fuse: fix sparse warnings
[safe/jmp/linux-2.6] / fs / fuse / dev.c
index c510533..87250b6 100644 (file)
@@ -1,6 +1,6 @@
 /*
   FUSE: Filesystem in Userspace
-  Copyright (C) 2001-2005  Miklos Szeredi <miklos@szeredi.hu>
+  Copyright (C) 2001-2006  Miklos Szeredi <miklos@szeredi.hu>
 
   This program can be distributed under the terms of the GNU GPL.
   See the file COPYING.
@@ -19,7 +19,7 @@
 
 MODULE_ALIAS_MISCDEV(FUSE_MINOR);
 
-static kmem_cache_t *fuse_req_cachep;
+static struct kmem_cache *fuse_req_cachep;
 
 static struct fuse_conn *fuse_get_conn(struct file *file)
 {
@@ -34,13 +34,22 @@ static void fuse_request_init(struct fuse_req *req)
 {
        memset(req, 0, sizeof(*req));
        INIT_LIST_HEAD(&req->list);
+       INIT_LIST_HEAD(&req->intr_entry);
        init_waitqueue_head(&req->waitq);
        atomic_set(&req->count, 1);
 }
 
 struct fuse_req *fuse_request_alloc(void)
 {
-       struct fuse_req *req = kmem_cache_alloc(fuse_req_cachep, SLAB_KERNEL);
+       struct fuse_req *req = kmem_cache_alloc(fuse_req_cachep, GFP_KERNEL);
+       if (req)
+               fuse_request_init(req);
+       return req;
+}
+
+struct fuse_req *fuse_request_alloc_nofs(void)
+{
+       struct fuse_req *req = kmem_cache_alloc(fuse_req_cachep, GFP_NOFS);
        if (req)
                fuse_request_init(req);
        return req;
@@ -64,20 +73,6 @@ static void restore_sigs(sigset_t *oldset)
        sigprocmask(SIG_SETMASK, oldset, NULL);
 }
 
-/*
- * Reset request, so that it can be reused
- *
- * The caller must be _very_ careful to make sure, that it is holding
- * the only reference to req
- */
-void fuse_reset_request(struct fuse_req *req)
-{
-       int preallocated = req->preallocated;
-       BUG_ON(atomic_read(&req->count) != 1);
-       fuse_request_init(req);
-       req->preallocated = preallocated;
-}
-
 static void __fuse_get_request(struct fuse_req *req)
 {
        atomic_inc(&req->count);
@@ -90,196 +85,128 @@ static void __fuse_put_request(struct fuse_req *req)
        atomic_dec(&req->count);
 }
 
-static struct fuse_req *do_get_request(struct fuse_conn *fc)
+static void fuse_req_init_context(struct fuse_req *req)
 {
-       struct fuse_req *req;
-
-       spin_lock(&fuse_lock);
-       BUG_ON(list_empty(&fc->unused_list));
-       req = list_entry(fc->unused_list.next, struct fuse_req, list);
-       list_del_init(&req->list);
-       spin_unlock(&fuse_lock);
-       fuse_request_init(req);
-       req->preallocated = 1;
        req->in.h.uid = current->fsuid;
        req->in.h.gid = current->fsgid;
        req->in.h.pid = current->pid;
-       return req;
 }
 
-/* This can return NULL, but only in case it's interrupted by a SIGKILL */
-struct fuse_req *fuse_get_request(struct fuse_conn *fc)
+struct fuse_req *fuse_get_req(struct fuse_conn *fc)
 {
-       int intr;
+       struct fuse_req *req;
        sigset_t oldset;
+       int intr;
+       int err;
 
        atomic_inc(&fc->num_waiting);
        block_sigs(&oldset);
-       intr = down_interruptible(&fc->outstanding_sem);
+       intr = wait_event_interruptible(fc->blocked_waitq, !fc->blocked);
        restore_sigs(&oldset);
-       if (intr) {
-               atomic_dec(&fc->num_waiting);
-               return NULL;
-       }
-       return do_get_request(fc);
-}
+       err = -EINTR;
+       if (intr)
+               goto out;
 
-/* Must be called with fuse_lock held */
-static void fuse_putback_request(struct fuse_conn *fc, struct fuse_req *req)
-{
-       if (req->preallocated) {
-               atomic_dec(&fc->num_waiting);
-               list_add(&req->list, &fc->unused_list);
-       } else
-               fuse_request_free(req);
+       err = -ENOTCONN;
+       if (!fc->connected)
+               goto out;
 
-       /* If we are in debt decrease that first */
-       if (fc->outstanding_debt)
-               fc->outstanding_debt--;
-       else
-               up(&fc->outstanding_sem);
-}
+       req = fuse_request_alloc();
+       err = -ENOMEM;
+       if (!req)
+               goto out;
 
-void fuse_put_request(struct fuse_conn *fc, struct fuse_req *req)
-{
-       if (atomic_dec_and_test(&req->count)) {
-               spin_lock(&fuse_lock);
-               fuse_putback_request(fc, req);
-               spin_unlock(&fuse_lock);
-       }
-}
+       fuse_req_init_context(req);
+       req->waiting = 1;
+       return req;
 
