dlmfs.c

  1// SPDX-License-Identifier: GPL-2.0-or-later
  2/*
  3 * dlmfs.c
  4 *
  5 * Code which implements the kernel side of a minimal userspace
  6 * interface to our DLM. This file handles the virtual file system
  7 * used for communication with userspace. Credit should go to ramfs,
  8 * which was a template for the fs side of this module.
  9 *
 10 * Copyright (C) 2003, 2004 Oracle.  All rights reserved.
 11 */
 12
 13/* Simple VFS hooks based on: */
 14/*
 15 * Resizable simple ram filesystem for Linux.
 16 *
 17 * Copyright (C) 2000 Linus Torvalds.
 18 *               2000 Transmeta Corp.
 19 */
 20
 21#include <linux/module.h>
 22#include <linux/fs.h>
 23#include <linux/fs_context.h>
 24#include <linux/pagemap.h>
 25#include <linux/types.h>
 26#include <linux/slab.h>
 27#include <linux/highmem.h>
 28#include <linux/init.h>
 29#include <linux/string.h>
 30#include <linux/backing-dev.h>
 31#include <linux/poll.h>
 32
 33#include <linux/uaccess.h>
 34
 35#include "../stackglue.h"
 36#include "userdlm.h"
 37
 38#define MLOG_MASK_PREFIX ML_DLMFS
 39#include "../cluster/masklog.h"
 40
 41
 42static const struct super_operations dlmfs_ops;
 43static const struct file_operations dlmfs_file_operations;
 44static const struct inode_operations dlmfs_dir_inode_operations;
 45static const struct inode_operations dlmfs_root_inode_operations;
 46static const struct inode_operations dlmfs_file_inode_operations;
 47static struct kmem_cache *dlmfs_inode_cache;
 48
 49struct workqueue_struct *user_dlm_worker;
 50
 51
 52
 53/*
 54 * These are the ABI capabilities of dlmfs.
 55 *
 56 * Over time, dlmfs has added some features that were not part of the
 57 * initial ABI.  Unfortunately, some of these features are not detectable
 58 * via standard usage.  For example, Linux's default poll always returns
 59 * EPOLLIN, so there is no way for a caller of poll(2) to know when dlmfs
 60 * added poll support.  Instead, we provide this list of new capabilities.
 61 *
 62 * Capabilities is a read-only attribute.  We do it as a module parameter
 63 * so we can discover it whether dlmfs is built in, loaded, or even not
 64 * loaded.
 65 *
 66 * The ABI features are local to this machine's dlmfs mount.  This is
 67 * distinct from the locking protocol, which is concerned with inter-node
 68 * interaction.
 69 *
 70 * Capabilities:
 71 * - bast	: EPOLLIN against the file descriptor of a held lock
 72 *		  signifies a bast fired on the lock.
 73 */
 74#define DLMFS_CAPABILITIES "bast stackglue"
 75static int param_set_dlmfs_capabilities(const char *val,
 76					const struct kernel_param *kp)
 77{
 78	printk(KERN_ERR "%s: readonly parameter\n", kp->name);
 79	return -EINVAL;
 80}
 81static int param_get_dlmfs_capabilities(char *buffer,
 82					const struct kernel_param *kp)
 83{
 84	return sysfs_emit(buffer, DLMFS_CAPABILITIES);
 85}
 86module_param_call(capabilities, param_set_dlmfs_capabilities,
 87		  param_get_dlmfs_capabilities, NULL, 0444);
 88MODULE_PARM_DESC(capabilities, DLMFS_CAPABILITIES);
 89
 90
 91/*
 92 * decodes a set of open flags into a valid lock level and a set of flags.
