kernel/watch_queue.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /* Watch queue and general notification mechanism, built on pipes
   3  *
   4  * Copyright (C) 2020 Red Hat, Inc. All Rights Reserved.
   5  * Written by David Howells (dhowells@redhat.com)
   6  *
   7  * See Documentation/watch_queue.rst
   8  */
   9
  10 #define pr_fmt(fmt) "watchq: " fmt
  11 #include <linux/module.h>
  12 #include <linux/init.h>
  13 #include <linux/sched.h>
  14 #include <linux/slab.h>
  15 #include <linux/printk.h>
  16 #include <linux/miscdevice.h>
  17 #include <linux/fs.h>
  18 #include <linux/mm.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/poll.h>
  21 #include <linux/uaccess.h>
  22 #include <linux/vmalloc.h>
  23 #include <linux/file.h>
  24 #include <linux/security.h>
  25 #include <linux/cred.h>
  26 #include <linux/sched/signal.h>
  27 #include <linux/watch_queue.h>
  28 #include <linux/pipe_fs_i.h>
  29
  30 MODULE_DESCRIPTION("Watch queue");
  31 MODULE_AUTHOR("Red Hat, Inc.");
  32 MODULE_LICENSE("GPL");
  33
  34 #define WATCH_QUEUE_NOTE_SIZE 128
  35 #define WATCH_QUEUE_NOTES_PER_PAGE (PAGE_SIZE / WATCH_QUEUE_NOTE_SIZE)
  36
  37 static void watch_queue_pipe_buf_release(struct pipe_inode_info *pipe,
  38                                          struct pipe_buffer *buf)
  39 {
  40         struct watch_queue *wqueue = (struct watch_queue *)buf->private;
  41         struct page *page;
  42         unsigned int bit;
  43
  44         /* We need to work out which note within the page this refers to, but
  45          * the note might have been maximum size, so merely ANDing the offset
  46          * off doesn't work.  OTOH, the note must've been more than zero size.
  47          */
  48         bit = buf->offset + buf->len;
  49         if ((bit & (WATCH_QUEUE_NOTE_SIZE - 1)) == 0)
  50                 bit -= WATCH_QUEUE_NOTE_SIZE;
  51         bit /= WATCH_QUEUE_NOTE_SIZE;
  52
  53         page = buf->page;
  54         bit += page->index;
  55
  56         set_bit(bit, wqueue->notes_bitmap);
  57 }
  58
  59 // No try_steal function => no stealing
  60 #define watch_queue_pipe_buf_try_steal NULL
  61
  62 /* New data written to a pipe may be appended to a buffer with this type. */
  63 static const struct pipe_buf_operations watch_queue_pipe_buf_ops = {
  64         .release        = watch_queue_pipe_buf_release,
  65         .try_steal      = watch_queue_pipe_buf_try_steal,
  66         .get            = generic_pipe_buf_get,
  67 };
  68
  69 /*
  70  * Post a notification to a watch queue.
