[PATCH] aio: kiocb locking to serialise retry and cancel
authorBenjamin LaHaise <bcrl@linux.intel.com>
Fri, 9 Sep 2005 20:02:09 +0000 (13:02 -0700)
committerLinus Torvalds <torvalds@g5.osdl.org>
Fri, 9 Sep 2005 20:57:32 +0000 (13:57 -0700)
Implement a per-kiocb lock to serialise retry operations and cancel.  This
is done using wait_on_bit_lock() on the KIF_LOCKED bit of kiocb->ki_flags.
Also, make the cancellation path lock the kiocb and subsequently release
all references to it if the cancel was successful.  This version includes a
fix for the deadlock with __aio_run_iocbs.

Signed-off-by: Benjamin LaHaise <bcrl@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
fs/aio.c

index 769791df36b4f357bdcbbd2fa891cb23d6f477cd..201c1847fa07b694601a652e0e68db5f802eb042 100644 (file)
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -546,6 +546,24 @@ struct kioctx *lookup_ioctx(unsigned long ctx_id)
        return ioctx;
 }
 
+static int lock_kiocb_action(void *param)
+{
+       schedule();
+       return 0;
+}
+
+static inline void lock_kiocb(struct kiocb *iocb)
+{
+       wait_on_bit_lock(&iocb->ki_flags, KIF_LOCKED, lock_kiocb_action,
+                        TASK_UNINTERRUPTIBLE);
+}
+
+static inline void unlock_kiocb(struct kiocb *iocb)
+{
+       kiocbClearLocked(iocb);
+       wake_up_bit(&iocb->ki_flags, KIF_LOCKED);
+}
+
 /*
  * use_mm
  *     Makes the calling kernel thread take on the specified
@@ -786,7 +804,9 @@ static int __aio_run_iocbs(struct kioctx *ctx)
                 * Hold an extra reference while retrying i/o.
                 */
                iocb->ki_users++;       /* grab extra reference */
+               lock_kiocb(iocb);
                aio_run_iocb(iocb);
+               unlock_kiocb(iocb);
                if (__aio_put_req(ctx, iocb))  /* drop extra ref */
                        put_ioctx(ctx);
        }
@@ -1527,10 +1547,9 @@ int fastcall io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
                goto out_put_req;
 
        spin_lock_irq(&ctx->ctx_lock);
-       if (likely(list_empty(&ctx->run_list))) {
-               aio_run_iocb(req);
-       } else {
-               list_add_tail(&req->ki_run_list, &ctx->run_list);
+       aio_run_iocb(req);
+       unlock_kiocb(req);
+       if (!list_empty(&ctx->run_list)) {
                /* drain the run list */
                while (__aio_run_iocbs(ctx))
                        ;
@@ -1661,6 +1680,7 @@ asmlinkage long sys_io_cancel(aio_context_t ctx_id, struct iocb __user *iocb,
        if (NULL != cancel) {
                struct io_event tmp;
                pr_debug("calling cancel\n");
+               lock_kiocb(kiocb);
                memset(&tmp, 0, sizeof(tmp));
                tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
                tmp.data = kiocb->ki_user_data;
@@ -1672,6 +1692,7 @@ asmlinkage long sys_io_cancel(aio_context_t ctx_id, struct iocb __user *iocb,
                        if (copy_to_user(result, &tmp, sizeof(tmp)))
                                ret = -EFAULT;
                }
+               unlock_kiocb(kiocb);
        } else
                ret = -EINVAL;