-static void fuse_put_request_locked(struct fuse_conn *fc, struct fuse_req *req)
-{
-       if (atomic_dec_and_test(&req->count))
-               fuse_putback_request(fc, req);
+ out:
+       atomic_dec(&fc->num_waiting);
+       return ERR_PTR(err);
 }
 
-void fuse_release_background(struct fuse_req *req)
+/*
+ * Return request in fuse_file->reserved_req.  However that may
+ * currently be in use.  If that is the case, wait for it to become
+ * available.
+ */
+static struct fuse_req *get_reserved_req(struct fuse_conn *fc,
+                                        struct file *file)
 {
-       iput(req->inode);
-       iput(req->inode2);
-       if (req->file)
-               fput(req->file);
-       spin_lock(&fuse_lock);
-       list_del(&req->bg_entry);
-       spin_unlock(&fuse_lock);
+       struct fuse_req *req = NULL;
+       struct fuse_file *ff = file->private_data;
+
+       do {
+               wait_event(fc->reserved_req_waitq, ff->reserved_req);
+               spin_lock(&fc->lock);
+               if (ff->reserved_req) {
+                       req = ff->reserved_req;
+                       ff->reserved_req = NULL;
+                       get_file(file);
+                       req->stolen_file = file;
+               }
+               spin_unlock(&fc->lock);
+       } while (!req);
+
+       return req;
 }
 
 /*
- * This function is called when a request is finished.  Either a reply
- * has arrived or it was interrupted (and not yet sent) or some error
- * occurred during communication with userspace, or the device file
- * was closed.  In case of a background request the reference to the
- * stored objects are released.  The requester thread is woken up (if
- * still waiting), the 'end' callback is called if given, else the
- * reference to the request is released
- *
- * Releasing extra reference for foreground requests must be done
- * within the same locked region as setting state to finished.  This
- * is because fuse_reset_request() may be called after request is
- * finished and it must be the sole possessor.  If request is
- * interrupted and put in the background, it will return with an error
- * and hence never be reset and reused.
- *
- * Called with fuse_lock, unlocks it
+ * Put stolen request back into fuse_file->reserved_req
  */
-static void request_end(struct fuse_conn *fc, struct fuse_req *req)
+static void put_reserved_req(struct fuse_conn *fc, struct fuse_req *req)
 {
-       list_del(&req->list);
-       req->state = FUSE_REQ_FINISHED;
-       if (!req->background) {
-               wake_up(&req->waitq);
-               fuse_put_request_locked(fc, req);
-               spin_unlock(&fuse_lock);
-       } else {
-               void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
-               req->end = NULL;
-               spin_unlock(&fuse_lock);
-               down_read(&fc->sbput_sem);
-               if (fc->mounted)
-                       fuse_release_background(req);
-               up_read(&fc->sbput_sem);
-               if (end)
-                       end(fc, req);
-               else
-                       fuse_put_request(fc, req);
-       }
+       struct file *file = req->stolen_file;
+       struct fuse_file *ff = file->private_data;
+
+       spin_lock(&fc->lock);
+       fuse_request_init(req);
+       BUG_ON(ff->reserved_req);
+       ff->reserved_req = req;
+       wake_up_all(&fc->reserved_req_waitq);
+       spin_unlock(&fc->lock);
+       fput(file);
 }
 
 /*
- * Unfortunately request interruption not just solves the deadlock
- * problem, it causes problems too.  These stem from the fact, that an
- * interrupted request is continued to be processed in userspace,
- * while all the locks and object references (inode and file) held
- * during the operation are released.
- *
- * To release the locks is exactly why there's a need to interrupt the
- * request, so there's not a lot that can be done about this, except
- * introduce additional locking in userspace.
+ * Gets a requests for a file operation, always succeeds
  *
- * More important is to keep inode and file references until userspace
- * has replied, otherwise FORGET and RELEASE could be sent while the
- * inode/file is still used by the filesystem.
+ * This is used for sending the FLUSH request, which must get to
+ * userspace, due to POSIX locks which may need to be unlocked.
  *
- * For this reason the concept of "background" request is introduced.
- * An interrupted request is backgrounded if it has been already sent
- * to userspace.  Backgrounding involves getting an extra reference to
- * inode(s) or file used in the request, and adding the request to
- * fc->background list.  When a reply is received for a background
- * request, the object references are released, and the request is
- * removed from the list.  If the filesystem is unmounted while there
- * are still background requests, the list is walked and references
- * are released as if a reply was received.
+ * If allocation fails due to OOM, use the reserved request in
+ * fuse_file.
  *
- * There's one more use for a background request.  The RELEASE message is
- * always sent as background, since it doesn't return an error or
- * data.
+ * This is very unlikely to deadlock accidentally, since the
+ * filesystem should not have it's own file open.  If deadlock is
+ * intentional, it can still be broken by "aborting" the filesystem.
  */
-static void background_request(struct fuse_conn *fc, struct fuse_req *req)
+struct fuse_req *fuse_get_req_nofail(struct fuse_conn *fc, struct file *file)
 {
-       req->background = 1;
-       list_add(&req->bg_entry, &fc->background);
-       if (req->inode)
-               req->inode = igrab(req->inode);
-       if (req->inode2)
-               req->inode2 = igrab(req->inode2);
-       if (req->file)
-               get_file(req->file);
+       struct fuse_req *req;
+
+       atomic_inc(&fc->num_waiting);
+       wait_event(fc->blocked_waitq, !fc->blocked);
+       req = fuse_request_alloc();
+       if (!req)
+               req = get_reserved_req(fc, file);
+
+       fuse_req_init_context(req);
+       req->waiting = 1;
+       return req;
 }
 
