ftl_write() unconditionally routes writes through rwb_write() when
FTL_HAVE_RWBUFFER is defined, ignoring the O_DIRECT flag stored in
dev->oflags. This means callers that open with O_DIRECT expecting
unbuffered writes still get their data buffered in RAM.
The O_DIRECT flag is already checked in ftl_flush() to select
between direct and read-modify-write paths, but the buffering
decision in ftl_write() never consults it.
Add O_DIRECT checks in both ftl_write() and ftl_read() to bypass
rwbuffer entirely when the flag is set. For writes, go directly to
ftl_flush() and normalize the return value to match the block
driver convention (nsectors on success). For reads, go directly to
ftl_reload(), bypassing any read-ahead cache.
This gives O_DIRECT callers a complete bypass of the rwbuffer layer
for both reads and writes, matching the documented contract.
Signed-off-by: Neil Berkman <neil@xuku.com>
c1c6e26395