]> git.ipfire.org Git - thirdparty/kernel/linux.git/commitdiff
Merge tag 'notifications-20200601' of git://git.kernel.org/pub/scm/linux/kernel/git...
authorLinus Torvalds <torvalds@linux-foundation.org>
Sat, 13 Jun 2020 16:56:21 +0000 (09:56 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Sat, 13 Jun 2020 16:56:21 +0000 (09:56 -0700)
Pull notification queue from David Howells:
 "This adds a general notification queue concept and adds an event
  source for keys/keyrings, such as linking and unlinking keys and
  changing their attributes.

  Thanks to Debarshi Ray, we do have a pull request to use this to fix a
  problem with gnome-online-accounts - as mentioned last time:

     https://gitlab.gnome.org/GNOME/gnome-online-accounts/merge_requests/47

  Without this, g-o-a has to constantly poll a keyring-based kerberos
  cache to find out if kinit has changed anything.

  [ There are other notification pending: mount/sb fsinfo notifications
    for libmount that Karel Zak and Ian Kent have been working on, and
    Christian Brauner would like to use them in lxc, but let's see how
    this one works first ]

  LSM hooks are included:

   - A set of hooks are provided that allow an LSM to rule on whether or
     not a watch may be set. Each of these hooks takes a different
     "watched object" parameter, so they're not really shareable. The
     LSM should use current's credentials. [Wanted by SELinux & Smack]

   - A hook is provided to allow an LSM to rule on whether or not a
     particular message may be posted to a particular queue. This is
     given the credentials from the event generator (which may be the
     system) and the watch setter. [Wanted by Smack]

  I've provided SELinux and Smack with implementations of some of these
  hooks.

  WHY
  ===

  Key/keyring notifications are desirable because if you have your
  kerberos tickets in a file/directory, your Gnome desktop will monitor
  that using something like fanotify and tell you if your credentials
  cache changes.

  However, we also have the ability to cache your kerberos tickets in
  the session, user or persistent keyring so that it isn't left around
  on disk across a reboot or logout. Keyrings, however, cannot currently
  be monitored asynchronously, so the desktop has to poll for it - not
  so good on a laptop. This facility will allow the desktop to avoid the
  need to poll.

  DESIGN DECISIONS
  ================

   - The notification queue is built on top of a standard pipe. Messages
     are effectively spliced in. The pipe is opened with a special flag:

        pipe2(fds, O_NOTIFICATION_PIPE);

     The special flag has the same value as O_EXCL (which doesn't seem
     like it will ever be applicable in this context)[?]. It is given up
     front to make it a lot easier to prohibit splice&co from accessing
     the pipe.

     [?] Should this be done some other way?  I'd rather not use up a new
         O_* flag if I can avoid it - should I add a pipe3() system call
         instead?

     The pipe is then configured::

        ioctl(fds[1], IOC_WATCH_QUEUE_SET_SIZE, queue_depth);
        ioctl(fds[1], IOC_WATCH_QUEUE_SET_FILTER, &filter);

     Messages are then read out of the pipe using read().

   - It should be possible to allow write() to insert data into the
     notification pipes too, but this is currently disabled as the
     kernel has to be able to insert messages into the pipe *without*
     holding pipe->mutex and the code to make this work needs careful
     auditing.

   - sendfile(), splice() and vmsplice() are disabled on notification
     pipes because of the pipe->mutex issue and also because they
     sometimes want to revert what they just did - but one or more
     notification messages might've been interleaved in the ring.

   - The kernel inserts messages with the wait queue spinlock held. This
     means that pipe_read() and pipe_write() have to take the spinlock
     to update the queue pointers.

   - Records in the buffer are binary, typed and have a length so that
     they can be of varying size.

     This allows multiple heterogeneous sources to share a common
     buffer; there are 16 million types available, of which I've used
     just a few, so there is scope for others to be used. Tags may be
     specified when a watchpoint is created to help distinguish the
     sources.

   - Records are filterable as types have up to 256 subtypes that can be
     individually filtered. Other filtration is also available.

   - Notification pipes don't interfere with each other; each may be
     bound to a different set of watches. Any particular notification
     will be copied to all the queues that are currently watching for it
     - and only those that are watching for it.

   - When recording a notification, the kernel will not sleep, but will
     rather mark a queue as having lost a message if there's
     insufficient space. read() will fabricate a loss notification
     message at an appropriate point later.

