lkml.org 
[lkml]   [2006]   [Sep]   [11]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    From
    SubjectRe: Opinion on ordering of writel vs. stores to RAM
    Date
    On Sunday, September 10, 2006 6:00 pm, Benjamin Herrenschmidt wrote:
    > > If we accept this, I don't think we're much better off than we are
    > > currently (not that I have a problem with that). That is, many
    > > drivers
    > > would still need to be fixed up.
    >
    > Not necessarily if you introduce the trick of doing the mmiowb() in
    > spin_unlock when a per-cpu flag has been set previously by writel... not
    > sure if it's worth tho.

    True, though again this would add a branch to writeX.

    > > Depends on what you mean by "ordered between MMIO and MMIO".
    > > mmiowb()
    > > was initially introduced to allow ordering of writes between CPUs,
    > > and
    > > really didn't say anything about memory vs. I/O, so the semantics you
    > > describe here would also be different (and more expensive) than what
    > > we
    > > have now.
    >
    > No. What I mean is that two consecutive MMIO writes on the same CPU stay
    > in order, and reads can't cross writes. The relaxed versions would still
    > require mmiowb() (or another name) for synchronisation against
    > spinlocks. As I told you before, I much prefer to sync of mmiowb() as a
    > sync with locks than a sync with "other MMIOs on anotehr CPU" since that
    > doesn't mean anything outside of the context of a lock.

    Sure, that's where one would typically use it, but it really is a memory
    barrier...

    >
    > > This is what mmiowb() is supposed to be, though only for writes.
    > > I.e.
    > > two writes from different CPUs may arrive out of order if you don't
    > > use
    > > mmiowb() carefully. Do you also need a mmiorb() macro or just a
    > > stronger mmiob()?
    >
    > No, you misunderstand me. That's the main problem with mmiowb() and
    > that's why it's so not clear to so many people: the way you keep
    > presenting it as synchronizing MMIO vs. MMIO. I think it's a lot more
    > clear if you present it as synchronizing MMIOs with locks. MMIO vs. MMIO
    > is anohter matter.

    That's because it *is* a barrier. I don't think it's any harder to understand
    then regular memory barriers for example. It's just that you'd typically use
    it in conjunction with locks to ensure proper device access.

    > It's wether consecutive MMIO writes can be
    > re-ordered, wether MMIO loads can cross a write (either because the load
    > is performed late, only when the value is actually used, or because the
    > load can be exucuted before a preceeding write). That's what current
    > __raw_* versions on PowerPC will allow, in addition to not doing endian
    > swap. My proposal was that __writel/__readl, however, would keep MMIO
    > vs. MMIO ordering (wouldn't allow that sort of re-ordering), however,
    > they wouldn't order vs. spinlock (would still require mmiowb) nor vs.
    > main memory (cacheable storage).

    Ok, that's fine, though I think you'd only want the very weak semantics (as
    provided by your __raw* routines) on write combined memory typically?

    > > mmiowb() could be written as io_to_io_write_barrier() if we wanted to be
    > > extra verbose. AIUI it's the same thing as smb_wmb() but for MMIO
    > > space.
    >
    > I'm very much against your terminology. It's -not- an IO to IO barrier.
    > It's an IO to lock barrier. Really. IO to IO is something else. ordering
    > of IOs between CPUs has absolutely no meaning outside of the context of
    > locked regions in any case.

    But it *is* MMIO vs. MMIO. There's confusion because your __raw* routines
    don't even guarantee same CPU ordering, while mmiowb() is solely intended for
    inter-CPU ordering.

    But as you say, the most common (maybe only) use model for it is to make sure
    critical sections protecting device access behave correctly, so I don't have
    a problem tying it to locks somehow.

    Jesse
    -
    To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
    the body of a message to majordomo@vger.kernel.org
    More majordomo info at http://vger.kernel.org/majordomo-info.html
    Please read the FAQ at http://www.tux.org/lkml/

    \
     
     \ /
      Last update: 2006-09-11 20:11    [W:0.025 / U:0.196 seconds]
    ©2003-2016 Jasper Spaans. hosted at Digital OceanAdvertise on this site