[xserver-xorg][PATCH 1/1] xwayland: ftruncate if posix_fallocate fails

[Ian Ray]

On Mon, Apr 25, 2016 at 04:49:07PM +0300, Pekka Paalanen wrote:
> On Mon, 25 Apr 2016 15:47:09 +0300
> Ian Ray <ian.ray at ge.com> wrote:
> 
> > On a slow system that is configured with SMART_SCHEDULE_POSSIBLE, large
> > posix_fallocate() requests may be interrupted by the SmartScheduleTimer
> > (SIGALRM) continuously. Fallback to ftruncate if posix_fallocate fails.
> > 
> > Signed-off-by: Ian Ray <ian.ray at ge.com>
> > ---
> >  hw/xwayland/xwayland-shm.c | 9 ++++++---
> >  1 file changed, 6 insertions(+), 3 deletions(-)
> > 
> > diff --git a/hw/xwayland/xwayland-shm.c b/hw/xwayland/xwayland-shm.c
> > index e8545b3..342b723 100644
> > --- a/hw/xwayland/xwayland-shm.c
> > +++ b/hw/xwayland/xwayland-shm.c
> > @@ -142,9 +142,12 @@ os_create_anonymous_file(off_t size)
> >  #ifdef HAVE_POSIX_FALLOCATE
> >      ret = posix_fallocate(fd, 0, size);
> >      if (ret != 0) {
> > -        close(fd);
> > -        errno = ret;
> > -        return -1;
> > +        /* Fallback to ftruncate in case of failure. */
> > +        ret = ftruncate(fd, size);
> > +        if (ret < 0) {
> > +            close(fd);
> > +            return -1;
> > +        }
> >      }
> >  #else
> >      ret = ftruncate(fd, size);
> 
> Hi Ian,
> 
> I indeed think this is a bit harsh. The first EINTR may happen on any
> system, so the very least we should try twice before giving up.
> 
> I'd also like to see some comments on whether fallocate making no
> progress when repeatedly restarted is normal or not. Maybe that should
> be asked on a kernel list?
> 
> 
> Thanks,
> pq

Hi Pekka,

Referring to  http://lxr.free-electrons.com/source/mm/shmem.c#L2235 and 
http://lkml.iu.edu/hypermail/linux/kernel/1603.0/03523.html,  fallocate
is _not_ expected to make progress when repeatedly restarted.

Taking Yuriy's reply in to account, then there seems to be are a couple
of options:

1. try fallocate() a couple of times and in case of EINTR fallback
   to ftruncate

2. mask SIGALRM while calling fallocate

Any preference?


The time taken, in seconds, by fallocate() for various lengths is shown
below (measured on iMX6 dual-core system).  Average of 10 measurements.

len    1024, min 0.000090, max 0.000113, mean 0.000095, stddev 0.000006
len    2048, min 0.000090, max 0.000097, mean 0.000093, stddev 0.000003
len    4096, min 0.000091, max 0.000542, mean 0.000139, stddev 0.000134
len    8192, min 0.000134, max 0.000141, mean 0.000136, stddev 0.000002
len   16384, min 0.000207, max 0.001136, mean 0.000305, stddev 0.000277
len   32768, min 0.000189, max 0.001858, mean 0.000418, stddev 0.000488
len   65536, min 0.000353, max 0.000369, mean 0.000360, stddev 0.000006
len  131072, min 0.000672, max 0.000691, mean 0.000678, stddev 0.000005
len  262144, min 0.001314, max 0.001416, mean 0.001337, stddev 0.000028
len  524288, min 0.002621, max 0.003185, mean 0.002698, stddev 0.000164
len 1048576, min 0.004323, max 0.006196, mean 0.004679, stddev 0.000615
len 2097152, min 0.008710, max 0.009571, mean 0.008841, stddev 0.000246
len 4194304, min 0.017518, max 0.017709, mean 0.017625, stddev 0.000070

Blue skies,
Ian