root/fs/kernfs/file.c

DEFINITIONS

1. kernfs_of
2. kernfs_seq_stop_active
3. kernfs_seq_start
4. kernfs_seq_next
5. kernfs_seq_stop
6. kernfs_seq_show
7. kernfs_file_direct_read
8. kernfs_fop_read
9. kernfs_fop_write
10. kernfs_vma_open
11. kernfs_vma_fault
12. kernfs_vma_page_mkwrite
13. kernfs_vma_access
14. kernfs_vma_set_policy
15. kernfs_vma_get_policy
16. kernfs_fop_mmap
17. kernfs_get_open_node
18. kernfs_put_open_node
19. kernfs_fop_open
20. kernfs_release_file
21. kernfs_fop_release
22. kernfs_drain_open_files
23. kernfs_generic_poll
24. kernfs_fop_poll
25. kernfs_notify_workfn
26. kernfs_notify
27. __kernfs_create_file

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * fs/kernfs/file.c - kernfs file implementation
   4  *
   5  * Copyright (c) 2001-3 Patrick Mochel
   6  * Copyright (c) 2007 SUSE Linux Products GmbH
   7  * Copyright (c) 2007, 2013 Tejun Heo <tj@kernel.org>
   8  */
   9 
  10 #include <linux/fs.h>
  11 #include <linux/seq_file.h>
  12 #include <linux/slab.h>
  13 #include <linux/poll.h>
  14 #include <linux/pagemap.h>
  15 #include <linux/sched/mm.h>
  16 #include <linux/fsnotify.h>
  17 
  18 #include "kernfs-internal.h"
  19 
  20 /*
  21  * There's one kernfs_open_file for each open file and one kernfs_open_node
  22  * for each kernfs_node with one or more open files.
  23  *
  24  * kernfs_node->attr.open points to kernfs_open_node.  attr.open is
  25  * protected by kernfs_open_node_lock.
  26  *
  27  * filp->private_data points to seq_file whose ->private points to
  28  * kernfs_open_file.  kernfs_open_files are chained at
  29  * kernfs_open_node->files, which is protected by kernfs_open_file_mutex.
  30  */
  31 static DEFINE_SPINLOCK(kernfs_open_node_lock);
  32 static DEFINE_MUTEX(kernfs_open_file_mutex);
  33 
  34 struct kernfs_open_node {
  35         atomic_t                refcnt;
  36         atomic_t                event;
  37         wait_queue_head_t       poll;
  38         struct list_head        files; /* goes through kernfs_open_file.list */
  39 };
  40 
  41 /*
  42  * kernfs_notify() may be called from any context and bounces notifications
  43  * through a work item.  To minimize space overhead in kernfs_node, the
  44  * pending queue is implemented as a singly linked list of kernfs_nodes.
  45  * The list is terminated with the self pointer so that whether a
  46  * kernfs_node is on the list or not can be determined by testing the next
  47  * pointer for NULL.
  48  */
  49 #define KERNFS_NOTIFY_EOL                       ((void *)&kernfs_notify_list)
  50 
  51 static DEFINE_SPINLOCK(kernfs_notify_lock);
  52 static struct kernfs_node *kernfs_notify_list = KERNFS_NOTIFY_EOL;
  53 
  54 static struct kernfs_open_file *kernfs_of(struct file *file)
  55 {
  56         return ((struct seq_file *)file->private_data)->private;
  57 }
  58 
  59 /*
  60  * Determine the kernfs_ops for the given kernfs_node.  This function must
  61  * be called while holding an active reference.
  62  */
  63 static const struct kernfs_ops *kernfs_ops(struct kernfs_node *kn)
  64 {
  65         if (kn->flags & KERNFS_LOCKDEP)
  66                 lockdep_assert_held(kn);
  67         return kn->attr.ops;
  68 }
  69 
  70 /*
  71  * As kernfs_seq_stop() is also called after kernfs_seq_start() or
  72  * kernfs_seq_next() failure, it needs to distinguish whether it's stopping
  73  * a seq_file iteration which is fully initialized with an active reference
  74  * or an aborted kernfs_seq_start() due to get_active failure.  The
  75  * position pointer is the only context for each seq_file iteration and
  76  * thus the stop condition should be encoded in it.  As the return value is
  77  * directly visible to userland, ERR_PTR(-ENODEV) is the only acceptable
  78  * choice to indicate get_active failure.
  79  *
  80  * Unfortunately, this is complicated due to the optional custom seq_file
  81  * operations which may return ERR_PTR(-ENODEV) too.  kernfs_seq_stop()
  82  * can't distinguish whether ERR_PTR(-ENODEV) is from get_active failure or
  83  * custom seq_file operations and thus can't decide whether put_active
  84  * should be performed or not only on ERR_PTR(-ENODEV).
  85  *
  86  * This is worked around by factoring out the custom seq_stop() and
  87  * put_active part into kernfs_seq_stop_active(), skipping it from
  88  * kernfs_seq_stop() if ERR_PTR(-ENODEV) while invoking it directly after
  89  * custom seq_file operations fail with ERR_PTR(-ENODEV) - this ensures
  90  * that kernfs_seq_stop_active() is skipped only after get_active failure.
