On Thu, Aug 18, 2011 at 11:26:29PM +0800, Kautuk Consul wrote: > Please find my comments inline to the email below: > > On Thu, Aug 18, 2011 at 6:05 PM, Wu Fengguang <fengguang.wu@xxxxxxxxx> wrote: > > On Thu, Aug 18, 2011 at 08:14:57PM +0800, Artem Bityutskiy wrote: > >> On Thu, 2011-08-18 at 17:48 +0800, Wu Fengguang wrote: > >> > > For example, the user might want to write-back pages in smaller > >> > > intervals to a block device which has a > >> > > faster known writeback speed. > >> > > >> > That's not a complete rational. What does the user ultimately want by > >> > setting a smaller interval? What would be the problems to the other > >> > slow devices if the user does so by simply setting a small value > >> > _globally_? > >> > > >> > We need strong use cases for doing such user interface changes. > >> > Would you detail the problem and the pains that can only (or best) > >> > be addressed by this patch? > >> > >> Here is a real use-case we had when developing the N900 phone. We had > >> internal flash and external microSD slot. Internal flash is soldered in > >> and cannot be removed by the user. MicroSD, in contrast, can be removed > >> by the user. > > Yes, of course. I forgot this aspect also. > In fact I, too work on embedded platforms and I have faced this > problem with removable USB > disks. Our embedded applications don't even tell the user when it > would be a good time to remove > the USB stick. > Hence we run into data integrity problems for our filesystems when > some writebacks have not been > completed before removal of the USB disk. > Thanks for mentioning this as this adds to a use-case for this feature. For the removable USB disks, we can do a policy that set dirty_background_time = 0. This will work better than hacking the dirty intervals. For one thing, it's impractical to set the latter to tiny values so as to avoid excessive wakeups. And the intervals interface is never a guarantee. dirty_expire_interval only promises to _start_ writeback on the expired inodes in "best efforts" way. Only the dirty_ratio interface guarantees to keep the number of pages under the limit, hence limiting the most data that can be lost in hot removal events. > >> For the internal flash we wanted long intervals and relaxed limits to > >> gain better performance. > > > > Understand -- it's backed by the battery anyway. > > > > Yeah it's a practical way. It might even optimize away some of the > > writes if they are truncated some time later. It also allows possible > > optimization of deferring the writes to user inactive periods. > > > > However the ultimate optimization could be to prioritize READs over > > WRITEs in the IO scheduler, so that async WRITEs have minimal impact > > on normal operations. It's the only option for the MicroSD case, > > anyway. > > > >> For MicroSD we wanted very short intervals and tough limits to make sure > >> that if the user suddenly removes his microSD (users do this all the > >> time) - we do not lose data. > > > > Pretty reasonable. > > > >> The discussed capability would be very useful in that case, AFAICS. > > > > Agreed. > > > >> IOW, this is not only about fast/slow devices and how quickly you want > >> to be able to sync the FS, this is also about data integrity guarantees. > > > > In fact I never think it would matter for fast/slow devices. It's the > > As I mentioned, if there is a comparitively faster device, you might want to set > smaller intervals in which your pages are synced with disk for quicker > memory reclamation > purposes. This can be used on servers that run apps that have high > disk accesses as > well as need a lot of memory. As I explained before, in that case, the > direct reclamation > procedure will cause the usermode apps to sleep while trying to free > up pages by flushing > them to disk via the filesystem's writepage(). Here you want to limit the number of dirty pages for reducing the chances page reclaim run into them. Again, the right interface for doing this job is dirty_ratio. Or if you need to do it per-bdi, it will be some per-bdi dirty_time interface that works in parallel with dirty_ratio, whatever smaller value will take effect. > > dirty_ratio/dirty_bytes interfaces that ask for improvement if care > > about too many pages being cached. > > > > The dirty_ratio/dirty_bytes interface is good as a spatial approach in > terms of number of pages > to actually write after each interval. > This still cannot solve the problem Artem is mentioning, because the > time at which removable disks > can be detached is indeterminable as the user can do this anytime he wants. > Whatever algorithm you use, you will eventually run into some > situation where the user detaches a > disk before the writeback can really happen. > I think it is up to the user/admin to determine how much write-back > interval is actually required for his/her > specific application. You seem to mis-understand how the dirty intervals sysctl values are carried out and have rather high expectations for them... What can be best done for removable disk, in terms of data integrity, is to immediate start writeout IO for any newly dirtied pages. Which can only be provided by the provisioned per-bdi dirty_background_time interface (by setting it to 0 for USB disks). > > The intervals interfaces are intended for data integrity and nothing > > more. > > Yes. That is correct, but do you feel that this data integrity is > possible in this age of removable > disks ? > That said, I would say that your patches are a very nice spatial > approach to a part of the solution. > Do you feel that combining a temporal approach along with your spatial > pattern analysis technique would > be the best way to ensure data integrity along with proper bandwidth > estimation for specific applications ? There will be no spatial/temporal difference when dirty_background_time=0, which is exactly what we can best do to protect data for removable disks. Thanks, Fengguang -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