   - The notification pipe is created and then watchpoints are attached
     to it, using one of:

        keyctl_watch_key(KEY_SPEC_SESSION_KEYRING, fds[1], 0x01);
        watch_mount(AT_FDCWD, "/", 0, fd, 0x02);
        watch_sb(AT_FDCWD, "/mnt", 0, fd, 0x03);

     where in both cases, fd indicates the queue and the number after is
     a tag between 0 and 255.

   - Watches are removed if either the notification pipe is destroyed or
     the watched object is destroyed. In the latter case, a message will
     be generated indicating the enforced watch removal.

  Things I want to avoid:

   - Introducing features that make the core VFS dependent on the
     network stack or networking namespaces (ie. usage of netlink).

   - Dumping all this stuff into dmesg and having a daemon that sits
     there parsing the output and distributing it as this then puts the
     responsibility for security into userspace and makes handling
     namespaces tricky. Further, dmesg might not exist or might be
     inaccessible inside a container.

   - Letting users see events they shouldn't be able to see.

  TESTING AND MANPAGES
  ====================

   - The keyutils tree has a pipe-watch branch that has keyctl commands
     for making use of notifications. Proposed manual pages can also be
     found on this branch, though a couple of them really need to go to
     the main manpages repository instead.

     If the kernel supports the watching of keys, then running "make
     test" on that branch will cause the testing infrastructure to spawn
     a monitoring process on the side that monitors a notifications pipe
     for all the key/keyring changes induced by the tests and they'll
     all be checked off to make sure they happened.

        https://git.kernel.org/pub/scm/linux/kernel/git/dhowells/keyutils.git/log/?h=pipe-watch

   - A test program is provided (samples/watch_queue/watch_test) that
     can be used to monitor for keyrings, mount and superblock events.
     Information on the notifications is simply logged to stdout"

* tag 'notifications-20200601' of git://git.kernel.org/pub/scm/linux/kernel/git/dhowells/linux-fs:
  smack: Implement the watch_key and post_notification hooks
  selinux: Implement the watch_key security hook
  keys: Make the KEY_NEED_* perms an enum rather than a mask
  pipe: Add notification lossage handling
  pipe: Allow buffers to be marked read-whole-or-error for notifications
  Add sample notification program
  watch_queue: Add a key/keyring notification facility
  security: Add hooks to rule on setting a watch
  pipe: Add general notification queue support
  pipe: Add O_NOTIFICATION_PIPE
  security: Add a hook for the point of notification insertion
  uapi: General notification queue definitions

19 files changed:
1  2 
Documentation/security/keys/core.rst
Documentation/userspace-api/ioctl/ioctl-number.rst
fs/pipe.c
fs/splice.c
include/linux/lsm_hook_defs.h
include/linux/lsm_hooks.h
include/linux/pipe_fs_i.h
include/linux/security.h
init/Kconfig
kernel/Makefile
kernel/watch_queue.c
samples/Kconfig
samples/Makefile
security/keys/Kconfig
security/keys/internal.h
security/keys/keyctl.c
security/security.c
security/selinux/hooks.c
security/smack/smack_lsm.c

Simple merge
diff --cc fs/pipe.c
Simple merge
diff --cc fs/splice.c
index 6b3c9a018a8e3e49d86e5a44c400ea80ffeb3856,50f3c0260c005b522ea1697410060cbf05bb9ca2..d7c8a7c4db07fff8cfc54a83adefba53207691a0
@@@ -1731,10 -1754,11 +1731,10 @@@ static int link_pipe(struct pipe_inode_
   * The 'flags' used are the SPLICE_F_* variants, currently the only
   * applicable one is SPLICE_F_NONBLOCK.
   */
 -static long do_tee(struct file *in, struct file *out, size_t len,
 -                 unsigned int flags)
 +long do_tee(struct file *in, struct file *out, size_t len, unsigned int flags)
  {
-       struct pipe_inode_info *ipipe = get_pipe_info(in);
-       struct pipe_inode_info *opipe = get_pipe_info(out);
+       struct pipe_inode_info *ipipe = get_pipe_info(in, true);
+       struct pipe_inode_info *opipe = get_pipe_info(out, true);
        int ret = -EINVAL;
  