  91  */
  92 static void kernfs_seq_stop_active(struct seq_file *sf, void *v)
  93 {
  94         struct kernfs_open_file *of = sf->private;
  95         const struct kernfs_ops *ops = kernfs_ops(of->kn);
  96 
  97         if (ops->seq_stop)
  98                 ops->seq_stop(sf, v);
  99         kernfs_put_active(of->kn);
 100 }
 101 
 102 static void *kernfs_seq_start(struct seq_file *sf, loff_t *ppos)
 103 {
 104         struct kernfs_open_file *of = sf->private;
 105         const struct kernfs_ops *ops;
 106 
 107         /*
 108          * @of->mutex nests outside active ref and is primarily to ensure that
 109          * the ops aren't called concurrently for the same open file.
 110          */
 111         mutex_lock(&of->mutex);
 112         if (!kernfs_get_active(of->kn))
 113                 return ERR_PTR(-ENODEV);
 114 
 115         ops = kernfs_ops(of->kn);
 116         if (ops->seq_start) {
 117                 void *next = ops->seq_start(sf, ppos);
 118                 /* see the comment above kernfs_seq_stop_active() */
 119                 if (next == ERR_PTR(-ENODEV))
 120                         kernfs_seq_stop_active(sf, next);
 121                 return next;
 122         } else {
 123                 /*
 124                  * The same behavior and code as single_open().  Returns
 125                  * !NULL if pos is at the beginning; otherwise, NULL.
 126                  */
 127                 return NULL + !*ppos;
 128         }
 129 }
 130 
 131 static void *kernfs_seq_next(struct seq_file *sf, void *v, loff_t *ppos)
 132 {
 133         struct kernfs_open_file *of = sf->private;
 134         const struct kernfs_ops *ops = kernfs_ops(of->kn);
 135 
 136         if (ops->seq_next) {
 137                 void *next = ops->seq_next(sf, v, ppos);
 138                 /* see the comment above kernfs_seq_stop_active() */
 139                 if (next == ERR_PTR(-ENODEV))
 140                         kernfs_seq_stop_active(sf, next);
 141                 return next;
 142         } else {
 143                 /*
 144                  * The same behavior and code as single_open(), always
 145                  * terminate after the initial read.
 146                  */
 147                 ++*ppos;
 148                 return NULL;
 149         }
 150 }
 151 
 152 static void kernfs_seq_stop(struct seq_file *sf, void *v)
 153 {
 154         struct kernfs_open_file *of = sf->private;
 155 
 156         if (v != ERR_PTR(-ENODEV))
 157                 kernfs_seq_stop_active(sf, v);
 158         mutex_unlock(&of->mutex);
 159 }
 160 
 161 static int kernfs_seq_show(struct seq_file *sf, void *v)
 162 {
 163         struct kernfs_open_file *of = sf->private;
 164 
 165         of->event = atomic_read(&of->kn->attr.open->event);
 166 
 167         return of->kn->attr.ops->seq_show(sf, v);
 168 }
 169 
 170 static const struct seq_operations kernfs_seq_ops = {
 171         .start = kernfs_seq_start,
 172         .next = kernfs_seq_next,
 173         .stop = kernfs_seq_stop,
 174         .show = kernfs_seq_show,
 175 };
 176 
 177 /*
 178  * As reading a bin file can have side-effects, the exact offset and bytes
 179  * specified in read(2) call should be passed to the read callback making
 180  * it difficult to use seq_file.  Implement simplistic custom buffering for
 181  * bin files.
 182  */
 183 static ssize_t kernfs_file_direct_read(struct kernfs_open_file *of,
 184                                        char __user *user_buf, size_t count,
 185                                        loff_t *ppos)
 186 {
 187         ssize_t len = min_t(size_t, count, PAGE_SIZE);
 188         const struct kernfs_ops *ops;
 189         char *buf;
 190 
 191         buf = of->prealloc_buf;
 192         if (buf)
 193                 mutex_lock(&of->prealloc_mutex);
 194         else
 195                 buf = kmalloc(len, GFP_KERNEL);
 196         if (!buf)
 197                 return -ENOMEM;
 198 
 199         /*
 200          * @of->mutex nests outside active ref and is used both to ensure that
 201          * the ops aren't called concurrently for the same open file.
