On Fri, Sep 1, 2017 at 7:24 AM, <sa514164@xxxxxxxxxxxxxxxx> wrote: > Hi: > I want to ask a question about CEPH_IOC_SYNCIO flag. > I know that when using O_SYNC flag or O_DIRECT flag, write call executes in other two code paths different than using CEPH_IOC_SYNCIO flag. > And I find the comments about CEPH_IOC_SYNCIO here: > > /* > * CEPH_IOC_SYNCIO - force synchronous IO > * > * This ioctl sets a file flag that forces the synchronous IO that > * bypasses the page cache, even if it is not necessary. This is > * essentially the opposite behavior of IOC_LAZYIO. This forces the > * same read/write path as a file opened by multiple clients when one > * or more of those clients is opened for write. > * > * Note that this type of sync IO takes a different path than a file > * opened with O_SYNC/D_SYNC (writes hit the page cache and are > * immediately flushed on page boundaries). It is very similar to > * O_DIRECT (writes bypass the page cache) except that O_DIRECT writes > * are not copied (user page must remain stable) and O_DIRECT writes > * have alignment restrictions (on the buffer and file offset). > */ > #define CEPH_IOC_SYNCIO _IO(CEPH_IOCTL_MAGIC, 5) > > My question is: > 1."This forces the same read/write path as a file opened by multiple clients when one or more of those clients is opened for write." -- Does this mean multiple clients can execute in the same code path when they all use the CEPH_IOC_SYNCIO flag? Will the use of CEPH_IOC_SYNCIO in all clients bring effects such as coherency and performance? If you're just using the normal interfaces, you don't need to play around with this. I *think* this ioctl is only so that if you are using lazyio (which disables the usual cache coherence), you can still get get data IO which is coordinated with other clients. > 2."...except that O_DIRECT writes are not copied (user page must remain stable)" -- As I know when threads write with CEPH_IOC_SYNCIO flag, the write call will block until ceph osd and mds send back responses. So even with CEPH_IOC_SYNCIO flag(the user pages are not locked here, I guess), but the user cannot use these pages. How can the use of CEPH_IOC_SYNCIO flag make better use of user space memory? I'm not very familiar with these mechanisms, but I think it's saying that if you use CEPH_IOC_SYNCIO in an async IO interface, once the async write returns then it will have done an internal copy and can use the pages again? Not really sure... -Greg -- To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
