This patch fixes a race condition in lseek. While it is expected that
unpredictable behaviour may result while repositioning the offset of a
file descriptor concurrently with reading/writing to the same file
descriptor, this should not happen when merely *reading* the file
descriptor's offset.
Unfortunately, the only portable way in Unix to read a file
descriptor's offset is lseek(fd, 0, SEEK_CUR); however executing this
concurrently with read/write may mess up the position, as shown by the
testcase below:
#include <sys/types.h>
#include <stdio.h>
#include <pthread.h>
#include <unistd.h>
#include <errno.h>
#include <string.h>
#include <pthread.h>
void *loop(void *ptr)
{
fprintf(stderr, "Starting seek thread\n");
while(1) {
if(lseek(0, 0LL, SEEK_CUR) < 0LL)
perror("seek");
}
}
int main(int argc, char **argv) {
long long l=0;
int r;
char buf[4096];
pthread_t thread;
pthread_create(&thread, 0, loop, 0);
for(r=0; 1 ; r++) {
int n = read(0, buf, 4096);
if(n == 0)
break;
if(n < 4096) {
fprintf(stderr, "Short read %d %s\n", n, strerror(errno));
}
l+= n;
}
fprintf(stderr, "Read %lld bytes\n", l);
return 0;
}
Compile this and run it on a multi-processor machine as
./a.out <bigFile
where bigFile is a 1 Gigabyte file. It should print 1073741824.
However, on a buggy kernel, it usually produces a bigger number. The
problem only happens on a multiprocessor machine. This is because an
lseek(fd, 0, SEEK_CUR) running concurrently with a read() or write()
will reset the position back to what it used to be when the read()
started.
This behavior was observed "in the wild" when using udpcast which uses
lseek to monitor progress of reading/writing the uncompressed data.
The patch below fixes the issue by "special-casing" the lseek(fd, 0,
SEEK_CUR) pattern.
Apparently, an attempt was already made to fix the issue by the
following code:
if (offset != file->f_pos) {
file->f_pos = offset;
file->f_version = 0;
}
However, this doesn't work if file->f_pos was changed (by read() or
write()) between the time offset was computed, and the time where it
considers writing it back.
Signed-off-by: Alain Knaff <[email protected]>
---
diff -pur kernel.orig/fs/read_write.c kernel/fs/read_write.c
--- kernel.orig/fs/read_write.c 2008-10-11 14:12:07.000000000 +0200
+++ kernel/fs/read_write.c 2008-11-06 19:55:59.000000000 +0100
@@ -42,6 +42,8 @@ generic_file_llseek_unlocked(struct file
offset += inode->i_size;
break;
case SEEK_CUR:
+ if(offset == 0)
+ return file->f_pos;
offset += file->f_pos;
}
retval = -EINVAL;
On Thu, 6 Nov 2008 21:21:12 +0100
Alain Knaff <[email protected]> wrote:
> This patch fixes a race condition in lseek. While it is expected that
> unpredictable behaviour may result while repositioning the offset of a
> file descriptor concurrently with reading/writing to the same file
> descriptor, this should not happen when merely *reading* the file
> descriptor's offset.
>
> Unfortunately, the only portable way in Unix to read a file
> descriptor's offset is lseek(fd, 0, SEEK_CUR); however executing this
> concurrently with read/write may mess up the position, as shown by the
> testcase below:
>
> #include <sys/types.h>
> #include <stdio.h>
> #include <pthread.h>
> #include <unistd.h>
> #include <errno.h>
> #include <string.h>
> #include <pthread.h>
>
>
> void *loop(void *ptr)
> {
> fprintf(stderr, "Starting seek thread\n");
> while(1) {
> if(lseek(0, 0LL, SEEK_CUR) < 0LL)
> perror("seek");
> }
> }
>
> int main(int argc, char **argv) {
> long long l=0;
> int r;
> char buf[4096];
>
> pthread_t thread;
> pthread_create(&thread, 0, loop, 0);
>
> for(r=0; 1 ; r++) {
> int n = read(0, buf, 4096);
> if(n == 0)
> break;
> if(n < 4096) {
> fprintf(stderr, "Short read %d %s\n", n, strerror(errno));
> }
> l+= n;
> }
> fprintf(stderr, "Read %lld bytes\n", l);
>
> return 0;
> }
>
> Compile this and run it on a multi-processor machine as
> ./a.out <bigFile
>
> where bigFile is a 1 Gigabyte file. It should print 1073741824.
