xref: /openbmc/linux/fs/orangefs/devorangefs-req.c (revision 430ff791)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * (C) 2001 Clemson University and The University of Chicago
4  *
5  * Changes by Acxiom Corporation to add protocol version to kernel
6  * communication, Copyright Acxiom Corporation, 2005.
7  *
8  * See COPYING in top-level directory.
9  */
10 
11 #include "protocol.h"
12 #include "orangefs-kernel.h"
13 #include "orangefs-dev-proto.h"
14 #include "orangefs-bufmap.h"
15 #include "orangefs-debugfs.h"
16 
17 #include <linux/debugfs.h>
18 #include <linux/slab.h>
19 
20 /* this file implements the /dev/pvfs2-req device node */
21 
22 uint32_t orangefs_userspace_version;
23 
24 static int open_access_count;
25 
26 static DEFINE_MUTEX(devreq_mutex);
27 
28 #define DUMP_DEVICE_ERROR()                                                   \
29 do {                                                                          \
30 	gossip_err("*****************************************************\n");\
31 	gossip_err("ORANGEFS Device Error:  You cannot open the device file ");  \
32 	gossip_err("\n/dev/%s more than once.  Please make sure that\nthere " \
33 		   "are no ", ORANGEFS_REQDEVICE_NAME);                          \
34 	gossip_err("instances of a program using this device\ncurrently "     \
35 		   "running. (You must verify this!)\n");                     \
36 	gossip_err("For example, you can use the lsof program as follows:\n");\
37 	gossip_err("'lsof | grep %s' (run this as root)\n",                   \
38 		   ORANGEFS_REQDEVICE_NAME);                                     \
39 	gossip_err("  open_access_count = %d\n", open_access_count);          \
40 	gossip_err("*****************************************************\n");\
41 } while (0)
42 
hash_func(__u64 tag,int table_size)43 static int hash_func(__u64 tag, int table_size)
44 {
45 	return do_div(tag, (unsigned int)table_size);
46 }
47 
orangefs_devreq_add_op(struct orangefs_kernel_op_s * op)48 static void orangefs_devreq_add_op(struct orangefs_kernel_op_s *op)
49 {
50 	int index = hash_func(op->tag, hash_table_size);
51 
52 	list_add_tail(&op->list, &orangefs_htable_ops_in_progress[index]);
53 }
54 
55 /*
56  * find the op with this tag and remove it from the in progress
57  * hash table.
58  */
orangefs_devreq_remove_op(__u64 tag)59 static struct orangefs_kernel_op_s *orangefs_devreq_remove_op(__u64 tag)
60 {
61 	struct orangefs_kernel_op_s *op, *next;
62 	int index;
63 
64 	index = hash_func(tag, hash_table_size);
65 
66 	spin_lock(&orangefs_htable_ops_in_progress_lock);
67 	list_for_each_entry_safe(op,
68 				 next,
69 				 &orangefs_htable_ops_in_progress[index],
70 				 list) {
71 		if (op->tag == tag && !op_state_purged(op) &&
72 		    !op_state_given_up(op)) {
73 			list_del_init(&op->list);
74 			spin_unlock(&orangefs_htable_ops_in_progress_lock);
75 			return op;
76 		}
77 	}
78 
79 	spin_unlock(&orangefs_htable_ops_in_progress_lock);
80 	return NULL;
81 }
82 
83 /* Returns whether any FS are still pending remounted */
mark_all_pending_mounts(void)84 static int mark_all_pending_mounts(void)
85 {
86 	int unmounted = 1;
87 	struct orangefs_sb_info_s *orangefs_sb = NULL;
88 
89 	spin_lock(&orangefs_superblocks_lock);
90 	list_for_each_entry(orangefs_sb, &orangefs_superblocks, list) {
91 		/* All of these file system require a remount */
92 		orangefs_sb->mount_pending = 1;
93 		unmounted = 0;
94 	}
95 	spin_unlock(&orangefs_superblocks_lock);
96 	return unmounted;
97 }
98 
99 /*
100  * Determine if a given file system needs to be remounted or not
101  *  Returns -1 on error
102  *           0 if already mounted
103  *           1 if needs remount
104  */
fs_mount_pending(__s32 fsid)105 static int fs_mount_pending(__s32 fsid)
