svn commit: r244793 - in user/attilio/membarclean/dev: drm drm2 netmap virtio
Bryan Venteicher
bryanv at freebsd.org
Sun Dec 30 20:01:02 UTC 2012
On Sat, Dec 29, 2012 at 10:37 AM, Attilio Rao <attilio at freebsd.org> wrote:
> On Sat, Dec 29, 2012 at 5:03 PM, Konstantin Belousov
> <kostikbel at gmail.com> wrote:
>> On Sat, Dec 29, 2012 at 07:13:57AM -0800, Attilio Rao wrote:
>>> On Sat, Dec 29, 2012 at 7:01 AM, Konstantin Belousov
>>> <kostikbel at gmail.com> wrote:
>>> > On Fri, Dec 28, 2012 at 11:34:39PM +0000, Attilio Rao wrote:
>>> >> On Fri, Dec 28, 2012 at 11:16 PM, Konstantin Belousov
>>> >> <kostikbel at gmail.com> wrote:
>>> >> > On Fri, Dec 28, 2012 at 10:55:47PM +0000, Attilio Rao wrote:
>>> >> >> On Fri, Dec 28, 2012 at 10:32 PM, Konstantin Belousov
>>> >> >> <kostikbel at gmail.com> wrote:
>>> >> >> > On Fri, Dec 28, 2012 at 10:18:41PM +0000, Attilio Rao wrote:
>>> >> >> >> Author: attilio
>>> >> >> >> Date: Fri Dec 28 22:18:41 2012
>>> >> >> >> New Revision: 244793
>>> >> >> >> URL: http://svnweb.freebsd.org/changeset/base/244793
>>> >> >> >>
>>> >> >> >> Log:
>>> >> >> >> - Remove rmb() usage from netmap and replace it with intended operation
>>> >> >> >> to do actual memory fetching reads.
>>> >> >> >> - GC unused DRM_WRITEMEMORYBARRIER() from drm and drm2.
>>> >> >> >> - Use atomic_load_acq_*() in virtio and drm2 in places that don't need
>>> >> >> >> to use rmb().
>>> >> >> >>
>>> >> >> >> All these changes remove completely rmb() from MI code, with the
>>> >> >> >> exception of cxgbe which will be hammered in a followup commit.
>>> >> >> >>
>>> >> >> >> Modified:
>>> >> >> >> user/attilio/membarclean/dev/drm/drmP.h
>>> >> >> >> user/attilio/membarclean/dev/drm2/drmP.h
>>> >> >> >> user/attilio/membarclean/dev/drm2/drm_atomic.h
>>> >> >> >> user/attilio/membarclean/dev/drm2/drm_irq.c
>>> >> >> >> user/attilio/membarclean/dev/netmap/netmap.c
>>> >> >> >> user/attilio/membarclean/dev/virtio/virtqueue.c
>>> >> >> >>
>>> >> >> >> Modified: user/attilio/membarclean/dev/drm/drmP.h
>>> >> >> >> ==============================================================================
>>> >> >> >> --- user/attilio/membarclean/dev/drm/drmP.h Fri Dec 28 22:06:50 2012 (r244792)
>>> >> >> >> +++ user/attilio/membarclean/dev/drm/drmP.h Fri Dec 28 22:18:41 2012 (r244793)
>>> >> >> >> @@ -241,11 +241,9 @@ typedef u_int32_t u32;
>>> >> >> >> typedef u_int16_t u16;
>>> >> >> >> typedef u_int8_t u8;
>>> >> >> >>
>>> >> >> >> -/* DRM_READMEMORYBARRIER() prevents reordering of reads.
>>> >> >> >> - * DRM_WRITEMEMORYBARRIER() prevents reordering of writes.
>>> >> >> >> +/* DRM_WRITEMEMORYBARRIER() prevents reordering of writes.
>>> >> >> >> * DRM_MEMORYBARRIER() prevents reordering of reads and writes.
