acct: new lifetime rules
authorAl Viro <viro@zeniv.linux.org.uk>
Thu, 7 Aug 2014 11:51:03 +0000 (07:51 -0400)
committerAl Viro <viro@zeniv.linux.org.uk>
Thu, 7 Aug 2014 18:40:08 +0000 (14:40 -0400)
Do not reuse bsd_acct_struct after closing the damn thing.
Structure lifetime is controlled by refcount now.  We also
have a mutex in there, held over closing and writing (the
file is O_APPEND, so we are not losing any concurrency).

As the result, we do not need to bother with get_file()/fput()
on log write anymore.  Moreover, do_acct_process() only needs
acct itself; file and pidns are picked from it.

Killed instances are distinguished by having NULL ->ns.
Refcount is protected by acct_lock; anybody taking the
mutex needs to grab a reference first.

The things will get a lot simpler in the next commits - this
is just the minimal chunk switching to the new lifetime rules.

Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
kernel/acct.c

index 08963a29287864fb4d466fcc440e0bc4e3d5538b..f9ef9db55c0e4dc4c98c66fdb71ee07614c4edba 100644 (file)
@@ -75,15 +75,11 @@ int acct_parm[3] = {4, 2, 30};
 /*
  * External references and all of the globals.
  */
-static void do_acct_process(struct bsd_acct_struct *acct,
-               struct pid_namespace *ns, struct file *);
+static void do_acct_process(struct bsd_acct_struct *acct);
 
-/*
- * This structure is used so that all the data protected by lock
- * can be placed in the same cache line as the lock.  This primes
- * the cache line to have the data after getting the lock.
- */
 struct bsd_acct_struct {
+       long                    count;
+       struct mutex            lock;
        int                     active;
        unsigned long           needcheck;
        struct file             *file;
@@ -157,39 +153,59 @@ out:
        return res;
 }
 
-/*
- * Close the old accounting file (if currently open) and then replace
- * it with file (if non-NULL).
- *
- * NOTE: acct_lock MUST be held on entry and exit.
- */
-static void acct_file_reopen(struct bsd_acct_struct *acct, struct file *file,
-               struct pid_namespace *ns)
+static void acct_put(struct bsd_acct_struct *p)
 {
-       struct file *old_acct = NULL;
-       struct pid_namespace *old_ns = NULL;
-
-       if (acct->file) {
-               old_acct = acct->file;
-               old_ns = acct->ns;
-               acct->active = 0;
-               acct->file = NULL;
-               acct->ns = NULL;
-               list_del(&acct->list);
-       }
-       if (file) {
-               acct->file = file;
-               acct->ns = ns;
-               acct->needcheck = jiffies;
-               acct->active = 0;
-               list_add(&acct->list, &acct_list);
+       spin_lock(&acct_lock);
+       if (!--p->count)
+               kfree(p);
+       spin_unlock(&acct_lock);
+}
+
+static struct bsd_acct_struct *acct_get(struct bsd_acct_struct **p)
+{
+       struct bsd_acct_struct *res;
+       spin_lock(&acct_lock);
+again:
+       res = *p;
+       if (res)
+               res->count++;
+       spin_unlock(&acct_lock);
+       if (res) {
+               mutex_lock(&res->lock);
+               if (!res->ns) {
+                       mutex_unlock(&res->lock);
+                       spin_lock(&acct_lock);
+                       if (!--res->count)
+                               kfree(res);
+                       goto again;
+               }
        }
-       if (old_acct) {
-               mnt_unpin(old_acct->f_path.mnt);
+       return res;
+}
+
+static void acct_kill(struct bsd_acct_struct *acct,
+                     struct bsd_acct_struct *new)
+{
+       if (acct) {
+               struct file *file = acct->file;
+               struct pid_namespace *ns = acct->ns;
+               spin_lock(&acct_lock);
+               list_del(&acct->list);
+               mnt_unpin(file->f_path.mnt);
                spin_unlock(&acct_lock);
-               do_acct_process(acct, old_ns, old_acct);
-               filp_close(old_acct, NULL);
+               do_acct_process(acct);
+               filp_close(file, NULL);
                spin_lock(&acct_lock);
+               ns->bacct = new;
+               if (new) {
+                       mnt_pin(new->file->f_path.mnt);
+                       list_add(&new->list, &acct_list);
+               }
+               acct->ns = NULL;
+               mutex_unlock(&acct->lock);
+               if (!(acct->count -= 2))
+                       kfree(acct);
+               spin_unlock(&acct_lock);
        }
 }
 
