1 // SPDX-License-Identifier: GPL-2.0
3 * (C) 2001 Clemson University and The University of Chicago
5 * Changes by Acxiom Corporation to add protocol version to kernel
6 * communication, Copyright Acxiom Corporation, 2005.
8 * See COPYING in top-level directory.
12 #include "orangefs-kernel.h"
13 #include "orangefs-dev-proto.h"
14 #include "orangefs-bufmap.h"
15 #include "orangefs-debugfs.h"
17 #include <linux/debugfs.h>
18 #include <linux/slab.h>
20 /* this file implements the /dev/pvfs2-req device node */
22 uint32_t orangefs_userspace_version
;
24 static int open_access_count
;
26 static DEFINE_MUTEX(devreq_mutex
);
28 #define DUMP_DEVICE_ERROR() \
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");\
43 static int hash_func(__u64 tag
, int table_size
)
45 return do_div(tag
, (unsigned int)table_size
);
48 static void orangefs_devreq_add_op(struct orangefs_kernel_op_s
*op
)
50 int index
= hash_func(op
->tag
, hash_table_size
);
52 list_add_tail(&op
->list
, &orangefs_htable_ops_in_progress
[index
]);
56 * find the op with this tag and remove it from the in progress
59 static struct orangefs_kernel_op_s
*orangefs_devreq_remove_op(__u64 tag
)
61 struct orangefs_kernel_op_s
*op
, *next
;
64 index
= hash_func(tag
, hash_table_size
);
66 spin_lock(&orangefs_htable_ops_in_progress_lock
);
67 list_for_each_entry_safe(op
,
69 &orangefs_htable_ops_in_progress
[index
],
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
);
79 spin_unlock(&orangefs_htable_ops_in_progress_lock
);
83 /* Returns whether any FS are still pending remounted */
84 static int mark_all_pending_mounts(void)
87 struct orangefs_sb_info_s
*orangefs_sb
= NULL
;
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;
95 spin_unlock(&orangefs_superblocks_lock
);
100 * Determine if a given file system needs to be remounted or not
101 * Returns -1 on error
102 * 0 if already mounted
105 static int fs_mount_pending(__s32 fsid
)
107 int mount_pending
= -1;
108 struct orangefs_sb_info_s
*orangefs_sb
= NULL
;
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
;
117 spin_unlock(&orangefs_superblocks_lock
);
118 return mount_pending
;
121 static int orangefs_devreq_open(struct inode
*inode
, struct file
*file
)
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",
132 if (!(file
->f_flags
& O_NONBLOCK
)) {
133 gossip_err("%s: device cannot be opened in blocking mode\n",
138 gossip_debug(GOSSIP_DEV_DEBUG
, "client-core: opening device\n");
139 mutex_lock(&devreq_mutex
);
141 if (open_access_count
== 0) {
142 open_access_count
= 1;
147 mutex_unlock(&devreq_mutex
);
151 gossip_debug(GOSSIP_DEV_DEBUG
,
152 "pvfs2-client-core: open device complete (ret = %d)\n",
157 /* Function for read() callers into the device */
158 static ssize_t
orangefs_devreq_read(struct file
*file
,
160 size_t count
, loff_t
*offset
)
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
;
168 /* We do not support blocking IO. */
169 if (!(file
->f_flags
& O_NONBLOCK
)) {
170 gossip_err("%s: blocking read from client-core.\n",
176 * The client will do an ioctl to find MAX_DEV_REQ_UPSIZE, then
177 * always read with that size buffer.
179 if (count
!= MAX_DEV_REQ_UPSIZE
) {
180 gossip_err("orangefs: client-core tried to read wrong size\n");
184 /* Check for an empty list before locking. */
185 if (list_empty(&orangefs_request_list
))
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
) {
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
);
201 fsid
= fsid_of_op(op
);
202 if (fsid
!= ORANGEFS_FS_ID_NULL
) {
204 /* Skip ops whose filesystem needs to be mounted. */
205 ret
= fs_mount_pending(fsid
);
207 gossip_debug(GOSSIP_DEV_DEBUG
,
208 "%s: mount pending, skipping op tag "
212 get_opname_string(op
));
213 spin_unlock(&op
->lock
);
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.
222 /* XXX: is there a better way to detect this? */
223 } else if (ret
== -1 &&
225 ORANGEFS_VFS_OP_FS_MOUNT
||
227 ORANGEFS_VFS_OP_GETATTR
||
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
));
234 "orangefs: ERROR: fs_mount_pending %d\n",
236 spin_unlock(&op
->lock
);
241 * Either this op does not pertain to a filesystem, is mounting
242 * a filesystem, or pertains to a mounted filesystem. Let it
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.
