       free_event(event);
+       /*
+        * If event_file is set, the fput() above will have called ->release()
+        * and that will take care of freeing the event.
+        */
+       if (!event_file)
+               free_event(event);
+err_cred:
+       if (task)
+               mutex_unlock(&task->signal->cred_guard_mutex);
 err_cpus:
        put_online_cpus();
 err_task:
@@ -8641,8 +8649,6 @@ perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
        /* Mark owner so we could distinguish it from user events. */
        event->owner = EVENT_OWNER_KERNEL;
 
-       account_event(event);
-
        ctx = find_get_context(event->pmu, task, event);
        if (IS_ERR(ctx)) {
                err = PTR_ERR(ctx);
@@ -8879,6 +8885,9 @@ static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
 
 /*
  * When a child task exits, feed back event values to parent events.
+ *
+ * Can be called with cred_guard_mutex held when called from
+ * install_exec_creds().
  */
 void perf_event_exit_task(struct task_struct *child)
 {
@@ -9291,7 +9300,6 @@ static void perf_event_init_cpu(int cpu)
        struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
 
        mutex_lock(&swhash->hlist_mutex);
-       swhash->online = true;
        if (swhash->hlist_refcount > 0) {
                struct swevent_hlist *hlist;
 
@@ -9333,14 +9341,7 @@ static void perf_event_exit_cpu_context(int cpu)
 
 static void perf_event_exit_cpu(int cpu)
 {
-       struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
-
        perf_event_exit_cpu_context(cpu);
-
-       mutex_lock(&swhash->hlist_mutex);
-       swhash->online = false;
-       swevent_hlist_release(swhash);
-       mutex_unlock(&swhash->hlist_mutex);
 }
 #else
 static inline void perf_event_exit_cpu(int cpu) { }