-/* Called with fuse_lock held.  Releases, and then reacquires it. */
-static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
+void fuse_put_request(struct fuse_conn *fc, struct fuse_req *req)
 {
-       sigset_t oldset;
-
-       spin_unlock(&fuse_lock);
-       block_sigs(&oldset);
-       wait_event_interruptible(req->waitq, req->state == FUSE_REQ_FINISHED);
-       restore_sigs(&oldset);
-       spin_lock(&fuse_lock);
-       if (req->state == FUSE_REQ_FINISHED && !req->interrupted)
-               return;
+       if (atomic_dec_and_test(&req->count)) {
+               if (req->waiting)
+                       atomic_dec(&fc->num_waiting);
 
-       if (!req->interrupted) {
-               req->out.h.error = -EINTR;
-               req->interrupted = 1;
-       }
-       if (req->locked) {
-               /* This is uninterruptible sleep, because data is
-                  being copied to/from the buffers of req.  During
-                  locked state, there mustn't be any filesystem
-                  operation (e.g. page fault), since that could lead
-                  to deadlock */
-               spin_unlock(&fuse_lock);
-               wait_event(req->waitq, !req->locked);
-               spin_lock(&fuse_lock);
+               if (req->stolen_file)
+                       put_reserved_req(fc, req);
+               else
+                       fuse_request_free(req);
        }
-       if (req->state == FUSE_REQ_PENDING) {
-               list_del(&req->list);
-               __fuse_put_request(req);
-       } else if (req->state == FUSE_REQ_SENT)
-               background_request(fc, req);
 }
 
 static unsigned len_args(unsigned numargs, struct fuse_arg *args)
@@ -293,38 +220,170 @@ static unsigned len_args(unsigned numargs, struct fuse_arg *args)
        return nbytes;
 }
 
-static void queue_request(struct fuse_conn *fc, struct fuse_req *req)
+static u64 fuse_get_unique(struct fuse_conn *fc)
 {
        fc->reqctr++;
        /* zero is special */
        if (fc->reqctr == 0)
                fc->reqctr = 1;
-       req->in.h.unique = fc->reqctr;
+
+       return fc->reqctr;
+}
+
+static void queue_request(struct fuse_conn *fc, struct fuse_req *req)
+{
+       req->in.h.unique = fuse_get_unique(fc);
        req->in.h.len = sizeof(struct fuse_in_header) +
                len_args(req->in.numargs, (struct fuse_arg *) req->in.args);
-       if (!req->preallocated) {
-               /* If request is not preallocated (either FORGET or
-                  RELEASE), then still decrease outstanding_sem, so
-                  user can't open infinite number of files while not
-                  processing the RELEASE requests.  However for
-                  efficiency do it without blocking, so if down()
-                  would block, just increase the debt instead */
-               if (down_trylock(&fc->outstanding_sem))
-                       fc->outstanding_debt++;
-       }
        list_add_tail(&req->list, &fc->pending);
        req->state = FUSE_REQ_PENDING;
+       if (!req->waiting) {
+               req->waiting = 1;
+               atomic_inc(&fc->num_waiting);
+       }
        wake_up(&fc->waitq);
        kill_fasync(&fc->fasync, SIGIO, POLL_IN);
 }
 
