Re: [PATCH AUTOSEL 5.0 59/66] fs: stream_open - opener for stream-like files so that read and write can run simultaneously without deadlock

From: Kirill Smelkov
Date: Fri Apr 26 2019 - 14:35:28 EST

On Fri, Apr 26, 2019 at 11:00:01AM +0000, David Laight wrote:
> From: Kirill Smelkov
> > Sent: 26 April 2019 08:46
> ...
> > I'm not sure I understand your comment completely, but we convert to
> > stream_open only drivers that actually do _not_ use position at all, and
> > that were already using nonseekable_open, thus pread and pwrite were
> > already returning -ESPIPE for them (nonseekable_open clears
> > FMODE_{PREAD,PWRITE} and ksys_{pread,pwrite}64 check for that flag). We
> > also convert only drivers that use no_llseek for .llseek, so lseek
> > on those files is/was always returning -ESPIPE as well.
> >
> > If a driver uses position in its read and write and has support for
> > pread/pwrite (FMODE_PREAD and FMODE_PWRITE), pread and pwrite are
> > already working _without_ file->f_pos locking - because those system
> > calls do not semantically update file->f_pos at all and thus do not take
> > file->f_pos_lock - i.e. pread/pwrite can be run simultaneously already.
> Looks like I knew that once :-)
> Mind you, 'man pread' on my system is somewhat uninformative.
> Maybe pread() should always be allowed at offset 0.
> Then you wouldn't need all this extra logic.

I'm not sure I understand. Do you propose any change? If yes - what is
the change you are proposing?

> > If libc implements pread as lseek+read it will work for a single
> > user case (single thread, or fd not shared between processes), but it
> > will break because of lseek+read non-atomicity if multiple preads are
> > simultaneously used from several threads. And also for such emulation
> > for multiple users case there is a chance for pread vs pwrite deadlock,
> > since those system calls are using read and write and read and write
> > take file->f_pos_lock.
> I'd actually rather the pread() failed to compile.


> The actual implementation did 3 lseek()s (to save and restore the offset).
> A user level emulation could usually get away with one lseek().