 202          */
 203         mutex_lock(&of->mutex);
 204         if (!kernfs_get_active(of->kn)) {
 205                 len = -ENODEV;
 206                 mutex_unlock(&of->mutex);
 207                 goto out_free;
 208         }
 209 
 210         of->event = atomic_read(&of->kn->attr.open->event);
 211         ops = kernfs_ops(of->kn);
 212         if (ops->read)
 213                 len = ops->read(of, buf, len, *ppos);
 214         else
 215                 len = -EINVAL;
 216 
 217         kernfs_put_active(of->kn);
 218         mutex_unlock(&of->mutex);
 219 
 220         if (len < 0)
 221                 goto out_free;
 222 
 223         if (copy_to_user(user_buf, buf, len)) {
 224                 len = -EFAULT;
 225                 goto out_free;
 226         }
 227 
 228         *ppos += len;
 229 
 230  out_free:
 231         if (buf == of->prealloc_buf)
 232                 mutex_unlock(&of->prealloc_mutex);
 233         else
 234                 kfree(buf);
 235         return len;
 236 }
 237 
 238 /**
 239  * kernfs_fop_read - kernfs vfs read callback
 240  * @file: file pointer
 241  * @user_buf: data to write
 242  * @count: number of bytes
 243  * @ppos: starting offset
 244  */
 245 static ssize_t kernfs_fop_read(struct file *file, char __user *user_buf,
 246                                size_t count, loff_t *ppos)
 247 {
 248         struct kernfs_open_file *of = kernfs_of(file);
 249 
 250         if (of->kn->flags & KERNFS_HAS_SEQ_SHOW)
 251                 return seq_read(file, user_buf, count, ppos);
 252         else
 253                 return kernfs_file_direct_read(of, user_buf, count, ppos);
 254 }
 255 
 256 /**
 257  * kernfs_fop_write - kernfs vfs write callback
 258  * @file: file pointer
 259  * @user_buf: data to write
 260  * @count: number of bytes
 261  * @ppos: starting offset
 262  *
 263  * Copy data in from userland and pass it to the matching kernfs write
 264  * operation.
 265  *
 266  * There is no easy way for us to know if userspace is only doing a partial
 267  * write, so we don't support them. We expect the entire buffer to come on
 268  * the first write.  Hint: if you're writing a value, first read the file,
 269  * modify only the the value you're changing, then write entire buffer
 270  * back.
 271  */
 272 static ssize_t kernfs_fop_write(struct file *file, const char __user *user_buf,
 273                                 size_t count, loff_t *ppos)
 274 {
 275         struct kernfs_open_file *of = kernfs_of(file);
 276         const struct kernfs_ops *ops;
 277         ssize_t len;
 278         char *buf;
 279 
 280         if (of->atomic_write_len) {
 281                 len = count;
 282                 if (len > of->atomic_write_len)
 283                         return -E2BIG;
 284         } else {
 285                 len = min_t(size_t, count, PAGE_SIZE);
 286         }
 287 
 288         buf = of->prealloc_buf;
 289         if (buf)
 290                 mutex_lock(&of->prealloc_mutex);
 291         else
 292                 buf = kmalloc(len + 1, GFP_KERNEL);
 293         if (!buf)
 294                 return -ENOMEM;
 295 
 296         if (copy_from_user(buf, user_buf, len)) {
 297                 len = -EFAULT;
 298                 goto out_free;
 299         }
 300         buf[len] = '\0';        /* guarantee string termination */
 301 
 302         /*
 303          * @of->mutex nests outside active ref and is used both to ensure that
 304          * the ops aren't called concurrently for the same open file.
 305          */
 306         mutex_lock(&of->mutex);
 307         if (!kernfs_get_active(of->kn)) {
 308                 mutex_unlock(&of->mutex);
 309                 len = -ENODEV;
 310                 goto out_free;
 311         }
 312 
 313         ops = kernfs_ops(of->kn);
 314         if (ops->write)
 315                 len = ops->write(of, buf, len, *ppos);
 316         else
 317                 len = -EINVAL;
 318 
 319         kernfs_put_active(of->kn);
 320         mutex_unlock(&of->mutex);
 321 
 322         if (len > 0)
 323                 *ppos += len;
 324 
 325 out_free:
 326         if (buf == of->prealloc_buf)
 327                 mutex_unlock(&of->prealloc_mutex);
 328         else
 329                 kfree(buf);
 330         return len;
 331 }
 332 
 333 static void kernfs_vma_open(struct vm_area_struct *vma)
 334 {
 335         struct file *file = vma->vm_file;
 336         struct kernfs_open_file *of = kernfs_of(file);
 337 
 338         if (!of->vm_ops)
 339                 return;
 340 
 341         if (!kernfs_get_active(of->kn))
 342                 return;
 343 
 344         if (of->vm_ops->open)
 345                 of->vm_ops->open(vma);
 346 
 347         kernfs_put_active(of->kn);
 348 }
 349 
 350 static vm_fault_t kernfs_vma_fault(struct vm_fault *vmf)
 351 {
 352         struct file *file = vmf->vma->vm_file;
 353         struct kernfs_open_file *of = kernfs_of(file);
 354         vm_fault_t ret;
 355 
 356         if (!of->vm_ops)
 357                 return VM_FAULT_SIGBUS;