@@ -197,47 +213,50 @@ static int acct_on(struct filename *pathname)
 {
        struct file *file;
        struct vfsmount *mnt;
-       struct pid_namespace *ns;
-       struct bsd_acct_struct *acct = NULL;
+       struct pid_namespace *ns = task_active_pid_ns(current);
+       struct bsd_acct_struct *acct, *old;
+
+       acct = kzalloc(sizeof(struct bsd_acct_struct), GFP_KERNEL);
+       if (!acct)
+               return -ENOMEM;
 
        /* Difference from BSD - they don't do O_APPEND */
        file = file_open_name(pathname, O_WRONLY|O_APPEND|O_LARGEFILE, 0);
-       if (IS_ERR(file))
+       if (IS_ERR(file)) {
+               kfree(acct);
                return PTR_ERR(file);
+       }
 
        if (!S_ISREG(file_inode(file)->i_mode)) {
+               kfree(acct);
                filp_close(file, NULL);
                return -EACCES;
        }
 
        if (!file->f_op->write) {
+               kfree(acct);
                filp_close(file, NULL);
                return -EIO;
        }
 
-       ns = task_active_pid_ns(current);
-       if (ns->bacct == NULL) {
-               acct = kzalloc(sizeof(struct bsd_acct_struct), GFP_KERNEL);
-               if (acct == NULL) {
-                       filp_close(file, NULL);
-                       return -ENOMEM;
-               }
-       }
+       acct->count = 1;
+       acct->file = file;
+       acct->needcheck = jiffies;
+       acct->ns = ns;
+       mutex_init(&acct->lock);
+       mnt = file->f_path.mnt;
 
-       spin_lock(&acct_lock);
-       if (ns->bacct == NULL) {
+       old = acct_get(&ns->bacct);
+       if (old) {
+               acct_kill(old, acct);
+       } else {
+               spin_lock(&acct_lock);
                ns->bacct = acct;
-               acct = NULL;
+               mnt_pin(mnt);
+               list_add(&acct->list, &acct_list);
+               spin_unlock(&acct_lock);
        }
-
-       mnt = file->f_path.mnt;
-       mnt_pin(mnt);
-       acct_file_reopen(ns->bacct, file, ns);
-       spin_unlock(&acct_lock);
-
        mntput(mnt); /* it's pinned, now give up active reference */
-       kfree(acct);
-
        return 0;
 }
 
@@ -270,15 +289,7 @@ SYSCALL_DEFINE1(acct, const char __user *, name)
                mutex_unlock(&acct_on_mutex);
                putname(tmp);
        } else {
-               struct bsd_acct_struct *acct;
-
-               acct = task_active_pid_ns(current)->bacct;
-               if (acct == NULL)
-                       return 0;
-
-               spin_lock(&acct_lock);
-               acct_file_reopen(acct, NULL, NULL);
-               spin_unlock(&acct_lock);
+               acct_kill(acct_get(&task_active_pid_ns(current)->bacct), NULL);
        }
 