+static void flush_bg_queue(struct fuse_conn *fc)
+{
+       while (fc->active_background < FUSE_MAX_BACKGROUND &&
+              !list_empty(&fc->bg_queue)) {
+               struct fuse_req *req;
+
+               req = list_entry(fc->bg_queue.next, struct fuse_req, list);
+               list_del(&req->list);
+               fc->active_background++;
+               queue_request(fc, req);
+       }
+}
+
 /*
- * This can only be interrupted by a SIGKILL
+ * This function is called when a request is finished.  Either a reply
+ * has arrived or it was aborted (and not yet sent) or some error
+ * occurred during communication with userspace, or the device file
+ * was closed.  The requester thread is woken up (if still waiting),
+ * the 'end' callback is called if given, else the reference to the
+ * request is released
+ *
+ * Called with fc->lock, unlocks it
  */
+static void request_end(struct fuse_conn *fc, struct fuse_req *req)
+       __releases(fc->lock)
+{
+       void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
+       req->end = NULL;
+       list_del(&req->list);
+       list_del(&req->intr_entry);
+       req->state = FUSE_REQ_FINISHED;
+       if (req->background) {
+               if (fc->num_background == FUSE_MAX_BACKGROUND) {
+                       fc->blocked = 0;
+                       wake_up_all(&fc->blocked_waitq);
+               }
+               if (fc->num_background == FUSE_CONGESTION_THRESHOLD) {
+                       clear_bdi_congested(&fc->bdi, READ);
+                       clear_bdi_congested(&fc->bdi, WRITE);
+               }
+               fc->num_background--;
+               fc->active_background--;
+               flush_bg_queue(fc);
+       }
+       spin_unlock(&fc->lock);
+       wake_up(&req->waitq);
+       if (end)
+               end(fc, req);
+       else
+               fuse_put_request(fc, req);
+}
+
+static void wait_answer_interruptible(struct fuse_conn *fc,
+                                     struct fuse_req *req)
+       __releases(fc->lock) __acquires(fc->lock)
+{
+       if (signal_pending(current))
+               return;
+
+       spin_unlock(&fc->lock);
+       wait_event_interruptible(req->waitq, req->state == FUSE_REQ_FINISHED);
+       spin_lock(&fc->lock);
+}
+
+static void queue_interrupt(struct fuse_conn *fc, struct fuse_req *req)
+{
+       list_add_tail(&req->intr_entry, &fc->interrupts);
+       wake_up(&fc->waitq);
+       kill_fasync(&fc->fasync, SIGIO, POLL_IN);
+}
+
+static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
+       __releases(fc->lock) __acquires(fc->lock)
+{
+       if (!fc->no_interrupt) {
+               /* Any signal may interrupt this */
+               wait_answer_interruptible(fc, req);
+
+               if (req->aborted)
+                       goto aborted;
+               if (req->state == FUSE_REQ_FINISHED)
+                       return;
+
+               req->interrupted = 1;
+               if (req->state == FUSE_REQ_SENT)
+                       queue_interrupt(fc, req);
+       }
+
+       if (!req->force) {
+               sigset_t oldset;
+
+               /* Only fatal signals may interrupt this */
+               block_sigs(&oldset);
+               wait_answer_interruptible(fc, req);
+               restore_sigs(&oldset);
+
+               if (req->aborted)
+                       goto aborted;
+               if (req->state == FUSE_REQ_FINISHED)
+                       return;
+
+               /* Request is not yet in userspace, bail out */
+               if (req->state == FUSE_REQ_PENDING) {
+                       list_del(&req->list);
+                       __fuse_put_request(req);
+                       req->out.h.error = -EINTR;
+                       return;
+               }
+       }
+
+       /*
+        * Either request is already in userspace, or it was forced.
+        * Wait it out.
+        */
+       spin_unlock(&fc->lock);
+       wait_event(req->waitq, req->state == FUSE_REQ_FINISHED);
+       spin_lock(&fc->lock);
+
+       if (!req->aborted)
+               return;
+
+ aborted:
+       BUG_ON(req->state != FUSE_REQ_FINISHED);
+       if (req->locked) {
+               /* This is uninterruptible sleep, because data is
+                  being copied to/from the buffers of req.  During
+                  locked state, there mustn't be any filesystem
+                  operation (e.g. page fault), since that could lead
+                  to deadlock */
+               spin_unlock(&fc->lock);
+               wait_event(req->waitq, !req->locked);
+               spin_lock(&fc->lock);
+       }
+}
+
 void request_send(struct fuse_conn *fc, struct fuse_req *req)
 {
        req->isreply = 1;
-       spin_lock(&fuse_lock);
+       spin_lock(&fc->lock);
        if (!fc->connected)
                req->out.h.error = -ENOTCONN;
        else if (fc->conn_error)
@@ -337,15 +396,30 @@ void request_send(struct fuse_conn *fc, struct fuse_req *req)
 
                request_wait_answer(fc, req);
        }
-       spin_unlock(&fuse_lock);
+       spin_unlock(&fc->lock);
+}
+
+static void request_send_nowait_locked(struct fuse_conn *fc,
+                                      struct fuse_req *req)
+{
+       req->background = 1;
+       fc->num_background++;
+       if (fc->num_background == FUSE_MAX_BACKGROUND)
+               fc->blocked = 1;
+       if (fc->num_background == FUSE_CONGESTION_THRESHOLD) {
+               set_bdi_congested(&fc->bdi, READ);
+               set_bdi_congested(&fc->bdi, WRITE);
+       }
+       list_add_tail(&req->list, &fc->bg_queue);
+       flush_bg_queue(fc);
 }
 
 static void request_send_nowait(struct fuse_conn *fc, struct fuse_req *req)
 {
-       spin_lock(&fuse_lock);
+       spin_lock(&fc->lock);
        if (fc->connected) {
-               queue_request(fc, req);
-               spin_unlock(&fuse_lock);
+               request_send_nowait_locked(fc, req);
+               spin_unlock(&fc->lock);
        } else {
                req->out.h.error = -ENOTCONN;
                request_end(fc, req);
@@ -361,48 +435,57 @@ void request_send_noreply(struct fuse_conn *fc, struct fuse_req *req)
 void request_send_background(struct fuse_conn *fc, struct fuse_req *req)
 {
        req->isreply = 1;
-       spin_lock(&fuse_lock);
-       background_request(fc, req);
-       spin_unlock(&fuse_lock);
        request_send_nowait(fc, req);
 }
 