 358 
 359         if (!kernfs_get_active(of->kn))
 360                 return VM_FAULT_SIGBUS;
 361 
 362         ret = VM_FAULT_SIGBUS;
 363         if (of->vm_ops->fault)
 364                 ret = of->vm_ops->fault(vmf);
 365 
 366         kernfs_put_active(of->kn);
 367         return ret;
 368 }
 369 
 370 static vm_fault_t kernfs_vma_page_mkwrite(struct vm_fault *vmf)
 371 {
 372         struct file *file = vmf->vma->vm_file;
 373         struct kernfs_open_file *of = kernfs_of(file);
 374         vm_fault_t ret;
 375 
 376         if (!of->vm_ops)
 377                 return VM_FAULT_SIGBUS;
 378 
 379         if (!kernfs_get_active(of->kn))
 380                 return VM_FAULT_SIGBUS;
 381 
 382         ret = 0;
 383         if (of->vm_ops->page_mkwrite)
 384                 ret = of->vm_ops->page_mkwrite(vmf);
 385         else
 386                 file_update_time(file);
 387 
 388         kernfs_put_active(of->kn);
 389         return ret;
 390 }
 391 
 392 static int kernfs_vma_access(struct vm_area_struct *vma, unsigned long addr,
 393                              void *buf, int len, int write)
 394 {
 395         struct file *file = vma->vm_file;
 396         struct kernfs_open_file *of = kernfs_of(file);
 397         int ret;
 398 
 399         if (!of->vm_ops)
 400                 return -EINVAL;
 401 
 402         if (!kernfs_get_active(of->kn))
 403                 return -EINVAL;
 404 
 405         ret = -EINVAL;
 406         if (of->vm_ops->access)
 407                 ret = of->vm_ops->access(vma, addr, buf, len, write);
 408 
 409         kernfs_put_active(of->kn);
 410         return ret;
 411 }
 412 
 413 #ifdef CONFIG_NUMA
 414 static int kernfs_vma_set_policy(struct vm_area_struct *vma,
 415                                  struct mempolicy *new)
 416 {
 417         struct file *file = vma->vm_file;
 418         struct kernfs_open_file *of = kernfs_of(file);
 419         int ret;
 420 
 421         if (!of->vm_ops)
 422                 return 0;
 423 
 424         if (!kernfs_get_active(of->kn))
 425                 return -EINVAL;
 426 
 427         ret = 0;
 428         if (of->vm_ops->set_policy)
 429                 ret = of->vm_ops->set_policy(vma, new);
 430 
 431         kernfs_put_active(of->kn);
 432         return ret;
 433 }
 434 
 435 static struct mempolicy *kernfs_vma_get_policy(struct vm_area_struct *vma,
 436                                                unsigned long addr)
 437 {
 438         struct file *file = vma->vm_file;
 439         struct kernfs_open_file *of = kernfs_of(file);
 440         struct mempolicy *pol;
 441 
 442         if (!of->vm_ops)
 443                 return vma->vm_policy;
 444 
 445         if (!kernfs_get_active(of->kn))
 446                 return vma->vm_policy;
 447 
 448         pol = vma->vm_policy;
 449         if (of->vm_ops->get_policy)
 450                 pol = of->vm_ops->get_policy(vma, addr);
 451 
 452         kernfs_put_active(of->kn);
 453         return pol;
 454 }
 455 
 456 #endif
 457 
 458 static const struct vm_operations_struct kernfs_vm_ops = {
 459         .open           = kernfs_vma_open,
 460         .fault          = kernfs_vma_fault,
 461         .page_mkwrite   = kernfs_vma_page_mkwrite,
 462         .access         = kernfs_vma_access,
 463 #ifdef CONFIG_NUMA
 464         .set_policy     = kernfs_vma_set_policy,
 465         .get_policy     = kernfs_vma_get_policy,
 466 #endif
 467 };
 468 
 469 static int kernfs_fop_mmap(struct file *file, struct vm_area_struct *vma)
 470 {
 471         struct kernfs_open_file *of = kernfs_of(file);
 472         const struct kernfs_ops *ops;
 473         int rc;
 474 
 475         /*
 476          * mmap path and of->mutex are prone to triggering spurious lockdep
 477          * warnings and we don't want to add spurious locking dependency
 478          * between the two.  Check whether mmap is actually implemented
 479          * without grabbing @of->mutex by testing HAS_MMAP flag.  See the
 480          * comment in kernfs_file_open() for more details.
 481          */
 482         if (!(of->kn->flags & KERNFS_HAS_MMAP))
 483                 return -ENODEV;
 484 
 485         mutex_lock(&of->mutex);
 486 
 487         rc = -ENODEV;
 488         if (!kernfs_get_active(of->kn))
 489                 goto out_unlock;
 490 
 491         ops = kernfs_ops(of->kn);
 492         rc = ops->mmap(of, vma);
 493         if (rc)
 494                 goto out_put;
 495 
 496         /*
 497          * PowerPC's pci_mmap of legacy_mem uses shmem_zero_setup()
 498          * to satisfy versions of X which crash if the mmap fails: that
 499          * substitutes a new vm_file, and we don't then want bin_vm_ops.
 500          */
 501         if (vma->vm_file != file)
 502                 goto out_put;
 503 
 504         rc = -EINVAL;
 505         if (of->mmapped && of->vm_ops != vma->vm_ops)
 506                 goto out_put;
 507 
 508         /*
 509          * It is not possible to successfully wrap close.
 510          * So error if someone is trying to use close.
