2 * (C) 2001 Clemson University and The University of Chicago
4 * Changes by Acxiom Corporation to add protocol version to kernel
5 * communication, Copyright Acxiom Corporation, 2005.
7 * See COPYING in top-level directory.
11 #include "orangefs-kernel.h"
12 #include "orangefs-dev-proto.h"
13 #include "orangefs-bufmap.h"
14 #include "orangefs-debugfs.h"
16 #include <linux/debugfs.h>
17 #include <linux/slab.h>
19 /* this file implements the /dev/pvfs2-req device node */
21 uint32_t orangefs_userspace_version;
23 static int open_access_count;
25 static DEFINE_MUTEX(devreq_mutex);
27 #define DUMP_DEVICE_ERROR() \
29 gossip_err("*****************************************************\n");\
30 gossip_err("ORANGEFS Device Error: You cannot open the device file "); \
31 gossip_err("\n/dev/%s more than once. Please make sure that\nthere " \
32 "are no ", ORANGEFS_REQDEVICE_NAME); \
33 gossip_err("instances of a program using this device\ncurrently " \
34 "running. (You must verify this!)\n"); \
35 gossip_err("For example, you can use the lsof program as follows:\n");\
36 gossip_err("'lsof | grep %s' (run this as root)\n", \
37 ORANGEFS_REQDEVICE_NAME); \
38 gossip_err(" open_access_count = %d\n", open_access_count); \
39 gossip_err("*****************************************************\n");\
42 static int hash_func(__u64 tag, int table_size)
44 return do_div(tag, (unsigned int)table_size);
47 static void orangefs_devreq_add_op(struct orangefs_kernel_op_s *op)
49 int index = hash_func(op->tag, hash_table_size);
51 list_add_tail(&op->list, &orangefs_htable_ops_in_progress[index]);
55 * find the op with this tag and remove it from the in progress
58 static struct orangefs_kernel_op_s *orangefs_devreq_remove_op(__u64 tag)
60 struct orangefs_kernel_op_s *op, *next;
63 index = hash_func(tag, hash_table_size);
65 spin_lock(&orangefs_htable_ops_in_progress_lock);
66 list_for_each_entry_safe(op,
68 &orangefs_htable_ops_in_progress[index],
70 if (op->tag == tag && !op_state_purged(op) &&
71 !op_state_given_up(op)) {
72 list_del_init(&op->list);
73 spin_unlock(&orangefs_htable_ops_in_progress_lock);
78 spin_unlock(&orangefs_htable_ops_in_progress_lock);
82 /* Returns whether any FS are still pending remounted */
83 static int mark_all_pending_mounts(void)
86 struct orangefs_sb_info_s *orangefs_sb = NULL;
88 spin_lock(&orangefs_superblocks_lock);
89 list_for_each_entry(orangefs_sb, &orangefs_superblocks, list) {
90 /* All of these file system require a remount */
91 orangefs_sb->mount_pending = 1;
94 spin_unlock(&orangefs_superblocks_lock);
99 * Determine if a given file system needs to be remounted or not
100 * Returns -1 on error
101 * 0 if already mounted
104 static int fs_mount_pending(__s32 fsid)
106 int mount_pending = -1;
107 struct orangefs_sb_info_s *orangefs_sb = NULL;
109 spin_lock(&orangefs_superblocks_lock);
110 list_for_each_entry(orangefs_sb, &orangefs_superblocks, list) {
111 if (orangefs_sb->fs_id == fsid) {
112 mount_pending = orangefs_sb->mount_pending;
116 spin_unlock(&orangefs_superblocks_lock);
117 return mount_pending;
120 static int orangefs_devreq_open(struct inode *inode, struct file *file)
124 /* in order to ensure that the filesystem driver sees correct UIDs */
125 if (file->f_cred->user_ns != &init_user_ns) {
126 gossip_err("%s: device cannot be opened outside init_user_ns\n",
131 if (!(file->f_flags & O_NONBLOCK)) {
132 gossip_err("%s: device cannot be opened in blocking mode\n",
137 gossip_debug(GOSSIP_DEV_DEBUG, "client-core: opening device\n");
138 mutex_lock(&devreq_mutex);
140 if (open_access_count == 0) {
141 open_access_count = 1;
146 mutex_unlock(&devreq_mutex);
150 gossip_debug(GOSSIP_DEV_DEBUG,
151 "pvfs2-client-core: open device complete (ret = %d)\n",
156 /* Function for read() callers into the device */
157 static ssize_t orangefs_devreq_read(struct file *file,
159 size_t count, loff_t *offset)
161 struct orangefs_kernel_op_s *op, *temp;
162 __s32 proto_ver = ORANGEFS_KERNEL_PROTO_VERSION;
163 static __s32 magic = ORANGEFS_DEVREQ_MAGIC;
164 struct orangefs_kernel_op_s *cur_op;
167 /* We do not support blocking IO. */
168 if (!(file->f_flags & O_NONBLOCK)) {
169 gossip_err("%s: blocking read from client-core.\n",
175 * The client will do an ioctl to find MAX_DEV_REQ_UPSIZE, then
176 * always read with that size buffer.
