Getting rid of atomic_load_acq_int(&fdp->fd_nfiles)) from fget_unlocked
Mateusz Guzik
mjguzik at gmail.com
Wed Aug 13 01:00:52 UTC 2014
On Sun, Jul 13, 2014 at 04:34:21PM +0300, Konstantin Belousov wrote:
> On Sun, Jul 13, 2014 at 04:25:21PM +0300, Konstantin Belousov wrote:
> > On Sun, Jul 13, 2014 at 05:55:00AM +0200, Mateusz Guzik wrote:
> > > Currently:
> > > /*
> > > * Avoid reads reordering and then a first access to the
> > > * fdp->fd_ofiles table which could result in OOB operation.
> > > */
> > > if (fd < 0 || fd >= atomic_load_acq_int(&fdp->fd_nfiles))
> > > return (EBADF);
> > >
> > > However, if we put fd_nfiles and fd_otable into one atomically replaced
> > > structure the only need to:
> > > 1. make sure the pointer is read once
> > > 2. issue a data dependency barrier - this is a noop on all supported
> > > architectures and we don't even have approprate macro, so doing nothing
> > > seems fine
> > >
> > > The motivation is to boost performance to amortize for seqlock cost, in
> > > case it hits the tree.
> > >
> > > This has no impact on races with capability lookup.
> > >
> > > In a microbenchmark of 16 threads reading from the same pipe fd
> > > immediately returning EAGAIN the numbers are:
> > > x vanilla-readpipe-run-sum
> > > + noacq-readpipe-run-sum
> > > [..]
> > > N Min Max Median Avg Stddev
> > > x 20 13133671 14900364 13893331 13827075 471500.82
> > > + 20 59479718 59527286 59496714 59499504 13752.968
> > > Difference at 95.0% confidence
> > > 4.56724e+07 +/- 213483
> > > 330.312% +/- 1.54395%
> > >
> > > There are 3 steps:
> > > 1. tidy up capsicum to accept fde:
> > > http://people.freebsd.org/~mjg/patches/single-fdtable-read-capsicum.patch
> > > 2. add __READ_ONCE:
> > > http://people.freebsd.org/~mjg/patches/read-once.patch
> > > 3. put stuff into one structure:
> > > http://people.freebsd.org/~mjg/patches/filedescenttable.patch
> > >
> > > Comments?
> >
> > We use 4-space indent for the continuation lines. Look at the malloc(9)
> > call in the patch 3.
> >
> > The filedescenttable is really long name. Could it be, for instance,
> > fdescenttbl ?
> >
> > Other than that, I think that the patches 2 and 3 are fine. I did not
> > looked at the patch 1.
>
>
> As an afterthought, you do not need __READ_ONCE(), the __DEVOLATILE() alone
> would do what you need as well.
Turns out patch 2 was quite bad.
Reading http://www.open-std.org/jtc1/sc22/wg14/www/C99RationaleV5.10.pdf
(pdf page 77) reveals:
A cast of a value to a qualified type has no effect; the qualification
(volatile, say) can have no effect on the access since it has occurred
prior to the cast. If it is necessary to access a non-volatile object
using volatile semantics, the technique is to cast the address of the
object to the appropriate pointer-to-qualified type, then dereference
that pointer.
So how about we just follow the recomandation and also get the type
automagically like linux folks do (added to sys/param.h):
#define READ_ONCE(var) (*(volatile __typeof(var) *)&(var))
http://people.freebsd.org/~mjg/patches/read-once.patch
I incorporated suggested changes have overwritten old patches.
http://people.freebsd.org/~mjg/patches/filedescenttable.patch
I would like to commit these changes this week with 2 weeks mfc.
--
Mateusz Guzik <mjguzik gmail.com>
More information about the freebsd-arch
mailing list