        if (unlikely(!(in->f_mode & FMODE_READ) ||
Simple merge
Simple merge
index 0c31b94612623d33098eb93b63230195b4914462,6626f511de6fbade43d5eebf77fc0256c10236c0..50afd0d0084caf38fa314004c453a8d9a6695dc0
@@@ -8,7 -8,10 +8,11 @@@
  #define PIPE_BUF_FLAG_ATOMIC  0x02    /* was atomically mapped */
  #define PIPE_BUF_FLAG_GIFT    0x04    /* page is a gift */
  #define PIPE_BUF_FLAG_PACKET  0x08    /* read() as a packet */
 -#define PIPE_BUF_FLAG_WHOLE   0x10    /* read() must return entire buffer or error */
 +#define PIPE_BUF_FLAG_CAN_MERGE       0x10    /* can merge buffers */
++#define PIPE_BUF_FLAG_WHOLE   0x20    /* read() must return entire buffer or error */
+ #ifdef CONFIG_WATCH_QUEUE
 -#define PIPE_BUF_FLAG_LOSS    0x20    /* Message loss happened after this buffer */
++#define PIPE_BUF_FLAG_LOSS    0x40    /* Message loss happened after this buffer */
+ #endif
  
  /**
   *    struct pipe_buffer - a linux kernel pipe buffer
Simple merge
diff --cc init/Kconfig
Simple merge
diff --cc kernel/Makefile
Simple merge
index 0000000000000000000000000000000000000000,9a9699c06709b19602478f4a6754bb6ad8dfc496..f74020f6bd9d5d87df96099b8df2a1d4ab4aa98e
mode 000000,100644..100644
--- /dev/null
@@@ -1,0 -1,659 +1,655 @@@
 -static int watch_queue_pipe_buf_steal(struct pipe_inode_info *pipe,
 -                                    struct pipe_buffer *buf)
 -{
 -      return -1; /* No. */
 -}
+ // SPDX-License-Identifier: GPL-2.0
+ /* Watch queue and general notification mechanism, built on pipes
+  *
+  * Copyright (C) 2020 Red Hat, Inc. All Rights Reserved.
+  * Written by David Howells (dhowells@redhat.com)
+  *
+  * See Documentation/watch_queue.rst
+  */
+ #define pr_fmt(fmt) "watchq: " fmt
+ #include <linux/module.h>
+ #include <linux/init.h>
+ #include <linux/sched.h>
+ #include <linux/slab.h>
+ #include <linux/printk.h>
+ #include <linux/miscdevice.h>
+ #include <linux/fs.h>
+ #include <linux/mm.h>
+ #include <linux/pagemap.h>
+ #include <linux/poll.h>
+ #include <linux/uaccess.h>
+ #include <linux/vmalloc.h>
+ #include <linux/file.h>
+ #include <linux/security.h>
+ #include <linux/cred.h>
+ #include <linux/sched/signal.h>
+ #include <linux/watch_queue.h>
+ #include <linux/pipe_fs_i.h>
+ MODULE_DESCRIPTION("Watch queue");
+ MODULE_AUTHOR("Red Hat, Inc.");
+ MODULE_LICENSE("GPL");
+ #define WATCH_QUEUE_NOTE_SIZE 128
+ #define WATCH_QUEUE_NOTES_PER_PAGE (PAGE_SIZE / WATCH_QUEUE_NOTE_SIZE)
+ static void watch_queue_pipe_buf_release(struct pipe_inode_info *pipe,
+                                        struct pipe_buffer *buf)
+ {
+       struct watch_queue *wqueue = (struct watch_queue *)buf->private;
+       struct page *page;
+       unsigned int bit;
+       /* We need to work out which note within the page this refers to, but
+        * the note might have been maximum size, so merely ANDing the offset
+        * off doesn't work.  OTOH, the note must've been more than zero size.
+        */
+       bit = buf->offset + buf->len;
+       if ((bit & (WATCH_QUEUE_NOTE_SIZE - 1)) == 0)
+               bit -= WATCH_QUEUE_NOTE_SIZE;
+       bit /= WATCH_QUEUE_NOTE_SIZE;
+       page = buf->page;
+       bit += page->index;
+       set_bit(bit, wqueue->notes_bitmap);
+ }
 -      .confirm        = generic_pipe_buf_confirm,
++// No try_steal function => no stealing
++#define watch_queue_pipe_buf_try_steal NULL
+ /* New data written to a pipe may be appended to a buffer with this type. */
+ static const struct pipe_buf_operations watch_queue_pipe_buf_ops = {
 -      .steal          = watch_queue_pipe_buf_steal,
+       .release        = watch_queue_pipe_buf_release,
++      .try_steal      = watch_queue_pipe_buf_try_steal,
+       .get            = generic_pipe_buf_get,
+ };
+ /*
+  * Post a notification to a watch queue.
+  */
+ static bool post_one_notification(struct watch_queue *wqueue,
+                                 struct watch_notification *n)
+ {
+       void *p;
+       struct pipe_inode_info *pipe = wqueue->pipe;
+       struct pipe_buffer *buf;
+       struct page *page;
+       unsigned int head, tail, mask, note, offset, len;
+       bool done = false;
+       if (!pipe)
+               return false;
+       spin_lock_irq(&pipe->rd_wait.lock);
+       if (wqueue->defunct)
+               goto out;
+       mask = pipe->ring_size - 1;
+       head = pipe->head;
+       tail = pipe->tail;
+       if (pipe_full(head, tail, pipe->ring_size))
+               goto lost;
+       note = find_first_bit(wqueue->notes_bitmap, wqueue->nr_notes);
+       if (note >= wqueue->nr_notes)