106 {
107 	int mount_pending = -1;
108 	struct orangefs_sb_info_s *orangefs_sb = NULL;
109 
110 	spin_lock(&orangefs_superblocks_lock);
111 	list_for_each_entry(orangefs_sb, &orangefs_superblocks, list) {
112 		if (orangefs_sb->fs_id == fsid) {
113 			mount_pending = orangefs_sb->mount_pending;
114 			break;
115 		}
116 	}
117 	spin_unlock(&orangefs_superblocks_lock);
118 	return mount_pending;
119 }
120 
orangefs_devreq_open(struct inode * inode,struct file * file)121 static int orangefs_devreq_open(struct inode *inode, struct file *file)
122 {
123 	int ret = -EINVAL;
124 
125 	/* in order to ensure that the filesystem driver sees correct UIDs */
126 	if (file->f_cred->user_ns != &init_user_ns) {
127 		gossip_err("%s: device cannot be opened outside init_user_ns\n",
128 			   __func__);
129 		goto out;
130 	}
131 
132 	if (!(file->f_flags & O_NONBLOCK)) {
133 		gossip_err("%s: device cannot be opened in blocking mode\n",
134 			   __func__);
135 		goto out;
136 	}
137 	ret = -EACCES;
138 	gossip_debug(GOSSIP_DEV_DEBUG, "client-core: opening device\n");
139 	mutex_lock(&devreq_mutex);
140 
141 	if (open_access_count == 0) {
142 		open_access_count = 1;
143 		ret = 0;
144 	} else {
145 		DUMP_DEVICE_ERROR();
146 	}
147 	mutex_unlock(&devreq_mutex);
148 
149 out:
150 
151 	gossip_debug(GOSSIP_DEV_DEBUG,
152 		     "pvfs2-client-core: open device complete (ret = %d)\n",
153 		     ret);
154 	return ret;
155 }
156 
157 /* Function for read() callers into the device */
orangefs_devreq_read(struct file * file,char __user * buf,size_t count,loff_t * offset)158 static ssize_t orangefs_devreq_read(struct file *file,
159 				 char __user *buf,
160 				 size_t count, loff_t *offset)
161 {
162 	struct orangefs_kernel_op_s *op, *temp;
163 	__s32 proto_ver = ORANGEFS_KERNEL_PROTO_VERSION;
164 	static __s32 magic = ORANGEFS_DEVREQ_MAGIC;
165 	struct orangefs_kernel_op_s *cur_op;
166 	unsigned long ret;
167 
168 	/* We do not support blocking IO. */
169 	if (!(file->f_flags & O_NONBLOCK)) {
170 		gossip_err("%s: blocking read from client-core.\n",
171 			   __func__);
172 		return -EINVAL;
173 	}
174 
175 	/*
176 	 * The client will do an ioctl to find MAX_DEV_REQ_UPSIZE, then
177 	 * always read with that size buffer.
178 	 */
179 	if (count != MAX_DEV_REQ_UPSIZE) {
180 		gossip_err("orangefs: client-core tried to read wrong size\n");
181 		return -EINVAL;
182 	}
183 
184 	/* Check for an empty list before locking. */
185 	if (list_empty(&orangefs_request_list))
186 		return -EAGAIN;
187 
188 restart:
189 	cur_op = NULL;
190 	/* Get next op (if any) from top of list. */
191 	spin_lock(&orangefs_request_list_lock);
192 	list_for_each_entry_safe(op, temp, &orangefs_request_list, list) {
193 		__s32 fsid;
194 		/* This lock is held past the end of the loop when we break. */
195 		spin_lock(&op->lock);
196 		if (unlikely(op_state_purged(op) || op_state_given_up(op))) {
197 			spin_unlock(&op->lock);
198 			continue;
199 		}
200 
201 		fsid = fsid_of_op(op);
202 		if (fsid != ORANGEFS_FS_ID_NULL) {
203 			int ret;
204 			/* Skip ops whose filesystem needs to be mounted. */
205 			ret = fs_mount_pending(fsid);
206 			if (ret == 1) {
207 				gossip_debug(GOSSIP_DEV_DEBUG,
208 				    "%s: mount pending, skipping op tag "
209 				    "%llu %s\n",
210 				    __func__,
211 				    llu(op->tag),
212 				    get_opname_string(op));
213 				spin_unlock(&op->lock);
214 				continue;
215 			/*
216 			 * Skip ops whose filesystem we don't know about unless
217 			 * it is being mounted or unmounted.  It is possible for
218 			 * a filesystem we don't know about to be unmounted if
219 			 * it fails to mount in the kernel after userspace has
220 			 * been sent the mount request.