254 spin_unlock(&orangefs_request_list_lock
);
258 gossip_debug(GOSSIP_DEV_DEBUG
, "%s: reading op tag %llu %s\n",
261 get_opname_string(cur_op
));
264 * Such an op should never be on the list in the first place. If so, we
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
);
275 list_del_init(&cur_op
->list
);
276 spin_unlock(&orangefs_request_list_lock
);
278 spin_unlock(&cur_op
->lock
);
280 /* Push the upcall out. */
281 ret
= copy_to_user(buf
, &proto_ver
, sizeof(__s32
));
284 ret
= copy_to_user(buf
+ sizeof(__s32
), &magic
, sizeof(__s32
));
287 ret
= copy_to_user(buf
+ 2 * sizeof(__s32
),
292 ret
= copy_to_user(buf
+ 2 * sizeof(__s32
) + sizeof(__u64
),
294 sizeof(struct orangefs_upcall_s
));
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
);
308 * Set the operation to be in progress and move it between lists since
309 * it has been sent to the client.
311 set_op_state_inprogress(cur_op
);
312 gossip_debug(GOSSIP_DEV_DEBUG
,
313 "%s: 1 op:%s: op_state:%d: process:%s:\n",
315 get_opname_string(cur_op
),
318 orangefs_devreq_add_op(cur_op
);
319 spin_unlock(&cur_op
->lock
);
320 spin_unlock(&orangefs_htable_ops_in_progress_lock
);
322 /* The client only asks to read one size buffer. */
323 return MAX_DEV_REQ_UPSIZE
;
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.
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",
338 get_opname_string(cur_op
),
341 list_add(&cur_op
->list
, &orangefs_request_list
);
342 spin_unlock(&cur_op
->lock
);
344 spin_unlock(&cur_op
->lock
);
345 complete(&cur_op
->waitq
);
347 spin_unlock(&orangefs_request_list_lock
);
352 * Function for writev() callers into the device.
354 * Userspace should have written:
358 * - struct orangefs_downcall_s
359 * - trailer buffer (in the case of READDIR operations)
361 static ssize_t
orangefs_devreq_write_iter(struct kiocb
*iocb
,
362 struct iov_iter
*iter
)
365 struct orangefs_kernel_op_s
*op
= NULL
;
371 int total
= ret
= iov_iter_count(iter
);
372 int downcall_size
= sizeof(struct orangefs_downcall_s
);
373 int head_size
= sizeof(head
);
375 gossip_debug(GOSSIP_DEV_DEBUG
, "%s: total:%d: ret:%zd:\n",
380 if (total
< MAX_DEV_REQ_DOWNSIZE
) {
381 gossip_err("%s: total:%d: must be at least:%u:\n",
384 (unsigned int) MAX_DEV_REQ_DOWNSIZE
);
388 if (!copy_from_iter_full(&head
, head_size
, iter
)) {
389 gossip_err("%s: failed to copy head.\n", __func__
);
393 if (head
.version
< ORANGEFS_MINIMUM_USERSPACE_VERSION
) {
394 gossip_err("%s: userspace claims version"
395 "%d, minimum version required: %d.\n",
398 ORANGEFS_MINIMUM_USERSPACE_VERSION
);
402 if (head
.magic
!= ORANGEFS_DEVREQ_MAGIC
) {
403 gossip_err("Error: Device magic number does not match.\n");
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");
414 /* remove the op from the in progress hash table */
415 op
= orangefs_devreq_remove_op(head
.tag
);
417 gossip_debug(GOSSIP_DEV_DEBUG
,
418 "%s: No one's waiting for tag %llu\n",
419 __func__
, llu(head
.tag
));
423 if (!copy_from_iter_full(&op
->downcall
, downcall_size
, iter
)) {
424 gossip_err("%s: failed to copy downcall.\n", __func__
);
428 if (op
->downcall
.status
)
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.
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",
443 op
->downcall
.trailer_size
,
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.",
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.",
466 if (op
->downcall
.type
!= ORANGEFS_VFS_OP_READDIR
)
469 op
->downcall
.trailer_buf
= vzalloc(op
->downcall
.trailer_size
);
470 if (!op
->downcall
.trailer_buf
)
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
);
482 * Return to vfs waitqueue, and back to service_operation
483 * through wait_for_matching_downcall.
485 spin_lock(&op
->lock
);
486 if (unlikely(op_is_cancel(op
))) {
487 spin_unlock(&op
->lock
);
489 } else if (unlikely(op_state_given_up(op
))) {
490 spin_unlock(&op
->lock
);
491 complete(&op
->waitq
);
493 set_op_state_serviced(op
);
494 gossip_debug(GOSSIP_DEV_DEBUG
,
495 "%s: op:%s: op_state:%d: process:%s:\n",
497 get_opname_string(op
),
500 spin_unlock(&op
->lock
);
505 op
->downcall
.status
= -(ORANGEFS_ERROR_BIT
| 9);
510 op
->downcall
.status
= -(ORANGEFS_ERROR_BIT
| 8);
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
523 static int orangefs_devreq_release(struct inode
*inode
, struct file
*file
)
527 gossip_debug(GOSSIP_DEV_DEBUG
,
528 "%s:pvfs2-client-core: exiting, closing device\n",
531 mutex_lock(&devreq_mutex
);
532 orangefs_bufmap_finalize();
534 open_access_count
= -1;
536 unmounted
= mark_all_pending_mounts();
537 gossip_debug(GOSSIP_DEV_DEBUG
, "ORANGEFS Device Close: Filesystem(s) %s\n",
538 (unmounted
? "UNMOUNTED" : "MOUNTED"));
541 purge_inprogress_ops();
543 orangefs_bufmap_run_down();
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
);
553 int is_daemon_in_service(void)
558 * What this function does is checks if client-core is alive
559 * based on the access count we maintain on the device.
