]> git.kernelconcepts.de Git - karo-tx-linux.git/commitdiff
fs/file.c: don't acquire files->file_lock in fd_install()
authorEric Dumazet <edumazet@google.com>
Tue, 30 Jun 2015 13:54:08 +0000 (15:54 +0200)
committerAl Viro <viro@zeniv.linux.org.uk>
Wed, 1 Jul 2015 06:30:09 +0000 (02:30 -0400)
Mateusz Guzik reported :

 Currently obtaining a new file descriptor results in locking fdtable
 twice - once in order to reserve a slot and second time to fill it.

Holding the spinlock in __fd_install() is needed in case a resize is
done, or to prevent a resize.

Mateusz provided an RFC patch and a micro benchmark :
  http://people.redhat.com/~mguzik/pipebench.c

A resize is an unlikely operation in a process lifetime,
as table size is at least doubled at every resize.

We can use RCU instead of the spinlock.

__fd_install() must wait if a resize is in progress.

The resize must block new __fd_install() callers from starting,
and wait that ongoing install are finished (synchronize_sched())

resize should be attempted by a single thread to not waste resources.

rcu_sched variant is used, as __fd_install() and expand_fdtable() run
from process context.

It gives us a ~30% speedup using pipebench on a dual Intel(R) Xeon(R)
CPU E5-2696 v2 @ 2.50GHz

Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Mateusz Guzik <mguzik@redhat.com>
Acked-by: Mateusz Guzik <mguzik@redhat.com>
Tested-by: Mateusz Guzik <mguzik@redhat.com>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Documentation/filesystems/porting
fs/file.c
include/linux/fdtable.h

index 3eae250254d581d253dd035109d568f8ef0c7873..ec5456113072ab93240033a4737953bd753afd76 100644 (file)
@@ -500,3 +500,7 @@ in your dentry operations instead.
        dentry,  it does not get nameidata at all and it gets called only when cookie
        is non-NULL.  Note that link body isn't available anymore, so if you need it,
        store it as cookie.
+--
+[mandatory]
+       __fd_install() & fd_install() can now sleep. Callers should not
+       hold a spinlock or other resources that do not allow a schedule.
index 93c5f89c248b07b6fba50fd31f4bd36ba9520fef..3d2eb4c542a491e686d9455e190d3e8760cedaee 100644 (file)
--- a/fs/file.c
+++ b/fs/file.c
@@ -147,6 +147,13 @@ static int expand_fdtable(struct files_struct *files, int nr)
 
        spin_unlock(&files->file_lock);
        new_fdt = alloc_fdtable(nr);
+
+       /* make sure all __fd_install() have seen resize_in_progress
+        * or have finished their rcu_read_lock_sched() section.
+        */
+       if (atomic_read(&files->count) > 1)
+               synchronize_sched();
+
        spin_lock(&files->file_lock);
        if (!new_fdt)
                return -ENOMEM;
@@ -158,21 +165,14 @@ static int expand_fdtable(struct files_struct *files, int nr)
                __free_fdtable(new_fdt);
                return -EMFILE;
        }
-       /*
-        * Check again since another task may have expanded the fd table while
-        * we dropped the lock
-        */
        cur_fdt = files_fdtable(files);
-       if (nr >= cur_fdt->max_fds) {
-               /* Continue as planned */
-               copy_fdtable(new_fdt, cur_fdt);
-               rcu_assign_pointer(files->fdt, new_fdt);
-               if (cur_fdt != &files->fdtab)
-                       call_rcu(&cur_fdt->rcu, free_fdtable_rcu);
-       } else {
-               /* Somebody else expanded, so undo our attempt */
-               __free_fdtable(new_fdt);
-       }
+       BUG_ON(nr < cur_fdt->max_fds);
+       copy_fdtable(new_fdt, cur_fdt);
+       rcu_assign_pointer(files->fdt, new_fdt);
+       if (cur_fdt != &files->fdtab)
+               call_rcu(&cur_fdt->rcu, free_fdtable_rcu);
+       /* coupled with smp_rmb() in __fd_install() */
+       smp_wmb();
        return 1;
 }
 
@@ -185,21 +185,38 @@ static int expand_fdtable(struct files_struct *files, int nr)
  * The files->file_lock should be held on entry, and will be held on exit.
  */
 static int expand_files(struct files_struct *files, int nr)
+       __releases(files->file_lock)
+       __acquires(files->file_lock)
 {
        struct fdtable *fdt;
+       int expanded = 0;
 
+repeat:
        fdt = files_fdtable(files);
 
        /* Do we need to expand? */
        if (nr < fdt->max_fds)
-               return 0;
+               return expanded;
 
        /* Can we expand? */
        if (nr >= sysctl_nr_open)
                return -EMFILE;
 
+       if (unlikely(files->resize_in_progress)) {
+               spin_unlock(&files->file_lock);
+               expanded = 1;
+               wait_event(files->resize_wait, !files->resize_in_progress);
+               spin_lock(&files->file_lock);
+               goto repeat;
+       }
+
        /* All good, so we try */
-       return expand_fdtable(files, nr);
+       files->resize_in_progress = true;
+       expanded = expand_fdtable(files, nr);
+       files->resize_in_progress = false;
+
+       wake_up_all(&files->resize_wait);
+       return expanded;
 }
 
 static inline void __set_close_on_exec(int fd, struct fdtable *fdt)
@@ -256,6 +273,8 @@ struct files_struct *dup_fd(struct files_struct *oldf, int *errorp)
        atomic_set(&newf->count, 1);
 
        spin_lock_init(&newf->file_lock);
+       newf->resize_in_progress = false;
+       init_waitqueue_head(&newf->resize_wait);
        newf->next_fd = 0;
        new_fdt = &newf->fdtab;
        new_fdt->max_fds = NR_OPEN_DEFAULT;
@@ -553,11 +572,21 @@ void __fd_install(struct files_struct *files, unsigned int fd,
                struct file *file)
 {
        struct fdtable *fdt;
-       spin_lock(&files->file_lock);
-       fdt = files_fdtable(files);
+
+       might_sleep();
+       rcu_read_lock_sched();
+
+       while (unlikely(files->resize_in_progress)) {
+               rcu_read_unlock_sched();
+               wait_event(files->resize_wait, !files->resize_in_progress);
+               rcu_read_lock_sched();
+       }
+       /* coupled with smp_wmb() in expand_fdtable() */
+       smp_rmb();
+       fdt = rcu_dereference_sched(files->fdt);
        BUG_ON(fdt->fd[fd] != NULL);
        rcu_assign_pointer(fdt->fd[fd], file);
-       spin_unlock(&files->file_lock);
+       rcu_read_unlock_sched();
 }
 
 void fd_install(unsigned int fd, struct file *file)
index 230f87bdf5ad02008ff622e65bc761e41d4b22e0..fbb88740634af011af6cc1be098f0a6d20aba07f 100644 (file)
@@ -47,6 +47,9 @@ struct files_struct {
    * read mostly part
    */
        atomic_t count;
+       bool resize_in_progress;
+       wait_queue_head_t resize_wait;
+
        struct fdtable __rcu *fdt;
        struct fdtable fdtab;
   /*