221 			 */
222 			/* XXX: is there a better way to detect this? */
223 			} else if (ret == -1 &&
224 				   !(op->upcall.type ==
225 					ORANGEFS_VFS_OP_FS_MOUNT ||
226 				     op->upcall.type ==
227 					ORANGEFS_VFS_OP_GETATTR ||
228 				     op->upcall.type ==
229 					ORANGEFS_VFS_OP_FS_UMOUNT)) {
230 				gossip_debug(GOSSIP_DEV_DEBUG,
231 				    "orangefs: skipping op tag %llu %s\n",
232 				    llu(op->tag), get_opname_string(op));
233 				gossip_err(
234 				    "orangefs: ERROR: fs_mount_pending %d\n",
235 				    fsid);
236 				spin_unlock(&op->lock);
237 				continue;
238 			}
239 		}
240 		/*
241 		 * Either this op does not pertain to a filesystem, is mounting
242 		 * a filesystem, or pertains to a mounted filesystem. Let it
243 		 * through.
244 		 */
245 		cur_op = op;
246 		break;
247 	}
248 
249 	/*
250 	 * At this point we either have a valid op and can continue or have not
251 	 * found an op and must ask the client to try again later.
252 	 */
253 	if (!cur_op) {
254 		spin_unlock(&orangefs_request_list_lock);
255 		return -EAGAIN;
256 	}
257 
258 	gossip_debug(GOSSIP_DEV_DEBUG, "%s: reading op tag %llu %s\n",
259 		     __func__,
260 		     llu(cur_op->tag),
261 		     get_opname_string(cur_op));
262 
263 	/*
264 	 * Such an op should never be on the list in the first place. If so, we
265 	 * will abort.
266 	 */
267 	if (op_state_in_progress(cur_op) || op_state_serviced(cur_op)) {
268 		gossip_err("orangefs: ERROR: Current op already queued.\n");
269 		list_del_init(&cur_op->list);
270 		spin_unlock(&cur_op->lock);
271 		spin_unlock(&orangefs_request_list_lock);
272 		return -EAGAIN;
273 	}
274 
275 	list_del_init(&cur_op->list);
276 	spin_unlock(&orangefs_request_list_lock);
277 
278 	spin_unlock(&cur_op->lock);
279 
280 	/* Push the upcall out. */
281 	ret = copy_to_user(buf, &proto_ver, sizeof(__s32));
282 	if (ret != 0)
283 		goto error;
284 	ret = copy_to_user(buf + sizeof(__s32), &magic, sizeof(__s32));
285 	if (ret != 0)
286 		goto error;
287 	ret = copy_to_user(buf + 2 * sizeof(__s32),
288 		&cur_op->tag,
289 		sizeof(__u64));
290 	if (ret != 0)
291 		goto error;
292 	ret = copy_to_user(buf + 2 * sizeof(__s32) + sizeof(__u64),
293 		&cur_op->upcall,
294 		sizeof(struct orangefs_upcall_s));
295 	if (ret != 0)
296 		goto error;
297 
298 	spin_lock(&orangefs_htable_ops_in_progress_lock);
299 	spin_lock(&cur_op->lock);
300 	if (unlikely(op_state_given_up(cur_op))) {
301 		spin_unlock(&cur_op->lock);
302 		spin_unlock(&orangefs_htable_ops_in_progress_lock);
303 		complete(&cur_op->waitq);
304 		goto restart;
305 	}
306 
307 	/*
308 	 * Set the operation to be in progress and move it between lists since
309 	 * it has been sent to the client.