  71  */
  72 static bool post_one_notification(struct watch_queue *wqueue,
  73                                   struct watch_notification *n)
  74 {
  75         void *p;
  76         struct pipe_inode_info *pipe = wqueue->pipe;
  77         struct pipe_buffer *buf;
  78         struct page *page;
  79         unsigned int head, tail, mask, note, offset, len;
  80         bool done = false;
  81
  82         if (!pipe)
  83                 return false;
  84
  85         spin_lock_irq(&pipe->rd_wait.lock);
  86
  87         if (wqueue->defunct)
  88                 goto out;
  89
  90         mask = pipe->ring_size - 1;
  91         head = pipe->head;
  92         tail = pipe->tail;
  93         if (pipe_full(head, tail, pipe->ring_size))
  94                 goto lost;
  95
  96         note = find_first_bit(wqueue->notes_bitmap, wqueue->nr_notes);
  97         if (note >= wqueue->nr_notes)
  98                 goto lost;
  99
 100         page = wqueue->notes[note / WATCH_QUEUE_NOTES_PER_PAGE];
 101         offset = note % WATCH_QUEUE_NOTES_PER_PAGE * WATCH_QUEUE_NOTE_SIZE;
 102         get_page(page);
 103         len = n->info & WATCH_INFO_LENGTH;
 104         p = kmap_atomic(page);
 105         memcpy(p + offset, n, len);
 106         kunmap_atomic(p);
 107
 108         buf = &pipe->bufs[head & mask];
 109         buf->page = page;
 110         buf->private = (unsigned long)wqueue;
 111         buf->ops = &watch_queue_pipe_buf_ops;
 112         buf->offset = offset;
 113         buf->len = len;
 114         buf->flags = PIPE_BUF_FLAG_WHOLE;
 115         pipe->head = head + 1;
 116
 117         if (!test_and_clear_bit(note, wqueue->notes_bitmap)) {
 118                 spin_unlock_irq(&pipe->rd_wait.lock);
 119                 BUG();
 120         }
 121         wake_up_interruptible_sync_poll_locked(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
 122         done = true;
 123
 124 out:
 125         spin_unlock_irq(&pipe->rd_wait.lock);
 126         if (done)
 127                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 128         return done;
 129
 130 lost:
 131         buf = &pipe->bufs[(head - 1) & mask];
 132         buf->flags |= PIPE_BUF_FLAG_LOSS;
 133         goto out;
 134 }
 135
 136 /*
 137  * Apply filter rules to a notification.
 138  */
 139 static bool filter_watch_notification(const struct watch_filter *wf,
 140                                       const struct watch_notification *n)
 141 {
 142         const struct watch_type_filter *wt;
 143         unsigned int st_bits = sizeof(wt->subtype_filter[0]) * 8;
 144         unsigned int st_index = n->subtype / st_bits;
 145         unsigned int st_bit = 1U << (n->subtype % st_bits);
 146         int i;
 147
 148         if (!test_bit(n->type, wf->type_filter))
 149                 return false;
 150
 151         for (i = 0; i < wf->nr_filters; i++) {
 152                 wt = &wf->filters[i];
 153                 if (n->type == wt->type &&
 154                     (wt->subtype_filter[st_index] & st_bit) &&
 155                     (n->info & wt->info_mask) == wt->info_filter)
 156                         return true;
 157         }
 158
 159         return false; /* If there is a filter, the default is to reject. */
 160 }
 161
 162 /**
 163  * __post_watch_notification - Post an event notification
 164  * @wlist: The watch list to post the event to.
 165  * @n: The notification record to post.
 166  * @cred: The creds of the process that triggered the notification.
 167  * @id: The ID to match on the watch.
 168  *
 169  * Post a notification of an event into a set of watch queues and let the users
 170  * know.
 171  *
 172  * The size of the notification should be set in n->info & WATCH_INFO_LENGTH and
 173  * should be in units of sizeof(*n).
 174  */
 175 void __post_watch_notification(struct watch_list *wlist,
 176                                struct watch_notification *n,
 177                                const struct cred *cred,
 178                                u64 id)
 179 {
 180         const struct watch_filter *wf;
 181         struct watch_queue *wqueue;
 182         struct watch *watch;
 183
 184         if (((n->info & WATCH_INFO_LENGTH) >> WATCH_INFO_LENGTH__SHIFT) == 0) {
 185                 WARN_ON(1);
 186                 return;
 187         }
 188
 189         rcu_read_lock();
 190
 191         hlist_for_each_entry_rcu(watch, &wlist->watchers, list_node) {
 192                 if (watch->id != id)
 193                         continue;
 194                 n->info &= ~WATCH_INFO_ID;
 195                 n->info |= watch->info_id;
 196
 197                 wqueue = rcu_dereference(watch->queue);
 198                 wf = rcu_dereference(wqueue->filter);
 199                 if (wf && !filter_watch_notification(wf, n))
 200                         continue;
 201
 202                 if (security_post_notification(watch->cred, cred, n) < 0)
 203                         continue;
 204
 205                 post_one_notification(wqueue, n);
 206         }
 207
 208         rcu_read_unlock();
 209 }
 210 EXPORT_SYMBOL(__post_watch_notification);
 211
 212 /*
 213  * Allocate sufficient pages to preallocation for the requested number of
 214  * notifications.