+               goto lost;
+       page = wqueue->notes[note / WATCH_QUEUE_NOTES_PER_PAGE];
+       offset = note % WATCH_QUEUE_NOTES_PER_PAGE * WATCH_QUEUE_NOTE_SIZE;
+       get_page(page);
+       len = n->info & WATCH_INFO_LENGTH;
+       p = kmap_atomic(page);
+       memcpy(p + offset, n, len);
+       kunmap_atomic(p);
+       buf = &pipe->bufs[head & mask];
+       buf->page = page;
+       buf->private = (unsigned long)wqueue;
+       buf->ops = &watch_queue_pipe_buf_ops;
+       buf->offset = offset;
+       buf->len = len;
+       buf->flags = PIPE_BUF_FLAG_WHOLE;
+       pipe->head = head + 1;
+       if (!test_and_clear_bit(note, wqueue->notes_bitmap)) {
+               spin_unlock_irq(&pipe->rd_wait.lock);
+               BUG();
+       }
+       wake_up_interruptible_sync_poll_locked(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
+       done = true;
+ out:
+       spin_unlock_irq(&pipe->rd_wait.lock);
+       if (done)
+               kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
+       return done;
+ lost:
+       buf = &pipe->bufs[(head - 1) & mask];
+       buf->flags |= PIPE_BUF_FLAG_LOSS;
+       goto out;
+ }
+ /*
+  * Apply filter rules to a notification.
+  */
+ static bool filter_watch_notification(const struct watch_filter *wf,
+                                     const struct watch_notification *n)
+ {
+       const struct watch_type_filter *wt;
+       unsigned int st_bits = sizeof(wt->subtype_filter[0]) * 8;
+       unsigned int st_index = n->subtype / st_bits;
+       unsigned int st_bit = 1U << (n->subtype % st_bits);
+       int i;
+       if (!test_bit(n->type, wf->type_filter))
+               return false;
+       for (i = 0; i < wf->nr_filters; i++) {
+               wt = &wf->filters[i];
+               if (n->type == wt->type &&
+                   (wt->subtype_filter[st_index] & st_bit) &&
+                   (n->info & wt->info_mask) == wt->info_filter)
+                       return true;
+       }
+       return false; /* If there is a filter, the default is to reject. */
+ }
+ /**
+  * __post_watch_notification - Post an event notification
+  * @wlist: The watch list to post the event to.
+  * @n: The notification record to post.
+  * @cred: The creds of the process that triggered the notification.
+  * @id: The ID to match on the watch.
+  *
+  * Post a notification of an event into a set of watch queues and let the users
+  * know.
+  *
+  * The size of the notification should be set in n->info & WATCH_INFO_LENGTH and
+  * should be in units of sizeof(*n).
+  */
+ void __post_watch_notification(struct watch_list *wlist,
+                              struct watch_notification *n,
+                              const struct cred *cred,
+                              u64 id)
+ {
+       const struct watch_filter *wf;
+       struct watch_queue *wqueue;
+       struct watch *watch;
+       if (((n->info & WATCH_INFO_LENGTH) >> WATCH_INFO_LENGTH__SHIFT) == 0) {
+               WARN_ON(1);
+               return;
+       }
+       rcu_read_lock();
+       hlist_for_each_entry_rcu(watch, &wlist->watchers, list_node) {
+               if (watch->id != id)
+                       continue;
+               n->info &= ~WATCH_INFO_ID;
+               n->info |= watch->info_id;
+               wqueue = rcu_dereference(watch->queue);
+               wf = rcu_dereference(wqueue->filter);
+               if (wf && !filter_watch_notification(wf, n))
+                       continue;
+               if (security_post_notification(watch->cred, cred, n) < 0)
+                       continue;
+               post_one_notification(wqueue, n);
+       }
+       rcu_read_unlock();
+ }
+ EXPORT_SYMBOL(__post_watch_notification);
+ /*
+  * Allocate sufficient pages to preallocation for the requested number of
+  * notifications.
+  */
+ long watch_queue_set_size(struct pipe_inode_info *pipe, unsigned int nr_notes)
+ {
+       struct watch_queue *wqueue = pipe->watch_queue;
+       struct page **pages;
+       unsigned long *bitmap;
+       unsigned long user_bufs;
+       unsigned int bmsize;
+       int ret, i, nr_pages;
+       if (!wqueue)
+               return -ENODEV;
+       if (wqueue->notes)
+               return -EBUSY;
+       if (nr_notes < 1 ||
+           nr_notes > 512) /* TODO: choose a better hard limit */
+               return -EINVAL;
+       nr_pages = (nr_notes + WATCH_QUEUE_NOTES_PER_PAGE - 1);
+       nr_pages /= WATCH_QUEUE_NOTES_PER_PAGE;
