OSDN Git Service

perf: Restructure perf syscall point of no return
authorPeter Zijlstra <peterz@infradead.org>
Wed, 9 Sep 2015 17:06:33 +0000 (19:06 +0200)
committerIngo Molnar <mingo@kernel.org>
Fri, 18 Sep 2015 07:20:24 +0000 (09:20 +0200)
The exclusive_event_installable() stuff only works because its
exclusive with the grouping bits.

Rework the code such that there is a sane place to error out before we
go do things we cannot undo.

Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
kernel/events/core.c

index f548f69..39679f7 100644 (file)
@@ -8297,13 +8297,30 @@ SYSCALL_DEFINE5(perf_event_open,
 
        if (move_group) {
                gctx = group_leader->ctx;
+               mutex_lock_double(&gctx->mutex, &ctx->mutex);
+       } else {
+               mutex_lock(&ctx->mutex);
+       }
+
+       /*
+        * Must be under the same ctx::mutex as perf_install_in_context(),
+        * because we need to serialize with concurrent event creation.
+        */
+       if (!exclusive_event_installable(event, ctx)) {
+               /* exclusive and group stuff are assumed mutually exclusive */
+               WARN_ON_ONCE(move_group);
+
+               err = -EBUSY;
+               goto err_locked;
+       }
 
+       WARN_ON_ONCE(ctx->parent_ctx);
+
+       if (move_group) {
                /*
                 * See perf_event_ctx_lock() for comments on the details
                 * of swizzling perf_event::ctx.
                 */
-               mutex_lock_double(&gctx->mutex, &ctx->mutex);
-
                perf_remove_from_context(group_leader, false);
 
                list_for_each_entry(sibling, &group_leader->sibling_list,
@@ -8311,13 +8328,7 @@ SYSCALL_DEFINE5(perf_event_open,
                        perf_remove_from_context(sibling, false);
                        put_ctx(gctx);
                }
-       } else {
-               mutex_lock(&ctx->mutex);
-       }
-
-       WARN_ON_ONCE(ctx->parent_ctx);
 
-       if (move_group) {
                /*
                 * Wait for everybody to stop referencing the events through
                 * the old lists, before installing it on new lists.
@@ -8349,22 +8360,20 @@ SYSCALL_DEFINE5(perf_event_open,
                perf_event__state_init(group_leader);
                perf_install_in_context(ctx, group_leader, group_leader->cpu);
                get_ctx(ctx);
-       }
 
-       if (!exclusive_event_installable(event, ctx)) {
-               err = -EBUSY;
-               mutex_unlock(&ctx->mutex);
-               fput(event_file);
-               goto err_context;
+               /*
+                * Now that all events are installed in @ctx, nothing
+                * references @gctx anymore, so drop the last reference we have
+                * on it.
+                */
+               put_ctx(gctx);
        }
 
        perf_install_in_context(ctx, event, event->cpu);
        perf_unpin_context(ctx);
 
-       if (move_group) {
+       if (move_group)
                mutex_unlock(&gctx->mutex);
-               put_ctx(gctx);
-       }
        mutex_unlock(&ctx->mutex);
 
        put_online_cpus();
@@ -8391,6 +8400,12 @@ SYSCALL_DEFINE5(perf_event_open,
        fd_install(event_fd, event_file);
        return event_fd;
 
+err_locked:
+       if (move_group)
+               mutex_unlock(&gctx->mutex);
+       mutex_unlock(&ctx->mutex);
+/* err_file: */
+       fput(event_file);
 err_context:
        perf_unpin_context(ctx);
        put_ctx(ctx);