178 if (count != MAX_DEV_REQ_UPSIZE) {
179 gossip_err("orangefs: client-core tried to read wrong size\n");
185 /* Get next op (if any) from top of list. */
186 spin_lock(&orangefs_request_list_lock);
187 list_for_each_entry_safe(op, temp, &orangefs_request_list, list) {
189 /* This lock is held past the end of the loop when we break. */
190 spin_lock(&op->lock);
191 if (unlikely(op_state_purged(op) || op_state_given_up(op))) {
192 spin_unlock(&op->lock);
196 fsid = fsid_of_op(op);
197 if (fsid != ORANGEFS_FS_ID_NULL) {
199 /* Skip ops whose filesystem needs to be mounted. */
200 ret = fs_mount_pending(fsid);
202 gossip_debug(GOSSIP_DEV_DEBUG,
203 "%s: mount pending, skipping op tag "
207 get_opname_string(op));
208 spin_unlock(&op->lock);
211 * Skip ops whose filesystem we don't know about unless
212 * it is being mounted or unmounted. It is possible for
213 * a filesystem we don't know about to be unmounted if
214 * it fails to mount in the kernel after userspace has
215 * been sent the mount request.
217 /* XXX: is there a better way to detect this? */
218 } else if (ret == -1 &&
220 ORANGEFS_VFS_OP_FS_MOUNT ||
222 ORANGEFS_VFS_OP_GETATTR ||
224 ORANGEFS_VFS_OP_FS_UMOUNT)) {
225 gossip_debug(GOSSIP_DEV_DEBUG,
226 "orangefs: skipping op tag %llu %s\n",
227 llu(op->tag), get_opname_string(op));
229 "orangefs: ERROR: fs_mount_pending %d\n",
231 spin_unlock(&op->lock);
236 * Either this op does not pertain to a filesystem, is mounting
237 * a filesystem, or pertains to a mounted filesystem. Let it
245 * At this point we either have a valid op and can continue or have not
246 * found an op and must ask the client to try again later.
249 spin_unlock(&orangefs_request_list_lock);
253 gossip_debug(GOSSIP_DEV_DEBUG, "%s: reading op tag %llu %s\n",
256 get_opname_string(cur_op));
259 * Such an op should never be on the list in the first place. If so, we
262 if (op_state_in_progress(cur_op) || op_state_serviced(cur_op)) {
263 gossip_err("orangefs: ERROR: Current op already queued.\n");
264 list_del_init(&cur_op->list);
265 spin_unlock(&cur_op->lock);
266 spin_unlock(&orangefs_request_list_lock);
270 list_del_init(&cur_op->list);
271 spin_unlock(&orangefs_request_list_lock);
273 spin_unlock(&cur_op->lock);
275 /* Push the upcall out. */
276 ret = copy_to_user(buf, &proto_ver, sizeof(__s32));
279 ret = copy_to_user(buf+sizeof(__s32), &magic, sizeof(__s32));
282 ret = copy_to_user(buf+2 * sizeof(__s32), &cur_op->tag, sizeof(__u64));
285 ret = copy_to_user(buf+2*sizeof(__s32)+sizeof(__u64), &cur_op->upcall,
286 sizeof(struct orangefs_upcall_s));
290 spin_lock(&orangefs_htable_ops_in_progress_lock);
291 spin_lock(&cur_op->lock);
292 if (unlikely(op_state_given_up(cur_op))) {
293 spin_unlock(&cur_op->lock);
294 spin_unlock(&orangefs_htable_ops_in_progress_lock);
295 complete(&cur_op->waitq);
300 * Set the operation to be in progress and move it between lists since
301 * it has been sent to the client.