+       user_bufs = account_pipe_buffers(pipe->user, pipe->nr_accounted, nr_pages);
+       if (nr_pages > pipe->max_usage &&
+           (too_many_pipe_buffers_hard(user_bufs) ||
+            too_many_pipe_buffers_soft(user_bufs)) &&
+           pipe_is_unprivileged_user()) {
+               ret = -EPERM;
+               goto error;
+       }
+       ret = pipe_resize_ring(pipe, nr_notes);
+       if (ret < 0)
+               goto error;
+       pages = kcalloc(sizeof(struct page *), nr_pages, GFP_KERNEL);
+       if (!pages)
+               goto error;
+       for (i = 0; i < nr_pages; i++) {
+               pages[i] = alloc_page(GFP_KERNEL);
+               if (!pages[i])
+                       goto error_p;
+               pages[i]->index = i * WATCH_QUEUE_NOTES_PER_PAGE;
+       }
+       bmsize = (nr_notes + BITS_PER_LONG - 1) / BITS_PER_LONG;
+       bmsize *= sizeof(unsigned long);
+       bitmap = kmalloc(bmsize, GFP_KERNEL);
+       if (!bitmap)
+               goto error_p;
+       memset(bitmap, 0xff, bmsize);
+       wqueue->notes = pages;
+       wqueue->notes_bitmap = bitmap;
+       wqueue->nr_pages = nr_pages;
+       wqueue->nr_notes = nr_pages * WATCH_QUEUE_NOTES_PER_PAGE;
+       return 0;
+ error_p:
+       for (i = 0; i < nr_pages; i++)
+               __free_page(pages[i]);
+       kfree(pages);
+ error:
+       (void) account_pipe_buffers(pipe->user, nr_pages, pipe->nr_accounted);
+       return ret;
+ }
+ /*
+  * Set the filter on a watch queue.
+  */
+ long watch_queue_set_filter(struct pipe_inode_info *pipe,
+                           struct watch_notification_filter __user *_filter)
+ {
+       struct watch_notification_type_filter *tf;
+       struct watch_notification_filter filter;
+       struct watch_type_filter *q;
+       struct watch_filter *wfilter;
+       struct watch_queue *wqueue = pipe->watch_queue;
+       int ret, nr_filter = 0, i;
+       if (!wqueue)
+               return -ENODEV;
+       if (!_filter) {
+               /* Remove the old filter */
+               wfilter = NULL;
+               goto set;
+       }
+       /* Grab the user's filter specification */
+       if (copy_from_user(&filter, _filter, sizeof(filter)) != 0)
+               return -EFAULT;
+       if (filter.nr_filters == 0 ||
+           filter.nr_filters > 16 ||
+           filter.__reserved != 0)
+               return -EINVAL;
+       tf = memdup_user(_filter->filters, filter.nr_filters * sizeof(*tf));
+       if (IS_ERR(tf))
+               return PTR_ERR(tf);
+       ret = -EINVAL;
+       for (i = 0; i < filter.nr_filters; i++) {
+               if ((tf[i].info_filter & ~tf[i].info_mask) ||
+                   tf[i].info_mask & WATCH_INFO_LENGTH)
+                       goto err_filter;
+               /* Ignore any unknown types */
+               if (tf[i].type >= sizeof(wfilter->type_filter) * 8)
+                       continue;
+               nr_filter++;
+       }
+       /* Now we need to build the internal filter from only the relevant
+        * user-specified filters.
+        */
+       ret = -ENOMEM;
+       wfilter = kzalloc(struct_size(wfilter, filters, nr_filter), GFP_KERNEL);
+       if (!wfilter)
+               goto err_filter;
+       wfilter->nr_filters = nr_filter;
+       q = wfilter->filters;
+       for (i = 0; i < filter.nr_filters; i++) {
+               if (tf[i].type >= sizeof(wfilter->type_filter) * BITS_PER_LONG)
+                       continue;
+               q->type                 = tf[i].type;
+               q->info_filter          = tf[i].info_filter;
+               q->info_mask            = tf[i].info_mask;
+               q->subtype_filter[0]    = tf[i].subtype_filter[0];
+               __set_bit(q->type, wfilter->type_filter);
+               q++;
+       }
+       kfree(tf);
+ set:
+       pipe_lock(pipe);
+       wfilter = rcu_replace_pointer(wqueue->filter, wfilter,
+                                     lockdep_is_held(&pipe->mutex));
+       pipe_unlock(pipe);
+       if (wfilter)
+               kfree_rcu(wfilter, rcu);
+       return 0;
+ err_filter:
+       kfree(tf);
+       return ret;
+ }
+ static void __put_watch_queue(struct kref *kref)
+ {
+       struct watch_queue *wqueue =
+               container_of(kref, struct watch_queue, usage);
+       struct watch_filter *wfilter;
+       int i;
+       for (i = 0; i < wqueue->nr_pages; i++)
+               __free_page(wqueue->notes[i]);
+       wfilter = rcu_access_pointer(wqueue->filter);
+       if (wfilter)
+               kfree_rcu(wfilter, rcu);
