fuse: implement ioctl support
authorTejun Heo <tj@kernel.org>
Wed, 26 Nov 2008 11:03:55 +0000 (12:03 +0100)
committerMiklos Szeredi <miklos@szeredi.hu>
Wed, 26 Nov 2008 11:03:55 +0000 (12:03 +0100)
Generic ioctl support is tricky to implement because only the ioctl
implementation itself knows which memory regions need to be read
and/or written.  To support this, fuse client can request retry of
ioctl specifying memory regions to read and write.  Deep copying
(nested pointers) can be implemented by retrying multiple times
resolving one depth of dereference at a time.

For security and cleanliness considerations, ioctl implementation has
restricted mode where the kernel determines data transfer directions
and sizes using the _IOC_*() macros on the ioctl command.  In this
mode, retry is not allowed.

For all FUSE servers, restricted mode is enforced.  Unrestricted ioctl
will be used by CUSE.

Plese read the comment on top of fs/fuse/file.c::fuse_file_do_ioctl()
for more information.

Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Miklos Szeredi <mszeredi@suse.cz>
fs/fuse/file.c
include/linux/fuse.h

index 61726980391340b1252c17d339cd10dc095d1b1b..baed06ea76227324b1778433b08e7f2675dfa0c2 100644 (file)
@@ -1469,6 +1469,282 @@ static loff_t fuse_file_llseek(struct file *file, loff_t offset, int origin)
        return retval;
 }
 