 215  */
 216 long watch_queue_set_size(struct pipe_inode_info *pipe, unsigned int nr_notes)
 217 {
 218         struct watch_queue *wqueue = pipe->watch_queue;
 219         struct page **pages;
 220         unsigned long *bitmap;
 221         unsigned long user_bufs;
 222         unsigned int bmsize;
 223         int ret, i, nr_pages;
 224
 225         if (!wqueue)
 226                 return -ENODEV;
 227         if (wqueue->notes)
 228                 return -EBUSY;
 229
 230         if (nr_notes < 1 ||
 231             nr_notes > 512) /* TODO: choose a better hard limit */
 232                 return -EINVAL;
 233
 234         nr_pages = (nr_notes + WATCH_QUEUE_NOTES_PER_PAGE - 1);
 235         nr_pages /= WATCH_QUEUE_NOTES_PER_PAGE;
 236         user_bufs = account_pipe_buffers(pipe->user, pipe->nr_accounted, nr_pages);
 237
 238         if (nr_pages > pipe->max_usage &&
 239             (too_many_pipe_buffers_hard(user_bufs) ||
 240              too_many_pipe_buffers_soft(user_bufs)) &&
 241             pipe_is_unprivileged_user()) {
 242                 ret = -EPERM;
 243                 goto error;
 244         }
 245
 246         ret = pipe_resize_ring(pipe, nr_notes);
 247         if (ret < 0)
 248                 goto error;
 249
 250         pages = kcalloc(sizeof(struct page *), nr_pages, GFP_KERNEL);
 251         if (!pages)
 252                 goto error;
 253
 254         for (i = 0; i < nr_pages; i++) {
 255                 pages[i] = alloc_page(GFP_KERNEL);
 256                 if (!pages[i])
 257                         goto error_p;
 258                 pages[i]->index = i * WATCH_QUEUE_NOTES_PER_PAGE;
 259         }
 260
 261         bmsize = (nr_notes + BITS_PER_LONG - 1) / BITS_PER_LONG;
 262         bmsize *= sizeof(unsigned long);
 263         bitmap = kmalloc(bmsize, GFP_KERNEL);
 264         if (!bitmap)
 265                 goto error_p;
 266
 267         memset(bitmap, 0xff, bmsize);
 268         wqueue->notes = pages;
 269         wqueue->notes_bitmap = bitmap;
 270         wqueue->nr_pages = nr_pages;
 271         wqueue->nr_notes = nr_pages * WATCH_QUEUE_NOTES_PER_PAGE;
 272         return 0;
 273
 274 error_p:
 275         for (i = 0; i < nr_pages; i++)
 276                 __free_page(pages[i]);
 277         kfree(pages);
 278 error:
 279         (void) account_pipe_buffers(pipe->user, nr_pages, pipe->nr_accounted);
 280         return ret;
 281 }
 282
 283 /*
 284  * Set the filter on a watch queue.
 285  */
 286 long watch_queue_set_filter(struct pipe_inode_info *pipe,
 287                             struct watch_notification_filter __user *_filter)
 288 {
 289         struct watch_notification_type_filter *tf;
 290         struct watch_notification_filter filter;
 291         struct watch_type_filter *q;
 292         struct watch_filter *wfilter;
 293         struct watch_queue *wqueue = pipe->watch_queue;
 294         int ret, nr_filter = 0, i;
 295
 296         if (!wqueue)
 297                 return -ENODEV;
 298
 299         if (!_filter) {
 300                 /* Remove the old filter */
 301                 wfilter = NULL;
 302                 goto set;
 303         }
 304
 305         /* Grab the user's filter specification */
 306         if (copy_from_user(&filter, _filter, sizeof(filter)) != 0)
 307                 return -EFAULT;
 308         if (filter.nr_filters == 0 ||
 309             filter.nr_filters > 16 ||
 310             filter.__reserved != 0)
 311                 return -EINVAL;
 312
 313         tf = memdup_user(_filter->filters, filter.nr_filters * sizeof(*tf));
 314         if (IS_ERR(tf))
 315                 return PTR_ERR(tf);
 316
 317         ret = -EINVAL;
 318         for (i = 0; i < filter.nr_filters; i++) {
 319                 if ((tf[i].info_filter & ~tf[i].info_mask) ||
 320                     tf[i].info_mask & WATCH_INFO_LENGTH)
 321                         goto err_filter;
 322                 /* Ignore any unknown types */
 323                 if (tf[i].type >= sizeof(wfilter->type_filter) * 8)
 324                         continue;
 325                 nr_filter++;
 326         }
 327
 328         /* Now we need to build the internal filter from only the relevant
 329          * user-specified filters.
