- 2.6.17 port work build breaks, but the patch set is relativly stable
[linux-flexiantxendom0-3.2.10.git] / fs / fuse / dev.c
index 0c9a2ee..104a62d 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.
@@ -23,13 +23,11 @@ static kmem_cache_t *fuse_req_cachep;
 
 static struct fuse_conn *fuse_get_conn(struct file *file)
 {
-       struct fuse_conn *fc;
-       spin_lock(&fuse_lock);
-       fc = file->private_data;
-       if (fc && !fc->connected)
-               fc = NULL;
-       spin_unlock(&fuse_lock);
-       return fc;
+       /*
+        * Lockless access is OK, because file->private data is set
+        * once during mount and is valid until the file is released.
+        */
+       return file->private_data;
 }
 
 static void fuse_request_init(struct fuse_req *req)
@@ -74,10 +72,8 @@ static void restore_sigs(sigset_t *oldset)
  */
 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)
@@ -92,80 +88,64 @@ static void __fuse_put_request(struct fuse_req *req)
        atomic_dec(&req->count);
 }
 
-static struct fuse_req *do_get_request(struct fuse_conn *fc)
+struct fuse_req *fuse_get_req(struct fuse_conn *fc)
 {
        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)
-{
-       int intr;
        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);
+       req = fuse_request_alloc();
+       err = -ENOMEM;
+       if (!req)
+               goto out;
 
-       /* If we are in debt decrease that first */
-       if (fc->outstanding_debt)
-               fc->outstanding_debt--;
-       else
-               up(&fc->outstanding_sem);
+       req->in.h.uid = current->fsuid;
+       req->in.h.gid = current->fsgid;
+       req->in.h.pid = current->pid;
+       req->waiting = 1;
+       return req;
+
+ out:
+       atomic_dec(&fc->num_waiting);
+       return ERR_PTR(err);
 }
 
 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);
+               if (req->waiting)
+                       atomic_dec(&fc->num_waiting);
+               fuse_request_free(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);
-}
-
-void fuse_release_background(struct fuse_req *req)
+/*
+ * Called with sbput_sem held for read (request_end) or write
+ * (fuse_put_super).  By the time fuse_put_super() is finished, all
+ * inodes belonging to background requests must be released, so the
+ * iputs have to be done within the locked region.
+ */
+void fuse_release_background(struct fuse_conn *fc, struct fuse_req *req)
 {
        iput(req->inode);
        iput(req->inode2);
-       if (req->file)
-               fput(req->file);
-       spin_lock(&fuse_lock);
+       spin_lock(&fc->lock);
        list_del(&req->bg_entry);
-       spin_unlock(&fuse_lock);
+       if (fc->num_background == FUSE_MAX_BACKGROUND) {
+               fc->blocked = 0;
+               wake_up_all(&fc->blocked_waitq);
+       }
+       fc->num_background--;
+       spin_unlock(&fc->lock);
 }
 
 /*
@@ -184,24 +164,29 @@ void fuse_release_background(struct fuse_req *req)
  * 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
+ * Called with fc->lock, unlocks it
  */
 static void request_end(struct fuse_conn *fc, struct fuse_req *req)
 {
        list_del(&req->list);
        req->state = FUSE_REQ_FINISHED;
        if (!req->background) {
+               spin_unlock(&fc->lock);
                wake_up(&req->waitq);
-               fuse_put_request_locked(fc, req);
-               spin_unlock(&fuse_lock);
+               fuse_put_request(fc, req);
        } else {
                void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
                req->end = NULL;
-               spin_unlock(&fuse_lock);
+               spin_unlock(&fc->lock);
                down_read(&fc->sbput_sem);
                if (fc->mounted)
-                       fuse_release_background(req);
+                       fuse_release_background(fc, req);
                up_read(&fc->sbput_sem);
+
+               /* fput must go outside sbput_sem, otherwise it can deadlock */
+               if (req->file)
+                       fput(req->file);
+
                if (end)
                        end(fc, req);
                else
@@ -242,6 +227,9 @@ static void background_request(struct fuse_conn *fc, struct fuse_req *req)
 {
        req->background = 1;
        list_add(&req->bg_entry, &fc->background);
+       fc->num_background++;
+       if (fc->num_background == FUSE_MAX_BACKGROUND)
+               fc->blocked = 1;
        if (req->inode)
                req->inode = igrab(req->inode);
        if (req->inode2)
@@ -250,16 +238,16 @@ static void background_request(struct fuse_conn *fc, struct fuse_req *req)
                get_file(req->file);
 }
 