+static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
+                       unsigned int nr_segs, size_t bytes, bool to_user)
+{
+       struct iov_iter ii;
+       int page_idx = 0;
+
+       if (!bytes)
+               return 0;
+
+       iov_iter_init(&ii, iov, nr_segs, bytes, 0);
+
+       while (iov_iter_count(&ii)) {
+               struct page *page = pages[page_idx++];
+               size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
+               void *kaddr, *map;
+
+               kaddr = map = kmap(page);
+
+               while (todo) {
+                       char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
+                       size_t iov_len = ii.iov->iov_len - ii.iov_offset;
+                       size_t copy = min(todo, iov_len);
+                       size_t left;
+
+                       if (!to_user)
+                               left = copy_from_user(kaddr, uaddr, copy);
+                       else
+                               left = copy_to_user(uaddr, kaddr, copy);
+
+                       if (unlikely(left))
+                               return -EFAULT;
+
+                       iov_iter_advance(&ii, copy);
+                       todo -= copy;
+                       kaddr += copy;
+               }
+
+               kunmap(map);
+       }
+
+       return 0;
+}
+
+/*
+ * For ioctls, there is no generic way to determine how much memory
+ * needs to be read and/or written.  Furthermore, ioctls are allowed
+ * to dereference the passed pointer, so the parameter requires deep
+ * copying but FUSE has no idea whatsoever about what to copy in or
+ * out.
+ *
+ * This is solved by allowing FUSE server to retry ioctl with
+ * necessary in/out iovecs.  Let's assume the ioctl implementation
+ * needs to read in the following structure.
+ *
+ * struct a {
+ *     char    *buf;
+ *     size_t  buflen;
+ * }
+ *
+ * On the first callout to FUSE server, inarg->in_size and
+ * inarg->out_size will be NULL; then, the server completes the ioctl
+ * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
+ * the actual iov array to
+ *
+ * { { .iov_base = inarg.arg,  .iov_len = sizeof(struct a) } }
+ *
+ * which tells FUSE to copy in the requested area and retry the ioctl.
+ * On the second round, the server has access to the structure and
+ * from that it can tell what to look for next, so on the invocation,
+ * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
+ *
+ * { { .iov_base = inarg.arg,  .iov_len = sizeof(struct a)     },
+ *   { .iov_base = a.buf,      .iov_len = a.buflen             } }
+ *
+ * FUSE will copy both struct a and the pointed buffer from the
+ * process doing the ioctl and retry ioctl with both struct a and the
+ * buffer.
+ *
+ * This time, FUSE server has everything it needs and completes ioctl
+ * without FUSE_IOCTL_RETRY which finishes the ioctl call.
+ *
+ * Copying data out works the same way.
+ *
+ * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
+ * automatically initializes in and out iovs by decoding @cmd with
+ * _IOC_* macros and the server is not allowed to request RETRY.  This
+ * limits ioctl data transfers to well-formed ioctls and is the forced
+ * behavior for all FUSE servers.
+ */
+static long fuse_file_do_ioctl(struct file *file, unsigned int cmd,
+                              unsigned long arg, unsigned int flags)
+{
+       struct inode *inode = file->f_dentry->d_inode;
+       struct fuse_file *ff = file->private_data;
+       struct fuse_conn *fc = get_fuse_conn(inode);
+       struct fuse_ioctl_in inarg = {
+               .fh = ff->fh,
+               .cmd = cmd,
+               .arg = arg,
+               .flags = flags
+       };
+       struct fuse_ioctl_out outarg;
+       struct fuse_req *req = NULL;
+       struct page **pages = NULL;
+       struct page *iov_page = NULL;
+       struct iovec *in_iov = NULL, *out_iov = NULL;
+       unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
+       size_t in_size, out_size, transferred;
+       int err;
+
+       /* assume all the iovs returned by client always fits in a page */
+       BUILD_BUG_ON(sizeof(struct iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
+
+       if (!fuse_allow_task(fc, current))
+               return -EACCES;
+
+       err = -EIO;
+       if (is_bad_inode(inode))
+               goto out;
+
+       err = -ENOMEM;
+       pages = kzalloc(sizeof(pages[0]) * FUSE_MAX_PAGES_PER_REQ, GFP_KERNEL);
+       iov_page = alloc_page(GFP_KERNEL);
+       if (!pages || !iov_page)
+               goto out;
+
+       /*
+        * If restricted, initialize IO parameters as encoded in @cmd.
+        * RETRY from server is not allowed.
+        */
+       if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
+               struct iovec *iov = page_address(iov_page);
+
+               iov->iov_base = (void *)arg;
+               iov->iov_len = _IOC_SIZE(cmd);
+
+               if (_IOC_DIR(cmd) & _IOC_WRITE) {
+                       in_iov = iov;
+                       in_iovs = 1;
+               }
+
+               if (_IOC_DIR(cmd) & _IOC_READ) {
+                       out_iov = iov;
+                       out_iovs = 1;
+               }
+       }
+
+ retry:
+       inarg.in_size = in_size = iov_length(in_iov, in_iovs);
+       inarg.out_size = out_size = iov_length(out_iov, out_iovs);
+
+       /*
+        * Out data can be used either for actual out data or iovs,
+        * make sure there always is at least one page.
+        */
+       out_size = max_t(size_t, out_size, PAGE_SIZE);
+       max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
+
+       /* make sure there are enough buffer pages and init request with them */
+       err = -ENOMEM;
+       if (max_pages > FUSE_MAX_PAGES_PER_REQ)
+               goto out;
+       while (num_pages < max_pages) {
+               pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
+               if (!pages[num_pages])
+                       goto out;
+               num_pages++;
+       }
+
+       req = fuse_get_req(fc);
+       if (IS_ERR(req)) {
+               err = PTR_ERR(req);
+               req = NULL;
+               goto out;
+       }
+       memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
+       req->num_pages = num_pages;
+
+       /* okay, let's send it to the client */
+       req->in.h.opcode = FUSE_IOCTL;
+       req->in.h.nodeid = get_node_id(inode);
+       req->in.numargs = 1;
+       req->in.args[0].size = sizeof(inarg);
+       req->in.args[0].value = &inarg;
+       if (in_size) {
+               req->in.numargs++;
+               req->in.args[1].size = in_size;
+               req->in.argpages = 1;
+
+               err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
+                                          false);
+               if (err)
+                       goto out;
+       }
+
+       req->out.numargs = 2;
+       req->out.args[0].size = sizeof(outarg);
+       req->out.args[0].value = &outarg;
+       req->out.args[1].size = out_size;
+       req->out.argpages = 1;
+       req->out.argvar = 1;
+
+       request_send(fc, req);
+       err = req->out.h.error;
+       transferred = req->out.args[1].size;
+       fuse_put_request(fc, req);
+       req = NULL;
+       if (err)
+               goto out;
+
+       /* did it ask for retry? */
+       if (outarg.flags & FUSE_IOCTL_RETRY) {
+               char *vaddr;
+
+               /* no retry if in restricted mode */
+               err = -EIO;
+               if (!(flags & FUSE_IOCTL_UNRESTRICTED))
+                       goto out;
+
+               in_iovs = outarg.in_iovs;
+               out_iovs = outarg.out_iovs;
+
+               /*
+                * Make sure things are in boundary, separate checks
+                * are to protect against overflow.
+                */
+               err = -ENOMEM;
+               if (in_iovs > FUSE_IOCTL_MAX_IOV ||
+                   out_iovs > FUSE_IOCTL_MAX_IOV ||
+                   in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
+                       goto out;
+
+               err = -EIO;
+               if ((in_iovs + out_iovs) * sizeof(struct iovec) != transferred)
+                       goto out;
+
+               /* okay, copy in iovs and retry */
+               vaddr = kmap_atomic(pages[0], KM_USER0);
+               memcpy(page_address(iov_page), vaddr, transferred);
+               kunmap_atomic(vaddr, KM_USER0);
+
+               in_iov = page_address(iov_page);
+               out_iov = in_iov + in_iovs;
+
+               goto retry;
+       }
+
+       err = -EIO;
+       if (transferred > inarg.out_size)
+               goto out;
+
+       err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
+ out:
+       if (req)
+               fuse_put_request(fc, req);
+       if (iov_page)
+               __free_page(iov_page);
+       while (num_pages)
+               __free_page(pages[--num_pages]);
+       kfree(pages);
+
+       return err ? err : outarg.result;
+}
+
+static long fuse_file_ioctl(struct file *file, unsigned int cmd,
+                           unsigned long arg)
+{
+       return fuse_file_do_ioctl(file, cmd, arg, 0);
+}
+
+static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
+                                  unsigned long arg)
+{
+       return fuse_file_do_ioctl(file, cmd, arg, FUSE_IOCTL_COMPAT);
+}
+
 static const struct file_operations fuse_file_operations = {
        .llseek         = fuse_file_llseek,
        .read           = do_sync_read,
@@ -1483,6 +1759,8 @@ static const struct file_operations fuse_file_operations = {
        .lock           = fuse_file_lock,
        .flock          = fuse_file_flock,
        .splice_read    = generic_file_splice_read,
+       .unlocked_ioctl = fuse_file_ioctl,
+       .compat_ioctl   = fuse_file_compat_ioctl,
 };
 