303 set_op_state_inprogress(cur_op);
304 gossip_debug(GOSSIP_DEV_DEBUG,
305 "%s: 1 op:%s: op_state:%d: process:%s:\n",
307 get_opname_string(cur_op),
310 orangefs_devreq_add_op(cur_op);
311 spin_unlock(&cur_op->lock);
312 spin_unlock(&orangefs_htable_ops_in_progress_lock);
314 /* The client only asks to read one size buffer. */
315 return MAX_DEV_REQ_UPSIZE;
318 * We were unable to copy the op data to the client. Put the op back in
319 * list. If client has crashed, the op will be purged later when the
320 * device is released.
322 gossip_err("orangefs: Failed to copy data to user space\n");
323 spin_lock(&orangefs_request_list_lock);
324 spin_lock(&cur_op->lock);
325 if (likely(!op_state_given_up(cur_op))) {
326 set_op_state_waiting(cur_op);
327 gossip_debug(GOSSIP_DEV_DEBUG,
328 "%s: 2 op:%s: op_state:%d: process:%s:\n",
330 get_opname_string(cur_op),
333 list_add(&cur_op->list, &orangefs_request_list);
334 spin_unlock(&cur_op->lock);
336 spin_unlock(&cur_op->lock);
337 complete(&cur_op->waitq);
339 spin_unlock(&orangefs_request_list_lock);
344 * Function for writev() callers into the device.
346 * Userspace should have written:
350 * - struct orangefs_downcall_s
351 * - trailer buffer (in the case of READDIR operations)
353 static ssize_t orangefs_devreq_write_iter(struct kiocb *iocb,
354 struct iov_iter *iter)
357 struct orangefs_kernel_op_s *op = NULL;
363 int total = ret = iov_iter_count(iter);
365 int downcall_size = sizeof(struct orangefs_downcall_s);
366 int head_size = sizeof(head);
368 gossip_debug(GOSSIP_DEV_DEBUG, "%s: total:%d: ret:%zd:\n",
373 if (total < MAX_DEV_REQ_DOWNSIZE) {
374 gossip_err("%s: total:%d: must be at least:%u:\n",
377 (unsigned int) MAX_DEV_REQ_DOWNSIZE);
381 n = copy_from_iter(&head, head_size, iter);
383 gossip_err("%s: failed to copy head.\n", __func__);
387 if (head.version < ORANGEFS_MINIMUM_USERSPACE_VERSION) {
388 gossip_err("%s: userspace claims version"
389 "%d, minimum version required: %d.\n",
392 ORANGEFS_MINIMUM_USERSPACE_VERSION);
396 if (head.magic != ORANGEFS_DEVREQ_MAGIC) {
397 gossip_err("Error: Device magic number does not match.\n");
401 if (!orangefs_userspace_version) {
402 orangefs_userspace_version = head.version;
403 } else if (orangefs_userspace_version != head.version) {
404 gossip_err("Error: userspace version changes\n");
408 /* remove the op from the in progress hash table */
409 op = orangefs_devreq_remove_op(head.tag);
411 gossip_debug(GOSSIP_DEV_DEBUG,
412 "%s: No one's waiting for tag %llu\n",
413 __func__, llu(head.tag));
417 n = copy_from_iter(&op->downcall, downcall_size, iter);
418 if (n != downcall_size) {
419 gossip_err("%s: failed to copy downcall.\n", __func__);
423 if (op->downcall.status)
427 * We've successfully peeled off the head and the downcall.
428 * Something has gone awry if total doesn't equal the
429 * sum of head_size, downcall_size and trailer_size.