 /*
+ * Called under fc->lock
+ *
+ * fc->connected must have been checked previously
+ */
+void request_send_background_locked(struct fuse_conn *fc, struct fuse_req *req)
+{
+       req->isreply = 1;
+       request_send_nowait_locked(fc, req);
+}
+
+/*
  * Lock the request.  Up to the next unlock_request() there mustn't be
  * anything that could cause a page-fault.  If the request was already
- * interrupted bail out.
+ * aborted bail out.
  */
-static int lock_request(struct fuse_req *req)
+static int lock_request(struct fuse_conn *fc, struct fuse_req *req)
 {
        int err = 0;
        if (req) {
-               spin_lock(&fuse_lock);
-               if (req->interrupted)
+               spin_lock(&fc->lock);
+               if (req->aborted)
                        err = -ENOENT;
                else
                        req->locked = 1;
-               spin_unlock(&fuse_lock);
+               spin_unlock(&fc->lock);
        }
        return err;
 }
 
 /*
- * Unlock request.  If it was interrupted during being locked, the
+ * Unlock request.  If it was aborted during being locked, the
  * requester thread is currently waiting for it to be unlocked, so
  * wake it up.
  */
-static void unlock_request(struct fuse_req *req)
+static void unlock_request(struct fuse_conn *fc, struct fuse_req *req)
 {
        if (req) {
-               spin_lock(&fuse_lock);
+               spin_lock(&fc->lock);
                req->locked = 0;
-               if (req->interrupted)
+               if (req->aborted)
                        wake_up(&req->waitq);
-               spin_unlock(&fuse_lock);
+               spin_unlock(&fc->lock);
        }
 }
 
 struct fuse_copy_state {
+       struct fuse_conn *fc;
        int write;
        struct fuse_req *req;
        const struct iovec *iov;
@@ -415,11 +498,12 @@ struct fuse_copy_state {
        unsigned len;
 };
 
-static void fuse_copy_init(struct fuse_copy_state *cs, int write,
-                          struct fuse_req *req, const struct iovec *iov,
-                          unsigned long nr_segs)
+static void fuse_copy_init(struct fuse_copy_state *cs, struct fuse_conn *fc,
+                          int write, struct fuse_req *req,
+                          const struct iovec *iov, unsigned long nr_segs)
 {
        memset(cs, 0, sizeof(*cs));
+       cs->fc = fc;
        cs->write = write;
        cs->req = req;
        cs->iov = iov;
@@ -449,7 +533,7 @@ static int fuse_copy_fill(struct fuse_copy_state *cs)
        unsigned long offset;
        int err;
 
-       unlock_request(cs->req);
+       unlock_request(cs->fc, cs->req);
        fuse_copy_finish(cs);
        if (!cs->seglen) {
                BUG_ON(!cs->nr_segs);
@@ -472,7 +556,7 @@ static int fuse_copy_fill(struct fuse_copy_state *cs)
        cs->seglen -= cs->len;
        cs->addr += cs->len;
 
-       return lock_request(cs->req);
+       return lock_request(cs->fc, cs->req);
 }
 
 /* Do as much copy to/from userspace buffer as we can */
@@ -573,51 +657,97 @@ static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
        return err;
 }
 
+static int request_pending(struct fuse_conn *fc)
+{
+       return !list_empty(&fc->pending) || !list_empty(&fc->interrupts);
+}
+
 /* Wait until a request is available on the pending list */
 static void request_wait(struct fuse_conn *fc)
 {
        DECLARE_WAITQUEUE(wait, current);
 
        add_wait_queue_exclusive(&fc->waitq, &wait);
-       while (fc->connected && list_empty(&fc->pending)) {
+       while (fc->connected && !request_pending(fc)) {
                set_current_state(TASK_INTERRUPTIBLE);
                if (signal_pending(current))
                        break;
 
-               spin_unlock(&fuse_lock);
+               spin_unlock(&fc->lock);
                schedule();
-               spin_lock(&fuse_lock);
+               spin_lock(&fc->lock);
        }
        set_current_state(TASK_RUNNING);
        remove_wait_queue(&fc->waitq, &wait);
 }
 
