orangefs: fix orangefs_superblock locking
authorAl Viro <viro@zeniv.linux.org.uk>
Fri, 25 Mar 2016 23:56:34 +0000 (19:56 -0400)
committerMike Marshall <hubcap@omnibond.com>
Sat, 26 Mar 2016 11:22:00 +0000 (07:22 -0400)
* switch orangefs_remount() to taking ORANGEFS_SB(sb) instead of sb
* remove from the list _before_ orangefs_unmount() - request_mutex
in the latter will make sure that nothing observed in the loop in
ORANGEFS_DEV_REMOUNT_ALL handling will get freed until the end
of loop
* on removal, keep the forward pointer and zero the back one.  That
way we can drop and regain the spinlock in the loop body (again,
ORANGEFS_DEV_REMOUNT_ALL one) and still be able to get to the
rest of the list.

Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Mike Marshall <hubcap@omnibond.com>
fs/orangefs/devorangefs-req.c
fs/orangefs/orangefs-kernel.h
fs/orangefs/super.c

index 35418d0b77bfc1378a5dd3f6efe3bc57f1a3d961..db170beba7974c3d5d4491df452f0d75ca15ea13 100644 (file)
@@ -572,8 +572,7 @@ static long dispatch_ioctl_command(unsigned int command, unsigned long arg)
        struct dev_mask_info_s mask_info = { 0 };
        struct dev_mask2_info_s mask2_info = { 0, 0 };
        int upstream_kmod = 1;
-       struct list_head *tmp = NULL;
-       struct orangefs_sb_info_s *orangefs_sb = NULL;
+       struct orangefs_sb_info_s *orangefs_sb;
 
        /* mtmoore: add locking here */
 
@@ -619,26 +618,32 @@ static long dispatch_ioctl_command(unsigned int command, unsigned long arg)
                gossip_debug(GOSSIP_DEV_DEBUG,
                             "%s: priority remount in progress\n",
                             __func__);
-               list_for_each(tmp, &orangefs_superblocks) {
-                       orangefs_sb =
-                               list_entry(tmp,
-                                          struct orangefs_sb_info_s,
-                                          list);
-                       if (orangefs_sb && (orangefs_sb->sb)) {
+               spin_lock(&orangefs_superblocks_lock);
+               list_for_each_entry(orangefs_sb, &orangefs_superblocks, list) {
+                       /*
+                        * We have to drop the spinlock, so entries can be
+                        * removed.  They can't be freed, though, so we just
+                        * keep the forward pointers and zero the back ones -
+                        * that way we can get to the rest of the list.
+                        */
+                       if (!orangefs_sb->list.prev)
+                               continue;
+                       gossip_debug(GOSSIP_DEV_DEBUG,
+                                    "%s: Remounting SB %p\n",
+                                    __func__,
+                                    orangefs_sb);
+
+                       spin_unlock(&orangefs_superblocks_lock);
+                       ret = orangefs_remount(orangefs_sb);
+                       spin_lock(&orangefs_superblocks_lock);
+                       if (ret) {
                                gossip_debug(GOSSIP_DEV_DEBUG,
-                                            "%s: Remounting SB %p\n",
-                                            __func__,
+                                            "SB %p remount failed\n",
                                             orangefs_sb);
-
-                               ret = orangefs_remount(orangefs_sb->sb);
-                               if (ret) {
-                                       gossip_debug(GOSSIP_DEV_DEBUG,
-                                                    "SB %p remount failed\n",
-                                                    orangefs_sb);
-                                       break;
-                               }
+                               break;
                        }
                }
+               spin_unlock(&orangefs_superblocks_lock);
                gossip_debug(GOSSIP_DEV_DEBUG,
                             "%s: priority remount complete\n",
                             __func__);
index db258d2ccc6a86528bfa63ebe698442d9de0c3cc..a9925e296ceb4c5f9b20fd47d37becd13941c90a 100644 (file)
@@ -462,7 +462,7 @@ struct dentry *orangefs_mount(struct file_system_type *fst,
                           void *data);
 
 void orangefs_kill_sb(struct super_block *sb);
-int orangefs_remount(struct super_block *sb);
+int orangefs_remount(struct orangefs_sb_info_s *);
 
 int fsid_key_table_initialize(void);
 void fsid_key_table_finalize(void);
@@ -598,38 +598,6 @@ int service_operation(struct orangefs_kernel_op_s *op,
        ((ORANGEFS_SB(inode->i_sb)->flags & ORANGEFS_OPT_INTR) ? \
                ORANGEFS_OP_INTERRUPTIBLE : 0)
 