 93 * returns < 0 if we have invalid flags
 94 * flags which mean something to us:
 95 * O_RDONLY -> PRMODE level
 96 * O_WRONLY -> EXMODE level
 97 *
 98 * O_NONBLOCK -> NOQUEUE
 99 */
100static int dlmfs_decode_open_flags(int open_flags,
101				   int *level,
102				   int *flags)
103{
104	if (open_flags & (O_WRONLY|O_RDWR))
105		*level = DLM_LOCK_EX;
106	else
107		*level = DLM_LOCK_PR;
108
109	*flags = 0;
110	if (open_flags & O_NONBLOCK)
111		*flags |= DLM_LKF_NOQUEUE;
112
113	return 0;
114}
115
116static int dlmfs_file_open(struct inode *inode,
117			   struct file *file)
118{
119	int status, level, flags;
120	struct dlmfs_filp_private *fp = NULL;
121	struct dlmfs_inode_private *ip;
122
123	if (S_ISDIR(inode->i_mode))
124		BUG();
125
126	mlog(0, "open called on inode %lu, flags 0x%x\n", inode->i_ino,
127		file->f_flags);
128
129	status = dlmfs_decode_open_flags(file->f_flags, &level, &flags);
130	if (status < 0)
131		goto bail;
132
133	/* We don't want to honor O_APPEND at read/write time as it
134	 * doesn't make sense for LVB writes. */
135	file->f_flags &= ~O_APPEND;
136
137	fp = kmalloc(sizeof(*fp), GFP_NOFS);
138	if (!fp) {
139		status = -ENOMEM;
140		goto bail;
141	}
142	fp->fp_lock_level = level;
143
144	ip = DLMFS_I(inode);
145
146	status = user_dlm_cluster_lock(&ip->ip_lockres, level, flags);
147	if (status < 0) {
148		/* this is a strange error to return here but I want
149		 * to be able userspace to be able to distinguish a
150		 * valid lock request from one that simply couldn't be
151		 * granted. */
152		if (flags & DLM_LKF_NOQUEUE && status == -EAGAIN)
153			status = -ETXTBSY;
154		kfree(fp);
155		goto bail;
156	}
157
158	file->private_data = fp;
159bail:
160	return status;
161}
162
163static int dlmfs_file_release(struct inode *inode,
164			      struct file *file)
165{
166	int level;
167	struct dlmfs_inode_private *ip = DLMFS_I(inode);
168	struct dlmfs_filp_private *fp = file->private_data;
169
170	if (S_ISDIR(inode->i_mode))
171		BUG();
172
173	mlog(0, "close called on inode %lu\n", inode->i_ino);
174
175	if (fp) {
176		level = fp->fp_lock_level;
177		if (level != DLM_LOCK_IV)
178			user_dlm_cluster_unlock(&ip->ip_lockres, level);
179
180		kfree(fp);
181		file->private_data = NULL;
182	}
183
184	return 0;
185}
186
187/*
188 * We do ->setattr() just to override size changes.  Our size is the size
189 * of the LVB and nothing else.
190 */
191static int dlmfs_file_setattr(struct mnt_idmap *idmap,
192			      struct dentry *dentry, struct iattr *attr)
193{
194	int error;
195	struct inode *inode = d_inode(dentry);
196
197	attr->ia_valid &= ~ATTR_SIZE;
198	error = setattr_prepare(&nop_mnt_idmap, dentry, attr);
199	if (error)
200		return error;
201
202	setattr_copy(&nop_mnt_idmap, inode, attr);
203	mark_inode_dirty(inode);
204	return 0;
205}
206
207static __poll_t dlmfs_file_poll(struct file *file, poll_table *wait)
208{
209	__poll_t event = 0;
210	struct inode *inode = file_inode(file);
211	struct dlmfs_inode_private *ip = DLMFS_I(inode);
212
213	poll_wait(file, &ip->ip_lockres.l_event, wait);
214
215	spin_lock(&ip->ip_lockres.l_lock);
216	if (ip->ip_lockres.l_flags & USER_LOCK_BLOCKED)
217		event = EPOLLIN | EPOLLRDNORM;
218	spin_unlock(&ip->ip_lockres.l_lock);
219
220	return event;
221}
222
223static ssize_t dlmfs_file_read(struct file *file,
224			       char __user *buf,
225			       size_t count,
226			       loff_t *ppos)
227{
228	char lvb[DLM_LVB_LEN];
229
230	if (!user_dlm_read_lvb(file_inode(file), lvb))
231		return 0;
232
233	return simple_read_from_buffer(buf, count, ppos, lvb, sizeof(lvb));
234}
235
236static ssize_t dlmfs_file_write(struct file *filp,
237				const char __user *buf,
238				size_t count,
239				loff_t *ppos)
240{
241	char lvb_buf[DLM_LVB_LEN];
242	int bytes_left;
243	struct inode *inode = file_inode(filp);
244
245	mlog(0, "inode %lu, count = %zu, *ppos = %llu\n",
246		inode->i_ino, count, *ppos);
247