 511          */
 512         rc = -EINVAL;
 513         if (vma->vm_ops && vma->vm_ops->close)
 514                 goto out_put;
 515 
 516         rc = 0;
 517         of->mmapped = true;
 518         of->vm_ops = vma->vm_ops;
 519         vma->vm_ops = &kernfs_vm_ops;
 520 out_put:
 521         kernfs_put_active(of->kn);
 522 out_unlock:
 523         mutex_unlock(&of->mutex);
 524 
 525         return rc;
 526 }
 527 
 528 /**
 529  *      kernfs_get_open_node - get or create kernfs_open_node
 530  *      @kn: target kernfs_node
 531  *      @of: kernfs_open_file for this instance of open
 532  *
 533  *      If @kn->attr.open exists, increment its reference count; otherwise,
 534  *      create one.  @of is chained to the files list.
 535  *
 536  *      LOCKING:
 537  *      Kernel thread context (may sleep).
 538  *
 539  *      RETURNS:
 540  *      0 on success, -errno on failure.
 541  */
 542 static int kernfs_get_open_node(struct kernfs_node *kn,
 543                                 struct kernfs_open_file *of)
 544 {
 545         struct kernfs_open_node *on, *new_on = NULL;
 546 
 547  retry:
 548         mutex_lock(&kernfs_open_file_mutex);
 549         spin_lock_irq(&kernfs_open_node_lock);
 550 
 551         if (!kn->attr.open && new_on) {
 552                 kn->attr.open = new_on;
 553                 new_on = NULL;
 554         }
 555 
 556         on = kn->attr.open;
 557         if (on) {
 558                 atomic_inc(&on->refcnt);
 559                 list_add_tail(&of->list, &on->files);
 560         }
 561 
 562         spin_unlock_irq(&kernfs_open_node_lock);
 563         mutex_unlock(&kernfs_open_file_mutex);
 564 
 565         if (on) {
 566                 kfree(new_on);
 567                 return 0;
 568         }
 569 
 570         /* not there, initialize a new one and retry */
 571         new_on = kmalloc(sizeof(*new_on), GFP_KERNEL);
 572         if (!new_on)
 573                 return -ENOMEM;
 574 
 575         atomic_set(&new_on->refcnt, 0);
 576         atomic_set(&new_on->event, 1);
 577         init_waitqueue_head(&new_on->poll);
 578         INIT_LIST_HEAD(&new_on->files);
 579         goto retry;
 580 }
 581 
 582 /**
 583  *      kernfs_put_open_node - put kernfs_open_node
 584  *      @kn: target kernfs_nodet
 585  *      @of: associated kernfs_open_file
 586  *
 587  *      Put @kn->attr.open and unlink @of from the files list.  If
 588  *      reference count reaches zero, disassociate and free it.
 589  *
 590  *      LOCKING:
 591  *      None.
 592  */
 593 static void kernfs_put_open_node(struct kernfs_node *kn,
 594                                  struct kernfs_open_file *of)
 595 {
 596         struct kernfs_open_node *on = kn->attr.open;
 597         unsigned long flags;
 598 
 599         mutex_lock(&kernfs_open_file_mutex);
 600         spin_lock_irqsave(&kernfs_open_node_lock, flags);
 601 
 602         if (of)
 603                 list_del(&of->list);
 604 
 605         if (atomic_dec_and_test(&on->refcnt))
 606                 kn->attr.open = NULL;
 607         else
 608                 on = NULL;
 609 
 610         spin_unlock_irqrestore(&kernfs_open_node_lock, flags);
 611         mutex_unlock(&kernfs_open_file_mutex);
 612 
 613         kfree(on);
 614 }
 615 
 616 static int kernfs_fop_open(struct inode *inode, struct file *file)
 617 {
 618         struct kernfs_node *kn = inode->i_private;
 619         struct kernfs_root *root = kernfs_root(kn);
 620         const struct kernfs_ops *ops;
 621         struct kernfs_open_file *of;
 622         bool has_read, has_write, has_mmap;
 623         int error = -EACCES;
 624 
 625         if (!kernfs_get_active(kn))
 626                 return -ENODEV;
 627 
 628         ops = kernfs_ops(kn);
 629 
 630         has_read = ops->seq_show || ops->read || ops->mmap;
 631         has_write = ops->write || ops->mmap;
 632         has_mmap = ops->mmap;
 633 
 634         /* see the flag definition for details */
 635         if (root->flags & KERNFS_ROOT_EXTRA_OPEN_PERM_CHECK) {
 636                 if ((file->f_mode & FMODE_WRITE) &&
 637                     (!(inode->i_mode & S_IWUGO) || !has_write))
 638                         goto err_out;
 639 
 640                 if ((file->f_mode & FMODE_READ) &&
 641                     (!(inode->i_mode & S_IRUGO) || !has_read))
 642                         goto err_out;
 643         }
 644 
 645         /* allocate a kernfs_open_file for the file */
 646         error = -ENOMEM;
 647         of = kzalloc(sizeof(struct kernfs_open_file), GFP_KERNEL);
 648         if (!of)
 649                 goto err_out;
 650 
 651         /*
 652          * The following is done to give a different lockdep key to
 653          * @of->mutex for files which implement mmap.  This is a rather
 654          * crude way to avoid false positive lockdep warning around
 655          * mm->mmap_sem - mmap nests @of->mutex under mm->mmap_sem and
 656          * reading /sys/block/sda/trace/act_mask grabs sr_mutex, under
 657          * which mm->mmap_sem nests, while holding @of->mutex.  As each
 658          * open file has a separate mutex, it's okay as long as those don't
 659          * happen on the same file.  At this point, we can't easily give
 660          * each file a separate locking class.  Let's differentiate on
 661          * whether the file has mmap or not for now.
