From: Zheng Liu <[email protected]>
This commit adds a new syscall called closefrom(2) which is almost the
same as closefrom(2) in FreeBSD. This syscall will close all open file
descriptors that is greater than or equal to the fd that is indicated
by caller. It is really useful for an application that has opened a
hundreds of thousands of files because it doesn't need to call close(2)
several times.
This version is just an proof-of-concept implementation. Now there are
two issues that need to be discussed. The first one is that the syscall
ignores all errors. It just returns '0'. Does it need to return some
errors? Another issue is that 'lowfd' argument is 'int' type but
fdtable->max_fds is 'unsigned int'. So that means that it could close
all fds if 'lowfd' is greater than 'MAX_INT'. Now it uses 'int' type
in order to keep consistent with closefrom(2) in FreeBSD. In FreeBSD
system will close all fds if 'lowfd' is less than 0.
The test program opens a lot of files and closes them, and then we
calculates the time (*unit: us*) of closing these files. Every cases
the program is run 3 times to compute the average number.
$ lscpu
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Byte Order: Little Endian
CPU(s): 16
On-line CPU(s) list: 0-15
Thread(s) per core: 2
Core(s) per socket: 4
CPU socket(s): 2
NUMA node(s): 2
Vendor ID: GenuineIntel
CPU family: 6
Model: 44
Stepping: 2
CPU MHz: 2400.000
BogoMIPS: 4799.88
Virtualization: VT-x
L1d cache: 32K
L1i cache: 32K
L2 cache: 256K
L3 cache: 12288K
NUMA node0 CPU(s): 0-3,8-11
NUMA node1 CPU(s): 4-7,12-15
*Unit: us*
1,000,000 files
===============
x vanilla
+ closefrom
+------------------------------------------------------------------------+
| ++ + xxx|
||_MA___| |A||
+------------------------------------------------------------------------+
N Min Max Median Avg Stddev
x 3 799552 810468 807615 805878.33 5661.4285
+ 3 474013 504391 476899 485101 16767.836
Difference at 95.0% confidence
-320777 +/- 28364.7
-39.8047% +/- 3.51972%
(Student's t, pooled s = 12514.2)
Improvement: 39.8047%
100,000 files
============
x vanilla
+ closefrom
+------------------------------------------------------------------------+
| + x |
|++ xx|
||A| A||
+------------------------------------------------------------------------+
N Min Max Median Avg Stddev
x 3 77815 78340 77975 78043.333 269.08797
+ 3 46586 47220 47008 46938 322.74448
Difference at 95.0% confidence
-31105.3 +/- 673.473
-39.8565% +/- 0.862947%
(Student's t, pooled s = 297.13)
Improvement: 39.8565%
10,000 files
============
x vanilla
+ closefrom
+------------------------------------------------------------------------+
| + |
|++ xxx|
||A |A||
+------------------------------------------------------------------------+
N Min Max Median Avg Stddev
x 3 6892 6979 6941 6937.3333 43.615746
+ 3 4347 4391 4367 4368.3333 22.030282
Difference at 95.0% confidence
-2569 +/- 78.3151
-37.0315% +/- 1.12889%
(Student's t, pooled s = 34.5519)
Improvement: 37.0315%
1,000 files
===========
x vanilla
+ closefrom
+------------------------------------------------------------------------+
| + x |
|+ + xx|
||_AM| A||
+------------------------------------------------------------------------+
N Min Max Median Avg Stddev
x 3 543 547 545 545 2
+ 3 303 314 312 309.66667 5.8594653
Difference at 95.0% confidence
-235.333 +/- 9.92309
-43.1804% +/- 1.82075%
(Student's t, pooled s = 4.37798)
Improvement: 43.1804%
100 files
=========
x vanilla
+ closefrom
+------------------------------------------------------------------------+
|+ |
|+ x |
|+ x x|
|A MA_||
+------------------------------------------------------------------------+
N Min Max Median Avg Stddev
x 3 57 58 57 57.333333 0.57735027
+ 3 32 32 32 32 0
Difference at 95.0% confidence
-25.3333 +/- 0.925333
-44.186% +/- 1.61395%
(Student's t, pooled s = 0.408248)
Improvement: 44.186%
10 files
========
x vanilla
+ closefrom
+------------------------------------------------------------------------+
| + x|
| + + x|
||_______M_________A__________________| A|
+------------------------------------------------------------------------+
N Min Max Median Avg Stddev
x 3 7 7 7 7 0
+ 3 5 6 5 5.3333333 0.57735027
Difference at 95.0% confidence
-1.66667 +/- 0.925333
-23.8095% +/- 13.219%
(Student's t, pooled s = 0.408248)
Improvement: 23.8095%
1 file
======
x vanilla
+ closefrom
+------------------------------------------------------------------------+
| * |
| * *|
||_____________M__________________A________________________________| |
+------------------------------------------------------------------------+
N Min Max Median Avg Stddev
x 3 2 3 2 2.3333333 0.57735027
+ 3 2 3 2 2.3333333 0.57735027
No difference proven at 95.0% confidence
Any comment or feedback are always welcome.