248	if (*ppos >= DLM_LVB_LEN)
249		return -ENOSPC;
250
251	/* don't write past the lvb */
252	if (count > DLM_LVB_LEN - *ppos)
253		count = DLM_LVB_LEN - *ppos;
254
255	if (!count)
256		return 0;
257
258	bytes_left = copy_from_user(lvb_buf, buf, count);
259	count -= bytes_left;
260	if (count)
261		user_dlm_write_lvb(inode, lvb_buf, count);
262
263	*ppos = *ppos + count;
264	mlog(0, "wrote %zu bytes\n", count);
265	return count;
266}
267
268static void dlmfs_init_once(void *foo)
269{
270	struct dlmfs_inode_private *ip =
271		(struct dlmfs_inode_private *) foo;
272
273	ip->ip_conn = NULL;
274	ip->ip_parent = NULL;
275
276	inode_init_once(&ip->ip_vfs_inode);
277}
278
279static struct inode *dlmfs_alloc_inode(struct super_block *sb)
280{
281	struct dlmfs_inode_private *ip;
282
283	ip = alloc_inode_sb(sb, dlmfs_inode_cache, GFP_NOFS);
284	if (!ip)
285		return NULL;
286
287	return &ip->ip_vfs_inode;
288}
289
290static void dlmfs_free_inode(struct inode *inode)
291{
292	kmem_cache_free(dlmfs_inode_cache, DLMFS_I(inode));
293}
294
295static void dlmfs_evict_inode(struct inode *inode)
296{
297	int status;
298	struct dlmfs_inode_private *ip;
299	struct user_lock_res *lockres;
300	int teardown;
301
302	clear_inode(inode);
303
304	mlog(0, "inode %lu\n", inode->i_ino);
305
306	ip = DLMFS_I(inode);
307	lockres = &ip->ip_lockres;
308
309	if (S_ISREG(inode->i_mode)) {
310		spin_lock(&lockres->l_lock);
311		teardown = !!(lockres->l_flags & USER_LOCK_IN_TEARDOWN);
312		spin_unlock(&lockres->l_lock);
313		if (!teardown) {
314			status = user_dlm_destroy_lock(lockres);
315			if (status < 0)
316				mlog_errno(status);
317		}
318		iput(ip->ip_parent);
319		goto clear_fields;
320	}
321
322	mlog(0, "we're a directory, ip->ip_conn = 0x%p\n", ip->ip_conn);
323	/* we must be a directory. If required, lets unregister the
324	 * dlm context now. */
325	if (ip->ip_conn)
326		user_dlm_unregister(ip->ip_conn);
327clear_fields:
328	ip->ip_parent = NULL;
329	ip->ip_conn = NULL;
330}
331
332static struct inode *dlmfs_get_root_inode(struct super_block *sb)
333{
334	struct inode *inode = new_inode(sb);
335	umode_t mode = S_IFDIR | 0755;
336
337	if (inode) {
338		inode->i_ino = get_next_ino();
339		inode_init_owner(&nop_mnt_idmap, inode, NULL, mode);
340		simple_inode_init_ts(inode);
341		inc_nlink(inode);
342
343		inode->i_fop = &simple_dir_operations;
344		inode->i_op = &dlmfs_root_inode_operations;
345	}
346
347	return inode;
348}
349
350static struct inode *dlmfs_get_inode(struct inode *parent,
351				     struct dentry *dentry,
352				     umode_t mode)
353{
354	struct super_block *sb = parent->i_sb;
355	struct inode * inode = new_inode(sb);
356	struct dlmfs_inode_private *ip;
357
358	if (!inode)
359		return NULL;
360
361	inode->i_ino = get_next_ino();
362	inode_init_owner(&nop_mnt_idmap, inode, parent, mode);
363	simple_inode_init_ts(inode);
364
365	ip = DLMFS_I(inode);
366	ip->ip_conn = DLMFS_I(parent)->ip_conn;
367
368	switch (mode & S_IFMT) {
369	default:
370		/* for now we don't support anything other than
371		 * directories and regular files. */
372		BUG();
373		break;
374	case S_IFREG:
375		inode->i_op = &dlmfs_file_inode_operations;
376		inode->i_fop = &dlmfs_file_operations;
377
378		i_size_write(inode,  DLM_LVB_LEN);
379
380		user_dlm_lock_res_init(&ip->ip_lockres, dentry);
381
382		/* released at clear_inode time, this insures that we
383		 * get to drop the dlm reference on each lock *before*
384		 * we call the unregister code for releasing parent
385		 * directories. */
386		ip->ip_parent = igrab(parent);
387		BUG_ON(!ip->ip_parent);
388		break;
389	case S_IFDIR:
390		inode->i_op = &dlmfs_dir_inode_operations;
391		inode->i_fop = &simple_dir_operations;
392
393		/* directory inodes start off with i_nlink ==
394		 * 2 (for "." entry) */
395		inc_nlink(inode);
396		break;
397	}
398	return inode;
399}
400
401/*
402 * File creation. Allocate an inode, and we're done..
