On 09/13/2012 02:56 PM, Paolo Bonzini wrote: > Il 13/09/2012 19:17, Jeff Cody ha scritto: >>>> >>>> Why does this matter? If raw-posix was opened directly (i.e. without >>>> the bs->file indirection) this would cause a writethrough file to be >>>> incorrectly opened as writeback. >>>> >>>> Paolo >>>> >> The problem this patch was trying to work around is that >> bdrv_open_common() forces BDRV_O_CACHE_WB (commit e1e9b0ac), but that >> setting is not preserved in bs->open_flags, so it is lost on a reopen. > > Perhaps we _should_ preserve that in bs->open_flags, while still using > the initial value of BDRV_O_CACHE_WB to initialize bs->enable_write_cache.
That would work. Part of the problem is that BDRV_O_CACHE_WB seems overloaded; maybe bdrv_parse_cache_flags() should set a new flag, called BDRV_O_CACHE_WCE, which can be used in lieu of enable_write_cache (similar to getting rid of keep_read_only in favor of BDRV_O_ALLOW_RDWR). And then bdrv_parse_cache_flags() would just not set BDRV_O_CACHE_WB, which can then be used exclusively by the lower layers for their parsing (and bdrv_open_common would just set bs->open_flags to always have BDRV_O_CACHE_WB). Then patch 2/16 would change to having bdrv_set_enable_write_cache() toggle BDRV_O_CACHE_WCE. > >> Is there a scenario currently that has raw-posix opened directly as a >> writethrough file, or were you more concerned with future use? > > Not for raw-posix, but IIRC some other protocol is opened directly > without a format on top. rbd perhaps? I'm concerned of having to work > around what seems like a bug elsewhere, in multiple protocols. > > Paolo >