>>> >> >> >> */
>>> >> >> >> -#define DRM_READMEMORYBARRIER() rmb()
>>> >> >> >> #define DRM_WRITEMEMORYBARRIER() wmb()
>>> >> >> >> #define DRM_MEMORYBARRIER() mb()
>>> >> >> >>
>>> >> >> >>
>>> >> >> >> Modified: user/attilio/membarclean/dev/drm2/drmP.h
>>> >> >> >> ==============================================================================
>>> >> >> >> --- user/attilio/membarclean/dev/drm2/drmP.h Fri Dec 28 22:06:50 2012 (r244792)
>>> >> >> >> +++ user/attilio/membarclean/dev/drm2/drmP.h Fri Dec 28 22:18:41 2012 (r244793)
>>> >> >> >> @@ -263,11 +263,9 @@ typedef int32_t s32;
>>> >> >> >> typedef int16_t s16;
>>> >> >> >> typedef int8_t s8;
>>> >> >> >>
>>> >> >> >> -/* DRM_READMEMORYBARRIER() prevents reordering of reads.
>>> >> >> >> - * DRM_WRITEMEMORYBARRIER() prevents reordering of writes.
>>> >> >> >> +/* DRM_WRITEMEMORYBARRIER() prevents reordering of writes.
>>> >> >> >> * DRM_MEMORYBARRIER() prevents reordering of reads and writes.
>>> >> >> >> */
>>> >> >> >> -#define DRM_READMEMORYBARRIER() rmb()
>>> >> >> >> #define DRM_WRITEMEMORYBARRIER() wmb()
>>> >> >> >> #define DRM_MEMORYBARRIER() mb()
>>> >> >> >>
>>> >> >> >>
>>> >> >> >> Modified: user/attilio/membarclean/dev/drm2/drm_atomic.h
>>> >> >> >> ==============================================================================
>>> >> >> >> --- user/attilio/membarclean/dev/drm2/drm_atomic.h Fri Dec 28 22:06:50 2012 (r244792)
>>> >> >> >> +++ user/attilio/membarclean/dev/drm2/drm_atomic.h Fri Dec 28 22:18:41 2012 (r244793)
>>> >> >> >> @@ -38,6 +38,7 @@ typedef u_int32_t atomic_t;
>>> >> >> >>
>>> >> >> >> #define atomic_set(p, v) (*(p) = (v))
>>> >> >> >> #define atomic_read(p) (*(p))
>>> >> >> >> +#define atomic_read_acq(p) atomic_load_acq_int(p)
>>> >> >> >> #define atomic_inc(p) atomic_add_int(p, 1)
>>> >> >> >> #define atomic_dec(p) atomic_subtract_int(p, 1)
>>> >> >> >> #define atomic_add(n, p) atomic_add_int(p, n)
>>> >> >> >>
>>> >> >> >> Modified: user/attilio/membarclean/dev/drm2/drm_irq.c
>>> >> >> >> ==============================================================================
>>> >> >> >> --- user/attilio/membarclean/dev/drm2/drm_irq.c Fri Dec 28 22:06:50 2012 (r244792)
>>> >> >> >> +++ user/attilio/membarclean/dev/drm2/drm_irq.c Fri Dec 28 22:18:41 2012 (r244793)
>>> >> >> >> @@ -701,8 +701,7 @@ u32 drm_vblank_count_and_time(struct drm
>>> >> >> >> do {
>>> >> >> >> cur_vblank = atomic_read(&dev->_vblank_count[crtc]);
>>> >> >> >> *vblanktime = vblanktimestamp(dev, crtc, cur_vblank);
>>> >> >> >> - rmb();
>>> >> >> >> - } while (cur_vblank != atomic_read(&dev->_vblank_count[crtc]));
>>> >> >> >> + } while (cur_vblank != atomic_read_acq(&dev->_vblank_count[crtc]));
>>> >> >> >>
>>> >> >> >> return cur_vblank;
>>> >> >> >> }
>>> >> >> >
>>> >> >> > The drm/drm2 is the contributed code, maintained outside the tree, and I
>>> >> >> > explicitely decided to keep the original approach. I will appreciate if
>>> >> >> > this would be left as it is.
