On Wed, Dec 17, 2014 at 10:03:13PM +0000, Al Viro wrote:
> On Wed, Dec 17, 2014 at 10:52:56AM -0800, Christoph Hellwig wrote:
> > On Wed, Dec 17, 2014 at 06:58:32AM -0800, Omar Sandoval wrote:
> > > See my previous message. If we use O_DIRECT on the original open, then
> > > filesystems that implement bmap but not direct_IO will no longer work.
> > > These are the ones that I found in my tree:
> >
> > In the long run I don't think they are worth keeping. But to keep you
> > out of that discussion you can just try an open without O_DIRECT if the
> > open with the flag failed.
>
> Umm... That's one possibility, of course (and if swapon(2) is on someone's
> hotpath, I really would like to see what the hell they are doing - it has
> to be interesting in a sick way).
If this is the approach you'd prefer, I'll go ahead and do that for v2.
I personally think it looks pretty kludgey, but I'm fine either way:
diff --git a/mm/swapfile.c b/mm/swapfile.c
index 63f55cc..c1b3073 100644
--- a/mm/swapfile.c
+++ b/mm/swapfile.c
@@ -2379,7 +2379,16 @@ SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
name = NULL;
goto bad_swap;
}
- swap_file = file_open_name(name, O_RDWR|O_LARGEFILE, 0);
+ swap_file = file_open_name(name, O_RDWR | O_LARGEFILE | O_DIRECT, 0);
+ if (IS_ERR(swap_file) && PTR_ERR(swap_file) == -EINVAL)
+ swap_file = file_open_name(name, O_RDWR | O_LARGEFILE, 0);
if (IS_ERR(swap_file)) {
error = PTR_ERR(swap_file);
swap_file = NULL;
> BTW, speaking of read/write vs. swap - what's the story with e.g. AFS
> write() checking IS_SWAPFILE() and failing with -EBUSY? Note that
> * it's done before acquiring i_mutex, so it isn't race-free
> * it's dubious from the POSIX POV - EBUSY isn't in the error
> list for write(2).
> * other filesystems generally don't have anything of that sort.
> NFS does, but local ones do not...
> Besides, do we even allow swapfiles on AFS?
AFS doesn't implement ->bmap or ->swap_activate, so that code is dead,
probably cargo-culted from the NFS code. It seems pretty pointless, not
only because it's inconsistent with the local filesystems like you
mentioned, but also because it's trivial to bypass with O_DIRECT on NFS:
ssize_t nfs_file_write(struct kiocb *iocb, struct iov_iter *from)
{
struct file *file = iocb->ki_filp;
struct inode *inode = file_inode(file);
unsigned long written = 0;
ssize_t result;
size_t count = iov_iter_count(from);
loff_t pos = iocb->ki_pos;
result = nfs_key_timeout_notify(file, inode);
if (result)
return result;
if (file->f_flags & O_DIRECT)
return nfs_file_direct_write(iocb, from, pos);
dprintk("NFS: write(%pD2, %zu@%Ld)\n",
file, count, (long long) pos);
result = -EBUSY;
if (IS_SWAPFILE(inode))
goto out_swapfile;
I think it's safe to scrap that code. However, this also led me to find that
NFS doesn't prevent truncates on an active swapfile. I'm submitting a patch for
that now.
--
Omar
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>