 662          *
 663          * Both paths of the branch look the same.  They're supposed to
 664          * look that way and give @of->mutex different static lockdep keys.
 665          */
 666         if (has_mmap)
 667                 mutex_init(&of->mutex);
 668         else
 669                 mutex_init(&of->mutex);
 670 
 671         of->kn = kn;
 672         of->file = file;
 673 
 674         /*
 675          * Write path needs to atomic_write_len outside active reference.
 676          * Cache it in open_file.  See kernfs_fop_write() for details.
 677          */
 678         of->atomic_write_len = ops->atomic_write_len;
 679 
 680         error = -EINVAL;
 681         /*
 682          * ->seq_show is incompatible with ->prealloc,
 683          * as seq_read does its own allocation.
 684          * ->read must be used instead.
 685          */
 686         if (ops->prealloc && ops->seq_show)
 687                 goto err_free;
 688         if (ops->prealloc) {
 689                 int len = of->atomic_write_len ?: PAGE_SIZE;
 690                 of->prealloc_buf = kmalloc(len + 1, GFP_KERNEL);
 691                 error = -ENOMEM;
 692                 if (!of->prealloc_buf)
 693                         goto err_free;
 694                 mutex_init(&of->prealloc_mutex);
 695         }
 696 
 697         /*
 698          * Always instantiate seq_file even if read access doesn't use
 699          * seq_file or is not requested.  This unifies private data access
 700          * and readable regular files are the vast majority anyway.
 701          */
 702         if (ops->seq_show)
 703                 error = seq_open(file, &kernfs_seq_ops);
 704         else
 705                 error = seq_open(file, NULL);
 706         if (error)
 707                 goto err_free;
 708 
 709         of->seq_file = file->private_data;
 710         of->seq_file->private = of;
 711 
 712         /* seq_file clears PWRITE unconditionally, restore it if WRITE */
 713         if (file->f_mode & FMODE_WRITE)
 714                 file->f_mode |= FMODE_PWRITE;
 715 
 716         /* make sure we have open node struct */
 717         error = kernfs_get_open_node(kn, of);
 718         if (error)
 719                 goto err_seq_release;
 720 
 721         if (ops->open) {
 722                 /* nobody has access to @of yet, skip @of->mutex */
 723                 error = ops->open(of);
 724                 if (error)
 725                         goto err_put_node;
 726         }
 727 
 728         /* open succeeded, put active references */
 729         kernfs_put_active(kn);
 730         return 0;
 731 
 732 err_put_node:
 733         kernfs_put_open_node(kn, of);
 734 err_seq_release:
 735         seq_release(inode, file);
 736 err_free:
 737         kfree(of->prealloc_buf);
 738         kfree(of);
 739 err_out:
 740         kernfs_put_active(kn);
 741         return error;
 742 }
 743 
 744 /* used from release/drain to ensure that ->release() is called exactly once */
 745 static void kernfs_release_file(struct kernfs_node *kn,
 746                                 struct kernfs_open_file *of)
 747 {
 748         /*
 749          * @of is guaranteed to have no other file operations in flight and
 750          * we just want to synchronize release and drain paths.
 751          * @kernfs_open_file_mutex is enough.  @of->mutex can't be used
 752          * here because drain path may be called from places which can
 753          * cause circular dependency.
 754          */
 755         lockdep_assert_held(&kernfs_open_file_mutex);
 756 
 757         if (!of->released) {
 758                 /*
 759                  * A file is never detached without being released and we
 760                  * need to be able to release files which are deactivated
 761                  * and being drained.  Don't use kernfs_ops().