310 	 */
311 	set_op_state_inprogress(cur_op);
312 	gossip_debug(GOSSIP_DEV_DEBUG,
313 		     "%s: 1 op:%s: op_state:%d: process:%s:\n",
314 		     __func__,
315 		     get_opname_string(cur_op),
316 		     cur_op->op_state,
317 		     current->comm);
318 	orangefs_devreq_add_op(cur_op);
319 	spin_unlock(&cur_op->lock);
320 	spin_unlock(&orangefs_htable_ops_in_progress_lock);
321 
322 	/* The client only asks to read one size buffer. */
323 	return MAX_DEV_REQ_UPSIZE;
324 error:
325 	/*
326 	 * We were unable to copy the op data to the client. Put the op back in
327 	 * list. If client has crashed, the op will be purged later when the
328 	 * device is released.
329 	 */
330 	gossip_err("orangefs: Failed to copy data to user space\n");
331 	spin_lock(&orangefs_request_list_lock);
332 	spin_lock(&cur_op->lock);
333 	if (likely(!op_state_given_up(cur_op))) {
334 		set_op_state_waiting(cur_op);
335 		gossip_debug(GOSSIP_DEV_DEBUG,
336 			     "%s: 2 op:%s: op_state:%d: process:%s:\n",
337 			     __func__,
338 			     get_opname_string(cur_op),
339 			     cur_op->op_state,
340 			     current->comm);
341 		list_add(&cur_op->list, &orangefs_request_list);
342 		spin_unlock(&cur_op->lock);
343 	} else {
344 		spin_unlock(&cur_op->lock);
345 		complete(&cur_op->waitq);
346 	}
347 	spin_unlock(&orangefs_request_list_lock);
348 	return -EFAULT;
349 }
350 
351 /*
352  * Function for writev() callers into the device.
353  *
354  * Userspace should have written:
355  *  - __u32 version
356  *  - __u32 magic
357  *  - __u64 tag
358  *  - struct orangefs_downcall_s
359  *  - trailer buffer (in the case of READDIR operations)
360  */
orangefs_devreq_write_iter(struct kiocb * iocb,struct iov_iter * iter)361 static ssize_t orangefs_devreq_write_iter(struct kiocb *iocb,
362 				      struct iov_iter *iter)
363 {
364 	ssize_t ret;
365 	struct orangefs_kernel_op_s *op = NULL;
366 	struct {
367 		__u32 version;
368 		__u32 magic;
369 		__u64 tag;
370 	} head;
371 	int total = ret = iov_iter_count(iter);
372 	int downcall_size = sizeof(struct orangefs_downcall_s);
373 	int head_size = sizeof(head);
374 
375 	gossip_debug(GOSSIP_DEV_DEBUG, "%s: total:%d: ret:%zd:\n",
376 		     __func__,
377 		     total,
378 		     ret);
379 
380         if (total < MAX_DEV_REQ_DOWNSIZE) {
381 		gossip_err("%s: total:%d: must be at least:%u:\n",
382 			   __func__,
383 			   total,
384 			   (unsigned int) MAX_DEV_REQ_DOWNSIZE);
385 		return -EFAULT;
386 	}
387 
388 	if (!copy_from_iter_full(&head, head_size, iter)) {
389 		gossip_err("%s: failed to copy head.\n", __func__);
390 		return -EFAULT;
391 	}
392 
393 	if (head.version < ORANGEFS_MINIMUM_USERSPACE_VERSION) {
394 		gossip_err("%s: userspace claims version"
395 			   "%d, minimum version required: %d.\n",
396 			   __func__,
397 			   head.version,
398 			   ORANGEFS_MINIMUM_USERSPACE_VERSION);
399 		return -EPROTO;
400 	}
401 
402 	if (head.magic != ORANGEFS_DEVREQ_MAGIC) {
403 		gossip_err("Error: Device magic number does not match.\n");
404 		return -EPROTO;
405 	}
406 
407 	if (!orangefs_userspace_version) {
408 		orangefs_userspace_version = head.version;
409 	} else if (orangefs_userspace_version != head.version) {
410 		gossip_err("Error: userspace version changes\n");
411 		return -EPROTO;
412 	}
413 
414 	/* remove the op from the in progress hash table */
415 	op = orangefs_devreq_remove_op(head.tag);
416 	if (!op) {
417 		gossip_debug(GOSSIP_DEV_DEBUG,
418 			     "%s: No one's waiting for tag %llu\n",
419 			     __func__, llu(head.tag));
420 		return ret;
421 	}
422 
423 	if (!copy_from_iter_full(&op->downcall, downcall_size, iter)) {
424 		gossip_err("%s: failed to copy downcall.\n", __func__);
425 		goto Efault;
426 	}
427 
428 	if (op->downcall.status)
429 		goto wakeup;
430 
431 	/*
432 	 * We've successfully peeled off the head and the downcall.
