Al Viro already made most vfs/write_iters fcntl safe, but pipe and splice
are still affected. Patchset against vfs.git#for-next f1d36c2d711166aaa84f37
TOC:
0001-pipe-fix-race-with-fcntl
0002-splice-fix-race-beween-splice_write-vs-fcntl
Fix other long standing issues caused by fcntl(,F_SETFL,):
- User can disable O_DIRECT for pipe[1] (paketized IO), but can not enable it again.
- Currently we do not set O_APPEND on pipe[1] (IMHO it is wrong, but let it be)
so it is reasonable to completely prohibit change O_APPEND flag on both
end's of pipe. Add ->check_flags method in order to diallow O_APPEND toggling.
Signed-off-by: Dmitry Monakhov <[email protected]>
---
fs/fcntl.c | 6 ++++--
fs/pipe.c | 16 +++++++++++++---
2 files changed, 17 insertions(+), 5 deletions(-)
diff --git a/fs/fcntl.c b/fs/fcntl.c
index ee85cd4..0bdc9c7 100644
--- a/fs/fcntl.c
+++ b/fs/fcntl.c
@@ -51,9 +51,11 @@ static int setfl(int fd, struct file * filp, unsigned long arg)
if (arg & O_NDELAY)
arg |= O_NONBLOCK;
+ /* allowed only for inodes with ->direct_io method or write pipe */
if (arg & O_DIRECT) {
- if (!filp->f_mapping || !filp->f_mapping->a_ops ||
- !filp->f_mapping->a_ops->direct_IO)
+ if ((!filp->f_mapping || !filp->f_mapping->a_ops ||
+ !filp->f_mapping->a_ops->direct_IO) &&
+ !(get_pipe_info(filp) && (filp->f_flags | O_WRONLY)))
return -EINVAL;
}
diff --git a/fs/pipe.c b/fs/pipe.c
index 8865f79..0c15647 100644
--- a/fs/pipe.c
+++ b/fs/pipe.c
@@ -329,9 +329,9 @@ pipe_read(struct kiocb *iocb, struct iov_iter *to)
return ret;
}
-static inline int is_packetized(struct file *file)
+static inline int is_packetized(struct kiocb *iocb)
{
- return (file->f_flags & O_DIRECT) != 0;
+ return (iocb->ki_flags & IOCB_DIRECT) != 0;
}
static ssize_t
@@ -427,7 +427,7 @@ pipe_write(struct kiocb *iocb, struct iov_iter *from)
buf->offset = 0;
buf->len = copied;
buf->flags = 0;
- if (is_packetized(filp)) {
+ if (is_packetized(iocb)) {
buf->ops = &packet_pipe_buf_ops;
buf->flags = PIPE_BUF_FLAG_PACKET;
}
@@ -943,6 +943,15 @@ err:
return ret;
}
+/* XXX: Currently it is not possible distinguish read side from write one */
+static int pipe_check_flags(int flags)
+{
+ if (flags & O_APPEND)
+ return -EINVAL;
+
+ return 0;
+}
+
const struct file_operations pipefifo_fops = {
.open = fifo_open,
.llseek = no_llseek,
@@ -952,6 +961,7 @@ const struct file_operations pipefifo_fops = {
.unlocked_ioctl = pipe_ioctl,
.release = pipe_release,
.fasync = pipe_fasync,
+ .check_flags = pipe_check_flags,
};
/*
--
1.7.1
file->f_flags & O_APPEND is checked twice
-> do_splice_direct or do_splice: return EINVAL if O_APPEND enabled
-> generic_write_checks: seek to end in case of O_APPEND
This is obviously whong and result in unpredictable behaviour if raced with
fcntl. It is reasonable to recheck append flag after kiocb was constructed
(where ->ki_flags is stable), for that reason we should use special
analog of vfs_write_iter() which asserts non-append behaviour.