-/* Called with fuse_lock held.  Releases, and then reacquires it. */
+/* Called with fc->lock held.  Releases, and then reacquires it. */
 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
 {
        sigset_t oldset;
 
-       spin_unlock(&fuse_lock);
+       spin_unlock(&fc->lock);
        block_sigs(&oldset);
        wait_event_interruptible(req->waitq, req->state == FUSE_REQ_FINISHED);
        restore_sigs(&oldset);
-       spin_lock(&fuse_lock);
+       spin_lock(&fc->lock);
        if (req->state == FUSE_REQ_FINISHED && !req->interrupted)
                return;
 
@@ -273,9 +261,9 @@ static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
                   locked state, there mustn't be any filesystem
                   operation (e.g. page fault), since that could lead
                   to deadlock */
-               spin_unlock(&fuse_lock);
+               spin_unlock(&fc->lock);
                wait_event(req->waitq, !req->locked);
-               spin_lock(&fuse_lock);
+               spin_lock(&fc->lock);
        }
        if (req->state == FUSE_REQ_PENDING) {
                list_del(&req->list);
@@ -304,19 +292,14 @@ static void queue_request(struct fuse_conn *fc, struct fuse_req *req)
        req->in.h.unique = fc->reqctr;
        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);
 }
 
 /*
@@ -325,7 +308,7 @@ static void queue_request(struct fuse_conn *fc, struct fuse_req *req)
 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)
@@ -338,15 +321,16 @@ 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(struct fuse_conn *fc, struct fuse_req *req)
 {
-       spin_lock(&fuse_lock);
+       spin_lock(&fc->lock);
+       background_request(fc, req);
        if (fc->connected) {
                queue_request(fc, req);
-               spin_unlock(&fuse_lock);
+               spin_unlock(&fc->lock);
        } else {
                req->out.h.error = -ENOTCONN;
                request_end(fc, req);
@@ -362,9 +346,6 @@ 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);
 }
 
@@ -373,16 +354,16 @@ void request_send_background(struct fuse_conn *fc, struct fuse_req *req)
  * anything that could cause a page-fault.  If the request was already
  * interrupted 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);
+               spin_lock(&fc->lock);
                if (req->interrupted)
                        err = -ENOENT;
                else
                        req->locked = 1;
-               spin_unlock(&fuse_lock);
+               spin_unlock(&fc->lock);
        }
        return err;
 }
@@ -392,18 +373,19 @@ static int lock_request(struct fuse_req *req)
  * 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)
                        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;
@@ -416,11 +398,12 @@ 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;
@@ -450,7 +433,7 @@ 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);
@@ -473,7 +456,7 @@ 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 */
@@ -585,9 +568,9 @@ static void request_wait(struct fuse_conn *fc)
                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);
@@ -606,18 +589,21 @@ static ssize_t fuse_dev_readv(struct file *file, const struct iovec *iov,
                              unsigned long nr_segs, loff_t *off)
 {
        int err;
-       struct fuse_conn *fc;
        struct fuse_req *req;
        struct fuse_in *in;
        struct fuse_copy_state cs;
        unsigned reqsize;
+       struct fuse_conn *fc = fuse_get_conn(file);
+       if (!fc)
+               return -EPERM;
 
  restart:
-       spin_lock(&fuse_lock);
-       fc = file->private_data;
-       err = -EPERM;
-       if (!fc)
+       spin_lock(&fc->lock);
+       err = -EAGAIN;
+       if ((file->f_flags & O_NONBLOCK) && fc->connected &&
+           list_empty(&fc->pending))
                goto err_unlock;
+
        request_wait(fc);
        err = -ENODEV;
        if (!fc->connected)
@@ -641,14 +627,14 @@ 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;
@@ -663,12 +649,12 @@ 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);
+               spin_unlock(&fc->lock);
        }
        return reqsize;
 
  err_unlock:
-       spin_unlock(&fuse_lock);
+       spin_unlock(&fc->lock);
        return err;
 }
 