433 	 * Something has gone awry if total doesn't equal the
434 	 * sum of head_size, downcall_size and trailer_size.
435 	 */
436 	if ((head_size + downcall_size + op->downcall.trailer_size) != total) {
437 		gossip_err("%s: funky write, head_size:%d"
438 			   ": downcall_size:%d: trailer_size:%lld"
439 			   ": total size:%d:\n",
440 			   __func__,
441 			   head_size,
442 			   downcall_size,
443 			   op->downcall.trailer_size,
444 			   total);
445 		goto Efault;
446 	}
447 
448 	/* Only READDIR operations should have trailers. */
449 	if ((op->downcall.type != ORANGEFS_VFS_OP_READDIR) &&
450 	    (op->downcall.trailer_size != 0)) {
451 		gossip_err("%s: %x operation with trailer.",
452 			   __func__,
453 			   op->downcall.type);
454 		goto Efault;
455 	}
456 
457 	/* READDIR operations should always have trailers. */
458 	if ((op->downcall.type == ORANGEFS_VFS_OP_READDIR) &&
459 	    (op->downcall.trailer_size == 0)) {
460 		gossip_err("%s: %x operation with no trailer.",
461 			   __func__,
462 			   op->downcall.type);
463 		goto Efault;
464 	}
465 
466 	if (op->downcall.type != ORANGEFS_VFS_OP_READDIR)
467 		goto wakeup;
468 
469 	op->downcall.trailer_buf = vzalloc(op->downcall.trailer_size);
470 	if (!op->downcall.trailer_buf)
471 		goto Enomem;
472 
473 	if (!copy_from_iter_full(op->downcall.trailer_buf,
474 			         op->downcall.trailer_size, iter)) {
475 		gossip_err("%s: failed to copy trailer.\n", __func__);
476 		vfree(op->downcall.trailer_buf);
477 		goto Efault;
478 	}
479 
480 wakeup:
481 	/*
482 	 * Return to vfs waitqueue, and back to service_operation
483 	 * through wait_for_matching_downcall.
484 	 */
485 	spin_lock(&op->lock);
486 	if (unlikely(op_is_cancel(op))) {
487 		spin_unlock(&op->lock);
488 		put_cancel(op);
489 	} else if (unlikely(op_state_given_up(op))) {
490 		spin_unlock(&op->lock);
491 		complete(&op->waitq);
492 	} else {
493 		set_op_state_serviced(op);
494 		gossip_debug(GOSSIP_DEV_DEBUG,
495 			     "%s: op:%s: op_state:%d: process:%s:\n",
496 			     __func__,
497 			     get_opname_string(op),
498 			     op->op_state,
499 			     current->comm);
500 		spin_unlock(&op->lock);
501 	}
502 	return ret;
503 
504 Efault:
505 	op->downcall.status = -(ORANGEFS_ERROR_BIT | 9);
506 	ret = -EFAULT;
507 	goto wakeup;
508 
509 Enomem:
510 	op->downcall.status = -(ORANGEFS_ERROR_BIT | 8);
511 	ret = -ENOMEM;
512 	goto wakeup;
513 }
514 
515 /*
516  * NOTE: gets called when the last reference to this device is dropped.
517  * Using the open_access_count variable, we enforce a reference count
518  * on this file so that it can be opened by only one process at a time.