Cc: Alexander Viro <[email protected]>
Signed-off-by: Zheng Liu <[email protected]>
---
arch/x86/syscalls/syscall_64.tbl | 1 +
fs/file.c | 30 ++++++++++++++++++++++++++++++
fs/open.c | 6 ++++++
include/linux/fdtable.h | 1 +
4 files changed, 38 insertions(+)
diff --git a/arch/x86/syscalls/syscall_64.tbl b/arch/x86/syscalls/syscall_64.tbl
index 04376ac3d9ef..33c71e4defc2 100644
--- a/arch/x86/syscalls/syscall_64.tbl
+++ b/arch/x86/syscalls/syscall_64.tbl
@@ -323,6 +323,7 @@
314 common sched_setattr sys_sched_setattr
315 common sched_getattr sys_sched_getattr
316 common renameat2 sys_renameat2
+317 common closefrom sys_closefrom
#
# x32-specific system call numbers start at 512 to avoid cache impact
diff --git a/fs/file.c b/fs/file.c
index b61293badfb1..15502ac4304c 100644
--- a/fs/file.c
+++ b/fs/file.c
@@ -601,6 +601,36 @@ out_unlock:
return -EBADF;
}
+int __close_fds(struct files_struct *files, int lowfd)
+{
+ struct file *file;
+ struct fdtable *fdt;
+ int fd;
+
+ if (lowfd < 0)
+ lowfd = 0;
+ spin_lock(&files->file_lock);
+ fdt = files_fdtable(files);
+ if (lowfd >= fdt->max_fds)
+ goto out_unlock;
+ for (fd = lowfd; fd < fdt->max_fds; fd++) {
+ file = fdt->fd[fd];
+ if (!file)
+ continue;
+
+ rcu_assign_pointer(fdt->fd[fd], NULL);
+ __clear_close_on_exec(fd, fdt);
+ __put_unused_fd(files, fd);
+ spin_unlock(&files->file_lock);
+ filp_close(file, files);
+ spin_lock(&files->file_lock);
+ }
+
+out_unlock:
+ spin_unlock(&files->file_lock);
+ return 0;
+}
+
void do_close_on_exec(struct files_struct *files)
{
unsigned i;
diff --git a/fs/open.c b/fs/open.c
index 631aea815def..afe3b631074a 100644
--- a/fs/open.c
+++ b/fs/open.c
@@ -1099,6 +1099,12 @@ SYSCALL_DEFINE1(close, unsigned int, fd)
}
EXPORT_SYMBOL(sys_close);
+SYSCALL_DEFINE1(closefrom, int, lowfd)
+{
+ return __close_fds(current->files, lowfd);
+}
+EXPORT_SYMBOL(sys_closefrom);
+
/*
* This routine simulates a hangup on the tty, to arrange that users
* are given clean terminals at login time.
diff --git a/include/linux/fdtable.h b/include/linux/fdtable.h
index 70e8e21c0a30..b0f9fc736481 100644
--- a/include/linux/fdtable.h
+++ b/include/linux/fdtable.h
@@ -114,6 +114,7 @@ extern void __fd_install(struct files_struct *files,
unsigned int fd, struct file *file);
extern int __close_fd(struct files_struct *files,
unsigned int fd);
+extern int __close_fds(struct files_struct *files, int lowfd);
extern struct kmem_cache *files_cachep;
--
1.7.9.7
On Tue, Apr 08, 2014 at 03:12:22PM +0800, Zheng Liu wrote:
>
> +int __close_fds(struct files_struct *files, int lowfd)
> +{
> + struct file *file;
> + struct fdtable *fdt;
> + int fd;
> +
> + if (lowfd < 0)
> + lowfd = 0;
> + spin_lock(&files->file_lock);
> + fdt = files_fdtable(files);
> + if (lowfd >= fdt->max_fds)
> + goto out_unlock;
> + for (fd = lowfd; fd < fdt->max_fds; fd++) {
> + file = fdt->fd[fd];
> + if (!file)
> + continue;
> +
> + rcu_assign_pointer(fdt->fd[fd], NULL);
> + __clear_close_on_exec(fd, fdt);
> + __put_unused_fd(files, fd);
> + spin_unlock(&files->file_lock);
> + filp_close(file, files);
> + spin_lock(&files->file_lock);
> + }
> +
> +out_unlock:
> + spin_unlock(&files->file_lock);
> + return 0;
> +}
> +
Can't comment on the usefulness of the patch, but I would like to note:
1. fdt could be freed after you drop the lock, but you never reload the
pointer, thus this looks like use-after-free
2. most of this looks like __close_fd, maybe some parts could be moved
to an inline function so that code duplication is reduced?
--
Mateusz Guzik
On Tue, Apr 08, 2014 at 10:21:37AM +0200, Mateusz Guzik wrote:
> On Tue, Apr 08, 2014 at 03:12:22PM +0800, Zheng Liu wrote:
> >
> > +int __close_fds(struct files_struct *files, int lowfd)
> > +{
> > + struct file *file;
> > + struct fdtable *fdt;
> > + int fd;
> > +
> > + if (lowfd < 0)
> > + lowfd = 0;
> > + spin_lock(&files->file_lock);
> > + fdt = files_fdtable(files);
> > + if (lowfd >= fdt->max_fds)
> > + goto out_unlock;
> > + for (fd = lowfd; fd < fdt->max_fds; fd++) {
> > + file = fdt->fd[fd];
> > + if (!file)
> > + continue;
> > +
> > + rcu_assign_pointer(fdt->fd[fd], NULL);
> > + __clear_close_on_exec(fd, fdt);
> > + __put_unused_fd(files, fd);
> > + spin_unlock(&files->file_lock);
> > + filp_close(file, files);
> > + spin_lock(&files->file_lock);
> > + }
> > +
> > +out_unlock:
> > + spin_unlock(&files->file_lock);
> > + return 0;
> > +}
> > +
>
> Can't comment on the usefulness of the patch, but I would like to note:
>
> 1. fdt could be freed after you drop the lock, but you never reload the
> pointer, thus this looks like use-after-free
> 2. most of this looks like __close_fd, maybe some parts could be moved
> to an inline function so that code duplication is reduced?
Ah, yes, my fault. I will fix them in next version. Thanks for
pointing it out.
Regards,
- Zheng