 /*
+ * Transfer an interrupt request to userspace
+ *
+ * Unlike other requests this is assembled on demand, without a need
+ * to allocate a separate fuse_req structure.
+ *
+ * Called with fc->lock held, releases it
+ */
+static int fuse_read_interrupt(struct fuse_conn *fc, struct fuse_req *req,
+                              const struct iovec *iov, unsigned long nr_segs)
+       __releases(fc->lock)
+{
+       struct fuse_copy_state cs;
+       struct fuse_in_header ih;
+       struct fuse_interrupt_in arg;
+       unsigned reqsize = sizeof(ih) + sizeof(arg);
+       int err;
+
+       list_del_init(&req->intr_entry);
+       req->intr_unique = fuse_get_unique(fc);
+       memset(&ih, 0, sizeof(ih));
+       memset(&arg, 0, sizeof(arg));
+       ih.len = reqsize;
+       ih.opcode = FUSE_INTERRUPT;
+       ih.unique = req->intr_unique;
+       arg.unique = req->in.h.unique;
+
+       spin_unlock(&fc->lock);
+       if (iov_length(iov, nr_segs) < reqsize)
+               return -EINVAL;
+
+       fuse_copy_init(&cs, fc, 1, NULL, iov, nr_segs);
+       err = fuse_copy_one(&cs, &ih, sizeof(ih));
+       if (!err)
+               err = fuse_copy_one(&cs, &arg, sizeof(arg));
+       fuse_copy_finish(&cs);
+
+       return err ? err : reqsize;
+}
+
+/*
  * Read a single request into the userspace filesystem's buffer.  This
  * function waits until a request is available, then removes it from
  * the pending list and copies request data to userspace buffer.  If
- * no reply is needed (FORGET) or request has been interrupted or
- * there was an error during the copying then it's finished by calling
+ * no reply is needed (FORGET) or request has been aborted or there
+ * was an error during the copying then it's finished by calling
  * request_end().  Otherwise add it to the processing list, and set
  * the 'sent' flag.
  */
-static ssize_t fuse_dev_readv(struct file *file, const struct iovec *iov,
-                             unsigned long nr_segs, loff_t *off)
+static ssize_t fuse_dev_read(struct kiocb *iocb, const struct iovec *iov,
+                             unsigned long nr_segs, loff_t pos)
 {
        int err;
        struct fuse_req *req;
        struct fuse_in *in;
        struct fuse_copy_state cs;
        unsigned reqsize;
+       struct file *file = iocb->ki_filp;
        struct fuse_conn *fc = fuse_get_conn(file);
        if (!fc)
                return -EPERM;
 
  restart:
-       spin_lock(&fuse_lock);
+       spin_lock(&fc->lock);
        err = -EAGAIN;
        if ((file->f_flags & O_NONBLOCK) && fc->connected &&
-           list_empty(&fc->pending))
+           !request_pending(fc))
                goto err_unlock;
 
        request_wait(fc);
@@ -625,9 +755,15 @@ static ssize_t fuse_dev_readv(struct file *file, const struct iovec *iov,
        if (!fc->connected)
                goto err_unlock;
        err = -ERESTARTSYS;
-       if (list_empty(&fc->pending))
+       if (!request_pending(fc))
                goto err_unlock;
 
+       if (!list_empty(&fc->interrupts)) {
+               req = list_entry(fc->interrupts.next, struct fuse_req,
+                                intr_entry);
+               return fuse_read_interrupt(fc, req, iov, nr_segs);
+       }
+
        req = list_entry(fc->pending.next, struct fuse_req, list);
        req->state = FUSE_REQ_READING;
        list_move(&req->list, &fc->io);
@@ -643,20 +779,21 @@ static ssize_t fuse_dev_readv(struct file *file, const struct iovec *iov,
                request_end(fc, req);
                goto restart;
        }
-       spin_unlock(&fuse_lock);
-       fuse_copy_init(&cs, 1, req, iov, nr_segs);
+       spin_unlock(&fc->lock);
+       fuse_copy_init(&cs, fc, 1, req, iov, nr_segs);
        err = fuse_copy_one(&cs, &in->h, sizeof(in->h));
        if (!err)
                err = fuse_copy_args(&cs, in->numargs, in->argpages,
                                     (struct fuse_arg *) in->args, 0);
        fuse_copy_finish(&cs);
-       spin_lock(&fuse_lock);
+       spin_lock(&fc->lock);
        req->locked = 0;
-       if (!err && req->interrupted)
-               err = -ENOENT;
+       if (req->aborted) {
+               request_end(fc, req);
+               return -ENODEV;
+       }
        if (err) {
-               if (!req->interrupted)
-                       req->out.h.error = -EIO;
+               req->out.h.error = -EIO;
                request_end(fc, req);
                return err;
        }
@@ -665,24 +802,17 @@ static ssize_t fuse_dev_readv(struct file *file, const struct iovec *iov,
        else {
                req->state = FUSE_REQ_SENT;
                list_move_tail(&req->list, &fc->processing);
-               spin_unlock(&fuse_lock);
+               if (req->interrupted)
+                       queue_interrupt(fc, req);
+               spin_unlock(&fc->lock);
        }
        return reqsize;
 
  err_unlock:
-       spin_unlock(&fuse_lock);
+       spin_unlock(&fc->lock);
        return err;
 }
 
