| Message ID | 5f9e8a7dcdf08bd2dd433f1a42690ab8e67e7915.1418618044.git.osandov@osandov.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Sun, Dec 14, 2014 at 09:26:57PM -0800, Omar Sandoval wrote: > The rw argument to direct_IO has some ill-defined semantics. Some > filesystems (e.g., ext4, FAT) decide whether they're doing a write with > rw == WRITE, but others (e.g., XFS) check rw & WRITE. Let's set a good > example in the swap file code and say ITER_BVEC belongs in > iov_iter->flags but not in rw. This caters to the least common > denominator and avoids a sweeping change of every direct_IO > implementation for now. Frankly, this is bogus. If anything, let's just kill the first argument completely - ->direct_IO() can always pick it from iter->type. As for catering to the least common denominator... To hell with the lowest common denominator. How many instances of ->direct_IO() do we have, anyway? 24 in the mainline (and we don't give a flying fuck for out-of-tree code, as a matter of policy). Moreover, several are of "do nothing" variety. FWIW, 'rw' is a mess. We used to have this: READ: O_DIRECT read WRITE: O_DIRECT write KERNEL_WRITE: swapout These days KERNEL_WRITE got replaced with ITER_BVEC | WRITE. The thing is, we have a bunch of places where we explicitly checked for being _equal_ to WRITE. I.e. the checks that gave a negative on swapouts. I suspect that most of them are wrong and should trigger on all writes, including swapouts, but I really didn't want to dig into that pile of fun back then. That's the main reason why 'rw' argument has survived at all... -- To unsubscribe from this list: send the line "unsubscribe linux-nfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Mon, Dec 15, 2014 at 06:16:02AM +0000, Al Viro wrote: > On Sun, Dec 14, 2014 at 09:26:57PM -0800, Omar Sandoval wrote: > > The rw argument to direct_IO has some ill-defined semantics. Some > > filesystems (e.g., ext4, FAT) decide whether they're doing a write with > > rw == WRITE, but others (e.g., XFS) check rw & WRITE. Let's set a good > > example in the swap file code and say ITER_BVEC belongs in > > iov_iter->flags but not in rw. This caters to the least common > > denominator and avoids a sweeping change of every direct_IO > > implementation for now. > > Frankly, this is bogus. If anything, let's just kill the first argument > completely - ->direct_IO() can always pick it from iter->type. > > As for catering to the least common denominator... To hell with the lowest > common denominator. How many instances of ->direct_IO() do we have, anyway? > 24 in the mainline (and we don't give a flying fuck for out-of-tree code, as > a matter of policy). Moreover, several are of "do nothing" variety. > > FWIW, 'rw' is a mess. We used to have this: > READ: O_DIRECT read > WRITE: O_DIRECT write > KERNEL_WRITE: swapout > > These days KERNEL_WRITE got replaced with ITER_BVEC | WRITE. The thing is, > we have a bunch of places where we explicitly checked for being _equal_ to > WRITE. I.e. the checks that gave a negative on swapouts. I suspect that most > of them are wrong and should trigger on all writes, including swapouts, but > I really didn't want to dig into that pile of fun back then. That's the > main reason why 'rw' argument has survived at all... > In that case, I'll take a stab at nuking rw. I'm almost certain that some of these are completely wrong (for example, of the form if (rw == WRITE) do_write(); else do_read();). This isn't an immediate problem for swap files on BTRFS, as __blockdev_direct_IO does a bitwise test, so I think I'll split it out into its own series. Thanks,
diff --git a/mm/page_io.c b/mm/page_io.c index 1630ac0..c229f88 100644 --- a/mm/page_io.c +++ b/mm/page_io.c @@ -285,8 +285,7 @@ int __swap_writepage(struct page *page, struct writeback_control *wbc, set_page_writeback(page); unlock_page(page); mutex_lock(&inode->i_mutex); - ret = mapping->a_ops->direct_IO(ITER_BVEC | WRITE, - &kiocb, &from, + ret = mapping->a_ops->direct_IO(WRITE, &kiocb, &from, kiocb.ki_pos); mutex_unlock(&inode->i_mutex); if (ret == PAGE_SIZE) {
The rw argument to direct_IO has some ill-defined semantics. Some filesystems (e.g., ext4, FAT) decide whether they're doing a write with rw == WRITE, but others (e.g., XFS) check rw & WRITE. Let's set a good example in the swap file code and say ITER_BVEC belongs in iov_iter->flags but not in rw. This caters to the least common denominator and avoids a sweeping change of every direct_IO implementation for now. Signed-off-by: Omar Sandoval <osandov@osandov.com> --- mm/page_io.c | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-)