On Wed, Jul 22, 2020 at 04:32:47PM -0700, Darrick J. Wong wrote: > On Wed, Jul 22, 2020 at 04:26:25PM -0700, Eric Biggers wrote: > > On Wed, Jul 22, 2020 at 03:34:04PM -0700, Eric Biggers wrote: > > > So, something like this: > > > > > > diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c > > > index 44bad4bb8831..2816194db46c 100644 > > > --- a/fs/ext4/inode.c > > > +++ b/fs/ext4/inode.c > > > @@ -3437,6 +3437,15 @@ static int ext4_iomap_begin(struct inode *inode, loff_t offset, loff_t length, > > > map.m_len = min_t(loff_t, (offset + length - 1) >> blkbits, > > > EXT4_MAX_LOGICAL_BLOCK) - map.m_lblk + 1; > > > > > > + /* > > > + * When inline encryption is enabled, sometimes I/O to an encrypted file > > > + * has to be broken up to guarantee DUN contiguity. Handle this by > > > + * limiting the length of the mapping returned. > > > + */ > > > + if (!(flags & IOMAP_REPORT)) > > > + map.m_len = fscrypt_limit_io_blocks(inode, map.m_lblk, > > > + map.m_len); > > > + > > > if (flags & IOMAP_WRITE) > > > ret = ext4_iomap_alloc(inode, &map, flags); > > > else > > > > > > > > > That also avoids any confusion between pages and blocks, which is nice. > > > > Correction: for fiemap, ext4 actually uses ext4_iomap_begin_report() instead of > > ext4_iomap_begin(). So we don't need to check for !IOMAP_REPORT. > > > > Also it could make sense to limit map.m_len after ext4_iomap_alloc() rather than > > before, so that we don't limit the length of the extent that gets allocated but > > rather just the length that gets returned to iomap. > > Naïve question here -- if the decision to truncate the bio depends on > the file block offset, can you achieve the same thing by capping the > length of the iovec prior to iomap_dio_rw? > > Granted that probably only makes sense if the LBLK IV thing is only > supposed to be used infrequently, and having to opencode a silly loop > might be more hassle than it's worth... > We *could* do the truncation there, but that would truncate the actual read() or write(). So, userspace would see a short read or write. And I understand that while applications are *supposed* to handle short reads and writes, many don't. I think Dave's suggestion makes more sense, since it would make this case be treated just like normal fragmentation of the file. - Eric
