[PATCH v2 2/7] block: take i_rwsem for the direct I/O write fallback

[Tal Zussman]

Commit c0e473a0d226 ("block: fix race between set_blocksize and read
paths") closed a race between set_blocksize() and block device I/O: with
large sector size support, set_blocksize() can change i_blkbits and the
mapping's minimum folio order while a concurrent reader still holds a
folio of the old, smaller order, leading to crashes. In particular, it
made blkdev_write_iter() wrap buffered writes in inode_lock_shared().

However, the direct I/O fallback path was missed in that conversion.
blkdev_write_iter() passes blkdev_buffered_write() as an argument to
direct_write_fallback() with no lock held. A direct write that completes
only partially then finishes as a buffered write with no protection.

This can cause a BUG by racing partial direct writes against
ioctl(BLKBSZSET). Writer threads issue O_DIRECT pwritev() with a
two-segment iovec whose second segment is an unreadable PROT_NONE
mapping. The direct path then writes the first segment, fails to pin the
second, and returns short, entering the fallback. A second thread keeps
toggling the second segment's protection so that some fallbacks get past
fault_in_iov_iter_readable() and reach the page cache, a third thread
populates the page cache with folios of the current block size via
pread() and readahead(), and a fourth thread toggles the block size
between 512 bytes and 64K with BLKBSZSET. The minimum folio order only
moves with block sizes above PAGE_SIZE, i.e. with
CONFIG_TRANSPARENT_HUGEPAGE raising BLK_MAX_BLOCK_SIZE to 64K.

On a CONFIG_DEBUG_VM kernel this yields the following BUG:

  page dumped because: VM_BUG_ON_FOLIO(folio_order(folio) < mapping_min_folio_order(mapping))
  kernel BUG at mm/filemap.c:858!
  Oops: invalid opcode: 0000 [#1] SMP KASAN NOPTI
  RIP: 0010:__filemap_add_folio+0x860/0x8d0
  Call Trace:
   filemap_add_folio+0xc9/0x1f0
   __filemap_get_folio_mpol+0x240/0x660
   iomap_write_begin+0xa87/0xd70
   iomap_file_buffered_write+0x304/0x6a0
   blkdev_write_iter+0x255/0x510
   do_iter_readv_writev+0x23d/0x3c0
   vfs_writev+0x211/0x7d0
   do_pwritev+0x121/0x190
   do_syscall_64+0x121/0x630
   entry_SYSCALL_64_after_hwframe+0x77/0x7f

The same workload also trips WARN_ON_ONCE(pos >= folio_pos(folio) +
fsize) in iomap_trim_folio_range().

Fix this by calling blkdev_buffered_write() in the fallback path under
inode_lock_shared(), matching the plain buffered-write branch. With the
fix the same workload runs clean.

A short IOCB_NOWAIT direct write reaches the same fallback. Taking
i_rwsem there can now block behind set_blocksize(), and the fallback
already blocks on writeback of the data it copied in
direct_write_fallback(). blkdev_write_iter() already rejects a purely
buffered IOCB_NOWAIT write with -EOPNOTSUPP, so do not enter the
fallback for IOCB_NOWAIT at all: return the bytes the direct path
already wrote, or -EAGAIN if none, and let the caller retry.

The reproducer used was written by an LLM, and is available at [1].

[1] https://gist.github.com/tzussman/69d06bc57d42a42989eb038b1b5aeb74

Fixes: 3c20917120ce ("block/bdev: enable large folio support for large logical block sizes")
Reported-by: Sashiko <sashiko-bot@xxxxxxxxxx>
Link: https://sashiko.dev/#/patchset/20260730-blk-dontcache-v7-0-3e8e6850068d%40columbia.edu?part=5
Assisted-by: Claude:claude-fable-5
Signed-off-by: Tal Zussman <tz2294@xxxxxxxxxxxx>
---
 block/fops.c | 23 ++++++++++++++++++++---
 1 file changed, 20 insertions(+), 3 deletions(-)

diff --git a/block/fops.c b/block/fops.c
index c57784773fe1..d5f569333f46 100644
--- a/block/fops.c
+++ b/block/fops.c
@@ -765,9 +765,26 @@ static ssize_t blkdev_write_iter(struct kiocb *iocb, struct iov_iter *from)
 
 	if (iocb->ki_flags & IOCB_DIRECT) {
 		ret = blkdev_direct_write(iocb, from);
-		if (ret >= 0 && iov_iter_count(from))
-			ret = direct_write_fallback(iocb, from, ret,
-					blkdev_buffered_write(iocb, from));
+		if (ret >= 0 && iov_iter_count(from)) {
+			if (iocb->ki_flags & IOCB_NOWAIT) {
+				/*
+				 * The buffered fallback blocks on i_rwsem and
+				 * on writeback of the data it copied: return
+				 * the short direct write instead and let the
+				 * caller retry.
+				 */
+				if (!ret)
+					ret = -EAGAIN;
+			} else {
+				ssize_t ret2;
+
+				inode_lock_shared(bd_inode);
+				ret2 = blkdev_buffered_write(iocb, from);
+				inode_unlock_shared(bd_inode);
+				ret = direct_write_fallback(iocb, from, ret,
+							    ret2);
+			}
+		}
 	} else {
 		/*
 		 * Take i_rwsem and invalidate_lock to avoid racing with

-- 
2.39.5