519  * the devreq_mutex is used to make sure all i/o has completed
520  * before we call orangefs_bufmap_finalize, and similar such tricky
521  * situations
522  */
orangefs_devreq_release(struct inode * inode,struct file * file)523 static int orangefs_devreq_release(struct inode *inode, struct file *file)
524 {
525 	int unmounted = 0;
526 
527 	gossip_debug(GOSSIP_DEV_DEBUG,
528 		     "%s:pvfs2-client-core: exiting, closing device\n",
529 		     __func__);
530 
531 	mutex_lock(&devreq_mutex);
532 	orangefs_bufmap_finalize();
533 
534 	open_access_count = -1;
535 
536 	unmounted = mark_all_pending_mounts();
537 	gossip_debug(GOSSIP_DEV_DEBUG, "ORANGEFS Device Close: Filesystem(s) %s\n",
538 		     (unmounted ? "UNMOUNTED" : "MOUNTED"));
539 
540 	purge_waiting_ops();
541 	purge_inprogress_ops();
542 
543 	orangefs_bufmap_run_down();
544 
545 	gossip_debug(GOSSIP_DEV_DEBUG,
546 		     "pvfs2-client-core: device close complete\n");
547 	open_access_count = 0;
548 	orangefs_userspace_version = 0;
549 	mutex_unlock(&devreq_mutex);
550 	return 0;
551 }
552 
is_daemon_in_service(void)553 int is_daemon_in_service(void)
554 {
555 	int in_service;
556 
557 	/*
558 	 * What this function does is checks if client-core is alive
559 	 * based on the access count we maintain on the device.
560 	 */
561 	mutex_lock(&devreq_mutex);
562 	in_service = open_access_count == 1 ? 0 : -EIO;
563 	mutex_unlock(&devreq_mutex);
564 	return in_service;
565 }
566 
__is_daemon_in_service(void)567 bool __is_daemon_in_service(void)
568 {
569 	return open_access_count == 1;
570 }
571 
check_ioctl_command(unsigned int command)572 static inline long check_ioctl_command(unsigned int command)
573 {
574 	/* Check for valid ioctl codes */
575 	if (_IOC_TYPE(command) != ORANGEFS_DEV_MAGIC) {
576 		gossip_err("device ioctl magic numbers don't match! Did you rebuild pvfs2-client-core/libpvfs2? [cmd %x, magic %x != %x]\n",
577 			command,
578 			_IOC_TYPE(command),
579 			ORANGEFS_DEV_MAGIC);
580 		return -EINVAL;
581 	}
582 	/* and valid ioctl commands */
583 	if (_IOC_NR(command) >= ORANGEFS_DEV_MAXNR || _IOC_NR(command) <= 0) {
584 		gossip_err("Invalid ioctl command number [%d >= %d]\n",
585 			   _IOC_NR(command), ORANGEFS_DEV_MAXNR);
586 		return -ENOIOCTLCMD;
587 	}
588 	return 0;
589 }
590 
dispatch_ioctl_command(unsigned int command,unsigned long arg)591 static long dispatch_ioctl_command(unsigned int command, unsigned long arg)
592 {
593 	static __s32 magic = ORANGEFS_DEVREQ_MAGIC;
594 	static __s32 max_up_size = MAX_DEV_REQ_UPSIZE;
595 	static __s32 max_down_size = MAX_DEV_REQ_DOWNSIZE;
596 	struct ORANGEFS_dev_map_desc user_desc;
597 	int ret = 0;
598 	int upstream_kmod = 1;
599 	struct orangefs_sb_info_s *orangefs_sb;
600 
601 	/* mtmoore: add locking here */
602 
603 	switch (command) {
604 	case ORANGEFS_DEV_GET_MAGIC:
605 		return ((put_user(magic, (__s32 __user *) arg) == -EFAULT) ?
606 			-EIO :
607 			0);
608 	case ORANGEFS_DEV_GET_MAX_UPSIZE:
609 		return ((put_user(max_up_size,
610 				  (__s32 __user *) arg) == -EFAULT) ?
611 					-EIO :
612 					0);
613 	case ORANGEFS_DEV_GET_MAX_DOWNSIZE:
614 		return ((put_user(max_down_size,
615 				  (__s32 __user *) arg) == -EFAULT) ?