 330          */
 331         ret = -ENOMEM;
 332         wfilter = kzalloc(struct_size(wfilter, filters, nr_filter), GFP_KERNEL);
 333         if (!wfilter)
 334                 goto err_filter;
 335         wfilter->nr_filters = nr_filter;
 336
 337         q = wfilter->filters;
 338         for (i = 0; i < filter.nr_filters; i++) {
 339                 if (tf[i].type >= sizeof(wfilter->type_filter) * BITS_PER_LONG)
 340                         continue;
 341
 342                 q->type                 = tf[i].type;
 343                 q->info_filter          = tf[i].info_filter;
 344                 q->info_mask            = tf[i].info_mask;
 345                 q->subtype_filter[0]    = tf[i].subtype_filter[0];
 346                 __set_bit(q->type, wfilter->type_filter);
 347                 q++;
 348         }
 349
 350         kfree(tf);
 351 set:
 352         pipe_lock(pipe);
 353         wfilter = rcu_replace_pointer(wqueue->filter, wfilter,
 354                                       lockdep_is_held(&pipe->mutex));
 355         pipe_unlock(pipe);
 356         if (wfilter)
 357                 kfree_rcu(wfilter, rcu);
 358         return 0;
 359
 360 err_filter:
 361         kfree(tf);
 362         return ret;
 363 }
 364
 365 static void __put_watch_queue(struct kref *kref)
 366 {
 367         struct watch_queue *wqueue =
 368                 container_of(kref, struct watch_queue, usage);
 369         struct watch_filter *wfilter;
 370         int i;
 371
 372         for (i = 0; i < wqueue->nr_pages; i++)
 373                 __free_page(wqueue->notes[i]);
 374
 375         wfilter = rcu_access_pointer(wqueue->filter);
 376         if (wfilter)
 377                 kfree_rcu(wfilter, rcu);
 378         kfree_rcu(wqueue, rcu);
 379 }
 380
 381 /**
 382  * put_watch_queue - Dispose of a ref on a watchqueue.
 383  * @wqueue: The watch queue to unref.
 384  */
 385 void put_watch_queue(struct watch_queue *wqueue)
 386 {
 387         kref_put(&wqueue->usage, __put_watch_queue);
 388 }
 389 EXPORT_SYMBOL(put_watch_queue);
 390
 391 static void free_watch(struct rcu_head *rcu)
 392 {
 393         struct watch *watch = container_of(rcu, struct watch, rcu);
 394
 395         put_watch_queue(rcu_access_pointer(watch->queue));
 396         put_cred(watch->cred);
 397 }
 398
 399 static void __put_watch(struct kref *kref)
 400 {
 401         struct watch *watch = container_of(kref, struct watch, usage);
 402
 403         call_rcu(&watch->rcu, free_watch);
 404 }
 405
 406 /*
 407  * Discard a watch.
 408  */
 409 static void put_watch(struct watch *watch)
 410 {
 411         kref_put(&watch->usage, __put_watch);
 412 }
 413
 414 /**
 415  * init_watch_queue - Initialise a watch
 416  * @watch: The watch to initialise.
 417  * @wqueue: The queue to assign.
 418  *
 419  * Initialise a watch and set the watch queue.
 420  */
 421 void init_watch(struct watch *watch, struct watch_queue *wqueue)
 422 {
 423         kref_init(&watch->usage);
 424         INIT_HLIST_NODE(&watch->list_node);
 425         INIT_HLIST_NODE(&watch->queue_node);
 426         rcu_assign_pointer(watch->queue, wqueue);
 427 }
 428
 429 /**
 430  * add_watch_to_object - Add a watch on an object to a watch list
 431  * @watch: The watch to add
 432  * @wlist: The watch list to add to
 433  *
 434  * @watch->queue must have been set to point to the queue to post notifications
 435  * to and the watch list of the object to be watched.  @watch->cred must also
 436  * have been set to the appropriate credentials and a ref taken on them.
