On Thu, Dec 24, 2020 at 01:16:50AM +0900, Minwoo Im wrote:

Hello,

On 20-12-23 16:49:04, Christoph Hellwig wrote:

quoted

set_blocksize just sets the block sise used for buffer heads and should
not be called by the driver.  blkdev_get updates the block size, so
you must already have the fd re-reading the partition table open?
I'm not entirely sure how we can work around this except by avoiding
buffer head I/O in the partition reread code.  Note that this affects
all block drivers where the block size could change at runtime.

Thank you Christoph for your comment on this.

Agreed.  BLKRRPART leads us to block_read_full_page which takes buffer
heads for I/O.

Yes, __blkdev_get() sets i_blkbits of block device inode via
set_init_blocksize.  And Yes again as nvme-cli already opened the block
device fd and requests the BLKRRPART with that fd.  Also, __bdev_get()
only updates the i_blkbits(blocksize) in case bdev->bd_openers == 0 which
is the first time to open this block device.

Then, how about having NVMe driver prevent underflow case for the
request->__data_len is smaller than the logical block size like:

Not sure this helps.  I think we need to fix this proper and in the
block layer.  The long term fix is to stop messing with i_blksize
at all, but that is going to take very long.

I think for now the only thing we can do is to set a flag in the
gendisk when the block size changes and then reject all I/O until
the next first open that sets the blocksize.