616 					-EIO :
617 					0);
618 	case ORANGEFS_DEV_MAP:
619 		ret = copy_from_user(&user_desc,
620 				     (struct ORANGEFS_dev_map_desc __user *)
621 				     arg,
622 				     sizeof(struct ORANGEFS_dev_map_desc));
623 		/* WTF -EIO and not -EFAULT? */
624 		return ret ? -EIO : orangefs_bufmap_initialize(&user_desc);
625 	case ORANGEFS_DEV_REMOUNT_ALL:
626 		gossip_debug(GOSSIP_DEV_DEBUG,
627 			     "%s: got ORANGEFS_DEV_REMOUNT_ALL\n",
628 			     __func__);
629 
630 		/*
631 		 * remount all mounted orangefs volumes to regain the lost
632 		 * dynamic mount tables (if any) -- NOTE: this is done
633 		 * without keeping the superblock list locked due to the
634 		 * upcall/downcall waiting.  also, the request mutex is
635 		 * used to ensure that no operations will be serviced until
636 		 * all of the remounts are serviced (to avoid ops between
637 		 * mounts to fail)
638 		 */
639 		ret = mutex_lock_interruptible(&orangefs_request_mutex);
640 		if (ret < 0)
641 			return ret;
642 		gossip_debug(GOSSIP_DEV_DEBUG,
643 			     "%s: priority remount in progress\n",
644 			     __func__);
645 		spin_lock(&orangefs_superblocks_lock);
646 		list_for_each_entry(orangefs_sb, &orangefs_superblocks, list) {
647 			/*
648 			 * We have to drop the spinlock, so entries can be
649 			 * removed.  They can't be freed, though, so we just
650 			 * keep the forward pointers and zero the back ones -
651 			 * that way we can get to the rest of the list.
652 			 */
653 			if (!orangefs_sb->list.prev)
654 				continue;
655 			gossip_debug(GOSSIP_DEV_DEBUG,
656 				     "%s: Remounting SB %p\n",
657 				     __func__,
658 				     orangefs_sb);
659 
660 			spin_unlock(&orangefs_superblocks_lock);
661 			ret = orangefs_remount(orangefs_sb);
662 			spin_lock(&orangefs_superblocks_lock);
663 			if (ret) {
664 				gossip_debug(GOSSIP_DEV_DEBUG,
665 					     "SB %p remount failed\n",
666 					     orangefs_sb);
667 				break;
668 			}
669 		}
670 		spin_unlock(&orangefs_superblocks_lock);
671 		gossip_debug(GOSSIP_DEV_DEBUG,
672 			     "%s: priority remount complete\n",
673 			     __func__);
674 		mutex_unlock(&orangefs_request_mutex);
675 		return ret;
676 
677 	case ORANGEFS_DEV_UPSTREAM:
678 		ret = copy_to_user((void __user *)arg,
679 				    &upstream_kmod,
680 				    sizeof(upstream_kmod));
681 
682 		if (ret != 0)
683 			return -EIO;
684 		else
685 			return ret;
686 
687 	case ORANGEFS_DEV_CLIENT_MASK:
688 		return orangefs_debugfs_new_client_mask((void __user *)arg);
689 	case ORANGEFS_DEV_CLIENT_STRING:
690 		return orangefs_debugfs_new_client_string((void __user *)arg);
691 	case ORANGEFS_DEV_DEBUG:
692 		return orangefs_debugfs_new_debug((void __user *)arg);
693 	default:
694 		return -ENOIOCTLCMD;
695 	}
696 	return -ENOIOCTLCMD;
697 }
698 
orangefs_devreq_ioctl(struct file * file,unsigned int command,unsigned long arg)699 static long orangefs_devreq_ioctl(struct file *file,
700 			       unsigned int command, unsigned long arg)
701 {
702 	long ret;
703 
704 	/* Check for properly constructed commands */
705 	ret = check_ioctl_command(command);
706 	if (ret < 0)
707 		return (int)ret;
708 
709 	return (int)dispatch_ioctl_command(command, arg);
710 }
711 
712 #ifdef CONFIG_COMPAT		/* CONFIG_COMPAT is in .config */
713 
714 /*  Compat structure for the ORANGEFS_DEV_MAP ioctl */