 762                  */
 763                 kn->attr.ops->release(of);
 764                 of->released = true;
 765         }
 766 }
 767 
 768 static int kernfs_fop_release(struct inode *inode, struct file *filp)
 769 {
 770         struct kernfs_node *kn = inode->i_private;
 771         struct kernfs_open_file *of = kernfs_of(filp);
 772 
 773         if (kn->flags & KERNFS_HAS_RELEASE) {
 774                 mutex_lock(&kernfs_open_file_mutex);
 775                 kernfs_release_file(kn, of);
 776                 mutex_unlock(&kernfs_open_file_mutex);
 777         }
 778 
 779         kernfs_put_open_node(kn, of);
 780         seq_release(inode, filp);
 781         kfree(of->prealloc_buf);
 782         kfree(of);
 783 
 784         return 0;
 785 }
 786 
 787 void kernfs_drain_open_files(struct kernfs_node *kn)
 788 {
 789         struct kernfs_open_node *on;
 790         struct kernfs_open_file *of;
 791 
 792         if (!(kn->flags & (KERNFS_HAS_MMAP | KERNFS_HAS_RELEASE)))
 793                 return;
 794 
 795         spin_lock_irq(&kernfs_open_node_lock);
 796         on = kn->attr.open;
 797         if (on)
 798                 atomic_inc(&on->refcnt);
 799         spin_unlock_irq(&kernfs_open_node_lock);
 800         if (!on)
 801                 return;
 802 
 803         mutex_lock(&kernfs_open_file_mutex);
 804 
 805         list_for_each_entry(of, &on->files, list) {
 806                 struct inode *inode = file_inode(of->file);
 807 
 808                 if (kn->flags & KERNFS_HAS_MMAP)
 809                         unmap_mapping_range(inode->i_mapping, 0, 0, 1);
 810 
 811                 if (kn->flags & KERNFS_HAS_RELEASE)
 812                         kernfs_release_file(kn, of);
 813         }
 814 
 815         mutex_unlock(&kernfs_open_file_mutex);
 816 
 817         kernfs_put_open_node(kn, NULL);
 818 }
 819 
 820 /*
 821  * Kernfs attribute files are pollable.  The idea is that you read
 822  * the content and then you use 'poll' or 'select' to wait for
 823  * the content to change.  When the content changes (assuming the
 824  * manager for the kobject supports notification), poll will
 825  * return EPOLLERR|EPOLLPRI, and select will return the fd whether
 826  * it is waiting for read, write, or exceptions.
 827  * Once poll/select indicates that the value has changed, you
 828  * need to close and re-open the file, or seek to 0 and read again.
 829  * Reminder: this only works for attributes which actively support
 830  * it, and it is not possible to test an attribute from userspace
 831  * to see if it supports poll (Neither 'poll' nor 'select' return
 832  * an appropriate error code).  When in doubt, set a suitable timeout value.
 833  */
 834 __poll_t kernfs_generic_poll(struct kernfs_open_file *of, poll_table *wait)
 835 {
 836         struct kernfs_node *kn = kernfs_dentry_node(of->file->f_path.dentry);
 837         struct kernfs_open_node *on = kn->attr.open;
 838 
 839         poll_wait(of->file, &on->poll, wait);
 840 
 841         if (of->event != atomic_read(&on->event))
 842                 return DEFAULT_POLLMASK|EPOLLERR|EPOLLPRI;
 843 
 844         return DEFAULT_POLLMASK;
 845 }
 846 
 847 static __poll_t kernfs_fop_poll(struct file *filp, poll_table *wait)
 848 {
 849         struct kernfs_open_file *of = kernfs_of(filp);
 850         struct kernfs_node *kn = kernfs_dentry_node(filp->f_path.dentry);
 851         __poll_t ret;
 852 
 853         if (!kernfs_get_active(kn))
 854                 return DEFAULT_POLLMASK|EPOLLERR|EPOLLPRI;
 855 
 856         if (kn->attr.ops->poll)
 857                 ret = kn->attr.ops->poll(of, wait);
 858         else
 859                 ret = kernfs_generic_poll(of, wait);
 860 
 861         kernfs_put_active(kn);
 862         return ret;
 863 }
 864 
 865 static void kernfs_notify_workfn(struct work_struct *work)
 866 {
 867         struct kernfs_node *kn;
 868         struct kernfs_super_info *info;
 869 repeat:
 870         /* pop one off the notify_list */
 871         spin_lock_irq(&kernfs_notify_lock);
 872         kn = kernfs_notify_list;
 873         if (kn == KERNFS_NOTIFY_EOL) {
 874                 spin_unlock_irq(&kernfs_notify_lock);
 875                 return;
 876         }
 877         kernfs_notify_list = kn->attr.notify_next;
 878         kn->attr.notify_next = NULL;
 879         spin_unlock_irq(&kernfs_notify_lock);
 880 
 881         /* kick fsnotify */
 882         mutex_lock(&kernfs_mutex);
 883 
 884         list_for_each_entry(info, &kernfs_root(kn)->supers, node) {
 885                 struct kernfs_node *parent;
 886                 struct inode *inode;
 887                 struct qstr name;
 888 
 889                 /*
 890                  * We want fsnotify_modify() on @kn but as the
 891                  * modifications aren't originating from userland don't
 892                  * have the matching @file available.  Look up the inodes
 893                  * and generate the events manually.