> However, on a buggy kernel, it usually produces a bigger number. The
> problem only happens on a multiprocessor machine. This is because an
> lseek(fd, 0, SEEK_CUR) running concurrently with a read() or write()
> will reset the position back to what it used to be when the read()
> started.
>
> This behavior was observed "in the wild" when using udpcast which uses
> lseek to monitor progress of reading/writing the uncompressed data.
>
> The patch below fixes the issue by "special-casing" the lseek(fd, 0,
> SEEK_CUR) pattern.
>
> Apparently, an attempt was already made to fix the issue by the
> following code:
>
> if (offset != file->f_pos) {
> file->f_pos = offset;
> file->f_version = 0;
> }
>
> However, this doesn't work if file->f_pos was changed (by read() or
> write()) between the time offset was computed, and the time where it
> considers writing it back.
>
> Signed-off-by: Alain Knaff <[email protected]>
>
> ---
>
> diff -pur kernel.orig/fs/read_write.c kernel/fs/read_write.c
> --- kernel.orig/fs/read_write.c 2008-10-11 14:12:07.000000000 +0200
> +++ kernel/fs/read_write.c 2008-11-06 19:55:59.000000000 +0100
> @@ -42,6 +42,8 @@ generic_file_llseek_unlocked(struct file
> offset += inode->i_size;
> break;
> case SEEK_CUR:
> + if(offset == 0)
> + return file->f_pos;
> offset += file->f_pos;
> }
> retval = -EINVAL;
OK, I think that a concurrent lseek(fd, 0, SEEK_CUR) is a sufficiently
sane operation that this is worth doing. As you point out, there is no
other way of userspace doing what is effectively a read-only operation
- userspace would be entitled to wonder "ytf did the kernel rewrite the
file offset for that?".
Do the below additions look OK?
From: Andrew Morton <[email protected]>
- fix coding-style
- fix default_llseek() as well
- add comments
Cc: Al Viro <[email protected]>
Cc: Alain Knaff <[email protected]>
Cc: Christoph Hellwig <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
---
fs/read_write.c | 19 +++++++++++++++++--
1 file changed, 17 insertions(+), 2 deletions(-)
diff -puN fs/read_write.c~vfs-lseekfd-0-seek_cur-race-condition-fix fs/read_write.c
--- a/fs/read_write.c~vfs-lseekfd-0-seek_cur-race-condition-fix
+++ a/fs/read_write.c
@@ -50,6 +50,14 @@ generic_file_llseek_unlocked(struct file
offset += inode->i_size;
break;
case SEEK_CUR:
+ /*
+ * Here we special-case the lseek(fd, 0, SEEK_CUR)
+ * position-querying operation. Avoid rewriting the "same"
+ * f_pos value back to the file because a concurrent read(),
+ * write() or lseek() might have altered it
+ */
+ if (offset == 0)
+ return file->f_pos;
offset += file->f_pos;
break;
}
@@ -105,8 +113,14 @@ loff_t default_llseek(struct file *file,
offset += i_size_read(file->f_path.dentry->d_inode);
break;
case SEEK_CUR:
- if(offset == 0)
- return file->f_pos;
+ /*
+ * See SEEK_CUR description in
+ * generic_file_llseek_unlocked()
+ */
+ if (offset == 0) {
+ retval = file->f_pos;
+ goto out;
+ }
offset += file->f_pos;
}
retval = -EINVAL;
@@ -117,6 +131,7 @@ loff_t default_llseek(struct file *file,
}
retval = offset;
}
+out:
unlock_kernel();
return retval;
}
_
Andrew Morton wrote:
[...]
> OK, I think that a concurrent lseek(fd, 0, SEEK_CUR) is a sufficiently
> sane operation that this is worth doing. As you point out, there is no
> other way of userspace doing what is effectively a read-only operation
> - userspace would be entitled to wonder "ytf did the kernel rewrite the
> file offset for that?".
>
>
> Do the below additions look OK?
Yes, these additions look ok to me. Thanks for your quick and easy reply.
> From: Andrew Morton <[email protected]>
>
> - fix coding-style
> - fix default_llseek() as well
Wow, that one was missing an unlock_kernel()... I'm wondering why this
hadn't let to some major SNAFU...
> - add comments
Thanks,
Alain