-static ssize_t fuse_dev_read(struct file *file, char __user *buf,
-                            size_t nbytes, loff_t *off)
-{
-       struct iovec iov;
-       iov.iov_len = nbytes;
-       iov.iov_base = buf;
-       return fuse_dev_readv(file, &iov, 1, off);
-}
-
 /* Look up request on processing list by unique ID */
 static struct fuse_req *request_find(struct fuse_conn *fc, u64 unique)
 {
@@ -691,7 +821,7 @@ static struct fuse_req *request_find(struct fuse_conn *fc, u64 unique)
        list_for_each(entry, &fc->processing) {
                struct fuse_req *req;
                req = list_entry(entry, struct fuse_req, list);
-               if (req->in.h.unique == unique)
+               if (req->in.h.unique == unique || req->intr_unique == unique)
                        return req;
        }
        return NULL;
@@ -727,19 +857,19 @@ static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
  * it from the list and copy the rest of the buffer to the request.
  * The request is finished by calling request_end()
  */
-static ssize_t fuse_dev_writev(struct file *file, const struct iovec *iov,
-                              unsigned long nr_segs, loff_t *off)
+static ssize_t fuse_dev_write(struct kiocb *iocb, const struct iovec *iov,
+                              unsigned long nr_segs, loff_t pos)
 {
        int err;
        unsigned nbytes = iov_length(iov, nr_segs);
        struct fuse_req *req;
        struct fuse_out_header oh;
        struct fuse_copy_state cs;
-       struct fuse_conn *fc = fuse_get_conn(file);
+       struct fuse_conn *fc = fuse_get_conn(iocb->ki_filp);
        if (!fc)
-               return -ENODEV;
+               return -EPERM;
 
-       fuse_copy_init(&cs, 0, NULL, iov, nr_segs);
+       fuse_copy_init(&cs, fc, 0, NULL, iov, nr_segs);
        if (nbytes < sizeof(struct fuse_out_header))
                return -EINVAL;
 
@@ -751,59 +881,66 @@ static ssize_t fuse_dev_writev(struct file *file, const struct iovec *iov,
            oh.len != nbytes)
                goto err_finish;
 
-       spin_lock(&fuse_lock);
+       spin_lock(&fc->lock);
        err = -ENOENT;
        if (!fc->connected)
                goto err_unlock;
 
        req = request_find(fc, oh.unique);
-       err = -EINVAL;
        if (!req)
                goto err_unlock;
 
-       if (req->interrupted) {
-               spin_unlock(&fuse_lock);
+       if (req->aborted) {
+               spin_unlock(&fc->lock);
                fuse_copy_finish(&cs);
-               spin_lock(&fuse_lock);
+               spin_lock(&fc->lock);
                request_end(fc, req);
                return -ENOENT;
        }
+       /* Is it an interrupt reply? */
+       if (req->intr_unique == oh.unique) {
+               err = -EINVAL;
+               if (nbytes != sizeof(struct fuse_out_header))
+                       goto err_unlock;
+
+               if (oh.error == -ENOSYS)
+                       fc->no_interrupt = 1;
+               else if (oh.error == -EAGAIN)
+                       queue_interrupt(fc, req);
+
+               spin_unlock(&fc->lock);
+               fuse_copy_finish(&cs);
+               return nbytes;
+       }
+
+       req->state = FUSE_REQ_WRITING;
        list_move(&req->list, &fc->io);
        req->out.h = oh;
        req->locked = 1;
        cs.req = req;
-       spin_unlock(&fuse_lock);
+       spin_unlock(&fc->lock);
 
        err = copy_out_args(&cs, &req->out, nbytes);
        fuse_copy_finish(&cs);
 
-       spin_lock(&fuse_lock);
+       spin_lock(&fc->lock);
        req->locked = 0;
        if (!err) {
-               if (req->interrupted)
+               if (req->aborted)
                        err = -ENOENT;
-       } else if (!req->interrupted)
+       } else if (!req->aborted)
                req->out.h.error = -EIO;
        request_end(fc, req);
 
        return err ? err : nbytes;
 
  err_unlock:
-       spin_unlock(&fuse_lock);
+       spin_unlock(&fc->lock);
  err_finish:
        fuse_copy_finish(&cs);
        return err;
 }
 
-static ssize_t fuse_dev_write(struct file *file, const char __user *buf,
-                             size_t nbytes, loff_t *off)
-{
-       struct iovec iov;
-       iov.iov_len = nbytes;
-       iov.iov_base = (char __user *) buf;
-       return fuse_dev_writev(file, &iov, 1, off);
-}
-
 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
 {
        unsigned mask = POLLOUT | POLLWRNORM;
@@ -813,12 +950,12 @@ static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
 
        poll_wait(file, &fc->waitq, wait);
 