 437  *
 438  * The caller must pin the queue and the list both and must hold the list
 439  * locked against racing watch additions/removals.
 440  */
 441 int add_watch_to_object(struct watch *watch, struct watch_list *wlist)
 442 {
 443         struct watch_queue *wqueue = rcu_access_pointer(watch->queue);
 444         struct watch *w;
 445
 446         hlist_for_each_entry(w, &wlist->watchers, list_node) {
 447                 struct watch_queue *wq = rcu_access_pointer(w->queue);
 448                 if (wqueue == wq && watch->id == w->id)
 449                         return -EBUSY;
 450         }
 451
 452         watch->cred = get_current_cred();
 453         rcu_assign_pointer(watch->watch_list, wlist);
 454
 455         spin_lock_bh(&wqueue->lock);
 456         kref_get(&wqueue->usage);
 457         kref_get(&watch->usage);
 458         hlist_add_head(&watch->queue_node, &wqueue->watches);
 459         spin_unlock_bh(&wqueue->lock);
 460
 461         hlist_add_head(&watch->list_node, &wlist->watchers);
 462         return 0;
 463 }
 464 EXPORT_SYMBOL(add_watch_to_object);
 465
 466 /**
 467  * remove_watch_from_object - Remove a watch or all watches from an object.
 468  * @wlist: The watch list to remove from
 469  * @wq: The watch queue of interest (ignored if @all is true)
 470  * @id: The ID of the watch to remove (ignored if @all is true)
 471  * @all: True to remove all objects
 472  *
 473  * Remove a specific watch or all watches from an object.  A notification is
 474  * sent to the watcher to tell them that this happened.
 475  */
 476 int remove_watch_from_object(struct watch_list *wlist, struct watch_queue *wq,
 477                              u64 id, bool all)
 478 {
 479         struct watch_notification_removal n;
 480         struct watch_queue *wqueue;
 481         struct watch *watch;
 482         int ret = -EBADSLT;
 483
 484         rcu_read_lock();
 485
 486 again:
 487         spin_lock(&wlist->lock);
 488         hlist_for_each_entry(watch, &wlist->watchers, list_node) {
 489                 if (all ||
 490                     (watch->id == id && rcu_access_pointer(watch->queue) == wq))
 491                         goto found;
 492         }
 493         spin_unlock(&wlist->lock);
 494         goto out;
 495
 496 found:
 497         ret = 0;
 498         hlist_del_init_rcu(&watch->list_node);
 499         rcu_assign_pointer(watch->watch_list, NULL);
 500         spin_unlock(&wlist->lock);
 501
 502         /* We now own the reference on watch that used to belong to wlist. */
 503
 504         n.watch.type = WATCH_TYPE_META;
 505         n.watch.subtype = WATCH_META_REMOVAL_NOTIFICATION;
 506         n.watch.info = watch->info_id | watch_sizeof(n.watch);
 507         n.id = id;
 508         if (id != 0)
 509                 n.watch.info = watch->info_id | watch_sizeof(n);
 510
 511         wqueue = rcu_dereference(watch->queue);
 512
 513         /* We don't need the watch list lock for the next bit as RCU is
 514          * protecting *wqueue from deallocation.
 515          */
 516         if (wqueue) {
 517                 post_one_notification(wqueue, &n.watch);
 518
 519                 spin_lock_bh(&wqueue->lock);
 520
 521                 if (!hlist_unhashed(&watch->queue_node)) {
 522                         hlist_del_init_rcu(&watch->queue_node);
 523                         put_watch(watch);
 524                 }
 525
 526                 spin_unlock_bh(&wqueue->lock);
 527         }
 528
 529         if (wlist->release_watch) {
 530                 void (*release_watch)(struct watch *);
 531
 532                 release_watch = wlist->release_watch;
 533                 rcu_read_unlock();
 534                 (*release_watch)(watch);
 535                 rcu_read_lock();
 536         }
 537         put_watch(watch);
 538
 539         if (all && !hlist_empty(&wlist->watchers))
 540                 goto again;
 541 out:
 542         rcu_read_unlock();
 543         return ret;
 544 }
 545 EXPORT_SYMBOL(remove_watch_from_object);
 546
 547 /*
 548  * Remove all the watches that are contributory to a queue.  This has the
 549  * potential to race with removal of the watches by the destruction of the
 550  * objects being watched or with the distribution of notifications.
