Review Request 130084: Add a pair of flags forcing fsync during copy loop

Sun Apr 16 02:01:18 UTC 2017

> On April 15, 2017, 8:30 a.m., David Faure wrote:
> > But doesn't this make copying much slower in the normal case? (copying onto a non-removable harddisk partition).
> > 
> > It sounds to me like this should be
> > 1) done internally in kio_file (no Job API for this)
> > 2) only when the destination is a removable device
> 
> Oswald Buddenhagen wrote:
>     i've read multiple times that fsync isn't a performance problem on modern file systems any more. whatever that may mean.
>     limiting this to cross-device isn't really sensible imo - a) one can have multiple internal disks and b) even if the disk stays in, at some point the flushing will commence and will be a major slowdown for subsequent operations.
>     in fact, this problem is bad enough that the linux kernel community realized it (which, in the area of disk i/o, never ceases to amaze) - https://lwn.net/Articles/682582/ (obvious followup question: what kernel do you use? this code seems to have landed in 4.10)

I'm using kernel `Linux kj-hedt-arch 4.10.8-1-ARCH #1 SMP PREEMPT Fri Mar 31 16:50:19 CEST 2017 x86_64 GNU/Linux`. My understanding is that the patchset you're talking about will not allow synchronous reads, such as faulting in application code, to get stuck behind a full writeback queue, by limiting how much work the MM subsystem can send to the disk - not by throttling how fast applications can dirty the device. I've not noticed any problems using the disk whilst writing with or without my patch to KIO here but I haven't really stressed it.

As to what the performance implications of fsync - I guess it depends how much you care about what you were planning to do with the file after you copy it. I implemented the "fsync if source/dest are on different filesystems" logic because in that case, one of the things you might be wanting to do is unmount the disk. If you wanted to interact with the file on the destination system instead, this patch would make it take (much) longer. The reason I implemented this with a job flag is that I was envisiging making it an option in Dolphin - like the "move/copy" menu when you drop, you could also get "copy with fsync" perhaps for this reason - we don't know what the user plans to do with the file afterwards.

I'm happy enough to use "is the device removeable?" as a heuristic for "the users next desired operation on this file is probably to unmount it" instead and delete the Job API - this would address my use case at least. How do you suggest I get this information in `kio_file`? On Linux it looks like I can get this from sysfs `/sys/dev/block/maj:min/removeable`, but I don't know how to do this for other platforms and don't have them available to test. Would the patch be OK if I just added this for linux?

- KJ

-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://git.reviewboard.kde.org/r/130084/#review103048
-----------------------------------------------------------

On April 15, 2017, 8:28 a.m., KJ Tsanaktsidis wrote:
> 
> -----------------------------------------------------------
> This is an automatically generated e-mail. To reply, visit:
> https://git.reviewboard.kde.org/r/130084/
> -----------------------------------------------------------
> 
> (Updated April 15, 2017, 8:28 a.m.)
> 
> 
> Review request for KDE Frameworks, Oswald Buddenhagen and Thiago Macieira.
> 
> 
> Repository: kio
> 
> 
> Description
> -------
> 
> When copying a large-ish file (~1-2GB) from very fast storage to very slow storage (e.g. an NVME SSD to a cheap USB flash drive) on a machine with lots of RAM, Dolphin displays a progress bar which finishes in a fraction of a second (i.e. as fast as it takes to read the source file into the Linux page cache). Unmounting the drive then of course takes a long time, with only an indeterminate spinner.
> 
> This patch adds an option to force fsync during copy jobs, so that the copy progress bar measures how long it will take to actually copy the file to the destination.
> 
> I've added two flags - Fsync and FsyncCrossFilesystem - to the JobEnum flag. The former will cause all copy operations to fsync during the copy loop, whilst the latter will only fsync copies that are across different filesystems.
> 
> If this patch gets OK'd, I have another patch which adds support for this into the appropriate places in Dolphin. I would think that at least FsyncCrossFilesystem should be the default, but Fsync always might be a little heavy handed. At the least fsync'ing cross-filesystem copies ensures that the unmount won't take forever.
> 
> 
> Diffs
> -----
> 
>   src/core/copyjob.cpp 7c02cb50d7e9c11bbcd9264832357f3fd6dc8c16 
>   src/core/filecopyjob.cpp 301b7039158b7dc537b9004c14845b3d1d60f8eb 
>   src/core/job_base.h 0be9629f42277afc5f72d00d0cae5c9c1cd2b8bc 
>   src/core/slavebase.cpp 3778df813b8568657a2cbd9412c1244f94696a0c 
>   src/ioslaves/file/file_unix.cpp 3c1b9927e3dd2d0134f77caec6e6b24a0356d26f 
> 
> Diff: https://git.reviewboard.kde.org/r/130084/diff/
> 
> 
> Testing
> -------
> 
> Tested the patch with KDE/Dolphin on Arch Linux, which is version 5.32.0. The diff applies cleanly to master so I assume there shouldn't be any issues there, but I've not actually checked that. As advertised, copying a file to USB flash storage now displays an accurate progress bar.
> 
> I experimented with how often fsync should be called on my hardware, and I found calling it every ~1M copied caused no decrease in copy performance whilst still providing accurate progress info. That is the setting I've gone with in this patch. I'm open to suggestions on how this could be tuned better though.
> 
> 
> Thanks,
> 
> KJ Tsanaktsidis
> 
>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.kde.org/pipermail/kde-frameworks-devel/attachments/20170416/eb521f1f/attachment-0001.html>