-       spin_lock(&fuse_lock);
+       spin_lock(&fc->lock);
        if (!fc->connected)
                mask = POLLERR;
-       else if (!list_empty(&fc->pending))
+       else if (request_pending(fc))
                mask |= POLLIN | POLLRDNORM;
-       spin_unlock(&fuse_lock);
+       spin_unlock(&fc->lock);
 
        return mask;
 }
@@ -826,7 +963,7 @@ static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
 /*
  * Abort all requests on the given list (pending or processing)
  *
- * This function releases and reacquires fuse_lock
+ * This function releases and reacquires fc->lock
  */
 static void end_requests(struct fuse_conn *fc, struct list_head *head)
 {
@@ -835,14 +972,14 @@ static void end_requests(struct fuse_conn *fc, struct list_head *head)
                req = list_entry(head->next, struct fuse_req, list);
                req->out.h.error = -ECONNABORTED;
                request_end(fc, req);
-               spin_lock(&fuse_lock);
+               spin_lock(&fc->lock);
        }
 }
 
 /*
  * Abort requests under I/O
  *
- * The requests are set to interrupted and finished, and the request
+ * The requests are set to aborted and finished, and the request
  * waiter is woken up.  This will make request_wait_answer() wait
  * until the request is unlocked and then return.
  *
@@ -851,13 +988,14 @@ static void end_requests(struct fuse_conn *fc, struct list_head *head)
  * locked).
  */
 static void end_io_requests(struct fuse_conn *fc)
+       __releases(fc->lock) __acquires(fc->lock)
 {
        while (!list_empty(&fc->io)) {
                struct fuse_req *req =
                        list_entry(fc->io.next, struct fuse_req, list);
                void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
 
-               req->interrupted = 1;
+               req->aborted = 1;
                req->out.h.error = -ECONNABORTED;
                req->state = FUSE_REQ_FINISHED;
                list_del_init(&req->list);
@@ -866,10 +1004,10 @@ static void end_io_requests(struct fuse_conn *fc)
                        req->end = NULL;
                        /* The end function will consume this reference */
                        __fuse_get_request(req);
-                       spin_unlock(&fuse_lock);
+                       spin_unlock(&fc->lock);
                        wait_event(req->waitq, !req->locked);
                        end(fc, req);
-                       spin_lock(&fuse_lock);
+                       spin_lock(&fc->lock);
                }
        }
 }
@@ -890,35 +1028,36 @@ static void end_io_requests(struct fuse_conn *fc)
  * onto the pending list is prevented by req->connected being false.
  *
  * Progression of requests under I/O to the processing list is
- * prevented by the req->interrupted flag being true for these
- * requests.  For this reason requests on the io list must be aborted
- * first.
+ * prevented by the req->aborted flag being true for these requests.
+ * For this reason requests on the io list must be aborted first.
  */
 void fuse_abort_conn(struct fuse_conn *fc)
 {
-       spin_lock(&fuse_lock);
+       spin_lock(&fc->lock);
        if (fc->connected) {
                fc->connected = 0;
+               fc->blocked = 0;
                end_io_requests(fc);
                end_requests(fc, &fc->pending);
                end_requests(fc, &fc->processing);
                wake_up_all(&fc->waitq);
+               wake_up_all(&fc->blocked_waitq);
                kill_fasync(&fc->fasync, SIGIO, POLL_IN);
        }
-       spin_unlock(&fuse_lock);
+       spin_unlock(&fc->lock);
 }
 
 static int fuse_dev_release(struct inode *inode, struct file *file)
 {
        struct fuse_conn *fc = fuse_get_conn(file);
        if (fc) {
-               spin_lock(&fuse_lock);
+               spin_lock(&fc->lock);
                fc->connected = 0;
                end_requests(fc, &fc->pending);
                end_requests(fc, &fc->processing);
-               spin_unlock(&fuse_lock);
+               spin_unlock(&fc->lock);
                fasync_helper(-1, file, 0, &fc->fasync);
-               kobject_put(&fc->kobj);
+               fuse_conn_put(fc);
        }
 
        return 0;
@@ -928,7 +1067,7 @@ static int fuse_dev_fasync(int fd, struct file *file, int on)
 {
        struct fuse_conn *fc = fuse_get_conn(file);
        if (!fc)
-               return -ENODEV;
+               return -EPERM;
 
        /* No locking - fasync_helper does its own locking */
        return fasync_helper(fd, file, on, &fc->fasync);
@@ -937,10 +1076,10 @@ static int fuse_dev_fasync(int fd, struct file *file, int on)
 const struct file_operations fuse_dev_operations = {
        .owner          = THIS_MODULE,
        .llseek         = no_llseek,
-       .read           = fuse_dev_read,
-       .readv          = fuse_dev_readv,
-       .write          = fuse_dev_write,
-       .writev         = fuse_dev_writev,
+       .read           = do_sync_read,
+       .aio_read       = fuse_dev_read,
+       .write          = do_sync_write,
+       .aio_write      = fuse_dev_write,
        .poll           = fuse_dev_poll,
        .release        = fuse_dev_release,
        .fasync         = fuse_dev_fasync,
@@ -957,7 +1096,7 @@ int __init fuse_dev_init(void)
        int err = -ENOMEM;
        fuse_req_cachep = kmem_cache_create("fuse_request",
                                            sizeof(struct fuse_req),
-                                           0, 0, NULL, NULL);
+                                           0, 0, NULL);
        if (!fuse_req_cachep)
                goto out;