 static const struct file_operations fuse_direct_io_file_operations = {
@@ -1495,6 +1773,8 @@ static const struct file_operations fuse_direct_io_file_operations = {
        .fsync          = fuse_fsync,
        .lock           = fuse_file_lock,
        .flock          = fuse_file_flock,
+       .unlocked_ioctl = fuse_file_ioctl,
+       .compat_ioctl   = fuse_file_compat_ioctl,
        /* no mmap and splice_read */
 };
 
index 7caa473306e49aaf6b042c9de335c5633d681d19..608e300ae883d9289410f3cf8f431f6e2888e5d5 100644 (file)
@@ -148,6 +148,21 @@ struct fuse_file_lock {
  */
 #define FUSE_READ_LOCKOWNER    (1 << 1)
 
+/**
+ * Ioctl flags
+ *
+ * FUSE_IOCTL_COMPAT: 32bit compat ioctl on 64bit machine
+ * FUSE_IOCTL_UNRESTRICTED: not restricted to well-formed ioctls, retry allowed
+ * FUSE_IOCTL_RETRY: retry with new iovecs
+ *
+ * FUSE_IOCTL_MAX_IOV: maximum of in_iovecs + out_iovecs
+ */
+#define FUSE_IOCTL_COMPAT      (1 << 0)
+#define FUSE_IOCTL_UNRESTRICTED        (1 << 1)
+#define FUSE_IOCTL_RETRY       (1 << 2)
+
+#define FUSE_IOCTL_MAX_IOV     256
+
 enum fuse_opcode {
        FUSE_LOOKUP        = 1,
        FUSE_FORGET        = 2,  /* no reply */
@@ -185,6 +200,7 @@ enum fuse_opcode {
        FUSE_INTERRUPT     = 36,
        FUSE_BMAP          = 37,
        FUSE_DESTROY       = 38,
+       FUSE_IOCTL         = 39,
 };
 
 /* The read buffer is required to be at least 8k, but may be much larger */
@@ -385,6 +401,22 @@ struct fuse_bmap_out {
        __u64   block;
 };
 
+struct fuse_ioctl_in {
+       __u64   fh;
+       __u32   flags;
+       __u32   cmd;
+       __u64   arg;
+       __u32   in_size;
+       __u32   out_size;
+};
+
+struct fuse_ioctl_out {
+       __s32   result;
+       __u32   flags;
+       __u32   in_iovs;
+       __u32   out_iovs;
+};
+
 struct fuse_in_header {
        __u32   len;
        __u32   opcode;