+       kfree_rcu(wqueue, rcu);
+ }
+ /**
+  * put_watch_queue - Dispose of a ref on a watchqueue.
+  * @wqueue: The watch queue to unref.
+  */
+ void put_watch_queue(struct watch_queue *wqueue)
+ {
+       kref_put(&wqueue->usage, __put_watch_queue);
+ }
+ EXPORT_SYMBOL(put_watch_queue);
+ static void free_watch(struct rcu_head *rcu)
+ {
+       struct watch *watch = container_of(rcu, struct watch, rcu);
+       put_watch_queue(rcu_access_pointer(watch->queue));
+       put_cred(watch->cred);
+ }
+ static void __put_watch(struct kref *kref)
+ {
+       struct watch *watch = container_of(kref, struct watch, usage);
+       call_rcu(&watch->rcu, free_watch);
+ }
+ /*
+  * Discard a watch.
+  */
+ static void put_watch(struct watch *watch)
+ {
+       kref_put(&watch->usage, __put_watch);
+ }
+ /**
+  * init_watch_queue - Initialise a watch
+  * @watch: The watch to initialise.
+  * @wqueue: The queue to assign.
+  *
+  * Initialise a watch and set the watch queue.
+  */
+ void init_watch(struct watch *watch, struct watch_queue *wqueue)
+ {
+       kref_init(&watch->usage);
+       INIT_HLIST_NODE(&watch->list_node);
+       INIT_HLIST_NODE(&watch->queue_node);
+       rcu_assign_pointer(watch->queue, wqueue);
+ }
+ /**
+  * add_watch_to_object - Add a watch on an object to a watch list
+  * @watch: The watch to add
+  * @wlist: The watch list to add to
+  *
+  * @watch->queue must have been set to point to the queue to post notifications
+  * to and the watch list of the object to be watched.  @watch->cred must also
+  * have been set to the appropriate credentials and a ref taken on them.
+  *
+  * The caller must pin the queue and the list both and must hold the list
+  * locked against racing watch additions/removals.
+  */
+ int add_watch_to_object(struct watch *watch, struct watch_list *wlist)
+ {
+       struct watch_queue *wqueue = rcu_access_pointer(watch->queue);
+       struct watch *w;
+       hlist_for_each_entry(w, &wlist->watchers, list_node) {
+               struct watch_queue *wq = rcu_access_pointer(w->queue);
+               if (wqueue == wq && watch->id == w->id)
+                       return -EBUSY;
+       }
+       watch->cred = get_current_cred();
+       rcu_assign_pointer(watch->watch_list, wlist);
+       spin_lock_bh(&wqueue->lock);
+       kref_get(&wqueue->usage);
+       kref_get(&watch->usage);
+       hlist_add_head(&watch->queue_node, &wqueue->watches);
+       spin_unlock_bh(&wqueue->lock);
+       hlist_add_head(&watch->list_node, &wlist->watchers);
+       return 0;
+ }
+ EXPORT_SYMBOL(add_watch_to_object);
+ /**
+  * remove_watch_from_object - Remove a watch or all watches from an object.
+  * @wlist: The watch list to remove from
+  * @wq: The watch queue of interest (ignored if @all is true)
+  * @id: The ID of the watch to remove (ignored if @all is true)
+  * @all: True to remove all objects
+  *
+  * Remove a specific watch or all watches from an object.  A notification is
+  * sent to the watcher to tell them that this happened.
+  */
+ int remove_watch_from_object(struct watch_list *wlist, struct watch_queue *wq,
+                            u64 id, bool all)
+ {
+       struct watch_notification_removal n;
+       struct watch_queue *wqueue;
+       struct watch *watch;
+       int ret = -EBADSLT;
+       rcu_read_lock();
+ again:
+       spin_lock(&wlist->lock);
+       hlist_for_each_entry(watch, &wlist->watchers, list_node) {
+               if (all ||
+                   (watch->id == id && rcu_access_pointer(watch->queue) == wq))
+                       goto found;
+       }
+       spin_unlock(&wlist->lock);
+       goto out;
+ found:
+       ret = 0;
+       hlist_del_init_rcu(&watch->list_node);
+       rcu_assign_pointer(watch->watch_list, NULL);
+       spin_unlock(&wlist->lock);
+       /* We now own the reference on watch that used to belong to wlist. */
+       n.watch.type = WATCH_TYPE_META;
+       n.watch.subtype = WATCH_META_REMOVAL_NOTIFICATION;
+       n.watch.info = watch->info_id | watch_sizeof(n.watch);
+       n.id = id;
+       if (id != 0)
+               n.watch.info = watch->info_id | watch_sizeof(n);
+       wqueue = rcu_dereference(watch->queue);
+       /* We don't need the watch list lock for the next bit as RCU is
+        * protecting *wqueue from deallocation.
+        */
+       if (wqueue) {
+               post_one_notification(wqueue, &n.watch);
+               spin_lock_bh(&wqueue->lock);
+               if (!hlist_unhashed(&watch->queue_node)) {
+                       hlist_del_init_rcu(&watch->queue_node);