>>> >> >>
>>> >> >> Aren't the stubs in atomic FreeBSD specific?
>>> >> >> So it does rely on some sort of Linux layer emulation?
>>> >> >> Right now my target is to get rid of rmb() and wmb() and garbage
>>> >> >> collect them. If drm2 is Linux specific and needs to stay that way I
>>> >> >> will make a different patch.
>>> >> >
>>> >> > *mb() are the Linux KPI, and almost any non-trivial ported driver
>>> >> > depends on them. The mess which we already have in tree with the
>>> >> > duplicated and sometimes wrong definitions of the barries is because we
>>> >> > get this compat functions too late, and the driver authors added them
>>> >> > locally.
>>> >> >
>>> >> > The functions should stay there, and their usage in the ported drivers too.
>>> >> > Otherwise, the mess will reappear and continue to spread.
>>> >>
>>> >> I don't agree with "the functions should stay there" just for your
>>> >> premise: they are not a FreeBSD KPI and they should be defined in
>>> >> atomic.h. They violate our atomic/membar model and they are only
>>> >> useful for Linux-feature emulation purposes. More importantly, they
>>> >> are mostly wrong and the requirements are niether clear or precise
>>> >> because they makes just no sense in FreeBSD.
>>> > In what way they are wrong ?
>>> >
>>> > I am fine with removing them from the machine/atomic.h and putting
>>> > under any other namespace, even into machine/never_use_the_linuxisms.h.
>>> > But driver authors must not be forced to reimplement the same KPI
>>> > Nth time.
>>>
>>> This is fine.
>>> They are wrong because Linux requirements about write memory barriers
>>> are different than our _rel barriers and their requirements about read
>>> memory barriers are different than our _acq barriers.
>>> Per-Linux doc, infact, their write barriers only serialize over other
>>> pre-stores while their read barriers only serialize over post-loads.
>>> FreeBSD serializes over pre-stores/loads (for _rel) and
>>> post-stores/loads (for _acq).
>> Yes, but this does not make Linux-style rmb/wmb barriers incompatible with
>> the FreeBSD acq/rel barriers. They pairs have obviously different
>> semantic, they are not interchangable, but there is nothing new to state.
>> I do not see how the rmb/wmb can be claimed incompatible or wrong.
>> They are only different.
>
> As first thing they are undocumented in our world. Second thing they
> are different by freebsd philosophy. And I hardly believe all drivers
> developer understands this nit differences between freebsd _rel/_acq
> and wmb()/acq(), leaving alone architecture porters.
>
>>>
>>> Having the wmb() and rmb() in linux compat stub, that all the drivers
>>> can reuse, is a perfectly acceptable compromise.
>>> However such layer doesn't exist nowadays. I think there were some
>>> resistence in the past about implementing it, because people were
>>> freightened that drivers weren't going to be ported natively to
>>> FreeBSD anymore. However for simple stuff like membars this should not
>>> withstand.
>> Ok.
>>
>>>
>>> > BTW, did you looked at the lib/libc/sys/__vdso_gettimeofday.c:binuptime() ?
>>> >
>>> >>
>>> >> My current plan is to just keep mb() at the atomic-FreeBSD level and
>>> >> get rid of the others.
>> Still, there are situations where FreeBSD implementation does not work,
>> see above example from libc.
>
> I still need to look into it, but there are also other cases, like the
> virtio ones, where we want to get _rel or _acq semantic for smaller
> than int read and writes. I think I will make a followup soon on all
> the cases where the wmb() and rmb() are used, try to see where our
> initial model is failing, if any, and try to formalize the situation
> some more. This is a first attempt.
> But supporting wmb() and rmb() at the architecture level only because
> "Linux does it" is wrong. If they belong to a compat layer, it is
> acceptable.
>
For VirtIO, I think I would be fine with reverting r240427 to make
everything just straight mb()'s, and then #ifdef'ing atomic(9) for
those arch's with 16-bit support.
> Attilio
>
>
> --
> Peace can only be achieved by understanding - A. Einstein
More information about the svn-src-user
mailing list