Hello Jeff, On 9/8/20 9:44 PM, Jeff Layton wrote: > On Tue, 2020-09-08 at 13:27 +0200, Jan Kara wrote: >> Added Jeff to CC since he has written the code... >> >> On Mon 07-09-20 09:11:06, Michael Kerrisk (man-pages) wrote: >>> [Widening the CC to include Andrew and linux-fsdevel@] >>> [Milan: thanks for the patch, but it's unclear to me from your commit >>> message how/if you verified the details.] >>> >>> Andrew, maybe you (or someone else) can comment, since long ago your >>> >>> commit f79e2abb9bd452d97295f34376dedbec9686b986 >>> Author: Andrew Morton <akpm@xxxxxxxx> >>> Date: Fri Mar 31 02:30:42 2006 -0800 >>> >>> included a comment that is referred to in stackoverflow discussion >>> about this topic (that SO discussion is in turn referred to by >>> https://bugzilla.kernel.org/show_bug.cgi?id=194757). >>> >>> The essence as I understand it, is this: >>> (1) fsync() (and similar) may fail EIO or ENOSPC, at which point data >>> has not been synced. >>> (2) In this case, the EIO/ENOSPC setting is cleared so that... >>> (3) A subsequent fsync() might return success, but... >>> (4) That doesn't mean that the data in (1) landed on the disk. >> >> Correct. >> >>> The proposed manual page patch below wants to document this, but I'd >>> be happy to have an FS-knowledgeable person comment before I apply. >> >> Just a small comment below: >> >>> On Sat, 29 Aug 2020 at 09:13, <milan.opensource@xxxxxxxxx> wrote: >>>> From: Milan Shah <milan.opensource@xxxxxxxxx> >>>> >>>> This Fix addresses Bug 194757. >>>> Ref: https://bugzilla.kernel.org/show_bug.cgi?id=194757 >>>> --- >>>> man2/fsync.2 | 13 +++++++++++++ >>>> 1 file changed, 13 insertions(+) >>>> >>>> diff --git a/man2/fsync.2 b/man2/fsync.2 >>>> index 96401cd..f38b3e4 100644 >>>> --- a/man2/fsync.2 >>>> +++ b/man2/fsync.2 >>>> @@ -186,6 +186,19 @@ In these cases disk caches need to be disabled using >>>> or >>>> .BR sdparm (8) >>>> to guarantee safe operation. >>>> + >>>> +When >>>> +.BR fsync () >>>> +or >>>> +.BR fdatasync () >>>> +returns >>>> +.B EIO >>>> +or >>>> +.B ENOSPC >>>> +any error flags on pages in the file mapping are cleared, so subsequent synchronisation attempts >>>> +will return without error. It is >>>> +.I not >>>> +safe to retry synchronisation and assume that a non-error return means prior writes are now on disk. >>>> .SH SEE ALSO >>>> .BR sync (1), >>>> .BR bdflush (2), >> >> So the error state isn't really stored "on pages in the file mapping". >> Current implementation (since 4.14) is that error state is stored in struct >> file (I think this tends to be called "file description" in manpages) and >> so EIO / ENOSPC is reported once for each file description of the file that >> was open before the error happened. Not sure if we want to be so precise in >> the manpages or if it just confuses people. Anyway your takeway that no >> error on subsequent fsync() does not mean data was written is correct. >> >> > > Thinking about it more, I think we ought to spell this out explicitly as > we can in the manpage. This is a point of confusion for a lot of people > and not understanding this can lead to data integrity bugs. Maybe > something like this in the NOTES section? > > ''' > When fsync returns an error, the file is considered to be "clean". A > subsequent call to fsync will not result in a reattempt to write out the > data, unless that data has been rewritten. Applications that want to > reattempt writing to the file after a transient error must re-write > their data. > ''' Thanks. It's incredibly helpful when someone with the needed domain-specific knowledge suggest a wording! > To be clear: > > In practice, you'd only have to write enough to redirty each page in > most cases. Presumably, this could be accomplished by write(2)-ing exactly the same user space buffers again? So, I'd like to expand your text a little. How would the following be: [[ When fsync() or fdatasync() returns an error, the file is considered to be "clean", even though the corresponding modified ("dirty") buffer cache pages may not have been flushed to the storage device. A subsequent call to fsync() will not result in a reattempt to write out the data, unless that data has in the meantime been rewritten. Applications that want to reattempt writing to the file after a transient error--for example, EIO and ENOSPC can occur because of transient conditions--must rewrite their data. ]] How is that text? > Also, it is hard to claim that the above behavior is universally true. A > filesystem could opt to keep the pages dirty for some errors, but the > vast majority just toss out the data whenever there is a writeback > problem. I think I won't worry about trying to discuss such variations in the manual page. BTW, I added a loosely related question in a reply that I just sent to Jan. Maybe you have some thoughts there also. Thanks, Michael -- Michael Kerrisk Linux man-pages maintainer; http://www.kernel.org/doc/man-pages/ Linux/UNIX System Programming Training: http://man7.org/training/