561 mutex_lock(&devreq_mutex
);
562 in_service
= open_access_count
== 1 ? 0 : -EIO
;
563 mutex_unlock(&devreq_mutex
);
567 bool __is_daemon_in_service(void)
569 return open_access_count
== 1;
572 static inline long check_ioctl_command(unsigned int command
)
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",
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
);
591 static long dispatch_ioctl_command(unsigned int command
, unsigned long arg
)
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
;
598 int upstream_kmod
= 1;
599 struct orangefs_sb_info_s
*orangefs_sb
;
601 /* mtmoore: add locking here */
604 case ORANGEFS_DEV_GET_MAGIC
:
605 return ((put_user(magic
, (__s32 __user
*) arg
) == -EFAULT
) ?
608 case ORANGEFS_DEV_GET_MAX_UPSIZE
:
609 return ((put_user(max_up_size
,
610 (__s32 __user
*) arg
) == -EFAULT
) ?
613 case ORANGEFS_DEV_GET_MAX_DOWNSIZE
:
614 return ((put_user(max_down_size
,
615 (__s32 __user
*) arg
) == -EFAULT
) ?
618 case ORANGEFS_DEV_MAP
:
619 ret
= copy_from_user(&user_desc
,
620 (struct ORANGEFS_dev_map_desc __user
*)
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",
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
639 ret
= mutex_lock_interruptible(&orangefs_request_mutex
);
642 gossip_debug(GOSSIP_DEV_DEBUG
,
643 "%s: priority remount in progress\n",
645 spin_lock(&orangefs_superblocks_lock
);
646 list_for_each_entry(orangefs_sb
, &orangefs_superblocks
, list
) {
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.
653 if (!orangefs_sb
->list
.prev
)
655 gossip_debug(GOSSIP_DEV_DEBUG
,
656 "%s: Remounting SB %p\n",
660 spin_unlock(&orangefs_superblocks_lock
);
661 ret
= orangefs_remount(orangefs_sb
);
662 spin_lock(&orangefs_superblocks_lock
);
664 gossip_debug(GOSSIP_DEV_DEBUG
,
665 "SB %p remount failed\n",
670 spin_unlock(&orangefs_superblocks_lock
);
671 gossip_debug(GOSSIP_DEV_DEBUG
,
672 "%s: priority remount complete\n",
674 mutex_unlock(&orangefs_request_mutex
);
677 case ORANGEFS_DEV_UPSTREAM
:
678 ret
= copy_to_user((void __user
*)arg
,
680 sizeof(upstream_kmod
));
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
);
699 static long orangefs_devreq_ioctl(struct file
*file
,
700 unsigned int command
, unsigned long arg
)
704 /* Check for properly constructed commands */
705 ret
= check_ioctl_command(command
);
709 return (int)dispatch_ioctl_command(command
, arg
);
712 #ifdef CONFIG_COMPAT /* CONFIG_COMPAT is in .config */
714 /* Compat structure for the ORANGEFS_DEV_MAP ioctl */
715 struct ORANGEFS_dev_map_desc32
{
723 * 32 bit user-space apps' ioctl handlers when kernel modules
724 * is compiled as a 64 bit one
726 static long orangefs_devreq_compat_ioctl(struct file
*filp
, unsigned int cmd
,
731 /* Check for properly constructed commands */
732 ret
= check_ioctl_command(cmd
);
735 if (cmd
== ORANGEFS_DEV_MAP
) {
736 struct ORANGEFS_dev_map_desc desc
;
737 struct ORANGEFS_dev_map_desc32 d32
;
739 if (copy_from_user(&d32
, (void __user
*)args
, sizeof(d32
)))
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
);
748 /* no other ioctl requires translation */
749 return dispatch_ioctl_command(cmd
, args
);
752 #endif /* CONFIG_COMPAT is in .config */
754 static __poll_t
orangefs_devreq_poll(struct file
*file
,
755 struct poll_table_struct
*poll_table
)
757 __poll_t poll_revent_mask
= 0;
759 poll_wait(file
, &orangefs_request_list_waitq
, poll_table
);
761 if (!list_empty(&orangefs_request_list
))
762 poll_revent_mask
|= EPOLLIN
;
763 return poll_revent_mask
;
766 /* the assigned character device major number */
767 static int orangefs_dev_major
;
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
,
777 #ifdef CONFIG_COMPAT /* CONFIG_COMPAT is in .config */
778 .compat_ioctl
= orangefs_devreq_compat_ioctl
,
780 .poll
= orangefs_devreq_poll
784 * Initialize orangefs device specific state:
785 * Must be called at module load time only
787 int orangefs_dev_init(void)
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
;
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
);
808 void orangefs_dev_cleanup(void)
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
);