431 if ((head_size + downcall_size + op->downcall.trailer_size) != total) {
432 gossip_err("%s: funky write, head_size:%d"
433 ": downcall_size:%d: trailer_size:%lld"
434 ": total size:%d:\n",
438 op->downcall.trailer_size,
443 /* Only READDIR operations should have trailers. */
444 if ((op->downcall.type != ORANGEFS_VFS_OP_READDIR) &&
445 (op->downcall.trailer_size != 0)) {
446 gossip_err("%s: %x operation with trailer.",
452 /* READDIR operations should always have trailers. */
453 if ((op->downcall.type == ORANGEFS_VFS_OP_READDIR) &&
454 (op->downcall.trailer_size == 0)) {
455 gossip_err("%s: %x operation with no trailer.",
461 if (op->downcall.type != ORANGEFS_VFS_OP_READDIR)
464 op->downcall.trailer_buf =
465 vmalloc(op->downcall.trailer_size);
466 if (op->downcall.trailer_buf == NULL) {
467 gossip_err("%s: failed trailer vmalloc.\n",
471 memset(op->downcall.trailer_buf, 0, op->downcall.trailer_size);
472 n = copy_from_iter(op->downcall.trailer_buf,
473 op->downcall.trailer_size,
475 if (n != op->downcall.trailer_size) {
476 gossip_err("%s: failed to copy trailer.\n", __func__);
477 vfree(op->downcall.trailer_buf);
483 * Return to vfs waitqueue, and back to service_operation
484 * through wait_for_matching_downcall.
486 spin_lock(&op->lock);
487 if (unlikely(op_is_cancel(op))) {
488 spin_unlock(&op->lock);
490 } else if (unlikely(op_state_given_up(op))) {
491 spin_unlock(&op->lock);
492 complete(&op->waitq);
494 set_op_state_serviced(op);
495 gossip_debug(GOSSIP_DEV_DEBUG,
496 "%s: op:%s: op_state:%d: process:%s:\n",
498 get_opname_string(op),
501 spin_unlock(&op->lock);
506 op->downcall.status = -(ORANGEFS_ERROR_BIT | 9);
511 op->downcall.status = -(ORANGEFS_ERROR_BIT | 8);
517 * NOTE: gets called when the last reference to this device is dropped.
518 * Using the open_access_count variable, we enforce a reference count
519 * on this file so that it can be opened by only one process at a time.
520 * the devreq_mutex is used to make sure all i/o has completed
521 * before we call orangefs_bufmap_finalize, and similar such tricky
524 static int orangefs_devreq_release(struct inode *inode, struct file *file)
528 gossip_debug(GOSSIP_DEV_DEBUG,
529 "%s:pvfs2-client-core: exiting, closing device\n",
532 mutex_lock(&devreq_mutex);
533 orangefs_bufmap_finalize();
535 open_access_count = -1;
537 unmounted = mark_all_pending_mounts();
538 gossip_debug(GOSSIP_DEV_DEBUG, "ORANGEFS Device Close: Filesystem(s) %s\n",
539 (unmounted ? "UNMOUNTED" : "MOUNTED"));
542 purge_inprogress_ops();
544 orangefs_bufmap_run_down();
546 gossip_debug(GOSSIP_DEV_DEBUG,
547 "pvfs2-client-core: device close complete\n");
548 open_access_count = 0;
549 orangefs_userspace_version = 0;
550 mutex_unlock(&devreq_mutex);
554 int is_daemon_in_service(void)
559 * What this function does is checks if client-core is alive
560 * based on the access count we maintain on the device.