403 */
404/* SMP-safe */
405static struct dentry *dlmfs_mkdir(struct mnt_idmap * idmap,
406				  struct inode * dir,
407				  struct dentry * dentry,
408				  umode_t mode)
409{
410	int status;
411	struct inode *inode = NULL;
412	const struct qstr *domain = &dentry->d_name;
413	struct dlmfs_inode_private *ip;
414	struct ocfs2_cluster_connection *conn;
415
416	mlog(0, "mkdir %.*s\n", domain->len, domain->name);
417
418	/* verify that we have a proper domain */
419	if (domain->len >= GROUP_NAME_MAX) {
420		status = -EINVAL;
421		mlog(ML_ERROR, "invalid domain name for directory.\n");
422		goto bail;
423	}
424
425	inode = dlmfs_get_inode(dir, dentry, mode | S_IFDIR);
426	if (!inode) {
427		status = -ENOMEM;
428		mlog_errno(status);
429		goto bail;
430	}
431
432	ip = DLMFS_I(inode);
433
434	conn = user_dlm_register(domain);
435	if (IS_ERR(conn)) {
436		status = PTR_ERR(conn);
437		mlog(ML_ERROR, "Error %d could not register domain \"%.*s\"\n",
438		     status, domain->len, domain->name);
439		goto bail;
440	}
441	ip->ip_conn = conn;
442
443	inc_nlink(dir);
444	d_make_persistent(dentry, inode);
445
446	status = 0;
447bail:
448	if (status < 0)
449		iput(inode);
450	return ERR_PTR(status);
451}
452
453static int dlmfs_create(struct mnt_idmap *idmap,
454			struct inode *dir,
455			struct dentry *dentry,
456			umode_t mode,
457			bool excl)
458{
459	int status = 0;
460	struct inode *inode;
461	const struct qstr *name = &dentry->d_name;
462
463	mlog(0, "create %.*s\n", name->len, name->name);
464
465	/* verify name is valid and doesn't contain any dlm reserved
466	 * characters */
467	if (name->len >= USER_DLM_LOCK_ID_MAX_LEN ||
468	    name->name[0] == '$') {
469		status = -EINVAL;
470		mlog(ML_ERROR, "invalid lock name, %.*s\n", name->len,
471		     name->name);
472		goto bail;
473	}
474
475	inode = dlmfs_get_inode(dir, dentry, mode | S_IFREG);
476	if (!inode) {
477		status = -ENOMEM;
478		mlog_errno(status);
479		goto bail;
480	}
481
482	d_make_persistent(dentry, inode);
483bail:
484	return status;
485}
486
487static int dlmfs_unlink(struct inode *dir,
488			struct dentry *dentry)
489{
490	int status;
491	struct inode *inode = d_inode(dentry);
492
493	mlog(0, "unlink inode %lu\n", inode->i_ino);
494
495	/* if there are no current holders, or none that are waiting
496	 * to acquire a lock, this basically destroys our lockres. */
497	status = user_dlm_destroy_lock(&DLMFS_I(inode)->ip_lockres);
498	if (status < 0) {
499		mlog(ML_ERROR, "unlink %pd, error %d from destroy\n",
500		     dentry, status);
501		goto bail;
502	}
503	status = simple_unlink(dir, dentry);
504bail:
505	return status;
506}
507
508static int dlmfs_fill_super(struct super_block *sb, struct fs_context *fc)
509{
510	sb->s_maxbytes = MAX_LFS_FILESIZE;
511	sb->s_blocksize = PAGE_SIZE;
512	sb->s_blocksize_bits = PAGE_SHIFT;
513	sb->s_magic = DLMFS_MAGIC;
514	sb->s_op = &dlmfs_ops;
515	sb->s_root = d_make_root(dlmfs_get_root_inode(sb));
516	if (!sb->s_root)
517		return -ENOMEM;
518	return 0;
519}
520
521static const struct file_operations dlmfs_file_operations = {