Signed-off-by: Dmitry Monakhov <[email protected]>
---
fs/splice.c | 23 ++++++++++++++++++++++-
1 files changed, 22 insertions(+), 1 deletions(-)
diff --git a/fs/splice.c b/fs/splice.c
index 41cbb16..d49615d 100644
--- a/fs/splice.c
+++ b/fs/splice.c
@@ -922,6 +922,27 @@ ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
return ret;
}
+ssize_t splice_iter_write(struct file *file, struct iov_iter *iter, loff_t *ppos)
+{
+ struct kiocb kiocb;
+ ssize_t ret;
+
+ if (!file->f_op->write_iter)
+ return -EINVAL;
+
+ init_sync_kiocb(&kiocb, file);
+ if (kiocb.ki_flags & IOCB_APPEND)
+ return -EINVAL;
+
+ kiocb.ki_pos = *ppos;
+ iter->type |= WRITE;
+ ret = file->f_op->write_iter(&kiocb, iter);
+ BUG_ON(ret == -EIOCBQUEUED);
+ if (ret > 0)
+ *ppos = kiocb.ki_pos;
+ return ret;
+}
+
/**
* iter_file_splice_write - splice data from a pipe to a file
@@ -1005,7 +1026,7 @@ iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
iov_iter_bvec(&from, ITER_BVEC | WRITE, array, n,
sd.total_len - left);
- ret = vfs_iter_write(out, &from, &sd.pos);
+ ret = splice_iter_write(out, &from, &sd.pos);
if (ret <= 0)
break;
--
1.7.1
On Sun, Apr 12, 2015 at 09:08:21PM +0400, Dmitry Monakhov wrote:
> Fix other long standing issues caused by fcntl(,F_SETFL,):
> - User can disable O_DIRECT for pipe[1] (paketized IO), but can not enable it again.
> - Currently we do not set O_APPEND on pipe[1] (IMHO it is wrong, but let it be)
> so it is reasonable to completely prohibit change O_APPEND flag on both
> end's of pipe. Add ->check_flags method in order to diallow O_APPEND toggling.
>
> Signed-off-by: Dmitry Monakhov <[email protected]>
> ---
TBH, all those ->direct_IO != NULL checks seem to be a wrong approach.
If nothing else, it forces several filesystem into inventing a fake
->direct_IO just to fool those tests. How about we
* introduce FMODE_MAY_DIRECT and allow ->open() explicitly set it
* make open_check_o_direct() and fcntl.c check that instead of poking
in ->f_mapping->a_ops, etc.
* provide a variant of generic_file_open() that would set that
bit and use it on the filesystems that handle dio
Al Viro <[email protected]> writes:
> On Sun, Apr 12, 2015 at 09:08:21PM +0400, Dmitry Monakhov wrote:
>> Fix other long standing issues caused by fcntl(,F_SETFL,):
>> - User can disable O_DIRECT for pipe[1] (paketized IO), but can not enable it again.
>> - Currently we do not set O_APPEND on pipe[1] (IMHO it is wrong, but let it be)
>> so it is reasonable to completely prohibit change O_APPEND flag on both
>> end's of pipe. Add ->check_flags method in order to diallow O_APPEND toggling.
>>
>> Signed-off-by: Dmitry Monakhov <[email protected]>
>> ---
>
> TBH, all those ->direct_IO != NULL checks seem to be a wrong approach.
> If nothing else, it forces several filesystem into inventing a fake
> ->direct_IO just to fool those tests. How about we
> * introduce FMODE_MAY_DIRECT and allow ->open() explicitly set it
> * make open_check_o_direct() and fcntl.c check that instead of poking
> in ->f_mapping->a_ops, etc.
> * provide a variant of generic_file_open() that would set that
> bit and use it on the filesystems that handle dio
100% agree. FMODE is perfect place for that.
BTW: I always wondering: why we do not mark pipe[1]->f_flags with O_APPEND?
Probably the answer is that nobody care about ->f_flags since no_llseek
returns -ESPIPE, but f_flags are visiable via fcntl so IMHO it is
reasonable to fix that too.
> --
> To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
> the body of a message to [email protected]
> More majordomo info at http://vger.kernel.org/majordomo-info.html