+                       put_watch(watch);
+               }
+               spin_unlock_bh(&wqueue->lock);
+       }
+       if (wlist->release_watch) {
+               void (*release_watch)(struct watch *);
+               release_watch = wlist->release_watch;
+               rcu_read_unlock();
+               (*release_watch)(watch);
+               rcu_read_lock();
+       }
+       put_watch(watch);
+       if (all && !hlist_empty(&wlist->watchers))
+               goto again;
+ out:
+       rcu_read_unlock();
+       return ret;
+ }
+ EXPORT_SYMBOL(remove_watch_from_object);
+ /*
+  * Remove all the watches that are contributory to a queue.  This has the
+  * potential to race with removal of the watches by the destruction of the
+  * objects being watched or with the distribution of notifications.
+  */
+ void watch_queue_clear(struct watch_queue *wqueue)
+ {
+       struct watch_list *wlist;
+       struct watch *watch;
+       bool release;
+       rcu_read_lock();
+       spin_lock_bh(&wqueue->lock);
+       /* Prevent new additions and prevent notifications from happening */
+       wqueue->defunct = true;
+       while (!hlist_empty(&wqueue->watches)) {
+               watch = hlist_entry(wqueue->watches.first, struct watch, queue_node);
+               hlist_del_init_rcu(&watch->queue_node);
+               /* We now own a ref on the watch. */
+               spin_unlock_bh(&wqueue->lock);
+               /* We can't do the next bit under the queue lock as we need to
+                * get the list lock - which would cause a deadlock if someone
+                * was removing from the opposite direction at the same time or
+                * posting a notification.
+                */
+               wlist = rcu_dereference(watch->watch_list);
+               if (wlist) {
+                       void (*release_watch)(struct watch *);
+                       spin_lock(&wlist->lock);
+                       release = !hlist_unhashed(&watch->list_node);
+                       if (release) {
+                               hlist_del_init_rcu(&watch->list_node);
+                               rcu_assign_pointer(watch->watch_list, NULL);
+                               /* We now own a second ref on the watch. */
+                       }
+                       release_watch = wlist->release_watch;
+                       spin_unlock(&wlist->lock);
+                       if (release) {
+                               if (release_watch) {
+                                       rcu_read_unlock();
+                                       /* This might need to call dput(), so
+                                        * we have to drop all the locks.
+                                        */
+                                       (*release_watch)(watch);
+                                       rcu_read_lock();
+                               }
+                               put_watch(watch);
+                       }
+               }
+               put_watch(watch);
+               spin_lock_bh(&wqueue->lock);
+       }
+       spin_unlock_bh(&wqueue->lock);
+       rcu_read_unlock();
+ }
+ /**
+  * get_watch_queue - Get a watch queue from its file descriptor.
+  * @fd: The fd to query.
+  */
+ struct watch_queue *get_watch_queue(int fd)
+ {
+       struct pipe_inode_info *pipe;
+       struct watch_queue *wqueue = ERR_PTR(-EINVAL);
+       struct fd f;
+       f = fdget(fd);
+       if (f.file) {
+               pipe = get_pipe_info(f.file, false);
+               if (pipe && pipe->watch_queue) {
+                       wqueue = pipe->watch_queue;
+                       kref_get(&wqueue->usage);
+               }
+               fdput(f);
+       }
+       return wqueue;
+ }
+ EXPORT_SYMBOL(get_watch_queue);
+ /*
+  * Initialise a watch queue
+  */
+ int watch_queue_init(struct pipe_inode_info *pipe)
+ {
+       struct watch_queue *wqueue;
+       wqueue = kzalloc(sizeof(*wqueue), GFP_KERNEL);
+       if (!wqueue)
+               return -ENOMEM;
+       wqueue->pipe = pipe;
+       kref_init(&wqueue->usage);
+       spin_lock_init(&wqueue->lock);
+       INIT_HLIST_HEAD(&wqueue->watches);
+       pipe->watch_queue = wqueue;
+       return 0;
+ }
diff --cc samples/Kconfig
index 0cbb6146f3cf9a262e5834880b5b95d781bd0793,5c31971a5745c8339fd4bd925232918f1ba087bf..add431173bcca85130e664e213538c0135c183e2
@@@ -205,8 -190,11 +205,15 @@@ config SAMPLE_INTEL_ME
        help
          Build a sample program to work with mei device.
  
