On 04/01/2016 09:13 AM, Alex Bligh wrote: > > On 1 Apr 2016, at 16:12, Alex Bligh <alex@...872...> wrote: > >> as qemu doesn't use FUA on write and the kernel doesn't use FUA > > "as qemu doesn't use FUA other than on write" - sorry qemu's block/nbd-client.c currently sends NBD_CMD_FLAG_FUA on NBD_CMD_FLUSH (unconditionally, if NBD_FLAG_SEND_FUA was advertised), and on NBD_CMD_WRITE (conditionally: if NBD_FLAG_SEND_FUA was advertised AND the upper layer requested FUA). I don't see it being sent on reads, so on that front, you appear to be correct that no common clients are expecting any particular behavior on reads. qemu's nbd/server.c silently ignores NBD_CMD_FLAG_FUA on all commands except NBD_CMD_READ (conditionally calls blk_co_flush()), NBD_CMD_WRITE (conditionally calls blk_co_flush()). Ignoring the flag on NBD_CMD_FLUSH seems to make sense, since that code unconditionally calls blk_co_flush(), which is all the more the flag was enabling on read/write. And if you are correct that no real clients are sending FUA on READ, then deleting (or changing) the server support for FUA on read should have no negative impact. qemu doesn't yet implement WRITE_ZEROES. When qemu client is talking to a qemu server, there is no incompatibility between the two - all client commands that set FUA are sanely handled by the recipient server code. But the same is not true if you pair a current qemu client with current nbd.git server, as the qemu will unconditionally set FUA on flush, and the server now rejects it as an invalid flag (thus making it impossible for qemu to flush). So we probably ought to fix qemu client to quit unconditionally sending FUA (no change when qemu client talks to qemu server); and at the same time out to fix NBD server to silently ignore FUA on flush (fix the breakage between the pairing). -- Eric Blake eblake redhat com +1-919-301-3266 Libvirt virtualization library http://libvirt.org
Attachment:
signature.asc
Description: OpenPGP digital signature