        return error;
@@ -298,8 +309,19 @@ void acct_auto_close_mnt(struct vfsmount *m)
        spin_lock(&acct_lock);
 restart:
        list_for_each_entry(acct, &acct_list, list)
-               if (acct->file && acct->file->f_path.mnt == m) {
-                       acct_file_reopen(acct, NULL, NULL);
+               if (acct->file->f_path.mnt == m) {
+                       acct->count++;
+                       spin_unlock(&acct_lock);
+                       mutex_lock(&acct->lock);
+                       if (!acct->ns) {
+                               mutex_unlock(&acct->lock);
+                               spin_lock(&acct_lock);
+                               if (!--acct->count)
+                                       kfree(acct);
+                               goto restart;
+                       }
+                       acct_kill(acct, NULL);
+                       spin_lock(&acct_lock);
                        goto restart;
                }
        spin_unlock(&acct_lock);
@@ -319,8 +341,19 @@ void acct_auto_close(struct super_block *sb)
        spin_lock(&acct_lock);
 restart:
        list_for_each_entry(acct, &acct_list, list)
-               if (acct->file && acct->file->f_path.dentry->d_sb == sb) {
-                       acct_file_reopen(acct, NULL, NULL);
+               if (acct->file->f_path.dentry->d_sb == sb) {
+                       acct->count++;
+                       spin_unlock(&acct_lock);
+                       mutex_lock(&acct->lock);
+                       if (!acct->ns) {
+                               mutex_unlock(&acct->lock);
+                               spin_lock(&acct_lock);
+                               if (!--acct->count)
+                                       kfree(acct);
+                               goto restart;
+                       }
+                       acct_kill(acct, NULL);
+                       spin_lock(&acct_lock);
                        goto restart;
                }
        spin_unlock(&acct_lock);
@@ -328,17 +361,7 @@ restart:
 
 void acct_exit_ns(struct pid_namespace *ns)
 {
-       struct bsd_acct_struct *acct = ns->bacct;
-
-       if (acct == NULL)
-               return;
-
-       spin_lock(&acct_lock);
-       if (acct->file != NULL)
-               acct_file_reopen(acct, NULL, NULL);
-       spin_unlock(&acct_lock);
-
-       kfree(acct);
+       acct_kill(acct_get(&ns->bacct), NULL);
 }
 
 /*
@@ -507,12 +530,13 @@ static void fill_ac(acct_t *ac)
 /*
  *  do_acct_process does all actual work. Caller holds the reference to file.
  */
-static void do_acct_process(struct bsd_acct_struct *acct,
-               struct pid_namespace *ns, struct file *file)
+static void do_acct_process(struct bsd_acct_struct *acct)
 {
        acct_t ac;
        unsigned long flim;
        const struct cred *orig_cred;
+       struct pid_namespace *ns = acct->ns;
+       struct file *file = acct->file;
 
        /*
         * Accounting records are not subject to resource limits.
@@ -606,27 +630,12 @@ void acct_collect(long exitcode, int group_dead)
 static void slow_acct_process(struct pid_namespace *ns)
 {
        for ( ; ns; ns = ns->parent) {
-               struct file *file = NULL;
-               struct bsd_acct_struct *acct;
-
-               acct = ns->bacct;
-               /*
-                * accelerate the common fastpath:
-                */
-               if (!acct || !acct->file)
-                       continue;
-
-               spin_lock(&acct_lock);
-               file = acct->file;
-               if (unlikely(!file)) {
-                       spin_unlock(&acct_lock);
-                       continue;
+               struct bsd_acct_struct *acct = acct_get(&ns->bacct);
+               if (acct) {
+                       do_acct_process(acct);
+                       mutex_unlock(&acct->lock);
+                       acct_put(acct);
                }
-               get_file(file);
-               spin_unlock(&acct_lock);
-
-               do_acct_process(acct, ns, file);
-               fput(file);
        }
 }
 
@@ -645,8 +654,7 @@ void acct_process(void)
         * its parent.
         */
        for (ns = task_active_pid_ns(current); ns != NULL; ns = ns->parent) {
-               struct bsd_acct_struct *acct = ns->bacct;
-               if (acct && acct->file)
+               if (ns->bacct)
                        break;
        }
        if (unlikely(ns))