562 mutex_lock(&devreq_mutex);
563 in_service = open_access_count == 1 ? 0 : -EIO;
564 mutex_unlock(&devreq_mutex);
568 bool __is_daemon_in_service(void)
570 return open_access_count == 1;
573 static inline long check_ioctl_command(unsigned int command)
575 /* Check for valid ioctl codes */
576 if (_IOC_TYPE(command) != ORANGEFS_DEV_MAGIC) {
577 gossip_err("device ioctl magic numbers don't match! Did you rebuild pvfs2-client-core/libpvfs2? [cmd %x, magic %x != %x]\n",
583 /* and valid ioctl commands */
584 if (_IOC_NR(command) >= ORANGEFS_DEV_MAXNR || _IOC_NR(command) <= 0) {
585 gossip_err("Invalid ioctl command number [%d >= %d]\n",
586 _IOC_NR(command), ORANGEFS_DEV_MAXNR);
592 static long dispatch_ioctl_command(unsigned int command, unsigned long arg)
594 static __s32 magic = ORANGEFS_DEVREQ_MAGIC;
595 static __s32 max_up_size = MAX_DEV_REQ_UPSIZE;
596 static __s32 max_down_size = MAX_DEV_REQ_DOWNSIZE;
597 struct ORANGEFS_dev_map_desc user_desc;
599 int upstream_kmod = 1;
600 struct orangefs_sb_info_s *orangefs_sb;
602 /* mtmoore: add locking here */
605 case ORANGEFS_DEV_GET_MAGIC:
606 return ((put_user(magic, (__s32 __user *) arg) == -EFAULT) ?
609 case ORANGEFS_DEV_GET_MAX_UPSIZE:
610 return ((put_user(max_up_size,
611 (__s32 __user *) arg) == -EFAULT) ?
614 case ORANGEFS_DEV_GET_MAX_DOWNSIZE:
615 return ((put_user(max_down_size,
616 (__s32 __user *) arg) == -EFAULT) ?
619 case ORANGEFS_DEV_MAP:
620 ret = copy_from_user(&user_desc,
621 (struct ORANGEFS_dev_map_desc __user *)
623 sizeof(struct ORANGEFS_dev_map_desc));
624 /* WTF -EIO and not -EFAULT? */
625 return ret ? -EIO : orangefs_bufmap_initialize(&user_desc);
626 case ORANGEFS_DEV_REMOUNT_ALL:
627 gossip_debug(GOSSIP_DEV_DEBUG,
628 "%s: got ORANGEFS_DEV_REMOUNT_ALL\n",
632 * remount all mounted orangefs volumes to regain the lost
633 * dynamic mount tables (if any) -- NOTE: this is done
634 * without keeping the superblock list locked due to the
635 * upcall/downcall waiting. also, the request mutex is
636 * used to ensure that no operations will be serviced until
637 * all of the remounts are serviced (to avoid ops between
640 ret = mutex_lock_interruptible(&orangefs_request_mutex);
643 gossip_debug(GOSSIP_DEV_DEBUG,
644 "%s: priority remount in progress\n",
646 spin_lock(&orangefs_superblocks_lock);
647 list_for_each_entry(orangefs_sb, &orangefs_superblocks, list) {
649 * We have to drop the spinlock, so entries can be
650 * removed. They can't be freed, though, so we just
651 * keep the forward pointers and zero the back ones -
652 * that way we can get to the rest of the list.
654 if (!orangefs_sb->list.prev)
656 gossip_debug(GOSSIP_DEV_DEBUG,
657 "%s: Remounting SB %p\n",
661 spin_unlock(&orangefs_superblocks_lock);
662 ret = orangefs_remount(orangefs_sb);
663 spin_lock(&orangefs_superblocks_lock);
665 gossip_debug(GOSSIP_DEV_DEBUG,
666 "SB %p remount failed\n",
671 spin_unlock(&orangefs_superblocks_lock);
672 gossip_debug(GOSSIP_DEV_DEBUG,
673 "%s: priority remount complete\n",
675 mutex_unlock(&orangefs_request_mutex);
678 case ORANGEFS_DEV_UPSTREAM:
679 ret = copy_to_user((void __user *)arg,
681 sizeof(upstream_kmod));
688 case ORANGEFS_DEV_CLIENT_MASK:
689 return orangefs_debugfs_new_client_mask((void __user *)arg);
690 case ORANGEFS_DEV_CLIENT_STRING:
691 return orangefs_debugfs_new_client_string((void __user *)arg);
692 case ORANGEFS_DEV_DEBUG:
693 return orangefs_debugfs_new_debug((void __user *)arg);
700 static long orangefs_devreq_ioctl(struct file *file,
701 unsigned int command, unsigned long arg)
705 /* Check for properly constructed commands */
706 ret = check_ioctl_command(command);
710 return (int)dispatch_ioctl_command(command, arg);
713 #ifdef CONFIG_COMPAT /* CONFIG_COMPAT is in .config */
715 /* Compat structure for the ORANGEFS_DEV_MAP ioctl */
716 struct ORANGEFS_dev_map_desc32 {
723 static unsigned long translate_dev_map26(unsigned long args, long *error)
725 struct ORANGEFS_dev_map_desc32 __user *p32 = (void __user *)args;
727 * Depending on the architecture, allocate some space on the
728 * user-call-stack based on our expected layout.
