Hi Andy, On Fri, 2012-08-03 at 14:24 -0700, Andy Grover wrote: > [trimming CCs] > [resurrecting old thread] > Apologies for the slightly delayed response here.. > On 05/31/2012 01:25 AM, Nicholas A. Bellinger wrote: > > [PATCH-v2] target/file: Use O_DSYNC by default for FILEIO backends > > > > Convert to use O_DSYNC for all cases at FILEIO backend creation time to > > avoid the extra syncing of pure timestamp updates with legacy O_SYNC during > > default operation as recommended by hch. Continue to do this independently of > > Write Cache Enable (WCE) bit, as WCE=0 is currently the default for all backend > > devices and enabled by user on per device basis via attrib/emulate_write_cache. > > > > This patch drops the now unnecessary fd_buffered_io= token usage that was > > originally signaling when to explicitly disable O_SYNC at backend creation > > time for buffered I/O operation. This can end up being dangerous for a number > > of reasons during physical node failure, so go ahead and drop this option > > for now when O_DSYNC is used as the default. > > rtslib lets you specify buffered or unbuffered when creating backstores, > and would set fd_buffered_io=1. Should it just set > attrib/emulate_write_cache=1 now that fd_buffered_io=1 is gone? > So I've been thinking about this more, and am now considering to re-add an optional buffered FILEIO mode for people who understand what the risks are here wrt to data integrity + H/A failover, but still want the performance benefits of buffer cache for single node FILEIO backends w/ FUA WRITE + SYNCHRONIZE_CACHE. The one small bit that I'd like to also add is to force the setting of emulate_write_cache=1 (and cannot be disabled) when buffered FILEIO mode has been explicitly enabled by user. > Or to ask another way: overall, how should this option be presented to > the user? It seems rather important, and applies to both fileio and iblock. > So IBLOCK is different here because submit_bio() does not do any type of caching, unless your actual backend is doing it of course.. However, you'll typically still want to always set emulate_write_cache=1 here, otherwise SCSI clients (esp .32 based Linux distro ones) tend to perform much, much worse for almost every workload once they determine that write cache is disabled. --nab -- To unsubscribe from this list: send the line "unsubscribe target-devel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