 894                  */
 895                 inode = ilookup(info->sb, kn->id.ino);
 896                 if (!inode)
 897                         continue;
 898 
 899                 name = (struct qstr)QSTR_INIT(kn->name, strlen(kn->name));
 900                 parent = kernfs_get_parent(kn);
 901                 if (parent) {
 902                         struct inode *p_inode;
 903 
 904                         p_inode = ilookup(info->sb, parent->id.ino);
 905                         if (p_inode) {
 906                                 fsnotify(p_inode, FS_MODIFY | FS_EVENT_ON_CHILD,
 907                                          inode, FSNOTIFY_EVENT_INODE, &name, 0);
 908                                 iput(p_inode);
 909                         }
 910 
 911                         kernfs_put(parent);
 912                 }
 913 
 914                 fsnotify(inode, FS_MODIFY, inode, FSNOTIFY_EVENT_INODE,
 915                          &name, 0);
 916                 iput(inode);
 917         }
 918 
 919         mutex_unlock(&kernfs_mutex);
 920         kernfs_put(kn);
 921         goto repeat;
 922 }
 923 
 924 /**
 925  * kernfs_notify - notify a kernfs file
 926  * @kn: file to notify
 927  *
 928  * Notify @kn such that poll(2) on @kn wakes up.  Maybe be called from any
 929  * context.
 930  */
 931 void kernfs_notify(struct kernfs_node *kn)
 932 {
 933         static DECLARE_WORK(kernfs_notify_work, kernfs_notify_workfn);
 934         unsigned long flags;
 935         struct kernfs_open_node *on;
 936 
 937         if (WARN_ON(kernfs_type(kn) != KERNFS_FILE))
 938                 return;
 939 
 940         /* kick poll immediately */
 941         spin_lock_irqsave(&kernfs_open_node_lock, flags);
 942         on = kn->attr.open;
 943         if (on) {
 944                 atomic_inc(&on->event);
 945                 wake_up_interruptible(&on->poll);
 946         }
 947         spin_unlock_irqrestore(&kernfs_open_node_lock, flags);
 948 
 949         /* schedule work to kick fsnotify */
 950         spin_lock_irqsave(&kernfs_notify_lock, flags);
 951         if (!kn->attr.notify_next) {
 952                 kernfs_get(kn);
 953                 kn->attr.notify_next = kernfs_notify_list;
 954                 kernfs_notify_list = kn;
 955                 schedule_work(&kernfs_notify_work);
 956         }
 957         spin_unlock_irqrestore(&kernfs_notify_lock, flags);
 958 }
 959 EXPORT_SYMBOL_GPL(kernfs_notify);
 960 
 961 const struct file_operations kernfs_file_fops = {
 962         .read           = kernfs_fop_read,
 963         .write          = kernfs_fop_write,
 964         .llseek         = generic_file_llseek,
 965         .mmap           = kernfs_fop_mmap,
 966         .open           = kernfs_fop_open,
 967         .release        = kernfs_fop_release,
 968         .poll           = kernfs_fop_poll,
 969         .fsync          = noop_fsync,
 970 };
 971 
 972 /**
 973  * __kernfs_create_file - kernfs internal function to create a file
 974  * @parent: directory to create the file in
 975  * @name: name of the file
 976  * @mode: mode of the file
 977  * @uid: uid of the file
 978  * @gid: gid of the file
 979  * @size: size of the file
 980  * @ops: kernfs operations for the file
 981  * @priv: private data for the file
 982  * @ns: optional namespace tag of the file
 983  * @key: lockdep key for the file's active_ref, %NULL to disable lockdep
 984  *
 985  * Returns the created node on success, ERR_PTR() value on error.
 986  */
 987 struct kernfs_node *__kernfs_create_file(struct kernfs_node *parent,
 988                                          const char *name,
 989                                          umode_t mode, kuid_t uid, kgid_t gid,
 990                                          loff_t size,
 991                                          const struct kernfs_ops *ops,
 992                                          void *priv, const void *ns,
 993                                          struct lock_class_key *key)
 994 {
 995         struct kernfs_node *kn;
 996         unsigned flags;
 997         int rc;
 998 
 999         flags = KERNFS_FILE;
1000 
1001         kn = kernfs_new_node(parent, name, (mode & S_IALLUGO) | S_IFREG,
1002                              uid, gid, flags);
1003         if (!kn)
1004                 return ERR_PTR(-ENOMEM);
1005 
1006         kn->attr.ops = ops;
1007         kn->attr.size = size;
1008         kn->ns = ns;
1009         kn->priv = priv;
1010 
1011 #ifdef CONFIG_DEBUG_LOCK_ALLOC
1012         if (key) {
1013                 lockdep_init_map(&kn->dep_map, "kn->count", key, 0);
1014                 kn->flags |= KERNFS_LOCKDEP;
1015         }
1016 #endif
1017 
1018         /*
1019          * kn->attr.ops is accesible only while holding active ref.  We
1020          * need to know whether some ops are implemented outside active
1021          * ref.  Cache their existence in flags.
1022          */
1023         if (ops->seq_show)
1024                 kn->flags |= KERNFS_HAS_SEQ_SHOW;
1025         if (ops->mmap)
1026                 kn->flags |= KERNFS_HAS_MMAP;
1027         if (ops->release)
1028                 kn->flags |= KERNFS_HAS_RELEASE;
1029 
1030         rc = kernfs_add_one(kn);
1031         if (rc) {
1032                 kernfs_put(kn);
1033                 return ERR_PTR(rc);
1034         }
1035         return kn;
1036 }