-#define add_orangefs_sb(sb)                                            \
-do {                                                                   \
-       gossip_debug(GOSSIP_SUPER_DEBUG,                                \
-                    "Adding SB %p to orangefs superblocks\n",          \
-                    ORANGEFS_SB(sb));                                  \
-       spin_lock(&orangefs_superblocks_lock);                          \
-       list_add_tail(&ORANGEFS_SB(sb)->list, &orangefs_superblocks);           \
-       spin_unlock(&orangefs_superblocks_lock); \
-} while (0)
-
-#define remove_orangefs_sb(sb)                                         \
-do {                                                                   \
-       struct list_head *tmp = NULL;                                   \
-       struct list_head *tmp_safe = NULL;                              \
-       struct orangefs_sb_info_s *orangefs_sb = NULL;                  \
-                                                                       \
-       spin_lock(&orangefs_superblocks_lock);                          \
-       list_for_each_safe(tmp, tmp_safe, &orangefs_superblocks) {              \
-               orangefs_sb = list_entry(tmp,                           \
-                                     struct orangefs_sb_info_s,                \
-                                     list);                            \
-               if (orangefs_sb && (orangefs_sb->sb == sb)) {                   \
-                       gossip_debug(GOSSIP_SUPER_DEBUG,                \
-                           "Removing SB %p from orangefs superblocks\n",       \
-                       orangefs_sb);                                   \
-                       list_del(&orangefs_sb->list);                   \
-                       break;                                          \
-               }                                                       \
-       }                                                               \
-       spin_unlock(&orangefs_superblocks_lock);                                \
-} while (0)
-
 #define fill_default_sys_attrs(sys_attr, type, mode)                   \
 do {                                                                   \
        sys_attr.owner = from_kuid(current_user_ns(), current_fsuid()); \
index 5a89b808396679afc12105af3c836d1e0682485f..b9da9a0281c9b916e33d3d226abf36130a76b06c 100644 (file)
@@ -210,7 +210,7 @@ static int orangefs_remount_fs(struct super_block *sb, int *flags, char *data)
  * the client regains all of the mount information from us.
  * NOTE: this function assumes that the request_mutex is already acquired!
  */
-int orangefs_remount(struct super_block *sb)
+int orangefs_remount(struct orangefs_sb_info_s *orangefs_sb)
 {
        struct orangefs_kernel_op_s *new_op;
        int ret = -EINVAL;
@@ -221,7 +221,7 @@ int orangefs_remount(struct super_block *sb)
        if (!new_op)
                return -ENOMEM;
        strncpy(new_op->upcall.req.fs_mount.orangefs_config_server,
-               ORANGEFS_SB(sb)->devname,
+               orangefs_sb->devname,
                ORANGEFS_MAX_SERVER_ADDR_LEN);
 
        gossip_debug(GOSSIP_SUPER_DEBUG,
@@ -244,8 +244,8 @@ int orangefs_remount(struct super_block *sb)
                 * short-lived mapping that the system interface uses
                 * to map this superblock to a particular mount entry
                 */
-               ORANGEFS_SB(sb)->id = new_op->downcall.resp.fs_mount.id;
-               ORANGEFS_SB(sb)->mount_pending = 0;
+               orangefs_sb->id = new_op->downcall.resp.fs_mount.id;
+               orangefs_sb->mount_pending = 0;
        }
 
        op_release(new_op);
@@ -485,7 +485,12 @@ struct dentry *orangefs_mount(struct file_system_type *fst,
         * finally, add this sb to our list of known orangefs
         * sb's
         */
-       add_orangefs_sb(sb);
+       gossip_debug(GOSSIP_SUPER_DEBUG,
+                    "Adding SB %p to orangefs superblocks\n",
+                    ORANGEFS_SB(sb));
+       spin_lock(&orangefs_superblocks_lock);
+       list_add_tail(&ORANGEFS_SB(sb)->list, &orangefs_superblocks);
+       spin_unlock(&orangefs_superblocks_lock);
        op_release(new_op);
        return dget(sb->s_root);
 
@@ -512,10 +517,21 @@ void orangefs_kill_sb(struct super_block *sb)
         * issue the unmount to userspace to tell it to remove the
         * dynamic mount info it has for this superblock
         */
-       orangefs_unmount_sb(sb);
+        orangefs_unmount_sb(sb);
 
        /* remove the sb from our list of orangefs specific sb's */
-       remove_orangefs_sb(sb);
+
+       spin_lock(&orangefs_superblocks_lock);
+       __list_del_entry(&ORANGEFS_SB(sb)->list);       /* not list_del_init */
+       ORANGEFS_SB(sb)->list.prev = NULL;
+       spin_unlock(&orangefs_superblocks_lock);
+
+       /*
+        * make sure that ORANGEFS_DEV_REMOUNT_ALL loop that might've seen us
+        * gets completed before we free the dang thing.
+        */
+       mutex_lock(&request_mutex);
+       mutex_unlock(&request_mutex);
 
        /* free the orangefs superblock private data */
        kfree(ORANGEFS_SB(sb));