 551  */
 552 void watch_queue_clear(struct watch_queue *wqueue)
 553 {
 554         struct watch_list *wlist;
 555         struct watch *watch;
 556         bool release;
 557
 558         rcu_read_lock();
 559         spin_lock_bh(&wqueue->lock);
 560
 561         /* Prevent new additions and prevent notifications from happening */
 562         wqueue->defunct = true;
 563
 564         while (!hlist_empty(&wqueue->watches)) {
 565                 watch = hlist_entry(wqueue->watches.first, struct watch, queue_node);
 566                 hlist_del_init_rcu(&watch->queue_node);
 567                 /* We now own a ref on the watch. */
 568                 spin_unlock_bh(&wqueue->lock);
 569
 570                 /* We can't do the next bit under the queue lock as we need to
 571                  * get the list lock - which would cause a deadlock if someone
 572                  * was removing from the opposite direction at the same time or
 573                  * posting a notification.
 574                  */
 575                 wlist = rcu_dereference(watch->watch_list);
 576                 if (wlist) {
 577                         void (*release_watch)(struct watch *);
 578
 579                         spin_lock(&wlist->lock);
 580
 581                         release = !hlist_unhashed(&watch->list_node);
 582                         if (release) {
 583                                 hlist_del_init_rcu(&watch->list_node);
 584                                 rcu_assign_pointer(watch->watch_list, NULL);
 585
 586                                 /* We now own a second ref on the watch. */
 587                         }
 588
 589                         release_watch = wlist->release_watch;
 590                         spin_unlock(&wlist->lock);
 591
 592                         if (release) {
 593                                 if (release_watch) {
 594                                         rcu_read_unlock();
 595                                         /* This might need to call dput(), so
 596                                          * we have to drop all the locks.
 597                                          */
 598                                         (*release_watch)(watch);
 599                                         rcu_read_lock();
 600                                 }
 601                                 put_watch(watch);
 602                         }
 603                 }
 604
 605                 put_watch(watch);
 606                 spin_lock_bh(&wqueue->lock);
 607         }
 608
 609         spin_unlock_bh(&wqueue->lock);
 610         rcu_read_unlock();
 611 }
 612
 613 /**
 614  * get_watch_queue - Get a watch queue from its file descriptor.
 615  * @fd: The fd to query.
 616  */
 617 struct watch_queue *get_watch_queue(int fd)
 618 {
 619         struct pipe_inode_info *pipe;
 620         struct watch_queue *wqueue = ERR_PTR(-EINVAL);
 621         struct fd f;
 622
 623         f = fdget(fd);
 624         if (f.file) {
 625                 pipe = get_pipe_info(f.file, false);
 626                 if (pipe && pipe->watch_queue) {
 627                         wqueue = pipe->watch_queue;
 628                         kref_get(&wqueue->usage);
 629                 }
 630                 fdput(f);
 631         }
 632
 633         return wqueue;
 634 }
 635 EXPORT_SYMBOL(get_watch_queue);
 636
 637 /*
 638  * Initialise a watch queue
 639  */
 640 int watch_queue_init(struct pipe_inode_info *pipe)
 641 {
 642         struct watch_queue *wqueue;
 643
 644         wqueue = kzalloc(sizeof(*wqueue), GFP_KERNEL);
 645         if (!wqueue)
 646                 return -ENOMEM;
 647
 648         wqueue->pipe = pipe;
 649         kref_init(&wqueue->usage);
 650         spin_lock_init(&wqueue->lock);
 651         INIT_HLIST_HEAD(&wqueue->watches);
 652
 653         pipe->watch_queue = wqueue;
 654         return 0;
 655 }