compat: fs: Generic compat_sys_sendfile implementation
[firefly-linux-kernel-4.4.55.git] / fs / file_table.c
index a305d9e2d1b2aac05dcd456bdd23885652272439..c6780163bf3e911d347c7a553e028c7ba19ce463 100644 (file)
@@ -23,6 +23,8 @@
 #include <linux/lglock.h>
 #include <linux/percpu_counter.h>
 #include <linux/percpu.h>
+#include <linux/hardirq.h>
+#include <linux/task_work.h>
 #include <linux/ima.h>
 
 #include <linux/atomic.h>
@@ -41,7 +43,7 @@ static struct kmem_cache *filp_cachep __read_mostly;
 
 static struct percpu_counter nr_files __cacheline_aligned_in_smp;
 
-static inline void file_free_rcu(struct rcu_head *head)
+static void file_free_rcu(struct rcu_head *head)
 {
        struct file *f = container_of(head, struct file, f_u.fu_rcuhead);
 
@@ -215,7 +217,7 @@ static void drop_file_write_access(struct file *file)
                return;
        if (file_check_writeable(file) != 0)
                return;
-       mnt_drop_write(mnt);
+       __mnt_drop_write(mnt);
        file_release_write(file);
 }
 
@@ -251,7 +253,6 @@ static void __fput(struct file *file)
        }
        fops_put(file->f_op);
        put_pid(file->f_owner.pid);
-       file_sb_list_del(file);
        if ((file->f_mode & (FMODE_READ | FMODE_WRITE)) == FMODE_READ)
                i_readcount_dec(inode);
        if (file->f_mode & FMODE_WRITE)
@@ -263,120 +264,81 @@ static void __fput(struct file *file)
        mntput(mnt);
 }
 
-void fput(struct file *file)
+static DEFINE_SPINLOCK(delayed_fput_lock);
+static LIST_HEAD(delayed_fput_list);
+static void delayed_fput(struct work_struct *unused)
 {
-       if (atomic_long_dec_and_test(&file->f_count))
-               __fput(file);
+       LIST_HEAD(head);
+       spin_lock_irq(&delayed_fput_lock);
+       list_splice_init(&delayed_fput_list, &head);
+       spin_unlock_irq(&delayed_fput_lock);
+       while (!list_empty(&head)) {
+               struct file *f = list_first_entry(&head, struct file, f_u.fu_list);
+               list_del_init(&f->f_u.fu_list);
+               __fput(f);
+       }
 }
 
-EXPORT_SYMBOL(fput);
-
-struct file *fget(unsigned int fd)
+static void ____fput(struct callback_head *work)
 {
-       struct file *file;
-       struct files_struct *files = current->files;
-
-       rcu_read_lock();
-       file = fcheck_files(files, fd);
-       if (file) {
-               /* File object ref couldn't be taken */
-               if (file->f_mode & FMODE_PATH ||
-                   !atomic_long_inc_not_zero(&file->f_count))
-                       file = NULL;
-       }
-       rcu_read_unlock();
-
-       return file;
+       __fput(container_of(work, struct file, f_u.fu_rcuhead));
 }
 
-EXPORT_SYMBOL(fget);
-
-struct file *fget_raw(unsigned int fd)
+/*
+ * If kernel thread really needs to have the final fput() it has done
+ * to complete, call this.  The only user right now is the boot - we
+ * *do* need to make sure our writes to binaries on initramfs has
+ * not left us with opened struct file waiting for __fput() - execve()
+ * won't work without that.  Please, don't add more callers without
+ * very good reasons; in particular, never call that with locks
+ * held and never call that from a thread that might need to do
+ * some work on any kind of umount.
+ */
+void flush_delayed_fput(void)
 {
-       struct file *file;
-       struct files_struct *files = current->files;
-
-       rcu_read_lock();
-       file = fcheck_files(files, fd);
-       if (file) {
-               /* File object ref couldn't be taken */
-               if (!atomic_long_inc_not_zero(&file->f_count))
-                       file = NULL;
-       }
-       rcu_read_unlock();
-
-       return file;
+       delayed_fput(NULL);
 }
 
-EXPORT_SYMBOL(fget_raw);
+static DECLARE_WORK(delayed_fput_work, delayed_fput);
 