715 struct ORANGEFS_dev_map_desc32 {
716 	compat_uptr_t ptr;
717 	__s32 total_size;
718 	__s32 size;
719 	__s32 count;
720 };
721 
722 /*
723  * 32 bit user-space apps' ioctl handlers when kernel modules
724  * is compiled as a 64 bit one
725  */
orangefs_devreq_compat_ioctl(struct file * filp,unsigned int cmd,unsigned long args)726 static long orangefs_devreq_compat_ioctl(struct file *filp, unsigned int cmd,
727 				      unsigned long args)
728 {
729 	long ret;
730 
731 	/* Check for properly constructed commands */
732 	ret = check_ioctl_command(cmd);
733 	if (ret < 0)
734 		return ret;
735 	if (cmd == ORANGEFS_DEV_MAP) {
736 		struct ORANGEFS_dev_map_desc desc;
737 		struct ORANGEFS_dev_map_desc32 d32;
738 
739 		if (copy_from_user(&d32, (void __user *)args, sizeof(d32)))
740 			return -EFAULT;
741 
742 		desc.ptr = compat_ptr(d32.ptr);
743 		desc.total_size = d32.total_size;
744 		desc.size = d32.size;
745 		desc.count = d32.count;
746 		return orangefs_bufmap_initialize(&desc);
747 	}
748 	/* no other ioctl requires translation */
749 	return dispatch_ioctl_command(cmd, args);
750 }
751 
752 #endif /* CONFIG_COMPAT is in .config */
753 
orangefs_devreq_poll(struct file * file,struct poll_table_struct * poll_table)754 static __poll_t orangefs_devreq_poll(struct file *file,
755 				      struct poll_table_struct *poll_table)
756 {
757 	__poll_t poll_revent_mask = 0;
758 
759 	poll_wait(file, &orangefs_request_list_waitq, poll_table);
760 
761 	if (!list_empty(&orangefs_request_list))
762 		poll_revent_mask |= EPOLLIN;
763 	return poll_revent_mask;
764 }
765 
766 /* the assigned character device major number */
767 static int orangefs_dev_major;
768 
769 static const struct file_operations orangefs_devreq_file_operations = {
770 	.owner = THIS_MODULE,
771 	.read = orangefs_devreq_read,
772 	.write_iter = orangefs_devreq_write_iter,
773 	.open = orangefs_devreq_open,
774 	.release = orangefs_devreq_release,
775 	.unlocked_ioctl = orangefs_devreq_ioctl,
776 
777 #ifdef CONFIG_COMPAT		/* CONFIG_COMPAT is in .config */
778 	.compat_ioctl = orangefs_devreq_compat_ioctl,
779 #endif
780 	.poll = orangefs_devreq_poll
781 };
782 
783 /*
784  * Initialize orangefs device specific state:
785  * Must be called at module load time only
786  */
orangefs_dev_init(void)787 int orangefs_dev_init(void)
788 {
789 	/* register orangefs-req device  */
790 	orangefs_dev_major = register_chrdev(0,
791 					  ORANGEFS_REQDEVICE_NAME,
792 					  &orangefs_devreq_file_operations);
793 	if (orangefs_dev_major < 0) {
794 		gossip_debug(GOSSIP_DEV_DEBUG,
795 			     "Failed to register /dev/%s (error %d)\n",
796 			     ORANGEFS_REQDEVICE_NAME, orangefs_dev_major);
797 		return orangefs_dev_major;
798 	}
799 
800 	gossip_debug(GOSSIP_DEV_DEBUG,
801 		     "*** /dev/%s character device registered ***\n",
802 		     ORANGEFS_REQDEVICE_NAME);
803 	gossip_debug(GOSSIP_DEV_DEBUG, "'mknod /dev/%s c %d 0'.\n",
804 		     ORANGEFS_REQDEVICE_NAME, orangefs_dev_major);
805 	return 0;
806 }
807 
orangefs_dev_cleanup(void)808 void orangefs_dev_cleanup(void)
809 {
810 	unregister_chrdev(orangefs_dev_major, ORANGEFS_REQDEVICE_NAME);
811 	gossip_debug(GOSSIP_DEV_DEBUG,
812 		     "*** /dev/%s character device unregistered ***\n",
813 		     ORANGEFS_REQDEVICE_NAME);
814 }
815