730 struct ORANGEFS_dev_map_desc __user *p =
731 compat_alloc_user_space(sizeof(*p));
735 /* get the ptr from the 32 bit user-space */
736 if (get_user(addr, &p32->ptr))
738 /* try to put that into a 64-bit layout */
739 if (put_user(compat_ptr(addr), &p->ptr))
741 /* copy the remaining fields */
742 if (copy_in_user(&p->total_size, &p32->total_size, sizeof(__s32)))
744 if (copy_in_user(&p->size, &p32->size, sizeof(__s32)))
746 if (copy_in_user(&p->count, &p32->count, sizeof(__s32)))
748 return (unsigned long)p;
755 * 32 bit user-space apps' ioctl handlers when kernel modules
756 * is compiled as a 64 bit one
758 static long orangefs_devreq_compat_ioctl(struct file *filp, unsigned int cmd,
762 unsigned long arg = args;
764 /* Check for properly constructed commands */
765 ret = check_ioctl_command(cmd);
768 if (cmd == ORANGEFS_DEV_MAP) {
770 * convert the arguments to what we expect internally
773 arg = translate_dev_map26(args, &ret);
775 gossip_err("Could not translate dev map\n");
779 /* no other ioctl requires translation */
780 return dispatch_ioctl_command(cmd, arg);
783 #endif /* CONFIG_COMPAT is in .config */
785 /* the assigned character device major number */
786 static int orangefs_dev_major;
789 * Initialize orangefs device specific state:
790 * Must be called at module load time only
792 int orangefs_dev_init(void)
794 /* register orangefs-req device */
795 orangefs_dev_major = register_chrdev(0,
796 ORANGEFS_REQDEVICE_NAME,
797 &orangefs_devreq_file_operations);
798 if (orangefs_dev_major < 0) {
799 gossip_debug(GOSSIP_DEV_DEBUG,
800 "Failed to register /dev/%s (error %d)\n",
801 ORANGEFS_REQDEVICE_NAME, orangefs_dev_major);
802 return orangefs_dev_major;
805 gossip_debug(GOSSIP_DEV_DEBUG,
806 "*** /dev/%s character device registered ***\n",
807 ORANGEFS_REQDEVICE_NAME);
808 gossip_debug(GOSSIP_DEV_DEBUG, "'mknod /dev/%s c %d 0'.\n",
809 ORANGEFS_REQDEVICE_NAME, orangefs_dev_major);
813 void orangefs_dev_cleanup(void)
815 unregister_chrdev(orangefs_dev_major, ORANGEFS_REQDEVICE_NAME);
816 gossip_debug(GOSSIP_DEV_DEBUG,
817 "*** /dev/%s character device unregistered ***\n",
818 ORANGEFS_REQDEVICE_NAME);
821 static unsigned int orangefs_devreq_poll(struct file *file,
822 struct poll_table_struct *poll_table)
824 int poll_revent_mask = 0;
826 poll_wait(file, &orangefs_request_list_waitq, poll_table);
828 if (!list_empty(&orangefs_request_list))
829 poll_revent_mask |= POLL_IN;
830 return poll_revent_mask;
833 const struct file_operations orangefs_devreq_file_operations = {
834 .owner = THIS_MODULE,
835 .read = orangefs_devreq_read,
836 .write_iter = orangefs_devreq_write_iter,
837 .open = orangefs_devreq_open,
838 .release = orangefs_devreq_release,
839 .unlocked_ioctl = orangefs_devreq_ioctl,
841 #ifdef CONFIG_COMPAT /* CONFIG_COMPAT is in .config */
842 .compat_ioctl = orangefs_devreq_compat_ioctl,
844 .poll = orangefs_devreq_poll