-/*
- * Lightweight file lookup - no refcnt increment if fd table isn't shared.
- *
- * You can use this instead of fget if you satisfy all of the following
- * conditions:
- * 1) You must call fput_light before exiting the syscall and returning control
- *    to userspace (i.e. you cannot remember the returned struct file * after
- *    returning to userspace).
- * 2) You must not call filp_close on the returned struct file * in between
- *    calls to fget_light and fput_light.
- * 3) You must not clone the current task in between the calls to fget_light
- *    and fput_light.
- *
- * The fput_needed flag returned by fget_light should be passed to the
- * corresponding fput_light.
- */
-struct file *fget_light(unsigned int fd, int *fput_needed)
+void fput(struct file *file)
 {
-       struct file *file;
-       struct files_struct *files = current->files;
-
-       *fput_needed = 0;
-       if (atomic_read(&files->count) == 1) {
-               file = fcheck_files(files, fd);
-               if (file && (file->f_mode & FMODE_PATH))
-                       file = NULL;
-       } else {
-               rcu_read_lock();
-               file = fcheck_files(files, fd);
-               if (file) {
-                       if (!(file->f_mode & FMODE_PATH) &&
-                           atomic_long_inc_not_zero(&file->f_count))
-                               *fput_needed = 1;
-                       else
-                               /* Didn't get the reference, someone's freed */
-                               file = NULL;
+       if (atomic_long_dec_and_test(&file->f_count)) {
+               struct task_struct *task = current;
+               file_sb_list_del(file);
+               if (unlikely(in_interrupt() || task->flags & PF_KTHREAD)) {
+                       unsigned long flags;
+                       spin_lock_irqsave(&delayed_fput_lock, flags);
+                       list_add(&file->f_u.fu_list, &delayed_fput_list);
+                       schedule_work(&delayed_fput_work);
+                       spin_unlock_irqrestore(&delayed_fput_lock, flags);
+                       return;
                }
-               rcu_read_unlock();
+               init_task_work(&file->f_u.fu_rcuhead, ____fput);
+               task_work_add(task, &file->f_u.fu_rcuhead, true);
        }
-
-       return file;
 }
 
-struct file *fget_raw_light(unsigned int fd, int *fput_needed)
+/*
+ * synchronous analog of fput(); for kernel threads that might be needed
+ * in some umount() (and thus can't use flush_delayed_fput() without
+ * risking deadlocks), need to wait for completion of __fput() and know
+ * for this specific struct file it won't involve anything that would
+ * need them.  Use only if you really need it - at the very least,
+ * don't blindly convert fput() by kernel thread to that.
+ */
+void __fput_sync(struct file *file)
 {
-       struct file *file;
-       struct files_struct *files = current->files;
-
-       *fput_needed = 0;
-       if (atomic_read(&files->count) == 1) {
-               file = fcheck_files(files, fd);
-       } else {
-               rcu_read_lock();
-               file = fcheck_files(files, fd);
-               if (file) {
-                       if (atomic_long_inc_not_zero(&file->f_count))
-                               *fput_needed = 1;
-                       else
-                               /* Didn't get the reference, someone's freed */
-                               file = NULL;
-               }
-               rcu_read_unlock();
+       if (atomic_long_dec_and_test(&file->f_count)) {
+               struct task_struct *task = current;
+               file_sb_list_del(file);
+               BUG_ON(!(task->flags & PF_KTHREAD));
+               __fput(file);
        }
-
-       return file;
 }
 
+EXPORT_SYMBOL(fput);
+
 void put_filp(struct file *file)
 {
        if (atomic_long_dec_and_test(&file->f_count)) {
@@ -483,10 +445,8 @@ void mark_files_ro(struct super_block *sb)
 {
        struct file *f;
 
-retry:
        lg_global_lock(&files_lglock);
        do_file_list_for_each_entry(sb, f) {
-               struct vfsmount *mnt;
                if (!S_ISREG(f->f_path.dentry->d_inode->i_mode))
                       continue;
                if (!file_count(f))
@@ -499,12 +459,7 @@ retry:
                if (file_check_writeable(f) != 0)
                        continue;
                file_release_write(f);
-               mnt = mntget(f->f_path.mnt);
-               /* This can sleep, so we can't hold the spinlock. */
-               lg_global_unlock(&files_lglock);
-               mnt_drop_write(mnt);
-               mntput(mnt);
-               goto retry;
+               mnt_drop_write_file(f);
        } while_file_list_for_each_entry;
        lg_global_unlock(&files_lglock);
 }