 +config SAMPLE_WATCHDOG
 +      bool "watchdog sample"
 +      depends on CC_CAN_LINK
 +
+ config SAMPLE_WATCH_QUEUE
+       bool "Build example /dev/watch_queue notification consumer"
+       depends on HEADERS_INSTALL
+       help
+         Build example userspace program to use the new mount_notify(),
+         sb_notify() syscalls and the KEYCTL_WATCH_KEY keyctl() function.
  endif # SAMPLES
index 29c66aadd954af429ce40d4ef608b8e8e7a19a41,8617edf7f19a30e3d0f60e703da38bc392dfbc1d..5b2629778754131d4396fc211d022245b66db574
@@@ -26,4 -24,4 +26,5 @@@ obj-$(CONFIG_VIDEO_PCI_SKELETON)      += v4l
  obj-y                                 += vfio-mdev/
  subdir-$(CONFIG_SAMPLE_VFS)           += vfs
  obj-$(CONFIG_SAMPLE_INTEL_MEI)                += mei/
 +subdir-$(CONFIG_SAMPLE_WATCHDOG)      += watchdog
+ subdir-$(CONFIG_SAMPLE_WATCH_QUEUE)   += watch_queue
Simple merge
Simple merge
Simple merge
Simple merge
Simple merge
Simple merge