522	.open		= dlmfs_file_open,
523	.release	= dlmfs_file_release,
524	.poll		= dlmfs_file_poll,
525	.read		= dlmfs_file_read,
526	.write		= dlmfs_file_write,
527	.llseek		= default_llseek,
528};
529
530static const struct inode_operations dlmfs_dir_inode_operations = {
531	.create		= dlmfs_create,
532	.lookup		= simple_lookup,
533	.unlink		= dlmfs_unlink,
534};
535
536/* this way we can restrict mkdir to only the toplevel of the fs. */
537static const struct inode_operations dlmfs_root_inode_operations = {
538	.lookup		= simple_lookup,
539	.mkdir		= dlmfs_mkdir,
540	.rmdir		= simple_rmdir,
541};
542
543static const struct super_operations dlmfs_ops = {
544	.statfs		= simple_statfs,
545	.alloc_inode	= dlmfs_alloc_inode,
546	.free_inode	= dlmfs_free_inode,
547	.evict_inode	= dlmfs_evict_inode,
548	.drop_inode	= inode_just_drop,
549};
550
551static const struct inode_operations dlmfs_file_inode_operations = {
552	.getattr	= simple_getattr,
553	.setattr	= dlmfs_file_setattr,
554};
555
556static int dlmfs_get_tree(struct fs_context *fc)
557{
558	return get_tree_nodev(fc, dlmfs_fill_super);
559}
560
561static const struct fs_context_operations dlmfs_context_ops = {
562	.get_tree       = dlmfs_get_tree,
563};
564
565static int dlmfs_init_fs_context(struct fs_context *fc)
566{
567	fc->ops = &dlmfs_context_ops;
568
569	return 0;
570}
571
572static struct file_system_type dlmfs_fs_type = {
573	.owner		= THIS_MODULE,
574	.name		= "ocfs2_dlmfs",
575	.kill_sb	= kill_anon_super,
576	.init_fs_context = dlmfs_init_fs_context,
577};
578MODULE_ALIAS_FS("ocfs2_dlmfs");
579
580static int __init init_dlmfs_fs(void)
581{
582	int status;
583	int cleanup_inode = 0, cleanup_worker = 0;
584
585	dlmfs_inode_cache = kmem_cache_create("dlmfs_inode_cache",
586				sizeof(struct dlmfs_inode_private),
587				0, (SLAB_HWCACHE_ALIGN|SLAB_RECLAIM_ACCOUNT|
588					SLAB_ACCOUNT),
589				dlmfs_init_once);
590	if (!dlmfs_inode_cache) {
591		status = -ENOMEM;
592		goto bail;
593	}
594	cleanup_inode = 1;
595
596	user_dlm_worker = alloc_workqueue("user_dlm",
597					  WQ_MEM_RECLAIM | WQ_PERCPU, 0);
598	if (!user_dlm_worker) {
599		status = -ENOMEM;
600		goto bail;
601	}
602	cleanup_worker = 1;
603
604	user_dlm_set_locking_protocol();
605	status = register_filesystem(&dlmfs_fs_type);
606bail:
607	if (status) {
608		if (cleanup_inode)
609			kmem_cache_destroy(dlmfs_inode_cache);
610		if (cleanup_worker)
611			destroy_workqueue(user_dlm_worker);
612	} else
613		printk("OCFS2 User DLM kernel interface loaded\n");
614	return status;
615}
616
617static void __exit exit_dlmfs_fs(void)
618{
619	unregister_filesystem(&dlmfs_fs_type);
620
621	destroy_workqueue(user_dlm_worker);
622
623	/*
624	 * Make sure all delayed rcu free inodes are flushed before we
625	 * destroy cache.
626	 */
627	rcu_barrier();
628	kmem_cache_destroy(dlmfs_inode_cache);
629
630}
631
632MODULE_AUTHOR("Oracle");
633MODULE_LICENSE("GPL");
634MODULE_DESCRIPTION("OCFS2 DLM-Filesystem");
635
636module_init(init_dlmfs_fs)
637module_exit(exit_dlmfs_fs)