@@ -735,9 +721,9 @@ static ssize_t fuse_dev_writev(struct file *file, const struct iovec *iov,
        struct fuse_copy_state cs;
        struct fuse_conn *fc = fuse_get_conn(file);
        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;
 
@@ -749,7 +735,7 @@ 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;
@@ -760,9 +746,9 @@ static ssize_t fuse_dev_writev(struct file *file, const struct iovec *iov,
                goto err_unlock;
 
        if (req->interrupted) {
-               spin_unlock(&fuse_lock);
+               spin_unlock(&fc->lock);
                fuse_copy_finish(&cs);
-               spin_lock(&fuse_lock);
+               spin_lock(&fc->lock);
                request_end(fc, req);
                return -ENOENT;
        }
@@ -770,12 +756,12 @@ static ssize_t fuse_dev_writev(struct file *file, const struct iovec *iov,
        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)
@@ -787,7 +773,7 @@ static ssize_t fuse_dev_writev(struct file *file, const struct iovec *iov,
        return err ? err : nbytes;
 
  err_unlock:
-       spin_unlock(&fuse_lock);
+       spin_unlock(&fc->lock);
  err_finish:
        fuse_copy_finish(&cs);
        return err;
@@ -804,18 +790,19 @@ static ssize_t fuse_dev_write(struct file *file, const char __user *buf,
 
 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
 {
-       struct fuse_conn *fc = fuse_get_conn(file);
        unsigned mask = POLLOUT | POLLWRNORM;
-
+       struct fuse_conn *fc = fuse_get_conn(file);
        if (!fc)
-               return -ENODEV;
+               return POLLERR;
 
        poll_wait(file, &fc->waitq, wait);
 
-       spin_lock(&fuse_lock);
-       if (!list_empty(&fc->pending))
-                mask |= POLLIN | POLLRDNORM;
-       spin_unlock(&fuse_lock);
+       spin_lock(&fc->lock);
+       if (!fc->connected)
+               mask = POLLERR;
+       else if (!list_empty(&fc->pending))
+               mask |= POLLIN | POLLRDNORM;
+       spin_unlock(&fc->lock);
 
        return mask;
 }
@@ -823,7 +810,7 @@ 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)
 {
@@ -832,7 +819,7 @@ 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);
        }
 }
 
@@ -863,10 +850,10 @@ 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);
                }
        }
 }
@@ -893,36 +880,45 @@ static void end_io_requests(struct fuse_conn *fc)
  */
 void fuse_abort_conn(struct fuse_conn *fc)
 {
-       spin_lock(&fuse_lock);
+       spin_lock(&fc->lock);
        if (fc->connected) {
                fc->connected = 0;
                end_io_requests(fc);
                end_requests(fc, &fc->pending);
                end_requests(fc, &fc->processing);
                wake_up_all(&fc->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;
-
-       spin_lock(&fuse_lock);
-       fc = file->private_data;
+       struct fuse_conn *fc = fuse_get_conn(file);
        if (fc) {
+               spin_lock(&fc->lock);
                fc->connected = 0;
                end_requests(fc, &fc->pending);
                end_requests(fc, &fc->processing);
-       }
-       spin_unlock(&fuse_lock);
-       if (fc)
+               spin_unlock(&fc->lock);
+               fasync_helper(-1, file, 0, &fc->fasync);
                kobject_put(&fc->kobj);
+       }
 
        return 0;
 }
 
-struct file_operations fuse_dev_operations = {
+static int fuse_dev_fasync(int fd, struct file *file, int on)
+{
+       struct fuse_conn *fc = fuse_get_conn(file);
+       if (!fc)
+               return -EPERM;
+
+       /* No locking - fasync_helper does its own locking */
+       return fasync_helper(fd, file, on, &fc->fasync);
+}
+
+const struct file_operations fuse_dev_operations = {
        .owner          = THIS_MODULE,
        .llseek         = no_llseek,
        .read           = fuse_dev_read,
@@ -931,6 +927,7 @@ struct file_operations fuse_dev_operations = {
        .writev         = fuse_dev_writev,
        .poll           = fuse_dev_poll,
        .release        = fuse_dev_release,
+       .fasync         = fuse_dev_fasync,
 };
 
 static struct miscdevice fuse_miscdevice = {