Il 09/03/2012 11:31, Stefan Hajnoczi ha scritto:
>> > +#ifdef FALLOC_FL_PUNCH_HOLE
>> > +    retval = fallocate(s->fd, FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE,
>> > +                       sector_num << 9, (int64_t)nb_sectors << 9);
> I'm concerned about introducing blocking syscalls in coroutine code
> paths.  This needs to be done asynchronously.

Right; it is no worse than what is already there, except that XFS could
use paio_ioctl.  Alternatives are:

1) require a new-enough kernel and only use fallocate; return a NULL
aiocb if !has_discard and convert it to ENOTSUP.

2) extract now from my threads branch the work to generalize
posix-aio-compat into a more flexible threadpool, and move the AIO code
back to block/raw-posix.c.

Paolo